Browse the corpus

Introduction Increased sexual risk behaviour and high HIV prevalence in migrants compared with non-migrants in sub-Saharan Africa have led to research and prevention efforts that focus on migration as an individual risk factor and an important driver of HIV transmission.1,2 Additionally, studies have sought to estimate the effect of migration on population-level HIV prevalence.3,4 Common interpretations of the role of migration are centred around increased prevalence of sexual risk behaviours in migrants,5,6 and increased risk of HIV acquisition when destination communities have higher HIV prevalence than do origin communities.2,7 Differences in definitions and designs between studies present challenges in interpretation and comparison of empirical studies of migration and HIV risk, with some studies reporting no differences in HIV or sexual risk behaviours between migrants and non-migrants.8,9 The detailed demographic, migration, and HIV data available in several population-based cohorts and demographic surveillance systems have provided analysts with a valuable source of data for HIV risk studies.10 However, with a few exceptions, HIV surveys in these study populations have been restricted to adults who are residents in the study area.11–13

n, and HIV data available in several population-based cohorts and demographic surveillance systems have provided analysts with a valuable source of data for HIV risk studies.10 However, with a few exceptions, HIV surveys in these study populations have been restricted to adults who are residents in the study area.11–13 One exception is data available from the Africa Centre for Health and Population Studies Demographic Surveillance System (ACDIS) in rural KwaZulu-Natal, South Africa. In view of the very high levels of circular migration, in which adults migrate back and forth between rural and urban areas or other centres of employment,14 the ACDIS study population includes, and follows longitudinally, resident and non-resident members of households in the study area.15 Non-resident members are those deemed to belong to households in the rural areas despite being resident with another household either in the study area or outside. At each data collection round, information is recorded about new migrations of individuals and households into and out of the study area, and demographic data continues to be collected about non-resident members of households who are living elsewhere. Since 2003, ACDIS has administered HIV and sexual behaviour surveys every year, aiming to contact all adult residents. Between 2003 and 2011, the survey of individuals also included a sample of non-residents.16 Research in context Evidence before this study

One exception is data available from the Africa Centre for Health and Population Studies Demographic Surveillance System (ACDIS) in rural KwaZulu-Natal, South Africa. In view of the very high levels of circular migration, in which adults migrate back and forth between rural and urban areas or other centres of employment,14 the ACDIS study population includes, and follows longitudinally, resident and non-resident members of households in the study area.15 Non-resident members are those deemed to belong to households in the rural areas despite being resident with another household either in the study area or outside. At each data collection round, information is recorded about new migrations of individuals and households into and out of the study area, and demographic data continues to be collected about non-resident members of households who are living elsewhere. Since 2003, ACDIS has administered HIV and sexual behaviour surveys every year, aiming to contact all adult residents. Between 2003 and 2011, the survey of individuals also included a sample of non-residents.16 Research in context Evidence before this study Before public HIV treatment programmes in sub-Saharan Africa, migration was widely understood to be positively associated with increased sexual risk behaviours and consequently with increased HIV prevalence in migrants and their partners. However, evidence for differential HIV prevalence in migrants and non-migrants from analyses of longitudinal population-based data is more mixed. We updated findings from review papers about migration and HIV risk in sub-Saharan Africa. We identified additional studies in PubMed searches from Jan 1, 2013, to Aug 14, 2014, with the terms “HIV” and “migra*” or “mobility”, and “Africa”. We also searched PubMed and Google Scholar to identify recent reports from international agencies about priorities for HIV, internal migration, and population geographies.

e identified additional studies in PubMed searches from Jan 1, 2013, to Aug 14, 2014, with the terms “HIV” and “migra*” or “mobility”, and “Africa”. We also searched PubMed and Google Scholar to identify recent reports from international agencies about priorities for HIV, internal migration, and population geographies. Implications of all the available evidence Population-based studies to measure HIV infection and sexual behaviour need to take into account the social, behavioural, and situational risk factors present in the multiple environments to which migrants, especially those with circular migration patterns, are exposed. Our findings also suggest that a need exists to assess whether national HIV treatment and prevention programmes are providing services suitable for diverse, mobile populations. Added value of this study The updated data from the Africa Centre for Health and Population Studies Demographic Surveillance System are particularly informative because this was the first demographic surveillance system established in a highly mobile population with a severe HIV epidemic, and in which characterisation of migration and mobility was central to the conceptual and data model. We identified a positive association between recent migration and increased HIV prevalence for male and female residents. In our study we explicitly discussed the challenges of defining migration and the associated HIV risk in a rural African population with high levels of circular migration.

conceptual and data model. We identified a positive association between recent migration and increased HIV prevalence for male and female residents. In our study we explicitly discussed the challenges of defining migration and the associated HIV risk in a rural African population with high levels of circular migration. To examine the complexities of the association between migration and potential HIV risk, we used longitudinal data collected in ACDIS to compare population-level differences in sexual HIV risk behaviours and HIV prevalence with a range of indicators of migration and living arrangements.

conceptual and data model. We identified a positive association between recent migration and increased HIV prevalence for male and female residents. In our study we explicitly discussed the challenges of defining migration and the associated HIV risk in a rural African population with high levels of circular migration. To examine the complexities of the association between migration and potential HIV risk, we used longitudinal data collected in ACDIS to compare population-level differences in sexual HIV risk behaviours and HIV prevalence with a range of indicators of migration and living arrangements. Methods Study design and population Since January, 2000, longitudinal demographic and health data have been collected for roughly 90 000 household members from 12 000 households in a predominately rural 438 km2 demographic surveillance area (DSA) within the uMkhanyakude district of northern KwaZulu-Natal, South Africa.15,16 The DSA consists of Zulu tribal land with scattered households and a formal municipal township. Almost all the population speaks Zulu. The main sources of income for most households are formal employment and government grants, including pensions.17 Migrants move from the study area for various reasons, often related to employment or education, or to join partners or parents. Non-resident household members maintain social and physical connections with rural households through return visits and exchanges of material and physical support, including shared child care and financial support.15 Proportions of migrants are high, with 32% of women and 38% of men non-resident in 2008, and migration patterns differ by gender.17,18

maintain social and physical connections with rural households through return visits and exchanges of material and physical support, including shared child care and financial support.15 Proportions of migrants are high, with 32% of women and 38% of men non-resident in 2008, and migration patterns differ by gender.17,18 The prevalence of HIV in this DSA has increased during the past decade and, in 2011, 29% of resident adults aged 15–49 years were infected with HIV.19 Crude HIV incidence in residents is estimated to be 2·63 infections per 100 person-years (95% CI 2·50–2·77).20 The local public HIV Treatment and Care Programme was initiated in 2004 and expanded rapidly. By December, 2011, 20 598 adults had initiated treatment, which is estimated to be 31% of all resident adults with HIV infection aged 15–50 years, thus contributing to the increasing HIV prevalence seen between 2005 and 2011.19 During the same period, no evidence suggested any increase in sexual risk taking behaviour. Condom use during most recent sexual intercourse with a regular partner increased significantly for men by an average of 2·6% (95% CI 1·5–3·7%) points per year and 4·1% (3·0–5·3) per year for women. Condom use at most recent sexual intercourse with a casual partner did not increase over time; it was more than 50% in 2005 compared with less than 30% with regular partners for both men and women.21

sed significantly for men by an average of 2·6% (95% CI 1·5–3·7%) points per year and 4·1% (3·0–5·3) per year for women. Condom use at most recent sexual intercourse with a casual partner did not increase over time; it was more than 50% in 2005 compared with less than 30% with regular partners for both men and women.21 Procedures Demographic data (eg, births, deaths, and marriages) and information such as periods of absence and presence and migration events for all child and adult household members were collected twice per year from 2005 to 2011.16 Household membership was defined by key respondents and mainly related to perceptions of social connectedness and belonging. A household member was deemed a resident if they usually kept their day-to-day belongings and slept at the homestead; thus adults who are resident in the study area can be both inmigrants and people who have been residentially stable (always resident). Similarly, non-resident members included people who outmigrated from the study area and people who always lived in a place outside the study area.

belongings and slept at the homestead; thus adults who are resident in the study area can be both inmigrants and people who have been residentially stable (always resident). Similarly, non-resident members included people who outmigrated from the study area and people who always lived in a place outside the study area. Annual health surveys have been administered in the DSA since 2005, including collection of an anonymised blood sample for HIV testing and information about up to three most recent sexual partnerships in the past year.16,21 For the individual surveillance in each year, resident and non-resident eligibility lists were based on information from the household census in December of the previous year. All resident household members aged 15 years and older were eligible to participate, together with a stratified sample of non-resident household members (women aged 15–49 years and men aged 15–54 years). Sampling was stratified by sex and pattern of return visits to their household in the DSA, based on historically typical circular migration patterns (eg, monthly or annual return visits). Additionally, any non-resident individuals with a negative HIV test result in the HIV surveillance in the 2 years before the survey who were not selected in the current random sample were included in the non-resident sample. A special tracking fieldwork team established contact by telephone and arranged to interview non-residents at their home in the surveillance area during a return visit, or in their place of residence if outside the surveillance area. Tracking teams travelled as far as Gauteng province but not outside the country. The same questionnaires were used for resident and non-resident participants, and a dried blood spot sample obtained via fingerprick was collected from all consenting participants for HIV testing in a central laboratory.

surveillance area. Tracking teams travelled as far as Gauteng province but not outside the country. The same questionnaires were used for resident and non-resident participants, and a dried blood spot sample obtained via fingerprick was collected from all consenting participants for HIV testing in a central laboratory. The Nelson Mandela Medical School Research Ethics Committee of the University of KwaZulu-Natal (Durban, South Africa) gave ethics approval for all surveillance data collection activities. Statistical analysis We identified four conceptually distinct indicators relating to participants' residential status at the time of each HIV survey and recent experiences of migration (ie, residential change): current residential status (ie, resident [usually sleeps and keeps their day-to-day belongings at the homestead] vs non-resident household members); recent mobility, based solely on the number of nights spent in the homestead during the past 6 months (at home every night, at home most nights [ie, less than ten nights away], or more than ten nights away); recent migration (migration to a homestead in the DSA at least once within the past 2 years); and migration type (in recent migrants, whether the migration was into or out of the surveillance area versus from another homestead in the DSA).

home most nights [ie, less than ten nights away], or more than ten nights away); recent migration (migration to a homestead in the DSA at least once within the past 2 years); and migration type (in recent migrants, whether the migration was into or out of the surveillance area versus from another homestead in the DSA). We examined the associations between the four indicators in each survey round using the corr function in Stata 13 to calculate the product-moment correlation coefficient. We used logistic regression models adjusted for differences in age composition and survey year to assess differences with respect to each migration indicator for HIV prevalence and sexual behaviour indicators. We used seven indicators of sexual behaviour: proportion of participants reporting they ever had sex, proportion of participants who were sexually active in the past year, proportion of participants reporting multiple partnerships in the past year, proportion of participants reporting a casual partnership, point-prevalence of concurrent sexual partnerships,22 proportion of participants reporting condom use at most recent sexual intercourse with the most recent regular partner, and proportion of participants reporting condom use at most recent sexual intercourse with the most recent casual partner. We analysed the data separately for men aged 17–54 years and women aged 17–49 years. We selected the minimum age limit of 17 years because sexual behaviour data were not available for 15–16 year olds in 2009 and 2011. We used Wald tests to assess statistical significance. Finally, in view of the weak associations between some migration indicators, we examined the risk of sexual HIV risk behaviours and HIV prevalence with respect to migration indicators in residents and non-residents separately.

for 15–16 year olds in 2009 and 2011. We used Wald tests to assess statistical significance. Finally, in view of the weak associations between some migration indicators, we examined the risk of sexual HIV risk behaviours and HIV prevalence with respect to migration indicators in residents and non-residents separately. We adjusted estimates for study non-participation and non-response with a previously described approach.21 Briefly, we made adjustments for survey non-participation by use of inverse-probability weights in strata defined by year, sex, age group, residence location (rural, periurban, urban, or non-resident). We used multiple imputation to adjust for missing responses.21,23 We analysed the data with the statistical software R 3.0.2. We implemented multiple imputation by creating customised imputation models for the imputation framework in the R package mi.24 Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

We analysed the data with the statistical software R 3.0.2. We implemented multiple imputation by creating customised imputation models for the imputation framework in the R package mi.24 Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Between Jan 1, 2005, and Dec 31, 2011, many more eligible residents were contacted than were non-residents in all survey years (table 1). In both groups, contact generally decreased in the latter years of the survey. Differences in survey participation between residents and non-residents who were contacted were small, with no consistent trend across the years. Our ability to track and contact people differed between male and female non-residents (appendix). An increasingly large group of non-residents had return patterns (other return patterns) that were unpredictable or did not fall into the circular migration patterns that were common in the past. In any year, roughly a fifth of residents had migrated at least once in the past 2 years (table 2).

on-residents (appendix). An increasingly large group of non-residents had return patterns (other return patterns) that were unpredictable or did not fall into the circular migration patterns that were common in the past. In any year, roughly a fifth of residents had migrated at least once in the past 2 years (table 2). Current residential status was strongly associated with the number of nights spent in the DSA household in the past 6 months, but poorly associated with the indicators of recent migration and migration type. For women, the association between recent migration and residential status ranged from 0·13 to 0·35 during the study period (adjusted for sampling and response weights); association between the indicator of recent inmigration or outmigration and residential status ranged from 0·13 to 0·34; and association between the number of nights spent in the DSA household in the past 6 months and residential status ranged from −0·80 to −0·93. The pattern for men was similar: association between recent migration and residential status ranged from 0·07 to 0·29; association between the indicator of recent inmigration or outmigration and residential status ranged from 0·09 to 0·30; and association between the number of nights spent in the DSA household in the past 6 months and residential status ranged from −0·82 to −0·95.

ration and residential status ranged from 0·07 to 0·29; association between the indicator of recent inmigration or outmigration and residential status ranged from 0·09 to 0·30; and association between the number of nights spent in the DSA household in the past 6 months and residential status ranged from −0·82 to −0·95. When we adjusted for age and survey year, sexual risk behaviours were generally substantially higher in non-residents than in residents (table 3). This difference existed in both men and women, with some exceptions. The proportion of women who had a casual partner in the past year was significantly higher (p=0·0080) in non-residents than in residents, whereas the proportion of men who had casual partners did not differ between residential status groups. Condom use with regular partners was significantly higher in non-resident women than in resident women (p=0·0061), but the pattern was reversed in men (p=0·0047). Results for groups defined by the number of nights slept outside the DSA household in the past 6 months (recent mobility) were very similar to those for groups defined by current residential status (table 3). In women, results for groups defined by recent migration history were similar to those for groups defined by current residential status; whereas in men, significant differences only existed in comparisons of residential status and recent mobility, and not in comparisons of recent migration history. For both men and women, sexual risk behaviour was not significantly different when we compared recent migrants who migrated internally (ie, within the DSA) with those who migrated externally (ie, into or out of the DSA). HIV prevalence was not significantly different for any of the indicators.

recent migration history. For both men and women, sexual risk behaviour was not significantly different when we compared recent migrants who migrated internally (ie, within the DSA) with those who migrated externally (ie, into or out of the DSA). HIV prevalence was not significantly different for any of the indicators. In residents, differences existed with respect to migration indicators. Residents who had recently migrated and those who had spent more than ten nights away from their DSA household in the past 6 months had increased sexual risk behaviours, although most of the odds ratio estimates were not significant (table 4). Furthermore, HIV prevalence was significantly higher in male and female residents who had recently migrated than in residents who had not migrated in the past 2 years, and our data suggest that HIV prevalence is higher in recently mobile (spent more than ten nights away from DSA household in past 6 months) resident women than in resident women who were not recently mobile. In non-residents, no comparisons were significant, and in some cases confidence intervals were very wide.

years, and our data suggest that HIV prevalence is higher in recently mobile (spent more than ten nights away from DSA household in past 6 months) resident women than in resident women who were not recently mobile. In non-residents, no comparisons were significant, and in some cases confidence intervals were very wide. Discussion Our results show that the previously identified increased levels of sexual risk behaviours in non-residents and those who spend few nights in their household in this study population18,25,26 have persisted in the post-ART rollout period, 2005–11. However, when we assessed indicators of recent migration (eg, migration over the past 2 years) irrespective of residential status, we identified no differences in sexual risk behaviour patterns for men, although we still detected differences in some sexual risk behaviours for women. These gender differences, combined with previous findings of differential patterns of migration by gender in this study population25 suggest that it might be necessary to design interventions for migrant men and women separately.25 Camlin and colleagues,25 noted that women were less likely to be non-resident members of rural households and more likely to migrate within the study area (internally) than were men.25

by gender in this study population25 suggest that it might be necessary to design interventions for migrant men and women separately.25 Camlin and colleagues,25 noted that women were less likely to be non-resident members of rural households and more likely to migrate within the study area (internally) than were men.25 We noted that HIV prevalence in the population was not strongly related to any of our migration indicators (irrespective of residential status) or to residential status (regardless of migration history) in the post-ART era. However, when we examined HIV prevalence with respect to both migration and residential status, we saw that male and female residents who had migrated recently had higher HIV prevalence than did residents who had not migrated recently. This finding might be because HIV status depends on past risk behaviour and residential status is not strongly associated with migration history. Alternatively, this finding might result from the return of former non-residents to the study area upon ill-health or loss of employment or to access the local ART programme, or from a complex interplay between migration and local factors.27,28 Therefore, some residents might have acquired HIV while they were non-residents. This fact is crucial to discussions of targeted intervention strategies based on the identification of transmission hot spots,29 since the findings of previous studies showed localised spatial clustering of new HIV infections.30,31 Incident infections are those that occur in individuals who previously tested negative for HIV, but test positive for HIV at a later date. Dependent on the period of time between HIV tests, especially in a highly mobile population, not all incident cases identified in a transmission hot spot would necessarily have been acquired in that geographical location.

in individuals who previously tested negative for HIV, but test positive for HIV at a later date. Dependent on the period of time between HIV tests, especially in a highly mobile population, not all incident cases identified in a transmission hot spot would necessarily have been acquired in that geographical location. Residential status and indicators of recent migration and type of recent migration were poorly related to each other. We detected a strong, negative correlation between residential status and recent mobility, but this association varied between rounds, which might partly be explained by diverse domestic arrangements: even among the residents of the study area, roughly half of residents were reported not to sleep at home every night. These findings emphasise the importance and difficulties of identification of relevant migration indicators for HIV risk in this context and support recent suggestions from other researchers that knowledge about migration and HIV risk is incomplete.2 Efforts to appropriately conceptualise and measure migration in populations with generalised epidemics are timely in view of global initiatives that focus on migrants as one of the key populations with increased susceptibility to HIV.32 We propose that the emphasis on, and targeting of, labour migrants, who are often men, in HIV studies has tended to prioritise some migration types and flows, such as international and long-distance labour migration with infrequent return visits. By contrast, migration and mobility for many other people are poorly represented by the way that data are collected and analysed. For example, the increasing size of the subgroup of non-residents with return patterns classified as other during the period of this study (appendix), which restricts understanding of a much broader set of processes linked to residential instability.33,34 Our investigation of sexual risk behaviours and HIV prevalence in population subgroups defined by both migration indicators and residential status is a first step in development of a classification system of migration and mobility that further helps to identify individuals at risk of HIV. Our work supports the value of methodological innovations in the design of household-based studies, especially the use of a definition of household membership that includes resident and non-resident members.

of a classification system of migration and mobility that further helps to identify individuals at risk of HIV. Our work supports the value of methodological innovations in the design of household-based studies, especially the use of a definition of household membership that includes resident and non-resident members. Such approaches could be extended further to collect information about the so-called floating population in an area (ie, household visitors, and household residents who are absent for short periods). Updated information about the effect of migration on migrants' access to, and engagement with, HIV testing and HIV care and treatment is also urgently needed.

her to collect information about the so-called floating population in an area (ie, household visitors, and household residents who are absent for short periods). Updated information about the effect of migration on migrants' access to, and engagement with, HIV testing and HIV care and treatment is also urgently needed. HIV prevalence estimates from open population cohort studies show changes in population factors (eg, migration, deaths, and ageing), survey coverage, and HIV incidence.3 A strength of our study was our use of multiple imputation methods to try to ensure that the apparent changes were not artifacts of changes in the composition of respondents over time. In the multiple imputation, we used a wide range of variables collected as part of the demographic surveillance. However, selection might still have contributed to the results. The fact that we were only able to examine HIV prevalence rather than incidence is also a limitation, as was the fact that we had limited power to fully explore the different migration indicators for non-residents. From 2012, non-residents were no longer included in the Africa Centre individual surveillance and thus sexual behaviour or HIV data for this group were not available beyond 2011. However, no substantial changes have occurred in migration patterns since 2011 that would alter our interpretation of this study.

-residents. From 2012, non-residents were no longer included in the Africa Centre individual surveillance and thus sexual behaviour or HIV data for this group were not available beyond 2011. However, no substantial changes have occurred in migration patterns since 2011 that would alter our interpretation of this study. Our findings suggest a need to reassess whether the types of migration investigated as risk factors in the early studies of the HIV epidemic are still the most important in the ART era. Quantitative and qualitative studies and surveys to describe HIV infection and sexual behaviour in migrants, especially those with circular migration patterns, should examine the social, behavioural, and situational risk factors present in the multiple environments to which migrants are exposed (eg, both origin and destination communities). Such research is needed to inform HIV treatment and prevention programmes, and local information about migrants and highly mobile individuals will be crucial to the success of any place-based strategies to direct resources to transmission hotspots. Additionally, in HIV prevention approaches, including treatment as prevention, in which people with HIV infection are increasingly initiated on treatment early in the course of their infection, thought should be given to the widespread migration of this population. Thus, development of research and intervention strategies should proceed with caution unless the local context of migration and mobility is understood and can be incorporated. Supplementary Material Supplementary appendix

Our findings suggest a need to reassess whether the types of migration investigated as risk factors in the early studies of the HIV epidemic are still the most important in the ART era. Quantitative and qualitative studies and surveys to describe HIV infection and sexual behaviour in migrants, especially those with circular migration patterns, should examine the social, behavioural, and situational risk factors present in the multiple environments to which migrants are exposed (eg, both origin and destination communities). Such research is needed to inform HIV treatment and prevention programmes, and local information about migrants and highly mobile individuals will be crucial to the success of any place-based strategies to direct resources to transmission hotspots. Additionally, in HIV prevention approaches, including treatment as prevention, in which people with HIV infection are increasingly initiated on treatment early in the course of their infection, thought should be given to the widespread migration of this population. Thus, development of research and intervention strategies should proceed with caution unless the local context of migration and mobility is understood and can be incorporated. Supplementary Material Supplementary appendix Acknowledgments NM is supported by a Wellcome Trust fellowship (grant number WT083495MA). JWE received scholarship support from the British Marshall Aid and Commemoration Commission, and funding from the Bill & Melinda Gates Foundation through a grant to the HIV Modelling Consortium. VH's involvement in the Africa Centre research activities is supported by the ESRC (ES/J021202/1) and the Wellcome Trust (#082384/Z/07/Z). The Africa Centre receives core funding from the Wellcome Trust, including for the surveillance (grant 082384/Z/07/Z). We thank the community for their continued support and participation in ACDIS, and Africa Centre staff.

earch activities is supported by the ESRC (ES/J021202/1) and the Wellcome Trust (#082384/Z/07/Z). The Africa Centre receives core funding from the Wellcome Trust, including for the surveillance (grant 082384/Z/07/Z). We thank the community for their continued support and participation in ACDIS, and Africa Centre staff. Contributors NM, VH, and JWE designed the study and contributed to the statistical analysis. JWE created the customised imputation models. NM took primary responsibility for writing the manuscript. NM, JWE, M-LN, and VH contributed to data analysis interpretation and writing and critiquing of the manuscript. Declaration of interests We declare no competing interests. Table 1 Survey participation by year

Contributors NM, VH, and JWE designed the study and contributed to the statistical analysis. JWE created the customised imputation models. NM took primary responsibility for writing the manuscript. NM, JWE, M-LN, and VH contributed to data analysis interpretation and writing and critiquing of the manuscript. Declaration of interests We declare no competing interests. Table 1 Survey participation by year Women (aged 17–49 years) Men (aged 17–54 years) 2005 2006 2007 2008 2009 2010 2011 2005 2006 2007 2008 2009 2010 2011 Eligible population* 21 129 21 234 22 011 22 289 21 849 22 726 22 267 20 399 20 586 21 344 21 656 21 382 22 100 21 779 Residents Eligible 13 663 13 606 14 132 14 242 13 967 14 853 13 862 10 863 10 789 11 169 11 250 11 091 11 811 11 293 Contacted† 10 784 (94%) 10 973 (98%) 10 722 (94%) 11 034 (96%)  10362 (95%) 10 194 (87%) 9515 (85%) 7965 (93%) 7782 (94%) 7302 (87%) 7968 (93%) 7406 (90%) 7101 (80%) 6858 (77%) Participated 7711 (60%) 7450 (56%) 6181 (47%) 5539 (41%) 5806 (44%) 5769 (45%) 5897 (50%) 4793 (47%) 4586 (45%) 3261 (34%) 3118 (30%) 2937 (29%) 3152 (33%) 3493 (40%) Complete sexual behaviour data‡ 7403 (96%) 6892 (93%) 5429 (88%) 4359 (79%) 4403 (76%) 3834 (66%) 3608 (61%) 4527 (94%) 4249 (93%) 2800 (86%) 2434 (78%) 1988 (68%) 2113 (67%) 2203 (63%) HIV test result§ 4422 (57%) 4104 (55%) 3638 (59%) 3570 (64%) 3371 (58%) 4299 (75%) 3920 (66%) 2763 (58%) 2424 (53%) 1851 (57%) 1974 (63%) 1802 (61%) 1987 (63%) 2013 (58%) All non-residents Eligible 7466 7628 7879 8047 7882 7873 8405 9536 9797 10175 10406 10291 10289 10486 Sampled 1323 (18%) 1195 (16%) 1113 (14%) 1619 (20%) 1376 (17%) 1465 (19%) 1532 (18%) 1376 (14%) 1250 (13%) 1237 (12%) 1747 (17%) 1382 (13%) 1533 (15%) 1568 (15%) Contacted¶ 715 (56%) 573 (57%) 544 (66%) 926 (74%) 747 (73%) 749 (53%) 643 (43%) 692 (52%) 539 (53%) 520 (59%) 853 (67%) 660 (66%) 777 (53%) 599 (39%) Participated 422 (57%) 324 (48%) 330 (45%) 424 (35%) 476 (47%) 412 (53%) 461 (70%) 355 (50%) 284 (43%) 325 (45%) 322 (28%) 406 (45%) 405 (50%) 438 (72%) Complete sexual behaviour data‡ 392 (93%) 304 (94%) 306 (93%) 388 (92%) 438 (92%) 364 (88%) 399 (87%) 329 (93%) 272 (96%) 300 (92%) 298 (93%) 359 (88%) 359 (89%) 374 (85%) HIV test result§ 299 (71%) 196 (60%) 245 (74%) 309 (73%) 362 (76%) 328 (80%) 325 (70%) 260 (73%) 192 (68%) 264 (81%) 255 (79%) 317 (78%) 346 (85%) 318 (73%) Data are n or n (%).

a‡ 392 (93%) 304 (94%) 306 (93%) 388 (92%) 438 (92%) 364 (88%) 399 (87%) 329 (93%) 272 (96%) 300 (92%) 298 (93%) 359 (88%) 359 (89%) 374 (85%) HIV test result§ 299 (71%) 196 (60%) 245 (74%) 309 (73%) 362 (76%) 328 (80%) 325 (70%) 260 (73%) 192 (68%) 264 (81%) 255 (79%) 317 (78%) 346 (85%) 318 (73%) Data are n or n (%). * Resident and non-resident eligibility lists for the HIV surveillance were generated from a snapshot of the ACDIS database produced at the end of the previous year. † By the time of the scheduled survey visit, on average 22% of eligible resident men and 18% of eligible resident women had died, outmigrated (thus making them no longer eligible for the resident sample), migrated to an unknown destination, or were unable to complete the survey for other reasons; all other individuals on the eligibility list were deemed contactable and contribute to the denominator for the contact rate. ‡ Number and percentage of participants who answered all sexual behaviour questions. § Number and percentage of participants who agreed to an HIV test. ¶ By the time of the scheduled survey visit, on average 15% of sampled non-resident men and 14% of sampled non-resident women were uncontactable, had died, or were unable to complete the survey for other reasons; all other individuals on the eligibility list were deemed contactable and contributed to the denominator for the contact rate. Table 2 Indicators of recent experience of migration

¶ By the time of the scheduled survey visit, on average 15% of sampled non-resident men and 14% of sampled non-resident women were uncontactable, had died, or were unable to complete the survey for other reasons; all other individuals on the eligibility list were deemed contactable and contributed to the denominator for the contact rate. Table 2 Indicators of recent experience of migration Women Men 2005 2006 2007 2008 2009 2010 2011 2005 2006 2007 2008 2009 2010 2011 Residents Migrated at least once in past 2 years 20% 18% 20% 20% 19% 24% 21% 19% 19% 20% 21% 20% 25% 22% External migration events* 56% 61% 61% 58% 60% 57% 64% 63% 71% 68% 65% 71% 68% 73% Spent <10 nights away from DSA household in past 6 months 95% 96% 95% 97% 98% 98% 98% 97% 96% 96% 98% 97% 98% 98% Non-residents Migrated at least once in past 2 years 44% 45% 31% 45% 47% 38% 41% 39% 30% 31% 26% 44% 32% 32% External migration events* 65% 74% 69% 90% 71% 85% 86% 69% 81% 95% 64% 87% 96% 96% Spent <10 nights away from DSA household in past 6 months 17% 5% 6% 4% 3% 4% 7% 15% 3% 1% 5% 4% 1% 3% All percentage estimates are adjusted for sampling and response weights. Absolute numbers are not shown because of the stratified sample design for non-residents. The overall population size for each indicator is shown in table 1. DSA=demographic surveillance area. * We defined external migration as the migration of an individual or household within the DSA to a household outside the DSA; the proportion reported represents external migration in individuals who migrated at least once in the past 2 years (recent migrants).

Women Men 2005 2006 2007 2008 2009 2010 2011 2005 2006 2007 2008 2009 2010 2011 Residents Migrated at least once in past 2 years 20% 18% 20% 20% 19% 24% 21% 19% 19% 20% 21% 20% 25% 22% External migration events* 56% 61% 61% 58% 60% 57% 64% 63% 71% 68% 65% 71% 68% 73% Spent <10 nights away from DSA household in past 6 months 95% 96% 95% 97% 98% 98% 98% 97% 96% 96% 98% 97% 98% 98% Non-residents Migrated at least once in past 2 years 44% 45% 31% 45% 47% 38% 41% 39% 30% 31% 26% 44% 32% 32% External migration events* 65% 74% 69% 90% 71% 85% 86% 69% 81% 95% 64% 87% 96% 96% Spent <10 nights away from DSA household in past 6 months 17% 5% 6% 4% 3% 4% 7% 15% 3% 1% 5% 4% 1% 3% All percentage estimates are adjusted for sampling and response weights. Absolute numbers are not shown because of the stratified sample design for non-residents. The overall population size for each indicator is shown in table 1. DSA=demographic surveillance area. * We defined external migration as the migration of an individual or household within the DSA to a household outside the DSA; the proportion reported represents external migration in individuals who migrated at least once in the past 2 years (recent migrants). Table 3 Comparisons of sexual behaviour outcomes and HIV prevalence by migration indicators

* We defined external migration as the migration of an individual or household within the DSA to a household outside the DSA; the proportion reported represents external migration in individuals who migrated at least once in the past 2 years (recent migrants). Table 3 Comparisons of sexual behaviour outcomes and HIV prevalence by migration indicators Women Men Non-residents vs residents (n=19 020) Spent >10 nights away vs spent <10 nights away from DSA household in past 6 months (n=19 020) Migrated at least once in past 2 years vs did not migrate (n=19 020) External migration vs internal migration (n=7301)* Non-residents vs residents (n=14 031) Spent >10 nights away vs <10 nights away from DSA household in past 6 months (n=14 031) Migrated at least once in past 2 years vs did not migrate (n=14 031) External migration vs internal migration (n=4681)* Sexually active in past 12 months 1·19 (0·96–1·49) 1·24 (1·00–1·54) 1·22 (1·03–1·45)† 0·94 (0·69–1·29) 1·53 (1·14–2·05)† 1·55 (1·15–2·08)† 1·12 (0·84–1·49) 1·64 (0·94–2·87) Multiple partners in past year 3·10 (1·88–5·12)† 2·59 (1·53–4·40)† 2·04 (1·07–3·89)† 0·90 (0·38–2·12) 1·66 (1·32–2·09)† 1·65 (1·31–2·09)† 0·97 (0·73–1·30) 1·65 (0·89–3·06) Point prevalence of concurrency 3·90 (1·64–9·25)† 3·34 (1·35–8·25)† 2·53 (0·85–7·49) 2·35 (0·38–14·61) 1·80 (1·34–2·43)† 1·82 (1·35–2·46)† 0·87 (0·59–1·28) 1·41 (0·64–3·10) Had a casual partner in the past year 1·69 (1·15–2·50)† 1·55 (1·04–2·32)† 1·55 (1·03–2·33)† 0·85 (0·40–1·79) 1·11 (0·86–1·44) 1·10 (0·85–1·43) 0·93 (0·69–1·24) 1·25 (0·63–2·51) Condom use at most recent sexual intercourse with casual partner 2·06 (0·94–4·52) 1·87 (0·86–4·06) 1·69 (0·82–3·50) 0·33 (0·00–38·56) 1·18 (0·70–1·98) 1·12 (0·69–1·82) 0·74 (0·45–1·23) 0·82 (0·36–1·85) Condom use at most recent sexual intercourse with regular partner 1·38 (1·10–1·73)† 1·42 (1·13–1·77)† 1·19 (0·99–1·42) 1·16 (0·86–1·57) 0·70 (0·55–0·90)† 0·70 (0·54–0·89)† 0·90 (0·66–1·22) 0·84 (0·48–1·47) HIV prevalence 0·77 (0·55–1·06) 0·85 (0·64–1·14) 1·07 (0·90–1·27) 0·96 (0·70–1·32) 0·77 (0·55–1·10) 0·79 (0·58–1·09) 0·92 (0·68–1·24) 1·02 (0·53–1·96) Data are adjusted odds ratio (95% CI), adjusted for linear trend over time and 5 year age groups. Women were aged 17–49 years and men were aged 17–54 years. DSA=demographic surveillance area.

·55–1·06) 0·85 (0·64–1·14) 1·07 (0·90–1·27) 0·96 (0·70–1·32) 0·77 (0·55–1·10) 0·79 (0·58–1·09) 0·92 (0·68–1·24) 1·02 (0·53–1·96) Data are adjusted odds ratio (95% CI), adjusted for linear trend over time and 5 year age groups. Women were aged 17–49 years and men were aged 17–54 years. DSA=demographic surveillance area. * In individuals who migrated at least once in the past 2 years (recent migrants). † p<0·05. Table 4 Comparisons of sexual behaviour outcomes and HIV prevalence by migration groups for residents and non-residents

·55–1·06) 0·85 (0·64–1·14) 1·07 (0·90–1·27) 0·96 (0·70–1·32) 0·77 (0·55–1·10) 0·79 (0·58–1·09) 0·92 (0·68–1·24) 1·02 (0·53–1·96) Data are adjusted odds ratio (95% CI), adjusted for linear trend over time and 5 year age groups. Women were aged 17–49 years and men were aged 17–54 years. DSA=demographic surveillance area. * In individuals who migrated at least once in the past 2 years (recent migrants). † p<0·05. Table 4 Comparisons of sexual behaviour outcomes and HIV prevalence by migration groups for residents and non-residents Women who migrated at least once in past 2 years vs women who did not migrate Women who spent >10 nights away vs women who spent <10 nights away from DSA household in past 6 months Men who migrated at least once in past 2 years vs men who did not migrate Men who spent >10 nights away vs men who spent <10 nights away from DSA household in past 6 months Residents (n=17 895) Non-residents (n=2203) Residents (n=17 895) Non-residents (n=2203) Residents (n=12 921) Non-residents (n=1875) Residents (n=12 921) Non-residents (n=1875) Sexually active in the past 12 months 1·17 (1·09–1·26)* 1·26 (0·83–1·92) 1·20 (1·02–1·41)* 1·44 (0·79–2·62) 1·20 (1·09–1·32)* 0·98 (0·51–1·89) 1·36 (1·08–1·72)* 1·16 (0·53–2·54) Multiple partners in the past year 1·27 (0·97–1·66) 1·70 (0·56–5·18) 1·80 (1·14–2·84)* 0·86 (0·23–3·14) 1·15 (1·03–1·30)* 0·76 (0·46–1·25) 1·21 (0·95–1·55) 1·21 (0·62–2·36) Point prevalence of concurrency 1·26 (0·75–2·11) 1·99 (0·28–14·10) 2·28 (1·08–4·83)* 1·92 (0·06–62·20) 1·09 (0·94–1·27) 0·68 (0·36–1·29) 1·06 (0·76–1·48) 1·85 (0·73–4·72) Had a casual partner in the past year 1·18 (0·99–1·40) 1·79 (0·82–3·91) 1·08 (0·76–1·53) 0·77 (0·27–2·22) 1·03 (0·92–1·15) 0·76 (0·45–1·30) 1·12 (0·41–3·06) 1·12 (0·41–3·06) Condom use at most recent sexual intercourse with casual partner 1·08 (0·78–1·50) 1·56 (0·05–46·41) 1·04 (0·52–2·10) 1·02 (0·06–18·12) 0·90 (0·72–1·12) 0·60 (0·17–2·14) 1·32 (0·79–2·18) 1·12 (0·12–10·24) Condom use at most recent sexual intercourse with regular partner 1·08 (1·00–1·16)* 1·15 (0·78–1·69) 1·12 (0·97–1·29) 1·75 (0·88–3·46) 0·94 (0·84–1·05) 0·88 (0·52–1·50) 0·77 (0·61–0·97)* 0·90 (0·35–2·35) HIV prevalence 1·18 (1·10–1·26)* 1·06 (0·73–1·55) 1·16 (0·97–1·39) 1·45 (0·73–2·88) 1·19 (1·07–1·33)* 0·67 (0·31–1·49) 0·98 (0·75–1·29) 0·90 (0·38–2·11) Data are adjusted odds ratio (95% CI), adjusted for linear trend over time and 5 year age group. Women were aged 17–49 years and men were aged 17–54 years.

(0·35–2·35) HIV prevalence 1·18 (1·10–1·26)* 1·06 (0·73–1·55) 1·16 (0·97–1·39) 1·45 (0·73–2·88) 1·19 (1·07–1·33)* 0·67 (0·31–1·49) 0·98 (0·75–1·29) 0·90 (0·38–2·11) Data are adjusted odds ratio (95% CI), adjusted for linear trend over time and 5 year age group. Women were aged 17–49 years and men were aged 17–54 years. The indicator of external migration vs internal migration could only be assessed for resident men and women who migrated in the past 2 years because numbers were too small for non-residents; no estimates were statistically significant and data are not shown. DSA=demographic surveillance area. * p<0·05.

Introduction HIV/AIDS is a leading cause of death and disease burden, especially in sub-Saharan Africa.1, 2, 3, 4, 5 Introduction of antiretroviral therapy (ART) in 1996 greatly reduced HIV-related mortality.6, 7 Creation of the Joint United Nations Programme on HIV/AIDS (UNAIDS) in 1996; the Global Fund to Fight AIDS, Tuberculosis and Malaria in 2002; and the US President's Emergency Plan for AIDS Relief (PEPFAR) in 2003, galvanised the mobilisation of resources to combat the HIV epidemic. In the past 15 years, the global community has provided US$109·8 billion of development assistance to curb the HIV/AIDS epidemic.8 As a result, HIV mortality has declined overall in low-income and middle-income countries since 2004.1

(PEPFAR) in 2003, galvanised the mobilisation of resources to combat the HIV epidemic. In the past 15 years, the global community has provided US$109·8 billion of development assistance to curb the HIV/AIDS epidemic.8 As a result, HIV mortality has declined overall in low-income and middle-income countries since 2004.1 The success of ART and prevention of mother-to-child transmission programmes led to ambitious calls to eliminate HIV as a public health threat. However, maintenance and scale-up of sufficiently funded AIDS efforts will be crucial to realise the goal of ending the AIDS epidemic as a public health threat by 2030.9 Achievement of these goals, including the UNAIDS 90-90-90 targets, which aim to have 90% of people living with HIV know their status, 90% of those detected treated with ART, and 90% of those receiving treatment achieving viral load suppression,10 requires a coordinated global scale-up of prevention programmes, pre-exposure prophylaxis (PrEP), and detection and treatment programmes.11 However, development assistance for health targeted for HIV has stagnated since 2010, and, in many low-income countries, national resources for health are scarce and expected to grow slowly.12, 13 Research in context Evidence before this study

The success of ART and prevention of mother-to-child transmission programmes led to ambitious calls to eliminate HIV as a public health threat. However, maintenance and scale-up of sufficiently funded AIDS efforts will be crucial to realise the goal of ending the AIDS epidemic as a public health threat by 2030.9 Achievement of these goals, including the UNAIDS 90-90-90 targets, which aim to have 90% of people living with HIV know their status, 90% of those detected treated with ART, and 90% of those receiving treatment achieving viral load suppression,10 requires a coordinated global scale-up of prevention programmes, pre-exposure prophylaxis (PrEP), and detection and treatment programmes.11 However, development assistance for health targeted for HIV has stagnated since 2010, and, in many low-income countries, national resources for health are scarce and expected to grow slowly.12, 13 Research in context Evidence before this study We searched PubMed between Aug 18, 2015, and April 3, 2016, for studies that comprehensively assessed the burden of HIV/AIDS globally. Our search terms included “HIV” and “global” and “mortality” or “incidence” or “prevalence”, and searches were restricted to articles published in English up to April 1, 2016. To our knowledge through the search, Global Burden of Disease (GBD) and UNAIDS are the only two sources that provide comparable evaluations of levels and trends of the HIV/AIDS epidemic at both the global and country level. UNAIDS has provided global estimates on HIV/AIDS since 1997, and has developed two epidemiological programs to estimate incidence, prevalence, and mortality: Estimation and Projection Package (EPP) and Spectrum. GBD 2013 used improved versions of Spectrum to generate comprehensive, comparable estimates of levels and trends of HIV/AIDS incidence, prevalence, and mortality across geographies. Studies from both organisations have shown rapid changes in the HIV/AIDS epidemic worldwide and that up-to-date epidemiological and demographic information is needed to more accurately assess the burden of HIV at both the country and global level.

nds of HIV/AIDS incidence, prevalence, and mortality across geographies. Studies from both organisations have shown rapid changes in the HIV/AIDS epidemic worldwide and that up-to-date epidemiological and demographic information is needed to more accurately assess the burden of HIV at both the country and global level. Added value of this study

nds of HIV/AIDS incidence, prevalence, and mortality across geographies. Studies from both organisations have shown rapid changes in the HIV/AIDS epidemic worldwide and that up-to-date epidemiological and demographic information is needed to more accurately assess the burden of HIV at both the country and global level. Added value of this study For GBD 2015, we systematically updated the key inputs to our HIV/AIDS estimation process, which includes prevalence from national surveys and antenatal care clinics, demographic input on fertility and migration, mortality on and off antiretroviral therapy (ART), and background HIV-free mortality; updates to these inputs were concluded in April, 2016; October, 2015; December, 2015; and April, 2016, respectively. We also improved the integration of EPP, Spectrum, and the GBD all-cause mortality estimation process to make them internally consistent. For countries with high-quality vital registration data, we developed a new method to improve the accuracy of and consistency among estimates of HIV/AIDS incidence, prevalence, and mortality leveraging the number of deaths recorded each year as caused by HIV/AIDS. This method also allowed us to use vital registration data to generate plausible incidence curves in countries that are not part of UNAIDS' results, and in subnational units where we previously only had national-level data. We developed an ensemble model to reconcile HIV mortality estimates from EPP and Spectrum and from those indicated in GBD's all-cause mortality estimation process. Remarkable progress has been made in curbing the HIV/AIDS epidemic worldwide; however, our findings emphasise the need for continued efforts from governments and international agencies in the next 15 years to end AIDS by 2030, in view of the low ART coverage and stagnation in decline of annual new infections in the past decade.

e progress has been made in curbing the HIV/AIDS epidemic worldwide; however, our findings emphasise the need for continued efforts from governments and international agencies in the next 15 years to end AIDS by 2030, in view of the low ART coverage and stagnation in decline of annual new infections in the past decade. Implications of all available evidence Improving on existing models of HIV/AIDS burden estimates, this study provides the most comprehensive and internally consistent assessments of the levels and trends of HIV/AIDS incidence, prevalence, and mortality worldwide so far. This timely report provides much needed assessment of achievement of Millennium Development Goal 6, and lays out the challenges facing the global community in progress towards the HIV goals enshrined in Sustainable Development Goal 3 and the 90-90-90 UNAIDS targets. The ambitious goals set forth by the global community, and the few resources available to combat HIV/AIDS, emphasise the importance of understanding and monitoring the trends of each country's HIV/AIDS epidemic. Measurement of disease burden according to geographic units enables comparison with other major conditions, showing where the epidemic remains a dominant cause of health loss and where the burden is still rising in spite of national and global efforts. Such measurement also enables direct comparison of different HIV/AIDS metrics, emphasising the specific needs of each geographic region and allowing for a more targeted response to the epidemic.

pidemic remains a dominant cause of health loss and where the burden is still rising in spite of national and global efforts. Such measurement also enables direct comparison of different HIV/AIDS metrics, emphasising the specific needs of each geographic region and allowing for a more targeted response to the epidemic. UNAIDS produces a biannual assessment of incidence of infections, prevalence of people living with HIV, and deaths from HIV/AIDS;14 the Global Burden of Disease Study (GBD) provides an alternative assessment of these rates. UNAIDS and GBD estimates have increasingly converged at the global level.2 Nevertheless, estimates differ substantially in several countries, particularly in middle-income and high-income countries, where GBD estimates are based on data from vital registration systems and UNAIDS estimates are based on prevalence in high-risk groups and estimates of the fraction of the population in these groups. This report from GBD 2015 provides a unique perspective on the national-level epidemiology of HIV/AIDS, which includes a comprehensive assessment of HIV/AIDS incidence, prevalence, and deaths. Methods Study design GBD is a systematic, scientific effort to quantify all-cause mortality; cause-specific mortality; and disease incidence, prevalence, and burden attributable to risk factors by age, sex, and geography over time. GBD 2015 includes 195 countries and territories and covers the time span from 1980 to 2015. Additional details of the GBD cause hierarchy, data inputs and processing, and estimation methods have been published elsewhere.15

idence, prevalence, and burden attributable to risk factors by age, sex, and geography over time. GBD 2015 includes 195 countries and territories and covers the time span from 1980 to 2015. Additional details of the GBD cause hierarchy, data inputs and processing, and estimation methods have been published elsewhere.15 In brief, the GBD estimation framework for HIV/AIDS used the general natural history epidemiological models, Estimation and Projection Package (EPP) and Spectrum, developed by UNAIDS for estimation of the burden of HIV/AIDS for their biannual report on the state of the HIV/AIDS epidemic at the global and country levels.1 EPP uses HIV seroprevalence estimates from surveys and antenatal care clinics to estimate incidence curves that are consistent with the input data of prevalence and other factors, including on-ART and off-ART mortality and demographic information within the given population. Spectrum, a compartmental model, is used to generate age-specific and sex-specific incidence, prevalence, and mortality by use of the incidence curves generated in EPP and other key inputs, including program data on ART and prevention of mother-to-child transmission and other key assumptions of on-ART and off-ART mortality and HIV-free background mortality. Details of methods and parameters in EPP and Spectrum have been described previously.16, 17, 18, 19, 20, 21, 22, 23

rated in EPP and other key inputs, including program data on ART and prevention of mother-to-child transmission and other key assumptions of on-ART and off-ART mortality and HIV-free background mortality. Details of methods and parameters in EPP and Spectrum have been described previously.16, 17, 18, 19, 20, 21, 22, 23 In GBD 2015, we improved on UNAIDS' estimation procedures in four ways. First, we used additional data, both from vital registration systems and population health surveys, to measure seroprevalence. Second, we used consistent estimates of HIV-free mortality in both EPP and Spectrum, and in the estimation of on-ART and off-ART mortality—key inputs to both EPP and Spectrum. These HIV-free mortality rates, generated in GBD's all-cause mortality estimation process, have linked our HIV/AIDS estimation process and the all-cause mortality estimation process. Third, we developed an adjustment process—cohort incidence bias adjustment—to ensure that incidence and prevalence estimates formulated with Spectrum are consistent with HIV mortality estimates based on vital registration systems when available. Fourth, through an expanded literature search, we updated rates of on-ART mortality (appendix pp 6–10), particularly for developed countries, in close collaboration with the Antiretroviral Therapy Cohort Collaboration.24

rum are consistent with HIV mortality estimates based on vital registration systems when available. Fourth, through an expanded literature search, we updated rates of on-ART mortality (appendix pp 6–10), particularly for developed countries, in close collaboration with the Antiretroviral Therapy Cohort Collaboration.24 Due to the interconnected nature of the HIV modelling process and the process of estimation of mortality and causes of death, data and codes for the GBD 2015 HIV estimation process will be made available along with all the GBD 2015 results, in compliance with the Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER) developed by the WHO.25 Mortality estimation The GBD estimation framework contains three sources for estimates of HIV-specific mortality: estimated HIV mortality from Spectrum; estimated excess HIV/AIDS mortality in our all-cause mortality estimation process;15 and space–time Gaussian process regression smoothed cause-specific HIV/AIDS mortality from vital registration systems that were adjusted for incompleteness and misclassification of causes of death. We used tailored estimation methods to produce final estimates of mortality depending on age groups, and the availability and quality of data for mortality of HIV/AIDS.

othed cause-specific HIV/AIDS mortality from vital registration systems that were adjusted for incompleteness and misclassification of causes of death. We used tailored estimation methods to produce final estimates of mortality depending on age groups, and the availability and quality of data for mortality of HIV/AIDS. We assigned countries and territories to one of four groups, depending on data availability and quality. Group 1 included countries with prevalence data from either household surveys or antenatal care clinics, most of which have generalised epidemics. Group 2A referred to countries with high-quality vital registration systems, which in GBD 2015 included countries with more than 25 years of vital registration data with more than 95% completeness. Group 2B referred to countries with vital registration systems that were not in group 2A. Group 2C included countries for which we had no data from a vital registration system. Briefly, for adults in group 1 countries, we applied an ensemble model to average HIV/AIDS mortality rates from Spectrum and those implied by the all-cause mortality estimation process. This approach was based on the fact that our estimation processes (appendix pp 12–15) in EPP, Spectrum, and all-cause mortality models were intrinsically linked by the same HIV-free mortality rates at the draw level for group 1 countries. Because EPP and Spectrum are largely based on prevalence estimates from surveys and antenatal care clinics and various assumptions, and all-cause mortality estimation process in group 1 countries are mostly based on sibling survival data with various biases that need to be corrected for, we used our ensemble model to give equal weights to HIV mortality estimates from the two processes.

rom surveys and antenatal care clinics and various assumptions, and all-cause mortality estimation process in group 1 countries are mostly based on sibling survival data with various biases that need to be corrected for, we used our ensemble model to give equal weights to HIV mortality estimates from the two processes. For adults in group 2A countries, we used the results from space–time Gaussian process regression for age-specific HIV mortality. For adults in group 2B and 2C countries, we used the HIV-specific mortality rates from Spectrum with cohort incidence bias adjustment. For children younger than 5 years in group 1, we applied the proportion of all HIV deaths estimated within Spectrum to the age-specific all-cause mortality estimates. For children of this age in group 2A countries, we used space–time Gaussian process regression estimates of HIV mortality. For children aged 5–14 years from countries in group 1, we used the average of the HIV-specific mortality rates from Spectrum and the implied HIV mortality from the all-cause mortality process. For group 2A countries, we used estimates of HIV mortality from space–time Gaussian process regression. For groups 2B and 2C, we used the estimates of HIV-specific mortality from Spectrum. Incidence and prevalence estimation We generated incidence and prevalence estimates with the recoded Spectrum model with updated assumptions of on-ART and off-ART mortality and other program data from the UNAIDS country files.

For adults in group 2A countries, we used the results from space–time Gaussian process regression for age-specific HIV mortality. For adults in group 2B and 2C countries, we used the HIV-specific mortality rates from Spectrum with cohort incidence bias adjustment. For children younger than 5 years in group 1, we applied the proportion of all HIV deaths estimated within Spectrum to the age-specific all-cause mortality estimates. For children of this age in group 2A countries, we used space–time Gaussian process regression estimates of HIV mortality. For children aged 5–14 years from countries in group 1, we used the average of the HIV-specific mortality rates from Spectrum and the implied HIV mortality from the all-cause mortality process. For group 2A countries, we used estimates of HIV mortality from space–time Gaussian process regression. For groups 2B and 2C, we used the estimates of HIV-specific mortality from Spectrum. Incidence and prevalence estimation We generated incidence and prevalence estimates with the recoded Spectrum model with updated assumptions of on-ART and off-ART mortality and other program data from the UNAIDS country files. HIV cause-specific deaths from vital registration systems and sample registration systems are among the most reliable sources for estimation of the burden of HIV/AIDS. We used our cohort incidence bias adjustment method to scale the sizes of each incidence cohort on the basis of the raw estimates of HIV mortality from Spectrum, using unadjusted incidence curves and those observed in the vital registration system with proper incompleteness and cause misclassification adjustments.15 For this procedure, we first ran space–time Gaussian process regression on age-specific HIV/AIDS mortality rates after correcting for garbage codes, HIV misclassification, and under-registration by use of formal demographic methods to generate complete time-series estimates by location, sex, year, and age. We then restructured Spectrum by addition of another compartment such that it could follow groups of people living with HIV/AIDS who were infected in a specific year and age group. We then ran the modified program to produce 1000 draws of incidence, prevalence, and mortality for each location and sex combination. From this step, we were able to obtain the proportion of each infection cohort dying in each year and age cell after infection. We then used these proportions to weigh the ratio of the numbers of deaths based on the age-specific mortality rates from vital registration and processed by space–time Gaussian process regression, and the population estimated with Spectrum, and those directly from Spectrum. This process greatly improves both the model fit on mortality data, closer to what the adjusted vital registration suggests, and the incidence mortality ratio. Further details of the method are described in appendix pp 13–15.

cess regression, and the population estimated with Spectrum, and those directly from Spectrum. This process greatly improves both the model fit on mortality data, closer to what the adjusted vital registration suggests, and the incidence mortality ratio. Further details of the method are described in appendix pp 13–15. Uncertainty analysis We systematically propagated uncertainty across EPP, Spectrum, and the all-cause mortality estimation processes. We used 1000 draws of the quantities of interest throughout all the steps in the estimation process. Some key inputs to the HIV estimation process did not include uncertainty: these were estimates of fertility and population, HIV programme metrics (including coverage of ART and prevention of mother-to-child transmission), and behavioural factors. We present results with 95% uncertainty intervals (UIs). Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Uncertainty analysis We systematically propagated uncertainty across EPP, Spectrum, and the all-cause mortality estimation processes. We used 1000 draws of the quantities of interest throughout all the steps in the estimation process. Some key inputs to the HIV estimation process did not include uncertainty: these were estimates of fertility and population, HIV programme metrics (including coverage of ART and prevention of mother-to-child transmission), and behavioural factors. We present results with 95% uncertainty intervals (UIs). Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Global HIV incidence peaked in 1997, at 3·3 million new infections (95% UI 3·1–3·4 million), decreasing by 4·8% (4·0–5·5) per year to 2005 (figure 1A). From 2005 to 2015, the global incidence remained relatively stable, at about 2·5–2·6 million per year (figure 1A). Prevalence of people living with HIV increased rapidly, from 2·4 million (95% UI 2·1–2·8 million) in 1985, to 28·0 million (27·1–29·3 million) in 2000 (figure 1B). From 2000 to 2015, the number of people living with HIV increased by 0·8% (95% UI 0·6–1·0) per year, reaching 38·8 million (37·6–40·4 million) in 2015 (figure 1B). Global mortality peaked in 2005, at 1·8 million (95% UI 1·7–1·9 million) and subsequently fell by 5·5% (95% UI 5·0–5·9) per year to 1·2 million (1·1–1·3 million) in 2015 (figure 1C). The proportion of people living with HIV and receiving ART increased rapidly for both sexes between 2005 and 2015, from 6·4% (95% UI 5·6–7·4) to 38·6% (37·2–40·0) of men, and from 3·3% (3·0–3·6) to 42·4% (41·0–43·7) of women (figure 1D).

5% (95% UI 5·0–5·9) per year to 1·2 million (1·1–1·3 million) in 2015 (figure 1C). The proportion of people living with HIV and receiving ART increased rapidly for both sexes between 2005 and 2015, from 6·4% (95% UI 5·6–7·4) to 38·6% (37·2–40·0) of men, and from 3·3% (3·0–3·6) to 42·4% (41·0–43·7) of women (figure 1D). In 2015, 1·8 million (95% UI 1·7–2·1 million) new HIV infections, 75·4% (71·7–78·5) of new cases, were in sub-Saharan Africa, with large proportions in western, southern, and eastern sub-Saharan Africa (figure 2A). Outside sub-Saharan Africa, south Asia accounted for 206 830 (171 790–249 700), or 8·5% (7·0–10·0), of new infections per year (figure 2A). Southeast Asia accounted for 4·7% (95% UI 2·8–8·1) of global infections in 2015, and east Asia accounted for 2·3% (1·7–3·1; figure 2A). Distributions of new infections by sex were broadly similar (appendix pp 26–38); and prevalence and mortality have also been greatest in sub-Saharan Africa (appendix pp 59, 60). HIV infection rates varied tremendously across countries in 2015 (figure 2B; see appendix pp 64, 65 for incidence for 1990 and 2005). The highest rates of infection were in southern Africa, with more than 1% of the population per year becoming infected in Botswana, Lesotho, and Swaziland (figure 2B). Within sub-Saharan Africa, rates in excess of 150 per 100 000 people occurred in a cluster of countries from Nigeria to Tanzania, with the notable exceptions of the Democratic Republic of the Congo (42·0 per 100 000; 95% UI 12·3–101·7) and Ethiopia (39·4 per 100 000; 19·7–62·5; figure 2B). The highest estimated incidence rates in Europe were recorded in Russia, and in Asia were recorded in Cambodia (figure 2B). In the Americas, only Belize, Guyana, and Haiti had rates of more than 50 per 100 000 people (figure 2B). Among the countries in the highest quintile of sociodemographic index (a composite indicator based on equally weighted estimates of lag-distributed income per capita, average years of education among populations over 15 years, and total fertility rate),15 countries with incidence rates of more than 15 infections per 100 000 people included Antigua and Barbuda, the Bahamas, Bermuda, Trinidad and Tobago, and Russia. Annualised rates of change show that although incidence substantially declined globally from 2005 to 2015, rates increased in 74 countries (table).

y rate),15 countries with incidence rates of more than 15 infections per 100 000 people included Antigua and Barbuda, the Bahamas, Bermuda, Trinidad and Tobago, and Russia. Annualised rates of change show that although incidence substantially declined globally from 2005 to 2015, rates increased in 74 countries (table). Due to improved access to treatment, prevalence compared with incidence was higher in countries with a high sociodemographic index (table). Six countries (Botswana, Lesotho, Namibia, Swaziland, South Africa, and Zimbabwe) had a HIV prevalence of more than 10% of the entire population. Nine countries in sub-Saharan Africa (Central African Republic, Cameroon, Equatorial Guinea, Kenya, Mozambique, Malawi, Tanzania, Uganda, and Zambia) had a prevalence of more than 2·5% of the entire population. Outside sub-Saharan Africa, a further 11 countries (the Bahamas, Belize, Bermuda, Dominican Republic, Guyana, Haiti, Cambodia, Portugal, Suriname, Trinidad and Tobago, and Saint Vincent and the Grenadines) had prevalence rates between 0·5% and 2·5% (appendix p 57). In the past 10 years, global scale-up of ART has been extraordinary, especially in eastern and southern sub-Saharan Africa (figure 3A). However, despite these increases, the proportion of people living with HIV and receiving ART is highly variable and remains at very low levels in many countries, particularly in the Middle East and North Africa, eastern Europe, central Asia, east Asia, and some countries in southeast Asia (figure 3B). We recorded coverage in excess of 40% in North America, western Europe, Australasia; the arc of countries in eastern South America, from Guyana to Argentina and Chile; and the corridor of countries from Uganda to South Africa (figure 3B).

urope, central Asia, east Asia, and some countries in southeast Asia (figure 3B). We recorded coverage in excess of 40% in North America, western Europe, Australasia; the arc of countries in eastern South America, from Guyana to Argentina and Chile; and the corridor of countries from Uganda to South Africa (figure 3B). HIV death rates and recent time trends vary greatly across countries (table). Deaths vary substantially by age, showing both the patterns of incidence by age, differential rates of progression by sex and age, and differential ART coverage (figure 4A). More women than men died in people aged 15–29 years; after age 35 years, there were more deaths in men (figure 4A). Deaths in people aged 50 years and older account for 10% (95% UI 3·8–11·8) of deaths in men and 7·6% (1·5–9·8) of deaths in women (figure 4A). We recorded substantial heterogeneity in HIV mortality among countries in 2015 (appendix p 63). Among HIV/AIDS deaths in 2015, 17·8% were caused by HIV and tuberculosis co-infection, down from 19·6% in 2005 (figure 4A). We compared HIV deaths with the number of people living with HIV to provide a simple estimate of the annual excess mortality rate (figure 4B). This ratio is a function of the timing of the epidemic and the access to and quality of ART and other care. Of note, this ratio was much lower in GBD high-income regions than in other GBD super-regions.

th the number of people living with HIV to provide a simple estimate of the annual excess mortality rate (figure 4B). This ratio is a function of the timing of the epidemic and the access to and quality of ART and other care. Of note, this ratio was much lower in GBD high-income regions than in other GBD super-regions. At the time of writing, the latest available assessment from UNAIDS was published in 2016, at the global and regional level only.26 UNAIDS country level estimates are from their 2014 update for years up to 2014.27 GBD 2015 estimates of prevalence are in accordance with the UNAIDS estimates. For 2015, estimations of the people living with HIV were 38·8 million (95% UI 37·6–40·4 million) in GBD 2015, and 36·7 million (34·0–39·8 million) in UNAIDS 2016. Comparisons of prevalence estimates at the country level in 2005 and 2014, show strong concordance between the two estimate series, with an average intraclass correlation coefficient of 0·997 (figure 5A shows prevalence from both sources for 2014). The highest relative differences in prevalence among sub-Saharan African countries in 2014 were in Senegal, Burundi, Democratic Republic of the Congo, Congo, Kenya, Sierra Leone, Nigeria, and South Africa, where GBD 2015 estimates are at least 10% higher than those from UNAIDS 2014 (figure 5A). UNAIDS tends to have much higher estimates of mortality at the peak of the HIV/AIDS epidemic around 2005, and lower estimates in 2014, than GBD 2015; we noted a much higher level of heterogeneity at the country level (figure 5B). For countries in sub-Saharan Africa, GBD estimates of mortality are higher than those from UNAIDS 2014 for 26 countries. Among these countries, GBD estimates are more than 10% higher than UNAIDS 2014 estimates in 22 countries. For South Africa, our estimated deaths are 17·2% higher than UNAIDS 2014 estimates. The highest differences are in Swaziland and Democratic Republic of the Congo, where GBD 2015 estimates are more than 80% higher than UNAIDS' (figure 5B).

BD estimates are more than 10% higher than UNAIDS 2014 estimates in 22 countries. For South Africa, our estimated deaths are 17·2% higher than UNAIDS 2014 estimates. The highest differences are in Swaziland and Democratic Republic of the Congo, where GBD 2015 estimates are more than 80% higher than UNAIDS' (figure 5B). For estimates of annual new infections, UNAIDS 2014 has slightly higher estimates for years before 2000. Although the estimates are similar between the two series for most of the 2000s, the series have differed substantially since 2007 at the global level. UNAIDS 2014 estimated a much faster rate of decline in annual new infections than did GBD 2015. Globally, GBD 2015 estimates about 2·5 million new infections in 2014, whereas UNAIDS estimates about 2 million for the same year. UNAIDS 2016 has slightly higher estimates of incidence than the 2014 publication, at 2·1 million for 2015. In terms of annualised rate of decline in new infections between 2005 and 2014, GBD 2015 estimates about a 0·4% decline per year, whereas the UNAIDS estimates from 2014 show a 3·3% decline per year. In only seven countries (Madagascar, Democratic Republic of the Congo, Burkina Faso, Guinea-Bissau, Chad, Rwanda, and The Gambia) in sub-Saharan Africa, annualised rates of decline in new infections are faster in GBD 2015 than in UNAIDS 2014. In Côte d'Ivoire, Burundi, Eritrea, Zimbabwe, Lesotho, Nigeria, Botswana, and Kenya, GBD 2015 estimates an increase in numbers of new infections, whereas UNAIDS 2014 predicts a decline. The biggest difference is in Kenya, where results from GBD 2015 show an increase in annual new infections from 60 000 in 2005, to 146 700 in 2014, whereas UNAIDS shows a decrease from 73 000 to 56 000 during the same period.

015 estimates an increase in numbers of new infections, whereas UNAIDS 2014 predicts a decline. The biggest difference is in Kenya, where results from GBD 2015 show an increase in annual new infections from 60 000 in 2005, to 146 700 in 2014, whereas UNAIDS shows a decrease from 73 000 to 56 000 during the same period. Discussion Remarkable progress has been made in curbing the HIV/AIDS epidemic worldwide. HIV incidence reached its peak in 1997, and HIV deaths have been declining since the mid-2000s. However, annual incidence has stayed relatively constant since 2005, after a period of faster decline between 1997 and 2005. The number of people living with HIV/AIDS has been steadily increasing, and reached 38·8 million in 2015. At the country level, disparate levels and trends of the epidemic persist. These updated estimates at the global level are similar to those published in the GBD 2013 iteration for deaths; however, our present estimates for incidence and prevalence are lower for 2013 than in GBD 2013.1

and reached 38·8 million in 2015. At the country level, disparate levels and trends of the epidemic persist. These updated estimates at the global level are similar to those published in the GBD 2013 iteration for deaths; however, our present estimates for incidence and prevalence are lower for 2013 than in GBD 2013.1 The unfolding global HIV pandemic has advanced through three phases during which HIV/AIDS mortality has increased from 4·73 per 100 000 in 1995, the 39th-ranked cause of death, to 16·18 per 100 000 in 2015, the 11th-ranked cause of death worldwide. In the initial phase (1981–97), global HIV incidence and the number of people living with HIV increased, followed by huge increases in deaths related to the disease. From 1998 to 2005, incidence declined by 25·4%; however, because of the lag between infection and mortality, the number of deaths caused by HIV increased. In the third phase, mass scale-up of prevention of mother-to-child transmission and ART, particularly in low-income sub-Saharan Africa, led to a phase of declining HIV mortality, a decade of stagnation in the decline of global incidence rates, and steadily rising prevalence. These global patterns mask well documented but extraordinary heterogeneity across countries. Epidemics leading to more than 2·5% of the population being infected have happened largely in eastern, southern, and central sub-Saharan Africa. Although death rates and incidence declined in the past decade in many of these countries, they are increasing in many others where prevalence has been lower until now, such as Indonesia and the Philippines. The scale-up of ART, a key driver of the trends, has led to 41% of people living with HIV receiving ART worldwide.7

hough death rates and incidence declined in the past decade in many of these countries, they are increasing in many others where prevalence has been lower until now, such as Indonesia and the Philippines. The scale-up of ART, a key driver of the trends, has led to 41% of people living with HIV receiving ART worldwide.7 The scale-up of interventions for HIV/AIDS represents one of the great successes of global health collective action. This scale-up, particularly in low-income countries, has been fuelled by the increase in development assistance for HIV from $1·3 billion in 2000, to $10·8 billion in 2015.12, 13 The need for HIV programmes, particularly ART programmes, continues to grow because of both the sustained high incidence of infections and the success of ART in extending the lifespan of people living with HIV. However, since 2010, development assistance for HIV has remained nearly constant.12 This absence of additional funding is by stark contrast with the $36 billion needed annually to achieve the UN goal to end AIDS by 2030, as estimated by Piot and colleagues.9 UNAIDS and other international development agencies hope that the growing need for funding will be partly solved by expanded health spending by low-income countries.28, 29 However, Dieleman and colleagues12, 13 suggest that, on the basis of trends in the past few years, health spending in low-income countries will grow only slightly in the next 25 years. How will the impending financing gap be addressed? In middle-income countries, increased commitments to funding health programmes from national budgets could fill the gap. But in low-income countries, where, as in eastern and some countries in southern sub-Saharan Africa, HIV rates are the highest, domestic resources will not be sufficient. Dieleman and colleagues30 projected that government health expenditure is going to increase from $30·8 billion (95% UI 29·9–31·8 billion) in 2015, to $53·1 billion (47·5–57·9 billion) in 2030, in southern sub-Saharan Africa.

thern sub-Saharan Africa, HIV rates are the highest, domestic resources will not be sufficient. Dieleman and colleagues30 projected that government health expenditure is going to increase from $30·8 billion (95% UI 29·9–31·8 billion) in 2015, to $53·1 billion (47·5–57·9 billion) in 2030, in southern sub-Saharan Africa. Meeting the needs of people living with HIV will require a combination of concentrating development assistance for HIV on these low-income countries, improving the efficiency of HIV programmes, increasing domestic financing, lowering the cost of treatment (including prices of ART drugs), and reducing future incidence through more concerted efforts. Development assistance efforts will also need to be scaled up if the free flow of low-cost generic drugs is hampered. Additionally, public and private sectors need to be incentivised to continue research and development of new and better prevention and treatment strategies to combat the epidemic in the long term. Special efforts need to be made in high-risk populations in both concentrated and generalised epidemic settings in view of the continued high rate of transmission among these subpopulations, including men who have sex with men and injecting drug users. However, on the basis of the epidemiological and financial trends, there is a major risk of a substantial shortfall in necessary funds to sustain life-saving ART programmes. The scarcity of adequate funds to provide ART for people living with HIV, together with the possibility of increasing drug resistance to existing ART treatments, will make achievement of the goal to end AIDS by 2030 extremely difficult.

f a substantial shortfall in necessary funds to sustain life-saving ART programmes. The scarcity of adequate funds to provide ART for people living with HIV, together with the possibility of increasing drug resistance to existing ART treatments, will make achievement of the goal to end AIDS by 2030 extremely difficult. WHO now recommends universal ART for all people with HIV.31, 32 In 2015, only 41% of people living with HIV were receiving ART; however, the 90-90-90 goals imply that 81% should be receiving ART and 73% will have viral suppression, which no country has yet achieved. Achievement of 81% ART coverage would require extension of ART coverage to at least 15·5 million additional people living with HIV by 2020, which implies an addition of 3·1 million per year between 2015 and 2020, while ensuring complete treatment adherence. Concerted efforts will be needed to scale up detection of new infections to meet the target of 90% of people knowing their status. The targeted expansion in ART coverage would play an important part in reducing the still high number of individuals dying from HIV. However, such expansion has enormous cost implications in an era when even maintenance of coverage in some low-income settings could be at risk in the presence of declining development assistance for health. Increased ART coverage might also play a part in reducing population transmission of HIV and therefore incidence.33, 34 The quality of ART embodied in the third 90 target of the UNAIDS strategy remains a major issue, as does the potential role of other care in extending survival. The simple comparison between HIV deaths and HIV prevalence shows that death rates in HIV-positive individuals are much lower in high-income countries than elsewhere. In fact, probability of death from HIV/AIDS while on ART in sub-Saharan Africa is on average 6·5 times higher than the probability in high-income countries among different age groups and time since start of ART treatment.

at death rates in HIV-positive individuals are much lower in high-income countries than elsewhere. In fact, probability of death from HIV/AIDS while on ART in sub-Saharan Africa is on average 6·5 times higher than the probability in high-income countries among different age groups and time since start of ART treatment. Calls for the end of AIDS have captured the imagination of the global health community.35 Largely as a result of the course of the HIV epidemic itself and spreading awareness of HIV among the general population, incidence declined between 1997, the year with peaked incidence, and 2005. However, our present estimates of HIV incidence, albeit driven mostly by prevalence data, suggest that incidence might not have declined much in the past decade. Incidence remained high, despite that much development assistance for HIV was spent on prevention programmes. Once the notable success of scale-up of prevention of mother-to-child transmission and reductions in transmission to children is accounted for, adult incidence remained even more resistant to change in the past 10 years. Effective strategies, such as male circumcision and PrEP, are available to reduce transmission even without changing sexual behaviour.36, 37, 38 Barrier methods for HIV prevention are also effective in reducing risk for transmission, as are some interventions targeting high-risk groups, such as needle exchanges.39, 40 Despite the existence of these approaches, incidence has not changed substantially. Although incidence has declined from 40·2 to 33·2 per 100 000 people at an annualised rate of decline of 1·9%, annual new infections have stayed relatively constant at about 2·5 million a year for the past decade. This finding could be explained by many factors, including that viral load suppression might be lower than the estimated 70% in low-income and middle-income countries, that ART coverage might be exaggerated in some countries, or that the rate of unsafe sex could be increasing in settings where the perceived risk of HIV has been reduced.

be explained by many factors, including that viral load suppression might be lower than the estimated 70% in low-income and middle-income countries, that ART coverage might be exaggerated in some countries, or that the rate of unsafe sex could be increasing in settings where the perceived risk of HIV has been reduced. Worldwide, Millennium Development Goal 6, to halt and reverse the spread of HIV and provide universal access to treatment for those who need it by 2015, has not been achieved. Between 2005 and 2015, 102 of 195 countries have experienced an increase in annual new infections. In sub-Saharan Africa, 15 of 46 countries managed to decrease annual infection during the same period. No countries had achieved the 81% target in 2015, and most developing countries have important gaps to fill by 2020.1 Sustainable Development Goal 3 aims to end HIV/AIDS by 2030. Achievement of such an ambitious goal will require great improvements in prevention efforts. The PEPFAR pivot, with its focus on high-transmission areas, might provide one such strategy,35 but the effectiveness of this approach is unproven, and the planning and evaluation for such programmes needs more granular data on the epidemic level and trends at the subnational level, which are still largely missing in most countries. To further reduce mortality from HIV/AIDS, another priority should be towards prevention, detection, and treatment of tuberculosis among people living with HIV as part of a strategy to reduce HIV disease progression and transmission, because tuberculosis, although largely preventable and treatable, is one of the most common opportunistic infections and the leading cause of death among HIV-infected individuals, as our study has shown.

tuberculosis among people living with HIV as part of a strategy to reduce HIV disease progression and transmission, because tuberculosis, although largely preventable and treatable, is one of the most common opportunistic infections and the leading cause of death among HIV-infected individuals, as our study has shown. Our assessment of HIV incidence and mortality in countries without vital registration data is driven by prevalence surveys and surveillance data on the prevalence of HIV among individuals attending antenatal care clinics. Estimation of incidence from prevalence is based on a set of assumptions about CD4 progression rates, off-ART and on-ART HIV death rates, and ART coverage. Such statistical back-estimation is inherently uncertain for recent time periods for which changes in incidence will not have changed prevalence as quickly. So far, efforts to develop incidence assays that can differentiate new from old infections have not been sufficiently robust or widely enough deployed to include in our or UNAIDS' estimation efforts, and the necessary sample sizes to track incidence could be challenging to obtain.41, 42, 43 Repeated measurements, such as the Swaziland HIV Incidence Measurement Survey (SHIMS), provide information about incidence in very few settings.44 Compared with the prevalence-based calculations, the SHIMS results show 2·4 infections per 100 person-years (95% UI 2·06–2·75) for 2011, which is consistent with GBD 2015 estimations of 2·15 infections per 100 people (1·91–2·46) for the same year. In view of the heightened focus on reducing HIV incidence as part of the end-of-AIDS vision, more efforts are needed to systematically supplement the approach of estimating incidence with prevalence data by use of additional information about case notifications with CD4 status, HIV viral load, and alternative assays as they emerge.

e heightened focus on reducing HIV incidence as part of the end-of-AIDS vision, more efforts are needed to systematically supplement the approach of estimating incidence with prevalence data by use of additional information about case notifications with CD4 status, HIV viral load, and alternative assays as they emerge. Our models, in addition to UNAIDS models for estimating HIV incidence and mortality, depend heavily on estimates of prevention of mother-to-child transmission and ART coverage. These numbers are developed by UNAIDS in consultation with national governments, the Global Fund, and PEPFAR.45 However, the underlying data at the facility or provider level are not available for inspection, critical appraisal, or validation. Evidence from countries where survey data for use of ART are available, such as Kenya, suggests that national assessments of numbers of people on ART collated by UNAIDS might be too high.46, 47, 48 If these findings were true in other countries, our estimates of ART coverage could likewise be exaggerated, as could our estimates of deaths from HIV. Data transparency for models used in global health estimation is rapidly increasing. GBD have adopted the GATHER guidelines developed by WHO and other partners.25 In the future, having input data on ART and prevention of mother-to-child transmission meet the GATHER guideline bar of transparency would be highly beneficial. Political sensitivities in some countries have restricted the transparency of UNAIDS on this issue; even the basic incidence and prevalence estimates generated by UNAIDS cannot be released for some countries such as India and Russia because of such issues. Fostering a culture of greater transparency and accountability for HIV prevention and treatment programmes will benefit everyone concerned with tackling HIV more effectively in the future.

and prevalence estimates generated by UNAIDS cannot be released for some countries such as India and Russia because of such issues. Fostering a culture of greater transparency and accountability for HIV prevention and treatment programmes will benefit everyone concerned with tackling HIV more effectively in the future. Subnational assessments, when available, suggest much spatial heterogeneity of HIV incidence, prevalence, and death.49, 50, 51 Use of more disaggregated assessments of the HIV epidemic will hopefully improve the quality of the results and the relevance to HIV prevention and treatment programmes. Disaggregated assessments of prevalence derived from survey data and surveillance data from antenatal care clinics are feasible. Progress will be needed on the availability of data for prevention of mother-to-child transmission and ART at the local level. Perhaps even more challenging is the need for estimates of the various demographic inputs required for the modelling efforts, including migration, fertility, and HIV-free mortality. The push toward district-level or even more fine-grained estimation is one of the most promising directions for improved estimation of the epidemic overall.

challenging is the need for estimates of the various demographic inputs required for the modelling efforts, including migration, fertility, and HIV-free mortality. The push toward district-level or even more fine-grained estimation is one of the most promising directions for improved estimation of the epidemic overall. Substantial differences between men and women remain in many aspects of the HIV epidemic. Our analysis shows that the age pattern of HIV/AIDS mortality is younger in women than in men. This finding is largely thought to result from age-disparate relationships in which men tend to have sex with women younger than them.52, 53 Furthermore, more women use ART, as shown by the roughly 15·4% higher ART coverage for women than men in 2015. We also recorded a high level of heterogeneity at the country level, with female ART coverage 50% higher than male coverage in countries such as Gabon, The Gambia, Nigeria, and Sierra Leone; at the same time, in India, Lithuania, and Maldives, male ART coverage was 50% higher.

r ART coverage for women than men in 2015. We also recorded a high level of heterogeneity at the country level, with female ART coverage 50% higher than male coverage in countries such as Gabon, The Gambia, Nigeria, and Sierra Leone; at the same time, in India, Lithuania, and Maldives, male ART coverage was 50% higher. The GBD 2015 estimates of prevalence are in line with those from UNAIDS 2014. The high concordance of country-level prevalence estimates between the two series is unsurprising given that they are both based on similar input datasets for prevalence. Much of the difference between the two estimates is a result of different assumptions of on-ART and off-ART mortality, background HIV-free mortality rate, initial CD4 distribution, and CD4 progression ratios. Further studies and close collaboration between UNAIDS and GBD are needed to fully understand the relative contribution of each of the aforementioned factors to the difference in estimates of incidence, prevalence, and mortality at the country level.

mortality rate, initial CD4 distribution, and CD4 progression ratios. Further studies and close collaboration between UNAIDS and GBD are needed to fully understand the relative contribution of each of the aforementioned factors to the difference in estimates of incidence, prevalence, and mortality at the country level. This analysis of data for HIV has several further limitations. First, we have used data from the Antiretroviral Therapy Cohort Collaboration to improve the assessment of death rates of patients on ART in high-income countries. However, we have not been able to incorporate data for the variation in quality of ART programmes such as viral load suppression by country. This information is not widely available. Second, our novel cohort incidence bias adjustment method leverages cause-of-death data to correct past incidence for each cohort. Although our testing shows that the method works fairly well, the information content in the approach for adjustments in the most recent time period is much more restricted than for earlier periods. Estimates of incidence with this approach in the most recent time periods might be biased. Third, we have attempted to propagate multiple sources of uncertainty into our final estimates. Because we did not include uncertainty caused by variation in the quality of ART and might underestimate uncertainty in ART coverage, our final uncertainty intervals could be too narrow. At the global level, our uncertainty levels might also be too narrow because we assume uncertainty in each country is independent of other countries. Fourth, we have not used the surveillance data for new cases in our analysis. Integration of such information in the future will probably increase the accuracy of incidence and prevalence estimates. Fifth, prevalence and programme data are still sparse in most countries. Prevalence estimates are largely determined by adjusted antenatal clinic data and national surveys. To depict the epidemic in populations accurately, rigorous data collection is needed. For example, ART coverage data should be directly collected in surveys through viral load testing, such as in the Lesotho 2014 Demographic and Health Survey,54 and questions about ART. Sixth, our study focuses on deaths with HIV/AIDS as the underlying cause of death and does not account for excess mortality from other non-communicable causes of deaths among people living with HIV.

surveys through viral load testing, such as in the Lesotho 2014 Demographic and Health Survey,54 and questions about ART. Sixth, our study focuses on deaths with HIV/AIDS as the underlying cause of death and does not account for excess mortality from other non-communicable causes of deaths among people living with HIV. Seventh, input data tend to be sparse for the most recent time period, and our models might have not captured the recent progress and lack thereof in some countries. Eighth, our models have not directly used other important variables, such as prevalence of sexually transmitted infections or rates of ART adherence, ART treatment failure, and HIV testing, as used by Optima (appendix p 56).55 These variables should be included in future updates to improve the precision of our estimates. Finally, although we integrated HIV cause-specific mortality data in our modelling framework for a large group of countries with vital registration data, inclusion of additional data sources such as HIV surveillance and case report in our analytical framework could improve the accuracy of our incidence estimates.

ates. Finally, although we integrated HIV cause-specific mortality data in our modelling framework for a large group of countries with vital registration data, inclusion of additional data sources such as HIV surveillance and case report in our analytical framework could improve the accuracy of our incidence estimates. Enormous progress has been made in reducing HIV deaths, especially in low-income countries, through the expansion of prevention of mother-to-child transmission and ART programmes funded largely through development assistance for HIV. However, achievement of the UNAIDS 90-90-90 targets will require major changes in how programmes are delivered and financed. Global efforts have had less impact on the incidence of new infections than on HIV mortality. Ending the AIDS epidemic by 2030 will require a dramatic change in how HIV prevention is pursued. Correspondence to: Dr Haidong Wang, Institute for Health Metrics and Evaluation, University of Washington, Seattle, WA 98121, USA haidong@uw.edu This online publication has been corrected. The corrected version first appeared at thelancet.com/hiv on August 22, 2016 Supplementary Material Supplementary appendix

Enormous progress has been made in reducing HIV deaths, especially in low-income countries, through the expansion of prevention of mother-to-child transmission and ART programmes funded largely through development assistance for HIV. However, achievement of the UNAIDS 90-90-90 targets will require major changes in how programmes are delivered and financed. Global efforts have had less impact on the incidence of new infections than on HIV mortality. Ending the AIDS epidemic by 2030 will require a dramatic change in how HIV prevention is pursued. Correspondence to: Dr Haidong Wang, Institute for Health Metrics and Evaluation, University of Washington, Seattle, WA 98121, USA haidong@uw.edu This online publication has been corrected. The corrected version first appeared at thelancet.com/hiv on August 22, 2016 Supplementary Material Supplementary appendix Acknowledgments We thank the countless individuals who have contributed to the Global Burden of Disease (GBD) Study 2015 in various capacities. We specifically thank Jeffrey Eaton and John Stover. HW and CJLM received funding for this study from the Bill & Melinda Gates Foundation; the National Institute of Mental Health, National Institutes of Health (NIH; R01MH110163); and the National Institute on Aging, NIH (P30AG047845). LJAR acknowledges the support of Qatar National Research Fund (NPRP 04-924-3-251) who provided the main funding for generating the data provided to the GBD–Institute for Health Metrics and Evaluation effort. BPAQ acknowledges institutional support from PRONABEC (National Program of Scholarship and Educational Loan), provided by the Peruvian government. DB is supported by the Bill & Melinda Gates Foundation (grant number OPP1068048). JDN was supported in his contribution to this work by a Fellowship from Fundação para a Ciência e a Tecnologia, Portugal (SFRH/BPD/92934/2013). KD is supported by a Wellcome Trust Fellowship in Public Health and Tropical Medicine (grant number 099876). TF received financial support from the Swiss National Science Foundation (SNSF; project number P300P3-154634). AG acknowledges funding from Sistema Nacional de Investigadores de Panamá-SNI. PJ is supported by Wellcome Trust–DBT India Alliance Clinical and Public Health Intermediate Fellowship. MK receives research support from the Academy of Finland, the Swedish Research Council, Alzheimerfonden, Alzheimer's Research & Prevention Foundation, Center for Innovative Medicine (CIMED) at Karolinska Institutet South Campus, AXA Research Fund, Wallenberg Clinical Scholars Award from the Knut och Alice Wallenbergs Foundation, and the Sheika Salama Bint Hamdan Al Nahyan Foundation. AK's work was supported by the Miguel Servet contract financed by the CP13/00150 and PI15/00862 projects, integrated into the National R&D&I and funded by the ISCIII (General Branch Evaluation and Promotion of Health Research), and the European Regional Development Fund (ERDF-FEDER).

a Bint Hamdan Al Nahyan Foundation. AK's work was supported by the Miguel Servet contract financed by the CP13/00150 and PI15/00862 projects, integrated into the National R&D&I and funded by the ISCIII (General Branch Evaluation and Promotion of Health Research), and the European Regional Development Fund (ERDF-FEDER). SML is funded by a National Institute for Health Research (NIHR) Clinician Scientist Fellowship (grant number NIHR/CS/010/014). HJL reports grants from the NIHR, EU Innovative Medicines Initiative, Centre for Strategic & International Studies, and WHO. WM is Program analyst, Population and Development, in the Peru Country Office of the United Nations Population Fund, which does not necessarily endorse this study. For UOM, funding from the German National Cohort Consortium (O1ER1511D) is gratefully acknowledged. KR reports grants from NIHR Oxford Biomedical Research Centre, NIHR Career Development Fellowship, and Oxford Martin School during the conduct of the study. GR acknowledges that work related to this paper has been done on the behalf of the GBD Genitourinary Disease Expert Group supported by the International Society of Nephrology (ISN). ISS reports grants from FAPESP (Brazilian public agency). RSS receives institutional support from Universidad de Ciencias Aplicadas y Ambientales, UDCA, Bogotá Colombia. SS receives postdoctoral funding from the Fonds de la recherche en santé du Québec (FRSQ), including its renewal. RTS was supported in part by grant number PROMETEOII/2015/021 from Generalitat Valenciana and the national grant PI14/00894 from ISCIII-FEDER. PY acknowledges support from Strategic Public Policy Research (HKU7003-SPPR-12).

octoral funding from the Fonds de la recherche en santé du Québec (FRSQ), including its renewal. RTS was supported in part by grant number PROMETEOII/2015/021 from Generalitat Valenciana and the national grant PI14/00894 from ISCIII-FEDER. PY acknowledges support from Strategic Public Policy Research (HKU7003-SPPR-12). GBD 2015 HIV Collaborators Haidong Wang*, Tim M Wolock, Austin Carter, Grant Nguyen, Hmwe Hmwe Kyu, Emmanuela Gakidou, Simon I Hay, Edward J Mills, Adam Trickey, William Msemburi, Matthew M Coates, Meghan D Mooney, Maya S Fraser, Amber Sligar, Joshua Salomon, Heidi J Larson, Joseph Friedman, Amanuel Alemu Abajobir†, Kalkidan Hassen Abate†, Kaja M Abbas†, Mohamed Magdy Abd El Razek†, Foad Abd-Allah†, Abdishakur M Abdulle†, Semaw Ferede Abera†, Ibrahim Abubakar†, Laith J Abu-Raddad†, Niveen M E Abu-Rmeileh†, Gebre Yitayih Abyu†, Akindele Olupelumi Adebiyi†, Isaac Akinkunmi Adedeji†, Ademola Lukman Adelekan†, Koranteng Adofo†, Arsène Kouablan Adou†, Oluremi N Ajala†, Tomi F Akinyemiju†, Nadia Akseer†, Faris Hasan Al Lami†, Ziyad Al-Aly†, Khurshid Alam†, Noore K M Alam†, Deena Alasfoor†, Saleh Fahed S Aldhahri†, Robert William Aldridge†, Miguel Angel Alegretti†, Alicia V Aleman†, Zewdie Aderaw Alemu†, Rafael Alfonso-Cristancho†, Raghib Ali†, Ala'a Alkerwi†, François Alla†, Rajaa Mohammad Salem Al-Raddadi†, Ubai Alsharif†, Elena Alvarez†, Nelson Alvis-Guzman†, Azmeraw T Amare†, Alemayehu Amberbir†, Adeladza Kofi Amegah†, Walid Ammar†, Stephen Marc Amrock†, Carl Abelardo T Antonio†, Palwasha Anwari†, Johan Ärnlöv†, Al Artaman†, Hamid Asayesh†, Rana Jawad Asghar†, Reza Assadi†, Suleman Atique†, Lydia S Atkins†, Euripide Frinel G Arthur Avokpaho†, Ashish Awasthi†, Beatriz Paulina Ayala Quintanilla†, Umar Bacha†, Alaa Badawi†, Aleksandra Barac†, Till Bärnighausen†, Arindam Basu†, Tigist Assefa Bayou†, Yibeltal Tebekaw Bayou†, Shahrzad Bazargan-Hejazi†, Justin Beardsley†, Neeraj Bedi†, Derrick A Bennett†, Isabela M Bensenor†, Balem Demtsu Betsu†, Addisu Shunu Beyene†, Eesh Bhatia†, Zulfiqar A Bhutta†, Sibhatu Biadgilign†, Boris Bikbov†, Sait Mentes Birlik†, Donal Bisanzio†, Michael Brainin†, Alexandra Brazinova†, Nicholas J K Breitborde†, Alexandria Brown†, Michael Burch†, Zahid A Butt†, Julio Cesar Campuzano†, Rosario Cárdenas†, Juan Jesus Carrero†, Carlos A Castañeda-Orjuela†, Jacqueline Castillo Rivas†, Ferrán Catalá-López†, Hsing-Yi Chang†, Jung-chen Chang†, Laxmikant Chavan†, Wanqing Chen†, Peggy Pei-Chia Chiang†, Mirriam Chibalabala†, Vesper Hichilombwe Chisumpa†, J

Michael Burch†, Zahid A Butt†, Julio Cesar Campuzano†, Rosario Cárdenas†, Juan Jesus Carrero†, Carlos A Castañeda-Orjuela†, Jacqueline Castillo Rivas†, Ferrán Catalá-López†, Hsing-Yi Chang†, Jung-chen Chang†, Laxmikant Chavan†, Wanqing Chen†, Peggy Pei-Chia Chiang†, Mirriam Chibalabala†, Vesper Hichilombwe Chisumpa†, J ee-Young Jasmine Choi†, Devasahayam Jesudas Christopher†, Liliana G Ciobanu†, Cyrus Cooper†, Tukur Dahiru†, Solomon Abrha Damtew†, Lalit Dandona†, Rakhi Dandona†, José das Neves†, Pieter de Jager†, Diego De Leo†, Louisa Degenhardt†, Robert P Dellavalle†, Kebede Deribe†, Amare Deribew†, Don C Des Jarlais†, Samath D Dharmaratne†, Eric L Ding†, Pratik Pinal Doshi†, Kerrie E Doyle†, Tim R Driscoll†, Manisha Dubey†, Yousef Mohamed Elshrek†, Iqbal Elyazar†, Aman Yesuf Endries†, Sergey Petrovich Ermakov†, Babak Eshrati†, Alireza Esteghamati†, Imad D A Faghmous†, Carla Sofia e Sa Farinha†, Andre Faro†, Maryam S Farvid†, Farshad Farzadfar†, Seyed-Mohammad Fereshtehnejad†, Joao C Fernandes†, Florian Fischer†, Joseph Robert Anderson Fitchett†, Nataliya Foigt†, Nancy Fullman†, Thomas Fürst†, Fortuné Gbètoho Gankpé†, Teshome Gebre†, Amanuel Tesfay Gebremedhin†, Alemseged Aregay Gebru†, Johanna M Geleijnse†, Bradford D Gessner†, Peter W Gething†, Tsegaye Tewelde Ghiwot†, Maurice Giroud†, Melkamu Dedefo Gishu†, Elizabeth Glaser†, Shifalika Goenka†, Amador Goodridge†, Sameer Vali Gopalani†, Atsushi Goto†, Harish Chander Gugnani†, Mark D C Guimaraes†, Rahul Gupta†, Rajeev Gupta†, Vipin Gupta†, Juanita Haagsma†, Nima Hafezi-Nejad†, Holly Hagan†, Gessessew Bugssa Hailu†, Randah Ribhi Hamadeh†, Samer Hamidi†, Mouhanad Hammami†, Graeme J Hankey†, Yuantao Hao†, Hilda L Harb†, Sivadasanpillai Harikrishnan†, Josep Maria Haro†, Kimani M Harun†, Rasmus Havmoeller†, Mohammad T Hedayati†, Ileana Beatriz Heredia-Pi†, Hans W Hoek†, Masako Horino†, Nobuyuki Horita†, H Dean Hosgood†, Damian G Hoy†, Mohamed Hsairi†, Guoqing Hu†, Hsiang Huang†, John J Huang†, Kim Moesgaard Iburg†, Bulat T Idrisov†, Kaire Innos†, Veena J Iyer†, Kathryn H Jacobsen†, Nader Jahanmehr†, Mihajlo B Jakovljevic†, Mehdi Javanbakht†, Achala Upendra Jayatilleke†, Panniyammakal Jeemon†, Vivekanand Jha†, Guohong Jiang†, Ying Jiang†, Tariku Jibat†, Jost B Jonas†, Zubair Kabir†, Ritul Kamal†, Haidong Kan†, André Karch†, Corine Kakizi Karema†, Dimitris Karletsos†, Amir Kasaeian†, Anil Kaul†, Norito Kawakami†, Jeanne Françoise Kayibanda†, Peter Njenga Keiyoro†, Andrew Haddon Kemp†, Andre Pascal Kengne†, Chandrasekharan Nair Kesa

g Jiang†, Ying Jiang†, Tariku Jibat†, Jost B Jonas†, Zubair Kabir†, Ritul Kamal†, Haidong Kan†, André Karch†, Corine Kakizi Karema†, Dimitris Karletsos†, Amir Kasaeian†, Anil Kaul†, Norito Kawakami†, Jeanne Françoise Kayibanda†, Peter Njenga Keiyoro†, Andrew Haddon Kemp†, Andre Pascal Kengne†, Chandrasekharan Nair Kesa vachandran†, Yousef Saleh Khader†, Ibrahim Khalil†, Abdur Rahman Khan†, Ejaz Ahmad Khan†, Young-Ho Khang†, Jagdish Khubchandani†, Yun Jin Kim†, Yohannes Kinfu†, Miia Kivipelto†, Yoshihiro Kokubo†, Soewarta Kosen†, Parvaiz A Koul†, Ai Koyanagi†, Barthelemy Kuate Defo†, Burcu Kucuk Bicer†, Veena S Kulkarni†, G Anil Kumar†, Dharmesh Kumar Lal†, Hilton Lam†, Jennifer O Lam†, Sinead M Langan†, Van C Lansingh†, Anders Larsson†, James Leigh†, Ricky Leung†, Yongmei Li†, Stephen S Lim†, Steven E Lipshultz†, Shiwei Liu†, Belinda K Lloyd†, Giancarlo Logroscino†, Paulo A Lotufo†, Raimundas Lunevicius†, Hassan Magdy Abd El Razek†, Mahdi Mahdavi†, P A Mahesh†, Marek Majdan†, Azeem Majeed†, Carla Makhlouf†, Reza Malekzadeh†, Chabila C Mapoma†, Wagner Marcenes†, Jose Martinez-Raga†, Melvin Barrientos Marzan†, Felix Masiye†, Amanda J Mason-Jones†, Bongani M Mayosi†, Martin McKee†, Peter A Meaney†, Man Mohan Mehndiratta†, Alemayehu B Mekonnen†, Yohannes Adama Melaku†, Peter Memiah†, Ziad A Memish†, Walter Mendoza†, Atte Meretoja†, Tuomo J Meretoja†, Francis Apolinary Mhimbira†, Ted R Miller†, Joseph Mikesell†, Mojde Mirarefin†, Karzan Abdulmuhsin Mohammad†, Shafiu Mohammed†, Ali H Mokdad†, Lorenzo Monasta†, Maziar Moradi-Lakeh†, Rintaro Mori†, Ulrich O Mueller†, Brighton Murimira†, Gudlavalleti Venkata Satyanarayana Murthy†, Aliya Naheed†, Luigi Naldi†, Vinay Nangia†, Denis Nash†, Haseeb Nawaz†, Chakib Nejjari†, Frida Namnyak Ngalesoni†, Jean de Dieu Ngirabega†, Quyen Le Nguyen†, Muhammad Imran Nisar†, Ole F Norheim†, Rosana E Norman†, Luke Nyakarahuka†, Felix Akpojene Ogbo†, In-Hwan Oh†, Foluke Adetola Ojelabi†, Bolajoko Olubukunola Olusanya†, Jacob Olusegun Olusanya†, John Nelson Opio†, Eyal Oren†, Erika Ota†, Hye-Youn Park†, Jae-Hyun Park†, Snehal T Patil†, Scott B Patten†, Vinod K Paul†, Katherine Pearson†, Emmanuel Kwame Peprah†, David M Pereira†, Norberto Perico†, Konrad Pesudovs†, Max Petzold†, Michael Robert Phillips†, Julian David Pillay†, Dietrich Plass†, Suzanne Polinder†, Farshad Pourmalek†, David M Prokop†, Mostafa Qorbani†, Anwar Rafay†, Kazem Rahimi†, Vafa Rahimi-Movaghar†, Mahfuzar Rahman†, Mohammad Hifz Ur Rahman†, Sajjad Ur Rahman†, Rajesh Kumar Rai†, Sasa Rajsi

co†, Konrad Pesudovs†, Max Petzold†, Michael Robert Phillips†, Julian David Pillay†, Dietrich Plass†, Suzanne Polinder†, Farshad Pourmalek†, David M Prokop†, Mostafa Qorbani†, Anwar Rafay†, Kazem Rahimi†, Vafa Rahimi-Movaghar†, Mahfuzar Rahman†, Mohammad Hifz Ur Rahman†, Sajjad Ur Rahman†, Rajesh Kumar Rai†, Sasa Rajsi c†, Usha Ram†, Saleem M Rana†, Paturi Vishnupriya Rao†, Giuseppe Remuzzi†, David Rojas-Rueda†, Luca Ronfani†, Gholamreza Roshandel†, Ambuj Roy†, George Mugambage Ruhago†, Mohammad Yahya Saeedi†, Rajesh Sagar†, Muhammad Muhammad Saleh†, Juan R Sanabria†, Itamar S Santos†, Rodrigo Sarmiento-Suarez†, Benn Sartorius†, Monika Sawhney†, Aletta E Schutte†, David C Schwebel†, Soraya Seedat†, Sadaf G Sepanlou†, Edson E Servan-Mori†, Masood Ali Shaikh†, Rajesh Sharma†, Jun She†, Sara Sheikhbahaei†, Jiabin Shen†, Kenji Shibuya†, Hwashin Hyun Shin†, Inga Dora Sigfusdottir†, Naris Silpakit†, Diego Augusto Santos Silva†, Dayane Gabriele Alves Silveira†, Edgar P Simard†, Shireen Sindi†, Jasvinder A Singh†, Om Prakash Singh†, Prashant Kumar Singh†, Vegard Skirbekk†, Karen Sliwa†, Samir Soneji†, Reed J D Sorensen†, Joan B Soriano†, David O Soti†, Chandrashekhar T Sreeramareddy†, Vasiliki Stathopoulou†, Nicholas Steel†, Bruno F Sunguya†, Soumya Swaminathan†, Bryan L Sykes†, Rafael Tabarés-Seisdedos†, Roberto Tchio Talongwa†, Mohammad Tavakkoli†, Bineyam Taye†, Bemnet Amare Tedla†, Tesfaye Tekle†, Girma Temam Shifa†, Awoke Misganaw Temesgen†, Abdullah Sulieman Terkawi†, Fisaha Haile Tesfay†, Gizachew Assefa Tessema†, Kiran Thapa†, Alan J Thomson†, Andrew L Thorne-Lyman†, Ruoyan Tobe-Gai†, Roman Topor-Madry†, Jeffrey Allen Towbin†, Bach Xuan Tran†, Zacharie Tsala Dimbuene†, Nikolaos Tsilimparis†, Abera Kenay Tura†, Kingsley Nnanna Ukwaja†, Chigozie Jesse Uneke†, Olalekan A Uthman†, N Venketasubramanian†, Sergey K Vladimirov†, Vasiliy Victorovich Vlassov†, Stein Emil Vollset†, Linhong Wang†, Elisabete Weiderpass†, Robert G Weintraub†, Andrea Werdecker†, Ronny Westerman†, Tissa Wijeratne†, James D Wilkinson†, Charles Shey Wiysonge†, Charles D A Wolfe†, Sungho Won†, John Q Wong†, Gelin Xu†, Ajit Kumar Yadav†, Bereket Yakob†, Ayalnesh Zemene Yalew†, Yuichiro Yano†, Mehdi Yaseri†, Henock Gebremedhin Yebyo†, Paul Yip†, Naohiro Yonemoto†, Seok-Jun Yoon†, Mustafa Z Younis†, Chuanhua Yu†, Shicheng Yu†, Zoubida Zaidi†, Maysaa El Sayed Zaki†, Hajo Zeeb†, Hao Zhang†, Yong Zhao†, Sanjay Zodpey†, Leo Zoeckler†, Liesl Joanna Zuhlke†, Alan D Lopez‡, Christopher J L Murray‡

†, Yuichiro Yano†, Mehdi Yaseri†, Henock Gebremedhin Yebyo†, Paul Yip†, Naohiro Yonemoto†, Seok-Jun Yoon†, Mustafa Z Younis†, Chuanhua Yu†, Shicheng Yu†, Zoubida Zaidi†, Maysaa El Sayed Zaki†, Hajo Zeeb†, Hao Zhang†, Yong Zhao†, Sanjay Zodpey†, Leo Zoeckler†, Liesl Joanna Zuhlke†, Alan D Lopez‡, Christopher J L Murray‡ *Corresponding author. †Authors listed alphabetically. ‡Joint senior authors.

†, Yuichiro Yano†, Mehdi Yaseri†, Henock Gebremedhin Yebyo†, Paul Yip†, Naohiro Yonemoto†, Seok-Jun Yoon†, Mustafa Z Younis†, Chuanhua Yu†, Shicheng Yu†, Zoubida Zaidi†, Maysaa El Sayed Zaki†, Hajo Zeeb†, Hao Zhang†, Yong Zhao†, Sanjay Zodpey†, Leo Zoeckler†, Liesl Joanna Zuhlke†, Alan D Lopez‡, Christopher J L Murray‡ *Corresponding author. †Authors listed alphabetically. ‡Joint senior authors. Affiliations Institute for Health Metrics and Evaluation (H Wang PhD, T M Wolock BA, A Carter BS, G Nguyen BA, H H Kyu PhD, Prof E Gakidou PhD, S I Hay DSc, W Msemburi MPhil, M M Coates BS, M D Mooney BS, M S Fraser BA, A Sligar MPH, Prof H J Larson PhD, J Friedman BA, A Brown MA, Prof L Dandona MD, N Fullman MPH, J Haagsma PhD, I Khalil MD, Prof H J Larson PhD, S S Lim PhD, J Mikesell BS, A H Mokdad PhD, M Moradi-Lakeh MD, K Pearson BA, N Silpakit BS, R J D Sorensen MPH, A M Temesgen PhD, Prof S E Vollset DrPH, L Zoeckler BA, Prof C J L Murray DPhil), University of Washington, Seattle, WA, USA (R Alfonso-Cristancho PhD, K M Harun MPH, D M Prokop BS); University of Ottawa, Ottawa, ON, Canada (E J Mills PhD); School of Social and Community Medicine, University of Bristol, Bristol, UK (A Trickey MSc); Harvard T H Chan School of Public Health (O N Ajala MD, Prof T Bärnighausen MD, E L Ding ScD, M S Farvid PhD), Department of Nutrition, T H Chan School of Public Health (A L Thorne-Lyman ScD), Channing Division of Network Medicine, Brigham & Women's Hospital, Harvard Medical School (Prof S Won PhD), Harvard University, Boston, MA, USA (J R A Fitchett MBBS, Prof J Salomon PhD); School of Public Health (A A Abajobir MPH), School of Population Health (D G Hoy PhD), The University of Queensland, Brisbane, QLD, Australia (N K M Alam MPH); Jimma University, Jimma, Ethiopia (K H Abate MS, T T Ghiwot MPH, A T Gebremedhin MPH); Virginia Tech, Blacksburg, VA, USA (Prof K M Abbas PhD); Aswan University Hospital, Aswan, Egypt (M M Abd El Razek MBBCh); Department of Neurology, Cairo University, Cairo, Egypt (Prof F Abd-Allah MD); New York University Abu Dhabi, Abu Dhabi, United Arab Emirates (A M Abdulle PhD); School of Public Health, College of Health Sciences (S F Abera MSc), School of Public Health (Y A Melaku MPH), College of Health Sciences (F H Tesfay MPH), Mekelle University, Mekelle, Ethiopia (G Y Abyu MS, T A Bayou BS, B D Betsu MPH, A A Gebru MPH, G B Hailu MSc, T Tekle MPH, A Z Yalew MS, H G Yebyo MS); Kilte Awlaelo Health and Demographic Surveillance System, Mekelle, Ethiopia (S F Abera MSc, A A Gebru MPH, G B Hailu MSc); University College London, London, UK (Prof I

kelle University, Mekelle, Ethiopia (G Y Abyu MS, T A Bayou BS, B D Betsu MPH, A A Gebru MPH, G B Hailu MSc, T Tekle MPH, A Z Yalew MS, H G Yebyo MS); Kilte Awlaelo Health and Demographic Surveillance System, Mekelle, Ethiopia (S F Abera MSc, A A Gebru MPH, G B Hailu MSc); University College London, London, UK (Prof I Abubakar PhD, R W Aldridge MSc); Infectious Disease Epidemiology Group, Weill Cornell Medical College in Qatar, Doha, Qatar (L J Abu-Raddad PhD); Institute of Community and Public Health, Birzeit University, Ramallah, Palestine (N M E Abu-Rmeileh PhD); College of Medicine (A O Adebiyi MD), Department of Sociology (I A Adedeji MS), University of Ibadan, Ibadan, Nigeria (A L Adelekan MPH, F A Ojelabi MPH); University College Hospital, Ibadan, Nigeria (A O Adebiyi MD); Olabisi Onabanjo University, Ago-Iwoye, Nigeria (I A Adedeji MS); Public Health Promotion Alliance, Osogbo, Nigeria (A L Adelekan MPH); Kwame Nkrumah University of Science and Technology, Kumasi, Ghana (K Adofo PhD); Association Ivoirienne pour le Bien-Être Familial, Abidjan, Côte d’Ivoire (A K Adou MD); University of Pittsburgh Medical Center, McKeesport, PA, USA (O N Ajala MD); Department of Epidemiology (T F Akinyemiju PhD), University of Alabama at Birmingham, Birmingham, AL, USA (D C Schwebel PhD, J A Singh MD); Hospital for Sick Children, Toronto, ON, Canada (N Akseer MS); University of Toronto, Toronto, ON, Canada (N Akseer MS); Baghdad College of Medicine, Baghdad, Iraq (F H Al Lami PhD); Washington University in Saint Louis, St Louis, MO, USA (Z Al-Aly MD); Murdoch Childrens Research Institute, Melbourne, VIC, Australia (K Alam PhD, R G Weintraub MBBS); General Practice and Primary Health Care Academic Centre (P P Chiang PhD), Departments of Medicine and the Florey (A Meretoja PhD), Melbourne School of Population and Global Health (Prof A D Lopez PhD), The University of Melbourne, Melbourne, VIC, Australia (K Alam PhD, R G Weintraub MBBS, Prof T Wijeratne PhD); Sydney School of Public Health (Prof T R Driscoll PhD), The University of Sydney, Sydney, NSW, Australia (K Alam PhD, Prof A H Kemp PhD, J Leigh PhD, A B Mekonnen MS); Queensland Health, Herston, QLD, Australia (N K M Alam MPH); Malaria and Other Parasitic Diseases Division (C K Karema MSc), Ministry of Health, Al Khuwair, Oman (D Alasfoor MSc); King Saud University, Riyadh, Saudi Arabia (S F S Aldhahri MD); Department of Anesthesiology (A S Terkawi MD), King Fahad Medical City, Riyadh, Saudi Arabia (S F S Aldhahri MD); Department of

am MPH); Malaria and Other Parasitic Diseases Division (C K Karema MSc), Ministry of Health, Al Khuwair, Oman (D Alasfoor MSc); King Saud University, Riyadh, Saudi Arabia (S F S Aldhahri MD); Department of Anesthesiology (A S Terkawi MD), King Fahad Medical City, Riyadh, Saudi Arabia (S F S Aldhahri MD); Department of Preventive and Social Medicine (M A Alegretti MD), School of Medicine (A V Aleman MD), University of the Republic, Montevideo, Uruguay; Debre Markos University, Addis Ababa, Ethiopia (Z A Alemu MPH); NIHR Musculoskeletal Biomedical Research Centre (Prof C Cooper FMedSci), Nuffield Department of Medicine (A Deribew PhD), Department of Zoology (P W Gething PhD), University of Oxford, Oxford, UK (R Ali FRCP, D A Bennett PhD, D Bisanzio PhD, K Rahimi DM); Luxembourg Institute of Health, Strassen, Luxembourg (A Alkerwi PhD); School of Public Health, University of Lorraine, Nancy, France (Prof F Alla PhD); Malaria and Other Parasitic Diseases Division (C K Karema MSc), Ministry of Health, Riyadh, Saudi Arabia (R M S Al-Raddadi PhD, M Y Saeedi PhD); Charité Universitätsmedizin, Berlin, Germany (U Alsharif MPH); Government, Madrid, Spain (E Alvarez PhD); Universidad de Cartagena, Cartagena de Indias, Colombia (Prof N Alvis-Guzman PhD); Department of Epidemiology (A T Amare MPH), Department of Psychiatry, University Medical Center Groningen (Prof H W Hoek MD), University of Groningen, Groningen, Netherlands (A K Tura MPH); College of Medicine and Health Sciences, Bahir Dar University, Bahir Dar, Ethiopia (A T Amare MPH); Discipline of Psychiatry, School of Medicine (A T Amare MPH), School of Medicine (Y A Melaku MPH), University of Adelaide, Adelaide, SA, Australia (L G Ciobanu MS, G A Tessema MPH); Dignitas International, Zomba, Malawi (A Amberbir PhD); University of Cape Coast, Cape Coast, Ghana (A K Amegah PhD); Ministry of Public Health, Beirut, Lebanon (W Ammar PhD, H L Harb MPH); Oregon Health & Science University, Portland, OR, USA (S M Amrock MD); Department of Health Policy and Administration, College of Public Health, University of the Philippines Manila, Manila, Philippines (C A T Antonio MD); Self-employed, Kabul, Afghanistan (P Anwari MS); Department of Medical Sciences, Uppsala University, Uppsala, Sweden (Prof J Ärnlöv PhD, Prof A Larsson PhD); Dalarna University, Falun, Sweden (Prof J Ärnlöv PhD); Consultant, Windsor, ON, Canada (A Artaman PhD); Department of Medical Emergency, School of Paramedic, Qom University of Medical Sciences, Qom, Iran (H Asayesh

rtment of Medical Sciences, Uppsala University, Uppsala, Sweden (Prof J Ärnlöv PhD, Prof A Larsson PhD); Dalarna University, Falun, Sweden (Prof J Ärnlöv PhD); Consultant, Windsor, ON, Canada (A Artaman PhD); Department of Medical Emergency, School of Paramedic, Qom University of Medical Sciences, Qom, Iran (H Asayesh PhD); South Asian Public Health Forum, Islamabad, Pakistan (R J Asghar MD); Mashhad University of Medical Sciences, Mashhad, Iran (R Assadi PhD); Graduate Institute of Biomedical Informatics, Taipei Medical University, Taipei, Taiwan (S Atique MS); Wellness, Human Services and Gender Relations, Ministry of Health, Castries, Saint Lucia (L S Atkins MPH); Africare Benin, Cotonou, Benin (E F G A Avokpaho MPH); LERAS Afrique, Parakou, Benin (E F G A Avokpaho MPH); Sanjay Gandhi Postgraduate Institute of Medical Sciences, Lucknow, India (A Awasthi MSc, E Bhatia MD); The Judith Lumley Centre for Mother, Infant and Family Health Research, La Trobe University, Melbourne, VIC, Australia (B P Ayala Quintanilla PhD); Peruvian National Institute of Health, Lima, Peru (B P Ayala Quintanilla PhD); School of Health Sciences, University of Management and Technology, Lahore, Pakistan (U Bacha PhD); Public Health Agency of Canada, Toronto, ON, Canada (A Badawi PhD); Faculty of Medicine, University of Belgrade, Belgrade, Serbia (A Barac PhD); Wellcome Trust Africa Centre for Health and Population Studies, Somkhele, Mtubatuba, South Africa (Prof T Bärnighausen MD); School of Health Sciences, University of Canterbury, Christchurch, New Zealand (A Basu PhD); Jhpiego-Ethiopia, Addis Ababa, Ethiopia (Y T Bayou PhD); Charles R.

de, Belgrade, Serbia (A Barac PhD); Wellcome Trust Africa Centre for Health and Population Studies, Somkhele, Mtubatuba, South Africa (Prof T Bärnighausen MD); School of Health Sciences, University of Canterbury, Christchurch, New Zealand (A Basu PhD); Jhpiego-Ethiopia, Addis Ababa, Ethiopia (Y T Bayou PhD); Charles R. Drew University of Medicine and Science, Los Angeles, CA, USA (Prof S Bazargan-Hejazi PhD); University of Oxford, Ho Hi Minh City, Vietnam (J Beardsley MBChB); College of Public Health and Tropical Medicine, Jazan, Saudi Arabia (N Bedi MD); Internal Medicine Department (Prof I S Santos PhD), University of São Paulo, São Paulo, Brazil (I M Bensenor PhD, Prof A H Kemp PhD, Prof P A Lotufo DrPH); Haramaya University, Harar, Ethiopia (A S Beyene MPH); Medical Center (Z A Bhutta PhD), Aga Khan University, Karachi, Pakistan (M I Nisar MSc); The Hospital for Sick Children, Toronto, ON, Canada (Z A Bhutta PhD); Independent Public Health Consultants, Addis Ababa, Ethiopia (S Biadgilign MPH); Department of Nephrology Issues of Transplanted Kidney, Academician V I Shumakov Federal Research Center of Transplantology and Artificial Organs, Moscow, Russia (B Bikbov MD); GBS-CIDP International Foundation, Menemen, Turkey (S M Birlik BS); Danube-University Krems, Krems, Austria (Prof M Brainin PhD); Faculty of Health Sciences and Social Work (A Brazinova PhD), Faculty of Health Sciences and Social Work, Department of Public Health (M Majdan PhD), Trnava University, Trnava, Slovakia; International Neurotrama Research Organization, Vienna, Austria (A Brazinova PhD); The Ohio State University, Columbus, OH, USA (Prof N J K Breitborde PhD); University of Arizona, Tucson, AZ, USA (Prof N J K Breitborde PhD, Prof E Oren PhD); Great Ormond Street Hospital for Children, London, UK (M Burch MD); Al Shifa Trust Eye Hospital, Rawalpindi, Pakistan (Z A Butt PhD); National Institute of Public Health, Cuernavaca, Mexico (J C Campuzano PhD, I B Heredia-Pi PhD, Prof E E Servan-Mori MSc); Universidad Autonoma Metropolitana, Mexico City, Mexico (R Cárdenas ScD); Department of Neurobiology, Care Sciences and Society (S Fereshtehnejad MD), Aging Research Center (Prof M Kivipelto PhD), Department of Medical Epidemiology and Biostatistics (E Weiderpass PhD), Karolinska Institute, Stockholm, Sweden (Prof J J Carrero PhD, R Havmoeller PhD, S Sindi PhD); Colombian National Health Observatory, Instituto Nacional de Salud, Bogotá DC, Colombia (C A Castñeda-Orjuela MSc); Epidemiology and Public Health Eva

, Department of Medical Epidemiology and Biostatistics (E Weiderpass PhD), Karolinska Institute, Stockholm, Sweden (Prof J J Carrero PhD, R Havmoeller PhD, S Sindi PhD); Colombian National Health Observatory, Instituto Nacional de Salud, Bogotá DC, Colombia (C A Castñeda-Orjuela MSc); Epidemiology and Public Health Eva luation Group, Public Health Department, Universidad Nacional de Colombia, Bogota, Colombia (C A Castñeda-Orjuela MSc); Caja Costarricense de Seguro Social, San Jose, Costa Rica (Prof J Castillo Rivas MPH); Universidad de Costa Rica, San Pedro, Montes de Oca, Costa Rica (Prof J Castillo Rivas MPH); Department of Medicine, CIBERSAM (Prof R Tabarés-Seisdedos PhD) and INCLIVA Health Research Institute, University of Valencia Valencia, Spain (F Catalá-López PhD); Clinical Epidemiology Program, Ottawa Hospital Research Institute, Ottawa, ON, Canada (F Catalá-López PhD); National Health Research Institutes, Zhunan Town, Taiwan (Prof H Chang DrPH); National Yang-Ming University, Taipei, Taiwan (Prof H Chang DrPH); College of Medicine, National Taiwan University, Taipei, Taiwan (Prof J Chang PhD); World Health Organization, New Delhi, India (L Chavan MD); Cancer Institute, Chinese Academy of Medical Sciences, Beijing, China (W Chen PhD); Malaria Control and Elimination Partnership in Africa, PATH, Lusaka, Zambia (M Chibalabala BS); University of Zambia, Lusaka, Zambia (V H Chisumpa MPhil, C C Mapoma PhD, F Masiye PhD); School of Public Health (P de Jager FCPHM), University of the Witwatersrand, Johannesburg, South Africa (V H Chisumpa MPhil, Prof M Petzold PhD); Seoul National University Hospital (J J Choi PhD), College of Medicine (Prof Y Khang MD), Seoul National University, Seoul, South Korea (Prof S Won PhD); Seoul National University Medical Library, Seoul, South Korea (J J Choi PhD); Christian Medical College, Vellore, India (Prof D J Christopher MD); MRC Lifecourse Epidemiology Unit, University of Southampton, Southampton, UK (Prof C Cooper FMedSci); NIHR Biomedical Research Centre, University of Southampton and University Hospital Southampton NHS Foundation Trust, Southampton, UK (Prof C Cooper FMedSci); Health Systems and Policy Research Unit (S Mohammed PhD), Ahmadu Bello University, Zaria, Nigeria (T Dahiru MA); Wolaita Sodo University, Wolaita Sodo, Ethiopia (S A Damtew MPH); School of Public Health (K Deribe MPH), Addis Ababa University, Addis Ababa, Ethiopia (S A Damtew MPH, G Temam Shifa MPH); Centre for Control of Chronic Conditions (P Jeemon PhD), Publi

ohammed PhD), Ahmadu Bello University, Zaria, Nigeria (T Dahiru MA); Wolaita Sodo University, Wolaita Sodo, Ethiopia (S A Damtew MPH); School of Public Health (K Deribe MPH), Addis Ababa University, Addis Ababa, Ethiopia (S A Damtew MPH, G Temam Shifa MPH); Centre for Control of Chronic Conditions (P Jeemon PhD), Publi c Health Foundation of India, New Delhi, India (Prof L Dandona MD, R Dandona PhD, S Goenka PhD, G A Kumar PhD, D K Lal MD, Prof G V S Murthy MD, Prof S Zodpey PhD); Instituto de Engenharia Biomédica (J das Neves PhD), Instituto de Investigação e Inovação em Saúde (J das Neves PhD), REQUIMTE/LAQV, Laboratório de Farmacognosia, Departamento de Química, Faculdade de Farmácia (Prof D M Pereira PhD), University of Porto, Porto, Portugal; National Institute for Occupational Health, National Health Laboratory Service, Johannesburg, South Africa (P de Jager FCPHM); Griffith University, Brisbane, QLD, Australia (Prof D De Leo DSc); National Drug and Alcohol Research Centre, University of New South Wales, Sydney, NSW, Australia (Prof L Degenhardt PhD); University of Colorado School of Medicine and the Colorado School of Public Health, Aurora, CO, USA (R P Dellavalle MD); Brighton and Sussex Medical School, Brighton, UK (K Deribe MPH); KEMRI-Wellcome Trust Research Programme, Kilifi, Kenya (A Deribew PhD); Mount Sinai Beth Israel, New York, NY, USA (Prof D C Des Jarlais PhD); Icahn School of Medicine at Mount Sinai, New York City, NY, USA (Prof D C Des Jarlais PhD); Department of Community Medicine, Faculty of Medicine, University of Peradeniya, Peradeniya, Sri Lanka (S D Dharmaratne MD); University of Southern California, Los Angeles, CA, USA (P P Doshi BS); Australian National University, Canberra, ACT, Australia (Prof K E Doyle PhD); RMIT University, Bundoora, VIC, Australia (Prof K E Doyle PhD); International Institute for Population Sciences, Mumbai, India (M Dubey MPhil, M H U Rahman MPhil, Prof U Ram PhD, A K Yadav MPhil); Food Science Department, Faculty of Agriculture, University of Tripoli, Tripoli, Libya (Prof Y M Elshrek PhD); Eijkman-Oxford Clinical Research Unit, Jakarta, Indonesia (I Elyazar PhD); Arba Minch University, Arba Minch, Ethiopia (A Y Endries MPH, G Temam Shifa MPH); The Institute of Social and Economic Studies of Population, Russian Academy of Sciences, Moscow, Russia (Prof S P Ermakov DSc); Federal Research Institute for Health Organization and Informatics, Ministry of Health of the Russian Federation, Moscow, Russia (Prof S P Ermakov DSc); Mini

Y Endries MPH, G Temam Shifa MPH); The Institute of Social and Economic Studies of Population, Russian Academy of Sciences, Moscow, Russia (Prof S P Ermakov DSc); Federal Research Institute for Health Organization and Informatics, Ministry of Health of the Russian Federation, Moscow, Russia (Prof S P Ermakov DSc); Mini stry of Health and Medical Education, Tehran, Iran (B Eshrati PhD); Arak University of Medical Sciences, Arak, Iran (B Eshrati PhD); Endocrinology and Metabolism Research Center (Prof A Esteghamati MD, N Hafezi-Nejad MD, S Sheikhbahaei MD), Non-Communicable Diseases Research Center, Endocrinology and Metabolism Population Sciences Institute (F Farzadfar MD, A Kasaeian PhD), Hematology-Oncology and Stem Cell Transplantation Research Center (A Kasaeian PhD), Digestive Diseases Research Institute (Prof R Malekzadeh MD, G Roshandel PhD, S G Sepanlou PhD), Sina Trauma and Surgery Research Center (Prof V Rahimi-Movaghar MD), Tehran University of Medical Sciences, Terhan, Iran (M Yaseri PhD); Department of Infectious Disease Epidemiology, London School of Hygiene & Tropical Medicine, London, UK (I D A Faghmous MPH, S M Langan PhD, Prof M McKee DSc, Prof G V S Murthy MD); DGS Directorate General of Health, Lisbon, Portugal (C S Farinha MSc); Universidade Aberta, Lisbon, Portugal (C S Farinha MSc); Federal University of Sergipe, Aracaju, Brazil (A Faro PhD); Institute for Health Policy, Boston, MA, USA (M S Farvid PhD); Pharmacology and Experimental Therapeutics, Institute for Biomedical Imaging and Life Sciences, Faculty of Medicine, University of Coimbra, Coimbra, Portugal (J C Fernandes PhD); Bielefeld University, Bielefeld, Germany (F Fischer MPH); Institute of Gerontology, Academy of Medical Science, Kyiv, Ukraine (N Foigt PhD); Department of Infectious Disease Epidemiology (T Fürst PhD), Imperial College London, London, UK (Prof A Majeed MD); Leras Afrique, Cotonou, Benin (F G Gankpé MD); CHU Hassan II, Fès, Morocco (F G Gankpé MD); The Task Force for Global Health, Decatur, GA, USA (T Gebre PhD); Ludwig Maximilians University, Munich, Germany (A T Gebremedhin MPH); Division of Human Nutrition (J M Geleijnse PhD), Wageningen University, Wageningen, Netherlands (T Jibat MS); Agence de Medecine Preventive, Paris, France (B D Gessner MD); University Hospital of Dijon, Dijon, France (Prof M Giroud MD); Haramaya University, Dire Dawa, Ethiopia (M D Gishu MS, A K Tura MPH); Kersa Health and Demographic Surveillance System, Harar, Ethiopia (M D Gishu MS); Heller School fo

therlands (T Jibat MS); Agence de Medecine Preventive, Paris, France (B D Gessner MD); University Hospital of Dijon, Dijon, France (Prof M Giroud MD); Haramaya University, Dire Dawa, Ethiopia (M D Gishu MS, A K Tura MPH); Kersa Health and Demographic Surveillance System, Harar, Ethiopia (M D Gishu MS); Heller School fo r Social Policy and Management, Brandeis University, Waltham, MA, USA (E Glaser MS); Instituto de Investigaciones Cientificas y Servicios de Alta Tecnologia—INDICASAT-AIP, City of Knowledge, Panama (A Goodridge PhD); Department of Health and Social Affairs, Government of the Federated States of Micronesia, Palikir, Federated States of Micronesia (S V Gopalani MPH); Department of Public Health, Tokyo Women's Medical University, Tokyo, Japan (A Goto MD); Departments of Microbiology and Epidemiology & Biostatistics, Saint James School of Medicine, The Quarter, Anguilla (Prof H C Gugnani PhD); Federal University of Minas Gerais, Belo Horizonte, Brazil (Prof M D C Guimaraes MD); West Virginia Bureau for Public Health, Charleston, WV, USA (R Gupta MD); Eternal Heart Care Centre and Research Institute, Jaipur, India (R Gupta PhD); University of Delhi, Delhi, India (V Gupta PhD); New York University, New York, NY, USA (Prof H Hagan PhD); Arabian Gulf University, Manama, Bahrain (Prof R R Hamadeh DPhil); Hamdan Bin Mohammed Smart University, Dubai, United Arab Emirates (S Hamidi PhD); Wayne County Department of Health and Human Services, Detroit, MI, USA (M Hammami MD); School of Medicine and Pharmacology, The University of Western Australia, Perth, WA, Australia (Prof G J Hankey MD); Harry Perkins Institute of Medical Research, Nedlands, WA, Australia (Prof G J Hankey MD); Western Australian Neuroscience Research Institute, Nedlands, WA, Australia (Prof G J Hankey MD); School of Public Health, Sun Yat-Sen University, Guangzhou, China (Prof Y Hao PhD); Sree Chitra Tirunal Institute for Medical Sciences and Technology, Trivandrum, India (S Harikrishnan DM); Research and Development Unit (A Koyanagi MD), Parc Sanitari Sant Joan de Déu (CIBERSAM), Barcelona, Spain (J M Haro MD); Universitat de Barcelona, Barcelona, Spain (J M Haro MD); Kenyatta University, Nairobi, Kenya (K M Harun MPH); Mazandaran University of Medical Sciences, Sari, Iran (Prof M T Hedayati PhD); Department of Epidemiology (Prof H W Hoek MD), Columbia University, New York, NY, USA (Prof V Skirbekk PhD); Nevada Division of Behavior and Public Health, Carson City, NV, USA (M Horino MPH); Fielding School of

robi, Kenya (K M Harun MPH); Mazandaran University of Medical Sciences, Sari, Iran (Prof M T Hedayati PhD); Department of Epidemiology (Prof H W Hoek MD), Columbia University, New York, NY, USA (Prof V Skirbekk PhD); Nevada Division of Behavior and Public Health, Carson City, NV, USA (M Horino MPH); Fielding School of Public Health, University of California Los Angeles, Los Angeles, CA, USA (M Horino MPH); Yokohama City University School of Medicine, Yokohama, Japan (N Horita MD); Albert Einstein College of Medicine, Bronx, NY, USA (Prof H D Hosgood PhD); Public Health Division, Secretariat of the Pacific Community, Noumea, New Caledonia (D G Hoy PhD); National Institute of Public Health, Tunis, Tunisia (Prof M Hsairi MD); Department of Epidemiology and Health Statistics, School of Public Health, Central South University, Changsha, China (G Hu PhD); Cambridge Health Alliance, Cambridge, MA, USA (H Huang MD); Yale University, New Haven, CT, USA (J J Huang MD); Aarhus University, Aarhus, Denmark (K M Iburg PhD); Bashkir State Medical University, Ufa, Russia (B T Idrisov MD); Boston Medical Center, Boston University, Boston, MA, USA (B T Idrisov MD); National Institute for Health Development, Tallinn, Estonia (K Innos PhD); Indian Institute of Public Health Gandhinagar, Ahmedabad, India (V J Iyer MPH); Department of Global and Community Health, George Mason University, Fairfax, VA, USA (K H Jacobsen PhD); School of Public Health, Shahid Beheshti University of Medical Sciences, Tehran, Iran (N Jahanmehr PhD); University of Kragujevac Faculty of Medical Sciences, Kragujevac, Serbia (Prof M B Jakovljevic PhD); University of Aberdeen, Aberdeen, UK (M Javanbakht PhD); Postgraduate Institute of Medicine, Colombo, Sri Lanka (A U Jayatilleke PhD); Institute of Violence and Injury Prevention, Colombo, Sri Lanka (A U Jayatilleke PhD); Postgraduate Institute of Medical Education and Research, Chandigarh, India (Prof V Jha DM); Tianjin Centers for Disease Control and Prevention, Tianjin, China (G Jiang MD); Department of Health Development, Institute of Industrial Ecological Sciences, Department of Environmental Epidemiology, University of Occupational and Environmental Health, Kitakyushu, Japan (Y Jiang PhD); School of Public Health (K Deribe MPH), Addis Ababa University, Debre Zeit, Ethiopia (T Jibat MS); Department of Ophthalmology, Medical Faculty Mannheim, Ruprecht-Karls-University Heidelberg, Mannheim, Germany (Prof J B Jonas MD); University College Cork, Cork, Ireland (Z Kabir PhD);

, Kitakyushu, Japan (Y Jiang PhD); School of Public Health (K Deribe MPH), Addis Ababa University, Debre Zeit, Ethiopia (T Jibat MS); Department of Ophthalmology, Medical Faculty Mannheim, Ruprecht-Karls-University Heidelberg, Mannheim, Germany (Prof J B Jonas MD); University College Cork, Cork, Ireland (Z Kabir PhD); CSIR-Indian Institute of Toxicology Research, Lucknow, India (R Kamal MSc, C N Kesavachandran PhD); Zhongshan Hospital (J She MD), Department of Nephrology, Zhongshan Hospital (H Zhang PhD), Fudan University, Shanghai, China (H Kan MD); Epidemiological and Statistical Methods Research Group, Helmholtz Centre for Infection Research, Braunschweig, Germany (A Karch MD); Hannover-Braunschweig Site, German Center for Infection Research, Braunschweig, Germany (A Karch MD); Clinton Health Access Initiative, Boston, MA, USA (D Karletsos MS); Oklahoma State University, Tulsa, OK, USA (A Kaul MD); School of Public Health (Prof N Kawakami MD), The University of Tokyo, Tokyo, Japan (K Shibuya MD); Ottawa Health Research Institute, Ottawa, ON, Canada (J F Kayibanda PhD); Institute of tropical and infectious diseases, Nairobi, Kenya (P N Keiyoro PhD); School of Continuing and Distance Education, Nairobi, Kenya (P N Keiyoro PhD); South African Medical Research Council, Cape Town, South Africa (A P Kengne PhD); Faculty of Health Sciences, Hatter Institute for Cardiovascular Research in Africa (Prof K Sliwa PhD), University of Cape Town, Cape Town, South Africa (A P Kengne PhD, Prof B M Mayosi DPhil); Jordan University of Science and Technology, Irbid, Jordan (Prof Y S Khader ScD); University of Louisville, Louisville, KY, USA (A R Khan MD); Health Services Academy, Islamabad, Pakistan (E A Khan MPH); Expanded Programme on Immunization, Islamabad, Pakistan (E A Khan MPH); Ball State University, Muncie, IN, USA (J Khubchandani PhD); Southern University College, Skudai, Malaysia (Y J Kim PhD); University of Canberra, Canberra, ACT, Australia (Y Kinfu PhD); Department of Preventive Cardiology, National Cerebral and Cardiovascular Center, Suita, Japan (Y Kokubo PhD); Center for Community Empowerment, Health Policy and Humanities, NIHRD, Jakarta, Indonesia (S Kosen MD); Sherikashmir Institute of Medical Sciences, Srinagar, India (Prof P A Koul MD); Department of Demography and Public Health Research Institute (Prof B Kuate Defo PhD), Department of Social and Preventive Medicine, School of Public Health (Prof B Kuate Defo PhD), University of Montreal, Montreal, QC, Canada; Institute o

Sherikashmir Institute of Medical Sciences, Srinagar, India (Prof P A Koul MD); Department of Demography and Public Health Research Institute (Prof B Kuate Defo PhD), Department of Social and Preventive Medicine, School of Public Health (Prof B Kuate Defo PhD), University of Montreal, Montreal, QC, Canada; Institute o f Public Health, Hacettepe University, Ankara, Turkey (B Kucuk Bicer PhD); Arkansas State University, State University, AK, USA (V S Kulkarni PhD); Institute of Health Policy and Development Studies, National Institutes of Health, Manila, Philippines (Prof H Lam PhD); Bloomberg School of Public Health (J O Lam MPH), Johns Hopkins University, Baltimore, MD, USA (B X Tran PhD); Help Me See, Inc, New York, NY, USA (V C Lansingh PhD); Instituo Mexicano de Oftalmologia, Queretaro, Mexico (V C Lansingh PhD); State University of New York Albany, Rensselaer, NY, USA (R Leung PhD); San Francisco VA Medical Center, San Francisco, CA, USA (Y Li PhD); School of Medicine, Wayne State University, Miami, FL, USA (Prof S E Lipshultz MD); Children's Hospital of Michigan, Detroit, MI, USA (Prof S E Lipshultz MD, Prof J D Wilkinson MD); National Center for Chronic and Noncommunicable Disease Control and Prevention (S Liu PhD, Prof L Wang MD), Chinese Center for Disease Control and Prevention, Beijing, China (Prof S Yu PhD); Eastern Health Clinical School, Monash University, Fitzroy, VIC, Australia (B K Lloyd PhD); Turning Point, Eastern Health, Melbourne, VIC, Australia (B K Lloyd PhD); University of Bari, Bari, Italy (Prof G Logroscino PhD); Aintree University Hospital National Health Service Foundation Trust, Liverpool, UK (Prof R Lunevicius PhD); School of Medicine, University of Liverpool, Liverpool, UK (Prof R Lunevicius PhD); Mansoura Faculty of Medicine, Mansoura, Egypt (H Magdy Abd El Razek BA, Prof M E Zaki PhD); Department of Public Health, University Medical Center (S Polinder PhD), Erasmus University Rotterdam, Rotterdam, Netherlands (M Mahdavi PhD); Iranian Ministry of Health, Tehran, Iran (M Mahdavi PhD); JSS Medical College, JSS University, Mysore, India (Prof P A Mahesh DNB); Queen Mary University of London, London, UK (Prof W Marcenes PhD); Hospital Universitario Doctor Peset, Valencia, Spain (J Martinez-Raga MD); University CEU-UCH, Moncada, Spain (J Martinez-Raga MD); University of the East Ramon Magsaysay Memorial Medical Center, Quezon City, Philippines (M B Marzan MSc); Department of Health Sciences, University of York, York, UK (A J Mason-Jones PhD); Perelem

Universitario Doctor Peset, Valencia, Spain (J Martinez-Raga MD); University CEU-UCH, Moncada, Spain (J Martinez-Raga MD); University of the East Ramon Magsaysay Memorial Medical Center, Quezon City, Philippines (M B Marzan MSc); Department of Health Sciences, University of York, York, UK (A J Mason-Jones PhD); Perelem an School of Medicine, University of Pennsylvania, Philadelphia, PA, USA (P A Meaney MD); Children's Hospital of Philadelphia, Philadelphia, PA, USA (P A Meaney MD); Janakpuri Superspecialty Hospital, New Delhi, India (Prof M M Mehndiratta DM); University of Gondar, Gondar, Ethiopia (A B Mekonnen MS, B A Tedla BSc, G A Tessema MPH); University of West Florida, Pensacola, FL, USA (P Memiah PhD); Saudi Ministry of Health, Riyadh, Saudi Arabia (Prof Z A Memish MD); College of Medicine, Alfaisal University, Riyadh, Saudi Arabia (Prof Z A Memish MD); United Nations Population Fund, Lima, Peru (W Mendoza MD); Department of Neurology, Helsinki University Central Hospital, Helsinki, Finland (A Meretoja PhD); Helsinki University Hospital, Comprehensive Cancer Center, Breast Surgery Unit, Helsinki, Finland (T J Meretoja PhD); University of Helsinki, Helsinki, Finland (T J Meretoja PhD); Ifakara Health Institute, Bagamoyo, Tanzania (F A Mhimbira MS); Pacific Institute for Research & Evaluation, Calverton, MD, USA (T R Miller PhD); Curtin University Centre for Population Health, Perth, WA, Australia (T R Miller PhD); Hunger Action Los Angeles, Los Angeles, CA, USA (M Mirarefin MPH); University of Salahaddin, Erbil, Iraq (K A Mohammad PhD); Institute of Public Health, Heidelberg University, Heidelberg, Germany (S Mohammed PhD); Institute for Maternal and Child Health, IRCCS Burlo Garofolo, Trieste, Italy (L Monasta DSc, L Ronfani PhD); Department of Community Medicine, Gastrointestinal and Liver Disease Research Center, Iran University of Medical Sciences, Tehran, Iran (M Moradi-Lakeh MD); National Center for Child Health and Development, Setagaya, Japan (R Mori PhD); Competence Center Mortality-Follow-Up of the German National Cohort (A Werdecker PhD), Federal Institute for Population Research, Wiesbaden, Germany (Prof U O Mueller PhD, R Westerman PhD); Ministry of Health and Child Care (AIDS & TB Unit), Bindura, Zimbabwe (B Murimira MS); Zimbabwe National Family Planning Council, Harare, Zimbabwe (B Murimira MS); International Centre for Diarrhoeal Disease Research (ICDDR), Bangladesh, Dhaka, Bangladesh (A Naheed PhD); Azienda Ospedaliera papa Giovanni XXIII, Bergamo, Ital

ry of Health and Child Care (AIDS & TB Unit), Bindura, Zimbabwe (B Murimira MS); Zimbabwe National Family Planning Council, Harare, Zimbabwe (B Murimira MS); International Centre for Diarrhoeal Disease Research (ICDDR), Bangladesh, Dhaka, Bangladesh (A Naheed PhD); Azienda Ospedaliera papa Giovanni XXIII, Bergamo, Ital y (Prof L Naldi MD); Suraj Eye Institute, Nagpur, India (Prof V Nangia MD); School of Public Health, City University of New York, New York, NY, USA (D Nash PhD); Southern Illinois University, Springfield, IL, USA (H Nawaz MD); Faculty of Medicine, Fez, Morocco (Prof C Nejjari PhD); Ministry of Health and Social Welfare, Dar es Salaam, Tanzania (F N Ngalesoni MSc); East African Community Health Research Commission, Kigali, Rwanda (J D Ngirabega PhD); Institute for Global Health Innovations, Duy Tan University, Da Nang, Vietnam (Q L Nguyen MD); Department of Global Public Health and Primary Care (Prof S E Vollset DrPH), University of Bergen, Bergen, Norway (Prof O F Norheim PhD); Institute of Health and Biomedical Innovation, Queensland University of Technology, Brisbane, QLD, Australia (R E Norman PhD); Makerere University, Kampala, Uganda (L Nyakarahuka MPH); Center for Research on Population and Health, Faculty of Health Sciences, American University of Beirut, Beirut, Lebanon (Prof C M Obermeyer DSc); University of Western Sydney, Sydney, NSW, Australia (F A Ogbo MPH); Department of Preventive Medicine, School of Medicine, Kyung Hee University, Seoul, South Korea (Prof I Oh PhD); Center for Healthy Start Initiative, Ikoyi, Nigeria (B O Olusanya PhD); Centre for Healthy Start Initiative, Lagos, Nigeria (J O Olusanya MBA); Lira District Local Government, Lira Municipal Council, Uganda (J N Opio MPH); National Research Institute for Child Health and Development, Tokyo, Japan (E Ota PhD); California Air Resources Board, Sacramento, CA, USA (H Park PhD); Department of Social and Preventive Medicine, Samsung Biomedical Research Institute, School of Medicine, Sungkyunkwan University, Suwon, South Korea (Prof J Park MPH); School of Dental Sciences, Karad, India (S T Patil MDS); Department of Community Health Sciences, University of Calgary, Calgary, AB, Canada (Prof S B Patten PhD); All India Institute of Medical Sciences, New Delhi, India (V K Paul MD, A Roy DM, R Sagar MD); National Heart, Lung, and Blood Institute, Bethesda, MD, USA (E K Peprah PhD); Centro Anna Maria Astori, IRCCS Mario Negri Institute for Pharmacological Research, Bergamo, Italy (N Perico MD, Pro

Canada (Prof S B Patten PhD); All India Institute of Medical Sciences, New Delhi, India (V K Paul MD, A Roy DM, R Sagar MD); National Heart, Lung, and Blood Institute, Bethesda, MD, USA (E K Peprah PhD); Centro Anna Maria Astori, IRCCS Mario Negri Institute for Pharmacological Research, Bergamo, Italy (N Perico MD, Pro f G Remuzzi MD); Flinders University, Adelaide, SA, Australia (Prof K Pesudovs PhD, F H Tesfay MPH); Health Metrics Unit, Gothenburg, Sweden (Prof M Petzold PhD); Shanghai Jiao Tong University School of Medicine, Shanghai, China (Prof M R Phillips MD); Rollins School of Public Health (E P Simard PhD), Emory University, Atlanta, GA, USA (Prof M R Phillips MD); Durban University of Technology, Durban, South Africa (J D Pillay PhD); Section Exposure Assessment and Environmental Health Indicators Federal Environment Agency, Berlin, Germany (D Plass DrPH); University of British Columbia, Vancouver, BC, Canada (F Pourmalek PhD); Department of Community Medicine, School of Medicine, Alborz University of Medical Sciences, Karaj, Iran (M Qorbani PhD); Contech International Health Consultants, Lahore, Pakistan (A Rafay MS, Prof S M Rana PhD); Contech School of Public Health, Lahore, Pakistan (A Rafay MS, Prof S M Rana PhD); Research and Evaluation Division, BRAC, Dhaka, Bangladesh (M Rahman PhD); Hamad Medical Corporation, Doha, Qatar (S U Rahman FCPS); Society for Health and Demographic Surveillance, Suri, India (R K Rai MPH); ERAWEB Program, UMIT, Hall in Tirol, Austria (S Rajsic MD); Diabetes Research Society, Hyderabad, India (Prof P V Rao MD); Diabetes Research Center, Hyderabad, India (Prof P V Rao MD); Azienda Ospedaliera Papa Giovanni XXIII, Bergamo, Italy (Prof G Remuzzi MD); Centre for Research in Environmental Epidemiology, ISGlobal, Barcelona, Spain (D Rojas-Rueda PhD); Golestan Research Center of Gastroenterology and Hepatology, Golestan University of Medical Sciences, Gorgan, Iran (G Roshandel PhD); Muhimbili University of Health and Allied Sciences, Dar es Salaam, Tanzania (G M Ruhago PhD, B F Sunguya MSc); Development Research and Projects Center, Abuja, Nigeria (M M Saleh MPH); Case Western Reserve University, Cleveland, OH, USA (J R Sanabria MD); Cancer Treatment Centers of America, RFU Chicago Medical School, North Chicago, IL, USA (J R Sanabria MD); Universidad Ciencias Aplicadas y Ambientales, Bogotá DC, Colombia (R Sarmiento-Suarez MPH); University of KwaZulu-Natal, Durban, South Africa (Prof B Sartorius PhD, B Yakob MPH); Marshall University, Huntin

a MD); Cancer Treatment Centers of America, RFU Chicago Medical School, North Chicago, IL, USA (J R Sanabria MD); Universidad Ciencias Aplicadas y Ambientales, Bogotá DC, Colombia (R Sarmiento-Suarez MPH); University of KwaZulu-Natal, Durban, South Africa (Prof B Sartorius PhD, B Yakob MPH); Marshall University, Huntin gton, WV, USA (M Sawhney PhD); Hypertension in Africa Research Team, North-West University, Potchefstroom, South Africa (Prof A E Schutte PhD); South African Medical Research Council, Potchefstroom, South Africa (Prof A E Schutte PhD); Stellenbosch University, Cape Town, South Africa (Prof S Seedat PhD, Prof C S Wiysonge PhD); Independent Consultant, Karachi, Pakistan (M A Shaikh MD); Indian Institute of Technology Ropar, Roopnagar, India (R Sharma MA); Research Institute at Nationwide Children's Hospital, Columbus, OH, USA (J Shen PhD); The Ohio State University College of Medicine, Columbus, OH, USA (J Shen PhD); Health Canada, Ottawa, ON, Canada (H H Shin PhD); Reykjavik University, Reykjavik, Iceland (I D Sigfusdottir PhD); Federal University of Santa Catarina, Florianopolis, Brazil (D A S Silva PhD); Brasília University, Brasília, Brazil (D G A Silveira MD); Department of Medicine, Institute of Medical Sciences, Banaras Hindu University, Varanasi, India (O P Singh PhD); Institute for Human Development, New Delhi, India (P K Singh PhD); Norwegian Institute of Public Health, Oslo, Norway (Prof V Skirbekk PhD); Dartmouth College, Lebanon, NH, USA (S Soneji PhD); Instituto de Investigación Hospital Universitario de la Princesa, Universidad Autónoma de Madrid, Cátedra UAM-Linde, Palma de Mallorca, Spain (Prof J B Soriano PhD); Malaria and Other Parasitic Diseases Division (C K Karema MSc), Ministry of Health, Nairobi, Kenya ( D O Soti MPH); Department of Community Medicine, International Medical University, Kuala Lumpur, Malaysia (C T Sreeramareddy MD); Attikon University Hospital, Athens, Greece (V Stathopoulou PhD); University of East Anglia, Norwich, UK (Prof N Steel PhD); Public Health England, London, UK (Prof N Steel PhD); National Institute for Research in Tuberculosis, Chennai, India (S Swaminathan MD); Departments of Criminology, Law & Society, Sociology, and Public Health, University of California-Irvine, Irvine, CA, USA (Prof B L Sykes PhD); Ministry of Health, MINSANTE, Yaounde, Cameroon (R T Talongwa MD); Westchester Medical Center, Valhalla, NY, USA (M Tavakkoli MD); Addis Ababa university, Addis Ababa, Ethiopia (B Taye PhD); James Cook University,

iety, Sociology, and Public Health, University of California-Irvine, Irvine, CA, USA (Prof B L Sykes PhD); Ministry of Health, MINSANTE, Yaounde, Cameroon (R T Talongwa MD); Westchester Medical Center, Valhalla, NY, USA (M Tavakkoli MD); Addis Ababa university, Addis Ababa, Ethiopia (B Taye PhD); James Cook University, Cairns, QLD, Australia (B A Tedla BSc); Department of Anesthesiology, University of Virginia, Charlottesville, VA, USA (A S Terkawi MD); Outcomes Research Consortium, Cleveland Clinic, Cleveland, OH, USA (A S Terkawi MD); Institute of Medicine, Kathmandu, Nepal (K Thapa BPH); Adaptive Knowledge Management, Victoria, BC, Canada (A J Thomson PhD); WorldFish, Penang, Malaysia (A L Thorne-Lyman ScD); National Center for Child Health and Development, Tokyo, Japan (R Tobe-Gai PhD); Institute of Public Health, Jagiellonian University, Krakow, Poland (R Topor-Madry PhD); Casimir the Great Foundation of Innovation and Development, Krakow, Poland (R Topor-Madry PhD); Le Bonheur Children's Hospital, Memphis, TN, USA (Prof J A Towbin MD); University of Tennessee Health Science Center, Memphis, TN, USA (Prof J A Towbin MD); St Jude Children's Research Hospital, Memphis, TN, USA (Prof J A Towbin MD); Hanoi Medical University, Hanoi, Vietnam (B X Tran PhD); Department of Population Sciences and Development, Faculty of Economics and Management, University of Kinshasa, Kinshasa, Democratic Republic of the Congo (Z Tsala Dimbuene PhD); University Heart Center of Hamburg, Hamburg, Germany (N Tsilimparis PhD); Department of Internal Medicine, Federal Teaching Hospital, Abakaliki, Nigeria (K N Ukwaja MD); Ebonyi State University, Abakaliki, Nigeria (C J Uneke PhD); Warwick Medical School, University of Warwick, Coventry, UK (O A Uthman PhD); Raffles Neuroscience Centre, Raffles Hospital, Singapore, Singapore (N Venketasubramanian FRCP[Edin]); Federal Research Institute for Health Organization and Informatics, Moscow, Russia (S K Vladimirov PhD); National Research University Higher School of Economics, Moscow, Russia (Prof V V Vlassov MD); Norwegian Institute of Public Health, Bergen, Norway (Prof S E Vollset DrPH); Department of Research, Cancer Registry of Norway Institute of Population-Based Cancer Research, Oslo, Norway (E Weiderpass PhD); Department of Community Medicine, Faculty of Health Sciences, University of Tromsø, The Arctic University of Norway, Tromsø, Norway (E Weiderpass PhD); Genetic Epidemiology Group, Folkhälsan Research Center, Helsinki, Finland (E Weiderpass PhD

titute of Population-Based Cancer Research, Oslo, Norway (E Weiderpass PhD); Department of Community Medicine, Faculty of Health Sciences, University of Tromsø, The Arctic University of Norway, Tromsø, Norway (E Weiderpass PhD); Genetic Epidemiology Group, Folkhälsan Research Center, Helsinki, Finland (E Weiderpass PhD ); Royal Children's Hospital, Melbourne, VIC, Australia (R G Weintraub MBBS); German National Cohort Consortium, Heidelberg, Germany (R Westerman PhD); Western Health Footscray, VIC, Australia (Prof T Wijeratne MD); School of Medicine, Wayne State University, Detroit, MI, USA (Prof J D Wilkinson MD); Division of Health and Social Care Research, King's College London, London, UK (Prof C D A Wolfe MD); National Institute for Health Research Comprehensive Biomedical Research Centre, Guy's & St Thomas’ NHS Foundation Trust and King's College London, London, UK (Prof C D A Wolfe MD); Ateneo School of Medicine and Public Health, Manila University, Pasig City, Philippines (J Q Wong MD); Department of Neurology, Jinling Hospital, Nanjing University School of Medicine, Nanjing, China (Prof G Xu PhD); Department of Preventive Medicine, Northwestern University, Chicago, IL, USA (Y Yano MD); Ophthalmic Research Center, Tehran, Iran (M Yaseri PhD); University of Zurich, Zurich, Switzerland (H G Yebyo MS); Social Work and Social Administration Department (Prof P Yip PhD), The Hong Kong Jockey Club Centre for Suicide Research and Prevention (Prof P Yip PhD), University of Hong Kong, Hong Kong, China; Department of Biostatistics, School of Public Health, Kyoto University, Kyoto, Japan (N Yonemoto MPH); Department of Preventive Medicine, College of Medicine, Korea University, Seoul, South Korea (S Yoon PhD); Jackson State University, Jackson, MS, USA (Prof M Z Younis DrPH); Department of Epidemiology and Biostatistics, School of Public Health (Prof C Yu PhD), Global Health Institute (Prof C Yu PhD), Wuhan University, Wuhan, China; University Hospital, Setif, Algeria (Prof Z Zaidi PhD); Leibniz Institute for Prevention Research and Epidemiology, Bremen, Germany (Prof H Zeeb PhD); Shanghai Institute of Kidney Disease and Dialysis, Shanghai, China (H Zhang PhD); Chongqing Medical University, Chongqing, China (Prof Y Zhao MSc); Red Cross War Memorial Children's Hospital, Cape Town, South Africa (L J Zuhlke PhD).

Institute for Prevention Research and Epidemiology, Bremen, Germany (Prof H Zeeb PhD); Shanghai Institute of Kidney Disease and Dialysis, Shanghai, China (H Zhang PhD); Chongqing Medical University, Chongqing, China (Prof Y Zhao MSc); Red Cross War Memorial Children's Hospital, Cape Town, South Africa (L J Zuhlke PhD). Contributors CJLM and HW prepared the first draft and finalised the draft based on comments from other authors and reviewer feedback. CJLM conceived the study and provided overall guidance. All other authors provided data, developed models, reviewed results, initiated modelling infrastructure, and reviewed and contributed to the report.

and HW prepared the first draft and finalised the draft based on comments from other authors and reviewer feedback. CJLM conceived the study and provided overall guidance. All other authors provided data, developed models, reviewed results, initiated modelling infrastructure, and reviewed and contributed to the report. Declaration of interests RA-C has been a GlaxoSmithKline (GSK) employee and shareholder. CAA reports grants and personal fees from Johnson & Johnson (Philippines). CC reports a financial relationship with Alliance for Better Bone Health, Amgen, Eli Lilly, GSK, Medtronic, Merck, Novartis, Pfizer, Roche, Servier, Takeda, and UCB. LD reports grants from Mundipharma, Reckitt Benckiser. BDG reports grants from Crucell, GSK, Hilleman Labs, Novartis, Pfizer, Merck, and Sanofi Pasteur. JBJ reports personal fees from consultancy with Mundipharma (Cambridge, UK); and has a patent application with University of Heidelberg (Heidelberg, Germany; Agents for use in the therapeutic or prophylactic treatment of myopia or hyperopia, Europäische Patentanmeldung 15000 771.4), is a patent holder with Biocompatibles UK (Franham, Surrey, UK; Treatment of eye diseases by use of encapsulated cells encoding and secreting neuroprotective factor and/or anti-angiogenic factor; patent number 20120263794). HJL reports personal fees from GSK, service on the Vaccine Strategic Advisory Board from Merck Vaccines (honorarium for meetings goes to the London School of Hygiene & Tropical Medicine), and grants from Novartis. PAL reports honoraria for lectures from Abbvie (Brazil). DCS reports grants from Vipaar and Carr & Carr. JAS reports paid consultancy for Savient, Takeda, Regeneron, Iroko, Merz, Bioiberica, Crealta, Allergan, UBM, WebMD, and the American College of Rheumatology and grants or grants pending from Takeda and Savient. JAS serves as the principal investigator for an investigator-initiated study funded by Horizon Pharmaceuticals through a grant to DINORA, a 501c3 entity, and is on the steering committee of OMERACT, an international organisation that develops measures for clinical trials and receives “arm's length” funding from 36 pharmaceutical companies. RGW reports personal fees from Actelion Pharmaceuticals. All other authors declare no competing interests.

o DINORA, a 501c3 entity, and is on the steering committee of OMERACT, an international organisation that develops measures for clinical trials and receives “arm's length” funding from 36 pharmaceutical companies. RGW reports personal fees from Actelion Pharmaceuticals. All other authors declare no competing interests. Figure 1 Evolution of the HIV epidemic from 1980 to 2015 Global estimates of new HIV infections (A), people living with HIV/AIDS (B), HIV/AIDS deaths (C), and proportion of people living with HIV receiving ART (D). Shaded areas show 95% uncertainty intervals. ART=antiretroviral therapy. Figure 2 Incidence of new HIV infections from 1980 to 2015, and HIV incidence in 2015 Global number of new HIV infections by region (A). Bars show the mean number of estimated new infections within a given year. Error bars represent 95% uncertainty intervals. Each Global Burden of Disease region is represented by a separate colour. HIV incidence by country (B). We calculated incidence as cumulative new cases of HIV throughout the year divided by the total population at the mid-year. Rates are per 100 000 people. Colour bins correspond to the 0–50th, 50–70th, 70–80th, 80–90th, 90th–92nd, 92nd–94th, 96–98th, 98–99th, and 99–100th percentiles to highlight variation within sub-Saharan Africa. ATG=Antigua and Barbuda. VCT=Saint Vincent and the Grenadines. LCA=Saint Lucia. TTO=Trinidad and Tobago. TLS=Timor-Leste. FSM=Federated States of Micronesia. Figure 3 Number of people living with HIV receiving ART from 1995 to 2015, and the proportion living with HIV receiving ART in 2015

Global number of new HIV infections by region (A). Bars show the mean number of estimated new infections within a given year. Error bars represent 95% uncertainty intervals. Each Global Burden of Disease region is represented by a separate colour. HIV incidence by country (B). We calculated incidence as cumulative new cases of HIV throughout the year divided by the total population at the mid-year. Rates are per 100 000 people. Colour bins correspond to the 0–50th, 50–70th, 70–80th, 80–90th, 90th–92nd, 92nd–94th, 96–98th, 98–99th, and 99–100th percentiles to highlight variation within sub-Saharan Africa. ATG=Antigua and Barbuda. VCT=Saint Vincent and the Grenadines. LCA=Saint Lucia. TTO=Trinidad and Tobago. TLS=Timor-Leste. FSM=Federated States of Micronesia. Figure 3 Number of people living with HIV receiving ART from 1995 to 2015, and the proportion living with HIV receiving ART in 2015 Number of people living with HIV receiving ART by region (A). Bars represent the mean number of people living with HIV who received ART within a given year. Error bars represent 95% uncertainty intervals. Each Global Burden of Disease (GBD) region is represented by a separate colour. Proportion of people living with HIV receiving ART by country (B). The number of people living with HIV receiving ART and the total number of people living with HIV are year-end point prevalences. ART=antiretroviral therapy. ATG=Antigua and Barbuda. VCT=Saint Vincent and the Grenadines. LCA=Saint Lucia. TTO=Trinidad and Tobago. TLS=Timor-Leste. FSM=Federated States of Micronesia. Figure 4 Global HIV/AIDS deaths, 2005–15

Number of people living with HIV receiving ART by region (A). Bars represent the mean number of people living with HIV who received ART within a given year. Error bars represent 95% uncertainty intervals. Each Global Burden of Disease (GBD) region is represented by a separate colour. Proportion of people living with HIV receiving ART by country (B). The number of people living with HIV receiving ART and the total number of people living with HIV are year-end point prevalences. ART=antiretroviral therapy. ATG=Antigua and Barbuda. VCT=Saint Vincent and the Grenadines. LCA=Saint Lucia. TTO=Trinidad and Tobago. TLS=Timor-Leste. FSM=Federated States of Micronesia. Figure 4 Global HIV/AIDS deaths, 2005–15 Global deaths caused by HIV/AIDS resulting in either mycobacterial infection (tuberculosis) or other diseases, by age and sex in 2015 (A); dark shading indicates deaths caused by tuberculosis associated with HIV; light shading indicates deaths caused by other diseases resulting from HIV; error bars show 95% uncertainty intervals. Mean estimates of global and super-regional HIV/AIDS deaths per prevalent case fom 2005 to 2015 (B). Figure 5 Comparison of GBD 2015 and UNAIDS 2014 estimates

Global deaths caused by HIV/AIDS resulting in either mycobacterial infection (tuberculosis) or other diseases, by age and sex in 2015 (A); dark shading indicates deaths caused by tuberculosis associated with HIV; light shading indicates deaths caused by other diseases resulting from HIV; error bars show 95% uncertainty intervals. Mean estimates of global and super-regional HIV/AIDS deaths per prevalent case fom 2005 to 2015 (B). Figure 5 Comparison of GBD 2015 and UNAIDS 2014 estimates Adult HIV prevalence rate (A) and estimates of death caused by HIV/AIDS (B). UNAIDS' published prevalence values are limited to three decimal places. The x and y values of each point are the log transformation of the mean estimates from UNAIDS and GBD, respectively, enabling variation to be seen despite disparate values. Tick-mark labels on the x and y axes are the value of the mean estimate before log transformation (ie, the real value and not the log-transformed value is shown). Locations mentioned in the manuscript are highlighted by plotting the ISO 3 code of the location. Each location is plotted with a different colour by super-region. GBD=Global Burden of Disease. UNAIDS=the Joint United Nations Programme on HIV and AIDS. ZAF=South Africa. KEN=Kenya. NGA=Nigeria. COG=Democratic Republic of the Congo. SLE=Sierra Leone. BDI=Burkina Faso. COD=Congo. SEN=Senegal. SWZ=Swaziland. Table Country-specific estimates of new HIV infections, people living with HIV, HIV/AIDS deaths, and ART coverage in 2015, and ARCs of age-standardised incidence, prevalence, and mortality from 2005 to 2015

Adult HIV prevalence rate (A) and estimates of death caused by HIV/AIDS (B). UNAIDS' published prevalence values are limited to three decimal places. The x and y values of each point are the log transformation of the mean estimates from UNAIDS and GBD, respectively, enabling variation to be seen despite disparate values. Tick-mark labels on the x and y axes are the value of the mean estimate before log transformation (ie, the real value and not the log-transformed value is shown). Locations mentioned in the manuscript are highlighted by plotting the ISO 3 code of the location. Each location is plotted with a different colour by super-region. GBD=Global Burden of Disease. UNAIDS=the Joint United Nations Programme on HIV and AIDS. ZAF=South Africa. KEN=Kenya. NGA=Nigeria. COG=Democratic Republic of the Congo. SLE=Sierra Leone. BDI=Burkina Faso. COD=Congo. SEN=Senegal. SWZ=Swaziland. Table Country-specific estimates of new HIV infections, people living with HIV, HIV/AIDS deaths, and ART coverage in 2015, and ARCs of age-standardised incidence, prevalence, and mortality from 2005 to 2015 New infections (in thousands) People living with HIV (in thousands) HIV/AIDS deaths (in thousands) ART coverage per 100 people living with HIV (%) Age-standardised incidence ARC from 2005 to 2015 Age-standardised prevalence ARC from 2005 to 2015 Age-standardised mortality ARC from 2005 to 2015 Global 2450·92 (2236·13 to 2686·79) 38 802·50 (37 635·88 to 40 371·67) 1192·57 (1131·11 to 1270·05) 40·60 (39·36 to 41·80) −0·02 (−0·03 to −0·01) 0·01 (0·00 to 0·01) −0·05 (−0·06 to −0·05) High SDI 101·75 (75·18 to 146·96) 2204·18 (1751·36 to 2799·27) 33·51 (31·96 to 35·43) 51·49 (43·90 to 57·55) 0·01 (−0·01 to 0·04) 0·01 (0·00 to 0·02) −0·01 (−0·02 to −0·01) High-to-middle SDI 646·76 (557·07 to 748·55) 10 421·94 (9873·26 to 10 989·83) 240·15 (224·08 to 259·28) 48·01 (45·99 to 50·13) −0·02 (−0·03 to −0·01) 0·01 (0·01 to 0·02) −0·05 (−0·06 to −0·05) Middle SDI 298·33 (238·18 to 394·91) 4155·45 (3616·14 to 5163·64) 131·57 (111·27 to 183·00) 37·66 (32·68 to 40·83) 0·00 (−0·02 to 0·02) 0·02 (0·01 to 0·03) −0·02 (−0·04 to −0·00) Low-to-middle SDI 796·30 (655·25 to 951·76) 11 783·44 (11 251·57 to 12 472·97) 408·87 (368·10 to 457·42) 35·48 (33·62 to 37·52) −0·01 (−0·03 to 0·01) −0·00 (−0·01 to 0·00) −0·06 (−0·07 to −0·06) Low SDI 606·54 (510·14 to 707·36) 10 213·36 (9762·90 to 10 684·37) 377·68 (350·43 to 408·08) 37·89 (35·93 to 39·79) −0·05 (−0·07 to −0·03) −0·01 (−0·02 to −0·01) −0·08 (−0·09 to −0·07) High-income 45·67 (37·88 to 53·92) 1660·18 (1359·94 to 1997·98) 13·95 (13·79 to 14·11) 66·91 (64·76 to 69·43) −0·01 (−0·02 to −0·00) −0·00 (−0·01 to 0·00) −0·06 (−0·06 to −0·05) High-income North America 24·16 (18·76 to 31·10) 882·60 (692·93 to 1136·45) 7·89 (7·79 to 7·98) 69·86 (66·81 to 73·51) −0·02 (−0·04 to −0·01) −0·00 (−0·01 to 0·00) −0·07 (−0·07 to −0·07) Canada 1·11 (0·18 to 2·81) 49·25 (15·89 to 102·34) 0·31 (0·29 to 0·33) 64·14 (56·58 to 73·43) −0·03 (−0·14 to 0·02) −0·01 (−0·03 to 0·01) −0·06 (−0·07 to −0·05) Greenland 0·00 (0·00 to 0·01) 0·23 (0·06 to 0·55) 0·00 (0·00 to 0·00) 61·88 (52·84 to 69·43) −0·10 (−0·65 to −0·02) −0·01 (−0·03 to 0·00) −0·03 (−0·07 to 0·01) USA 23·04 (17·68 to 29·96) 833·03 (648·62 to 1078·06) 7·57 (7·48 to 7·67) 70·18 (67·09 to 74·00) −0·02 (−

to 0·02) −0·01 (−0·03 to 0·01) −0·06 (−0·07 to −0·05) Greenland 0·00 (0·00 to 0·01) 0·23 (0·06 to 0·55) 0·00 (0·00 to 0·00) 61·88 (52·84 to 69·43) −0·10 (−0·65 to −0·02) −0·01 (−0·03 to 0·00) −0·03 (−0·07 to 0·01) USA 23·04 (17·68 to 29·96) 833·03 (648·62 to 1078·06) 7·57 (7·48 to 7·67) 70·18 (67·09 to 74·00) −0·02 (− 0·04 to −0·01) −0·00 (−0·01 to 0·00) −0·07 (−0·07 to −0·07) Australasia 0·45 (0·19 to 0·89) 18·69 (7·37 to 37·10) 0·10 (0·09 to 0·10) 62·24 (57·73 to 67·54) −0·02 (−0·04 to −0·01) −0·00 (−0·01 to 0·00) −0·04 (−0·05 to −0·03) Australia 0·39 (0·15 to 0·84) 16·24 (5·20 to 34·28) 0·09 (0·08 to 0·09) 62·38 (57·07 to 68·35) −0·02 (−0·03 to −0·01) −0·01 (−0·02 to 0·00) −0·04 (−0·05 to −0·03) New Zealand 0·06 (0·02 to 0·13) 2·45 (0·71 to 5·38) 0·01 (0·01 to 0·01) 60·68 (54·45 to 68·06) −0·03 (−0·10 to −0·01) 0·00 (−0·01 to 0·01) −0·04 (−0·05 to −0·03) High-income Asia Pacific 0·75 (0·55 to 1·02) 22·06 (14·82 to 35·17) 0·32 (0·31 to 0·33) 49·98 (45·98 to 53·54) −0·03 (−0·09 to −0·00) 0·02 (0·01 to 0·03) −0·01 (−0·01 to −0·01) Brunei 0·01 (0·00 to 0·03) 0·26 (0·09 to 0·59) 0·00 (0·00 to 0·00) 37·66 (29·23 to 47·75) −0·03 (−0·16 to 0·02) 0·02 (−0·00 to 0·04) −0·02 (−0·06 to 0·01) Japan 0·50 (0·40 to 0·60) 10·41 (8·40 to 12·69) 0·17 (0·17 to 0·17) 57·43 (55·29 to 60·02) 0·01 (−0·00 to 0·02) 0·04 (0·03 to 0·04) −0·03 (−0·03 to −0·03) Singapore 0·05 (0·02 to 0·10) 1·85 (0·60 to 4·06) 0·01 (0·01 to 0·01) 54·61 (45·56 to 64·61) 0·01 (−0·05 to 0·05) 0·01 (−0·01 to 0·02) 0·12 (0·11 to 0·12) South Korea 0·19 (0·02 to 0·43) 9·54 (2·92 to 21·96) 0·14 (0·13 to 0·14) 39·34 (31·76 to 47·08) −0·12 (−0·32 to −0·04) 0·00 (−0·01 to 0·02) 0·01 (0·01 to 0·02) Western Europe 12·89 (9·48 to 16·95) 651·38 (448·53 to 896·75) 3·42 (3·35 to 3·50) 63·81 (60·91 to 67·06) −0·03 (−0·04 to −0·02) −0·01 (−0·01 to −0·00) −0·06 (−0·06 to −0·05) Andorra 0·00 (0·00 to 0·01) 0·21 (0·02 to 1·37) 0·00 (0·00 to 0·01) 57·49 (32·52 to 80·56) −0·04 (−0·72 to 0·10) 0·01 (−0·03 to 0·08) −0·01 (−0·08 to 0·08) Austria 0·31 (0·10 to 0·71) 11·65 (2·72 to 28·30) 0·04 (0·04 to 0·04) 55·15 (48·70 to 62·52) −0·04 (−0·09 to −0·01) 0·01 (−0·00 to 0·02) −0·06 (−0·06 to −0·05) Belgium 0·21 (0·06 to 0·47) 10·68 (2·90 to 25·23) 0·05 (0·05 to 0·05) 61·74 (55·20 to 68·73) −0·03 (−0·12 to 0·00) −0·00 (−0·02 to 0·01) −0·04 (−0·05 to −0·04) Cyprus 0·01 (0·00 to 0·03) 0·39 (0·11 to 0·88) 0·00 (0·00 to 0·00) 48·50 (40·52 to 58·86) −0·06 (−0·67 to 0·02) 0·01 (−0·01 to 0·03) 0·00 (−0·04 to 0·03) Denmark 0·13 (0·03 to 0·30) 7·67 (2·13 to 1

90 to 25·23) 0·05 (0·05 to 0·05) 61·74 (55·20 to 68·73) −0·03 (−0·12 to 0·00) −0·00 (−0·02 to 0·01) −0·04 (−0·05 to −0·04) Cyprus 0·01 (0·00 to 0·03) 0·39 (0·11 to 0·88) 0·00 (0·00 to 0·00) 48·50 (40·52 to 58·86) −0·06 (−0·67 to 0·02) 0·01 (−0·01 to 0·03) 0·00 (−0·04 to 0·03) Denmark 0·13 (0·03 to 0·30) 7·67 (2·13 to 1 5·27) 0·03 (0·02 to 0·03) 62·61 (55·63 to 70·23) −0·06 (−0·19 to −0·01) −0·00 (−0·02 to 0·01) −0·05 (−0·06 to −0·04) Finland 0·03 (0·01 to 0·08) 1·35 (0·36 to 3·09) 0·01 (0·01 to 0·01) 57·85 (51·54 to 64·82) −0·06 (−0·19 to −0·02) 0·00 (−0·01 to 0·02) −0·04 (−0·05 to −0·04) France 0·96 (0·36 to 2·04) 79·17 (23·19 to 175·70) 0·49 (0·46 to 0·52) 63·37 (54·81 to 71·07) −0·04 (−0·08 to −0·02) −0·02 (−0·04 to −0·01) −0·07 (−0·08 to −0·07) Germany 1·76 (0·65 to 3·66) 60·55 (17·98 to 129·32) 0·43 (0·41 to 0·46) 55·55 (47·85 to 64·54) −0·01 (−0·04 to 0·01) 0·01 (−0·00 to 0·03) −0·03 (−0·04 to −0·03) Greece 0·05 (0·03 to 0·09) 1·22 (0·59 to 2·18) 0·02 (0·02 to 0·02) 39·67 (30·76 to 49·54) 0·01 (−0·02 to 0·03) 0·01 (−0·00 to 0·02) −0·02 (−0·03 to −0·01) Iceland 0·01 (0·00 to 0·02) 0·18 (0·05 to 0·42) 0·00 (0·00 to 0·00) 50·06 (40·56 to 61·16) −0·01 (−0·18 to 0·03) 0·01 (−0·01 to 0·03) −0·04 (−0·05 to −0·04) Ireland 0·06 (0·01 to 0·14) 2·55 (0·70 to 5·85) 0·01 (0·01 to 0·01) 58·51 (51·76 to 66·14) −0·03 (−0·15 to −0·00) −0·00 (−0·02 to 0·01) −0·00 (−0·01 to 0·01) Israel 0·17 (0·05 to 0·35) 4·85 (1·39 to 10·45) 0·04 (0·03 to 0·04) 50·92 (44·16 to 58·93) −0·01 (−0·09 to 0·01) 0·01 (−0·01 to 0·02) −0·02 (−0·03 to −0·01) Italy 1·96 (0·76 to 4·19) 137·07 (43·51 to 276·32) 0·61 (0·57 to 0·64) 67·08 (62·08 to 72·04) −0·05 (−0·07 to −0·03) −0·01 (−0·02 to −0·00) −0·03 (−0·03 to −0·02) Luxembourg 0·01 (0·00 to 0·03) 0·41 (0·12 to 0·96) 0·00 (0·00 to 0·00) 52·04 (43·46 to 61·91) 0·00 (−0·09 to 0·03) 0·01 (−0·01 to 0·02) −0·05 (−0·05 to −0·04) Malta 0·01 (0·00 to 0·02) 0·26 (0·08 to 0·58) 0·00 (0·00 to 0·00) 48·30 (38·71 to 59·59) 0·01 (−0·08 to 0·04) 0·02 (−0·00 to 0·04) −0·03 (−0·03 to −0·02) Netherlands 0·20 (0·07 to 0·47) 14·56 (4·14 to 32·34) 0·05 (0·05 to 0·05) 69·53 (62·81 to 76·01) −0·02 (−0·07 to 0·01) −0·02 (−0·03 to −0·01) −0·07 (−0·07 to −0·06) Norway 0·05 (0·02 to 0·11) 2·77 (0·78 to 6·18) 0·01 (0·01 to 0·01) 63·51 (57·61 to 69·83) −0·02 (−0·09 to −0·00) −0·01 (−0·02 to 0·00) −0·08 (−0·09 to −0·07) Portugal 2·22 (0·53 to 4·91) 115·25 (32·31 to 263·86) 0·53 (0·50 to 0·56) 60·58 (54·02 to 66·88) −0·04 (−0·13 to −0·01) −0·01 (−0·02 to 0·00) −0·07 (−0·08 to −0·07) Spain 2·35

0·02 to 0·11) 2·77 (0·78 to 6·18) 0·01 (0·01 to 0·01) 63·51 (57·61 to 69·83) −0·02 (−0·09 to −0·00) −0·01 (−0·02 to 0·00) −0·08 (−0·09 to −0·07) Portugal 2·22 (0·53 to 4·91) 115·25 (32·31 to 263·86) 0·53 (0·50 to 0·56) 60·58 (54·02 to 66·88) −0·04 (−0·13 to −0·01) −0·01 (−0·02 to 0·00) −0·07 (−0·08 to −0·07) Spain 2·35 (0·99 to 4·76) 130·33 (39·66 to 281·12) 0·82 (0·77 to 0·87) 65·54 (56·76 to 73·75) 0·01 (−0·01 to 0·02) −0·02 (−0·03 to −0·01) −0·08 (−0·09 to −0·07) Sweden 0·08 (0·03 to 0·15) 3·69 (1·62 to 6·61) 0·02 (0·02 to 0·02) 76·01 (71·06 to 82·01) −0·01 (−0·06 to 0·01) −0·00 (−0·02 to 0·01) −0·05 (−0·05 to −0·04) Switzerland 0·20 (0·05 to 0·45) 13·03 (3·77 to 28·24) 0·04 (0·04 to 0·04) 69·48 (64·10 to 75·77) 0·00 (−0·09 to 0·03) −0·01 (−0·03 to −0·01) −0·06 (−0·07 to −0·06) UK 2·06 (1·66 to 2·54) 52·67 (41·67 to 66·15) 0·22 (0·21 to 0·22) 61·21 (58·45 to 64·08) −0·04 (−0·04 to −0·03) 0·02 (0·02 to 0·03) −0·03 (−0·04 to −0·03) Southern Latin America 7·42 (3·55 to 10·30) 85·45 (56·64 to 122·34) 2·23 (2·13 to 2·32) 63·83 (58·62 to 69·84) 0·04 (−0·05 to 0·07) 0·02 (0·00 to 0·03) −0·01 (−0·02 to −0·01) Argentina 6·32 (2·58 to 9·20) 62·94 (36·49 to 96·26) 1·60 (1·51 to 1·70) 69·73 (64·04 to 76·37) 0·07 (−0·05 to 0·09) 0·03 (0·00 to 0·04) −0·01 (−0·02 to −0·01) Chile 0·71 (0·43 to 1·15) 16·25 (7·33 to 32·90) 0·46 (0·43 to 0·49) 45·88 (33·79 to 57·54) −0·05 (−0·08 to −0·01) −0·01 (−0·02 to 0·01) −0·00 (−0·01 to 0·00) Uruguay 0·38 (0·20 to 0·64) 6·26 (2·83 to 11·94) 0·16 (0·15 to 0·18) 46·96 (39·24 to 56·35) −0·01 (−0·05 to 0·03) 0·01 (−0·01 to 0·02) −0·01 (−0·02 to −0·00) Eastern Europe, central Europe, and central Asia 78·25 (52·91 to 122·49) 940·86 (617·41 to 1490·53) 28·38 (26·94 to 30·12) 20·07 (16·88 to 24·38) 0·02 (−0·01 to 0·06) 0·03 (0·02 to 0·04) 0·01 (0·00 to 0·01) Eastern Europe 73·10 (48·14 to 117·64) 864·89 (547·01 to 1413·14) 26·09 (24·67 to 27·65) 18·69 (15·34 to 23·40) 0·02 (−0·01 to 0·06) 0·03 (0·02 to 0·05) 0·01 (0·01 to 0·02) Belarus 1·37 (0·76 to 2·29) 17·50 (8·74 to 29·52) 0·59 (0·41 to 0·96) 35·42 (27·69 to 46·87) 0·01 (−0·03 to 0·05) 0·04 (0·02 to 0·06) 0·02 (−0·00 to 0·05) Estonia 0·11 (0·06 to 0·19) 1·62 (0·81 to 2·93) 0·03 (0·03 to 0·04) 31·07 (25·55 to 36·79) −0·02 (−0·04 to 0·00) 0·05 (0·03 to 0·06) −0·01 (−0·02 to 0·01) Latvia 0·17 (0·05 to 0·35) 2·93 (1·42 to 5·82) 0·11 (0·10 to 0·12) 16·62 (11·92 to 23·32) −0·05 (−0·16 to 0·00) 0·00 (−0·03 to 0·04) 0·04 (0·03 to 0·05) Lithuania 0·08 (0·01 to 0·17) 1·67 (0·83 to 3·17) 0·06 (0·06 to 0·07) 22·13 (16

31·07 (25·55 to 36·79) −0·02 (−0·04 to 0·00) 0·05 (0·03 to 0·06) −0·01 (−0·02 to 0·01) Latvia 0·17 (0·05 to 0·35) 2·93 (1·42 to 5·82) 0·11 (0·10 to 0·12) 16·62 (11·92 to 23·32) −0·05 (−0·16 to 0·00) 0·00 (−0·03 to 0·04) 0·04 (0·03 to 0·05) Lithuania 0·08 (0·01 to 0·17) 1·67 (0·83 to 3·17) 0·06 (0·06 to 0·07) 22·13 (16 ·73 to 29·50) −0·06 (−0·21 to −0·01) 0·00 (−0·03 to 0·03) 0·01 (−0·01 to 0·02) Moldova 0·54 (0·31 to 0·92) 7·94 (3·67 to 14·59) 0·18 (0·16 to 0·21) 21·43 (15·09 to 30·21) −0·01 (−0·03 to 0·01) 0·03 (0·02 to 0·04) −0·03 (−0·04 to −0·02) Russia 57·34 (32·75 to 102·27) 607·05 (312·14 to 1107·70) 17·89 (16·58 to 19·33) 13·91 (10·90 to 17·43) 0·05 (0·01 to 0·10) 0·05 (0·03 to 0·06) 0·03 (0·02 to 0·04) Ukraine 13·49 (9·92 to 18·67) 226·16 (132·70 to 360·43) 7·22 (6·52 to 8·01) 28·19 (21·83 to 36·31) −0·04 (−0·06 to −0·02) 0·01 (−0·00 to 0·01) −0·01 (−0·03 to −0·00) Central Europe 1·19 (0·82 to 1·55) 19·79 (14·35 to 26·53) 0·42 (0·39 to 0·49) 46·47 (41·14 to 52·17) 0·00 (−0·03 to 0·02) 0·02 (0·00 to 0·03) −0·04 (−0·04 to −0·02) Albania 0·00 (0·00 to 0·01) 0·07 (0·02 to 0·14) 0·00 (0·00 to 0·00) 46·34 (32·88 to 63·31) −0·07 (−0·57 to 0·04) 0·00 (−0·04 to 0·04) −0·00 (−0·04 to 0·04) Bosnia and Herzegovina 0·00 (0·00 to 0·01) 0·10 (0·03 to 0·21) 0·00 (0·00 to 0·01) 48·18 (35·41 to 63·19) −0·06 (−0·50 to 0·04) 0·00 (−0·03 to 0·03) −0·00 (−0·04 to 0·06) Bulgaria 0·14 (0·06 to 0·26) 1·86 (0·83 to 4·00) 0·05 (0·05 to 0·06) 17·02 (12·25 to 22·96) −0·00 (−0·06 to 0·03) 0·01 (−0·02 to 0·03) −0·05 (−0·06 to −0·03) Croatia 0·02 (0·00 to 0·02) 0·34 (0·16 to 0·60) 0·01 (0·01 to 0·01) 53·33 (43·15 to 65·99) −0·02 (−0·12 to 0·01) 0·01 (−0·01 to 0·03) 0·01 (−0·00 to 0·02) Czech Republic 0·04 (0·01 to 0·07) 0·75 (0·37 to 1·21) 0·01 (0·01 to 0·02) 53·80 (46·23 to 62·90) 0·01 (−0·10 to 0·05) 0·02 (−0·00 to 0·04) 0·03 (0·01 to 0·04) Hungary 0·06 (0·04 to 0·08) 1·24 (0·66 to 1·99) 0·04 (0·04 to 0·05) 45·77 (37·00 to 56·61) −0·02 (−0·05 to 0·00) −0·02 (−0·03 to −0·00) −0·08 (−0·09 to −0·06) Macedonia 0·01 (0·00 to 0·01) 0·09 (0·03 to 0·18) 0·00 (0·00 to 0·00) 40·71 (29·03 to 59·30) −0·06 (−0·53 to 0·04) 0·03 (−0·02 to 0·07) 0·02 (−0·02 to 0·08) Montenegro 0·00 (0·00 to 0·01) 0·05 (0·02 to 0·10) 0·00 (0·00 to 0·00) 42·44 (30·59 to 60·43) −0·04 (−0·53 to 0·07) 0·02 (−0·02 to 0·06) 0·01 (−0·03 to 0·07) Poland 0·41 (0·13 to 0·67) 7·71 (3·88 to 12·56) 0·14 (0·13 to 0·15) 56·60 (48·73 to 67·12) −0·00 (−0·11 to 0·03) 0·01 (−0·01 to 0·03) −0·02 (−0·03 to −0·01) Romania 0·45 (0·15 to 0·69) 6·33 (3·17

05 (0·02 to 0·10) 0·00 (0·00 to 0·00) 42·44 (30·59 to 60·43) −0·04 (−0·53 to 0·07) 0·02 (−0·02 to 0·06) 0·01 (−0·03 to 0·07) Poland 0·41 (0·13 to 0·67) 7·71 (3·88 to 12·56) 0·14 (0·13 to 0·15) 56·60 (48·73 to 67·12) −0·00 (−0·11 to 0·03) 0·01 (−0·01 to 0·03) −0·02 (−0·03 to −0·01) Romania 0·45 (0·15 to 0·69) 6·33 (3·17 to 10·39) 0·09 (0·08 to 0·10) 43·39 (34·99 to 54·03) 0·03 (−0·07 to 0·06) 0·04 (0·01 to 0·06) −0·06 (−0·07 to −0·05) Serbia 0·04 (0·02 to 0·08) 0·87 (0·38 to 2·21) 0·05 (0·03 to 0·12) 26·76 (20·80 to 32·60) −0·08 (−0·11 to −0·05) −0·00 (−0·02 to 0·01) 0·08 (0·03 to 0·13) Slovakia 0·02 (0·01 to 0·03) 0·24 (0·11 to 0·42) 0·01 (0·00 to 0·01) 46·16 (36·95 to 55·76) 0·02 (0·00 to 0·04) 0·04 (0·01 to 0·06) 0·00 (−0·03 to 0·02) Slovenia 0·01 (0·00 to 0·01) 0·14 (0·06 to 0·27) 0·00 (0·00 to 0·00) 58·71 (45·13 to 71·99) 0·02 (−0·01 to 0·04) −0·00 (−0·02 to 0·02) −0·07 (−0·08 to −0·06) Central Asia 3·96 (2·64 to 5·58) 56·19 (39·49 to 79·69) 1·87 (1·52 to 2·57) 30·50 (25·59 to 36·68) −0·01 (−0·05 to 0·03) 0·01 (−0·01 to 0·02) −0·03 (−0·05 to −0·02) Armenia 0·07 (0·03 to 0·13) 0·57 (0·29 to 1·04) 0·02 (0·01 to 0·02) 21·59 (17·29 to 27·60) 0·07 (−0·03 to 0·26) 0·07 (0·02 to 0·12) 0·05 (0·00 to 0·09) Azerbaijan 0·36 (0·17 to 0·58) 4·06 (1·89 to 7·45) 0·11 (0·07 to 0·21) 32·83 (23·57 to 47·48) 0·04 (−0·07 to 0·09) 0·02 (−0·01 to 0·05) −0·06 (−0·09 to −0·02) Georgia 0·15 (0·08 to 0·25) 1·70 (0·98 to 2·56) 0·03 (0·03 to 0·04) 38·75 (33·41 to 44·83) 0·04 (−0·01 to 0·09) 0·13 (0·10 to 0·15) 0·12 (0·09 to 0·15) Kazakhstan 1·63 (0·88 to 2·64) 17·70 (8·32 to 31·95) 0·31 (0·27 to 0·36) 24·79 (19·36 to 32·88) 0·09 (0·06 to 0·14) 0·04 (0·02 to 0·05) −0·05 (−0·07 to −0·04) Kyrgyzstan 0·32 (0·16 to 0·62) 6·62 (3·03 to 13·08) 0·30 (0·21 to 0·46) 32·24 (26·28 to 38·78) −0·07 (−0·14 to −0·01) 0·03 (0·01 to 0·05) 0·02 (0·01 to 0·04) Mongolia 0·01 (0·00 to 0·01) 0·08 (0·03 to 0·17) 0·00 (0·00 to 0·01) 26·46 (17·72 to 41·13) 0·03 (−0·12 to 0·18) −0·01 (−0·06 to 0·02) −0·06 (−0·09 to −0·03) Tajikistan 0·32 (0·14 to 0·61) 4·64 (2·20 to 8·86) 0·18 (0·13 to 0·30) 27·27 (20·60 to 36·20) −0·03 (−0·12 to 0·06) −0·01 (−0·03 to 0·02) −0·05 (−0·08 to −0·02) Turkmenistan 0·79 (0·10 to 1·96) 9·22 (3·17 to 19·29) 0·35 (0·22 to 0·56) 21·76 (13·98 to 36·05) 0·01 (−0·18 to 0·16) 0·02 (−0·05 to 0·07) −0·01 (−0·05 to 0·04) Uzbekistan 0·31 (0·10 to 0·68) 11·59 (5·53 to 26·29) 0·57 (0·35 to 1·16) 40·92 (29·74 to 52·77) −0·17 (−0·29 to −0·08) −0·05 (−0·06 to −0·03) −0·06 (−0·09 to −0·02) Latin America and Caribbea

9·22 (3·17 to 19·29) 0·35 (0·22 to 0·56) 21·76 (13·98 to 36·05) 0·01 (−0·18 to 0·16) 0·02 (−0·05 to 0·07) −0·01 (−0·05 to 0·04) Uzbekistan 0·31 (0·10 to 0·68) 11·59 (5·53 to 26·29) 0·57 (0·35 to 1·16) 40·92 (29·74 to 52·77) −0·17 (−0·29 to −0·08) −0·05 (−0·06 to −0·03) −0·06 (−0·09 to −0·02) Latin America and Caribbea n 85·47 (77·62 to 94·22) 1322·07 (1194·38 to 1474·60) 46·81 (43·27 to 50·98) 45·10 (43·68 to 46·49) −0·00 (−0·01 to 0·00) 0·01 (0·00 to 0·01) −0·02 (−0·03 to −0·02) Central Latin America 29·38 (24·91 to 34·23) 394·06 (328·88 to 465·79) 12·31 (12·01 to 12·71) 40·01 (38·29 to 41·84) 0·01 (0·00 to 0·03) 0·02 (0·02 to 0·03) −0·02 (−0·02 to −0·02) Colombia 6·15 (3·42 to 10·00) 73·95 (36·96 to 131·07) 2·42 (2·30 to 2·56) 29·75 (24·34 to 36·98) 0·03 (0·00 to 0·07) 0·02 (0·01 to 0·04) −0·02 (−0·03 to −0·02) Costa Rica 0·35 (0·22 to 0·50) 6·66 (3·38 to 10·89) 0·15 (0·14 to 0·16) 50·08 (43·35 to 56·40) −0·03 (−0·05 to −0·02) 0·01 (0·01 to 0·02) −0·02 (−0·03 to −0·01) El Salvador 0·80 (0·47 to 1·21) 16·11 (8·09 to 27·60) 0·55 (0·42 to 0·77) 46·22 (40·88 to 50·96) −0·05 (−0·07 to −0·03) 0·01 (−0·00 to 0·02) 0·02 (−0·01 to 0·04) Guatemala 1·67 (0·84 to 2·96) 27·74 (12·99 to 49·27) 0·68 (0·65 to 0·71) 42·04 (36·62 to 47·33) −0·03 (−0·10 to 0·03) −0·00 (−0·02 to 0·02) −0·06 (−0·06 to −0·05) Honduras 1·48 (0·88 to 2·29) 19·82 (12·09 to 30·07) 0·59 (0·50 to 0·73) 40·41 (35·61 to 45·45) 0·01 (−0·04 to 0·05) 0·02 (−0·01 to 0·04) −0·03 (−0·06 to 0·00) Mexico 12·47 (11·23 to 13·83) 169·52 (147·48 to 194·76) 5·17 (5·11 to 5·24) 45·67 (43·88 to 47·67) 0·01 (0·00 to 0·02) 0·03 (0·02 to 0·03) −0·02 (−0·03 to −0·02) Nicaragua 0·97 (0·49 to 1·68) 7·93 (4·03 to 13·74) 0·19 (0·16 to 0·25) 22·22 (19·37 to 25·34) 0·09 (0·04 to 0·14) 0·09 (0·06 to 0·11) 0·02 (−0·00 to 0·05) Panama 1·80 (0·96 to 3·16) 18·92 (9·91 to 31·50) 0·51 (0·38 to 0·75) 38·52 (33·38 to 43·53) 0·12 (0·05 to 0·25) 0·02 (0·01 to 0·04) −0·02 (−0·05 to 0·01) Venezuela 3·68 (1·64 to 6·45) 53·41 (27·48 to 95·77) 2·04 (1·94 to 2·15) 33·10 (28·61 to 38·45) −0·01 (−0·10 to 0·03) 0·02 (−0·00 to 0·04) 0·01 (0·00 to 0·02) Andean Latin America 3·83 (2·68 to 5·33) 49·31 (32·11 to 71·86) 1·81 (1·54 to 2·20) 34·59 (31·27 to 38·93) 0·01 (−0·01 to 0·04) 0·02 (0·01 to 0·04) −0·01 (−0·03 to 0·00) Bolivia 0·19 (0·10 to 0·31) 2·23 (1·01 to 4·17) 0·10 (0·07 to 0·14) 18·69 (15·60 to 21·98) 0·02 (−0·02 to 0·07) 0·02 (−0·00 to 0·04) −0·01 (−0·03 to 0·02) Ecuador 2·02 (1·14 to 3·19) 23·39 (12·33 to 39·59) 0·79 (0·64 to 1·09) 34·40 (30·49 to 38·80) 0

38·93) 0·01 (−0·01 to 0·04) 0·02 (0·01 to 0·04) −0·01 (−0·03 to 0·00) Bolivia 0·19 (0·10 to 0·31) 2·23 (1·01 to 4·17) 0·10 (0·07 to 0·14) 18·69 (15·60 to 21·98) 0·02 (−0·02 to 0·07) 0·02 (−0·00 to 0·04) −0·01 (−0·03 to 0·02) Ecuador 2·02 (1·14 to 3·19) 23·39 (12·33 to 39·59) 0·79 (0·64 to 1·09) 34·40 (30·49 to 38·80) 0 ·02 (−0·01 to 0·06) 0·04 (0·01 to 0·06) 0·01 (−0·02 to 0·03) Peru 1·62 (0·96 to 2·50) 23·70 (12·56 to 41·07) 0·92 (0·75 to 1·22) 35·94 (30·57 to 42·07) 0·00 (−0·03 to 0·03) 0·01 (−0·00 to 0·02) −0·03 (−0·04 to −0·02) Caribbean 17·29 (12·72 to 23·36) 307·45 (272·69 to 342·05) 11·28 (9·73 to 12·89) 46·11 (42·49 to 49·62) −0·02 (−0·04 to 0·01) −0·01 (−0·02 to 0·00) −0·07 (−0·08 to −0·06) Antigua and Barbuda 0·02 (0·01 to 0·04) 0·34 (0·14 to 0·71) 0·01 (0·01 to 0·01) 39·68 (29·72 to 49·29) 0·01 (−0·07 to 0·14) 0·00 (−0·02 to 0·05) −0·03 (−0·04 to −0·03) The Bahamas 0·11 (0·07 to 0·18) 3·33 (1·75 to 5·39) 0·11 (0·07 to 0·17) 51·08 (42·61 to 61·04) −0·07 (−0·10 to −0·03) −0·01 (−0·02 to 0·01) −0·05 (−0·07 to −0·03) Barbados 0·06 (0·03 to 0·10) 1·07 (0·50 to 1·86) 0·03 (0·02 to 0·03) 46·43 (38·95 to 55·50) 0·02 (−0·02 to 0·05) 0·00 (−0·01 to 0·01) −0·04 (−0·05 to −0·03) Belize 0·19 (0·12 to 0·28) 3·03 (1·58 to 5·21) 0·10 (0·06 to 0·17) 58·12 (53·15 to 63·65) 0·02 (−0·02 to 0·05) 0·00 (−0·01 to 0·01) 0·00 (−0·04 to 0·04) Bermuda 0·02 (0·01 to 0·04) 0·36 (0·15 to 0·77) 0·01 (0·01 to 0·01) 40·56 (30·34 to 50·33) 0·01 (−0·08 to 0·14) −0·00 (−0·02 to 0·04) −0·03 (−0·04 to −0·02) Cuba 1·14 (0·70 to 1·78) 18·71 (9·53 to 30·35) 0·32 (0·31 to 0·34) 62·31 (56·48 to 68·60) −0·00 (−0·03 to 0·02) 0·09 (0·06 to 0·11) 0·07 (0·06 to 0·08) Dominica 0·01 (0·00 to 0·03) 0·20 (0·08 to 0·41) 0·01 (0·00 to 0·01) 37·59 (28·28 to 46·01) 0·01 (−0·08 to 0·16) 0·01 (−0·01 to 0·06) −0·01 (−0·04 to 0·04) Dominican Republic 2·82 (2·03 to 3·84) 55·93 (48·99 to 62·58) 1·99 (1·34 to 2·55) 39·80 (35·00 to 45·75) 0·01 (−0·05 to 0·12) −0·04 (−0·05 to −0·03) −0·11 (−0·15 to −0·08) Grenada 0·02 (0·01 to 0·05) 0·33 (0·14 to 0·70) 0·01 (0·01 to 0·02) 33·58 (25·44 to 41·30) 0·00 (−0·08 to 0·13) 0·01 (−0·01 to 0·06) −0·01 (−0·04 to 0·04) Guyana 1·14 (0·65 to 1·78) 18·89 (8·80 to 32·29) 0·47 (0·30 to 0·71) 62·25 (54·32 to 69·67) −0·01 (−0·03 to 0·01) 0·02 (0·00 to 0·05) 0·02 (−0·01 to 0·05) Haiti 9·49 (5·37 to 15·10) 157·01 (132·94 to 183·68) 6·72 (5·38 to 8·10) 44·09 (38·75 to 49·87) −0·04 (−0·08 to 0·01) −0·02 (−0·03 to −0·00) −0·09 (−0·10 to −0·07) Jamaica 0·66 (0·41 to 0·97) 11·08 (5·47 to 17·67) 0·42 (0·30

to 0·71) 62·25 (54·32 to 69·67) −0·01 (−0·03 to 0·01) 0·02 (0·00 to 0·05) 0·02 (−0·01 to 0·05) Haiti 9·49 (5·37 to 15·10) 157·01 (132·94 to 183·68) 6·72 (5·38 to 8·10) 44·09 (38·75 to 49·87) −0·04 (−0·08 to 0·01) −0·02 (−0·03 to −0·00) −0·09 (−0·10 to −0·07) Jamaica 0·66 (0·41 to 0·97) 11·08 (5·47 to 17·67) 0·42 (0·30 to 0·63) 39·55 (34·00 to 46·05) −0·00 (−0·03 to 0·02) −0·00 (−0·02 to 0·01) −0·04 (−0·06 to −0·01) Puerto Rico 0·44 (0·17 to 1·10) 12·01 (4·45 to 28·56) 0·28 (0·26 to 0·30) 50·14 (35·87 to 63·33) −0·01 (−0·10 to 0·14) −0·03 (−0·05 to 0·01) −0·07 (−0·08 to −0·06) Saint Lucia 0·02 (0·01 to 0·05) 0·39 (0·16 to 0·82) 0·01 (0·01 to 0·01) 37·89 (28·04 to 46·64) 0·01 (−0·08 to 0·15) 0·01 (−0·01 to 0·05) −0·04 (−0·05 to −0·03) Saint Vincent and the Grenadines 0·05 (0·02 to 0·11) 0·82 (0·34 to 1·70) 0·02 (0·02 to 0·02) 36·75 (27·32 to 45·46) 0·01 (−0·08 to 0·15) 0·01 (−0·01 to 0·06) −0·03 (−0·04 to −0·02) Suriname 0·19 (0·11 to 0·32) 3·81 (1·79 to 6·76) 0·14 (0·10 to 0·21) 44·05 (35·93 to 50·95) −0·01 (−0·04 to 0·02) −0·01 (−0·02 to 0·00) 0·00 (−0·03 to 0·03) Trinidad and Tobago 0·31 (0·19 to 0·50) 8·20 (3·80 to 15·61) 0·22 (0·20 to 0·23) 41·38 (33·87 to 48·39) −0·05 (−0·07 to −0·03) −0·01 (−0·02 to −0·00) −0·05 (−0·05 to −0·04) US Virgin Islands 0·01 (0·01 to 0·03) 0·27 (0·11 to 0·55) 0·01 (0·01 to 0·02) 43·23 (31·95 to 53·07) 0·01 (−0·08 to 0·16) 0·01 (−0·01 to 0·06) −0·01 (−0·05 to 0·04) Tropical Latin America 34·97 (31·13 to 38·94) 571·24 (468·17 to 701·82) 21·41 (18·29 to 25·43) 48·93 (47·39 to 50·56) −0·01 (−0·01 to −0·00) 0·01 (0·00 to 0·01) 0·02 (0·01 to 0·03) Brazil 33·76 (30·24 to 37·52) 558·84 (454·38 to 687·40) 21·05 (17·92 to 25·13) 49·37 (47·89 to 50·94) −0·01 (−0·01 to −0·00) 0·01 (0·00 to 0·01) 0·02 (0·01 to 0·02) Paraguay 1·21 (0·48 to 2·42) 12·40 (5·81 to 22·65) 0·36 (0·26 to 0·46) 29·48 (24·14 to 36·67) 0·02 (−0·04 to 0·06) 0·06 (0·03 to 0·08) 0·05 (0·01 to 0·07) Southeast Asia, east Asia, and Oceania 174·31 (121·70 to 266·48) 2335·91 (1716·35 to 3511·41) 101·62 (77·06 to 158·16) 25·88 (20·76 to 32·27) 0·00 (−0·03 to 0·03) 0·04 (0·01 to 0·06) 0·04 (−0·01 to 0·07) East Asia 56·50 (41·02 to 77·70) 796·14 (591·71 to 1043·33) 42·74 (38·28 to 47·27) 17·90 (15·89 to 20·29) −0·01 (−0·04 to 0·00) 0·04 (0·03 to 0·05) 0·08 (0·07 to 0·09) China 55·20 (39·82 to 75·98) 779·48 (573·77 to 1024·76) 41·82 (37·43 to 46·26) 17·90 (15·84 to 20·33) −0·01 (−0·04 to 0·00) 0·04 (0·03 to 0·05) 0·08 (0·07 to 0·09) North Korea 0·91 (0·10 to 3·08) 11·16 (2·27 to 39·91) 0·62 (0·1

47·27) 17·90 (15·89 to 20·29) −0·01 (−0·04 to 0·00) 0·04 (0·03 to 0·05) 0·08 (0·07 to 0·09) China 55·20 (39·82 to 75·98) 779·48 (573·77 to 1024·76) 41·82 (37·43 to 46·26) 17·90 (15·84 to 20·33) −0·01 (−0·04 to 0·00) 0·04 (0·03 to 0·05) 0·08 (0·07 to 0·09) North Korea 0·91 (0·10 to 3·08) 11·16 (2·27 to 39·91) 0·62 (0·1 2 to 2·43) 17·20 (6·82 to 33·39) 0·01 (−0·14 to 0·11) 0·05 (−0·02 to 0·13) 0·08 (−0·01 to 0·17) Taiwan 0·39 (0·12 to 0·84) 5·51 (2·23 to 11·10) 0·29 (0·21 to 0·38) 17·08 (11·69 to 23·37) −0·02 (−0·11 to 0·03) 0·04 (0·01 to 0·07) 0·09 (0·06 to 0·11) Southeast Asia 116·19 (66·38 to 204·40) 1510·91 (925·80 to 2625·43) 57·90 (34·09 to 115·40) 29·77 (21·90 to 38·69) 0·00 (−0·04 to 0·05) 0·03 (−0·01 to 0·06) 0·01 (−0·05 to 0·06) Cambodia 7·83 (3·55 to 14·42) 82·97 (32·84 to 153·51) 2·60 (1·80 to 3·64) 29·52 (22·70 to 36·25) 0·04 (0·00 to 0·07) 0·01 (−0·00 to 0·03) 0·01 (−0·02 to 0·03) Indonesia 43·39 (8·51 to 123·93) 440·51 (90·19 to 1391·78) 18·56 (3·60 to 68·98) 11·67 (8·08 to 15·97) 0·02 (−0·04 to 0·09) 0·10 (0·06 to 0·16) 0·17 (0·11 to 0·23) Laos 0·51 (0·12 to 1·58) 6·93 (1·68 to 23·47) 0·18 (0·04 to 0·69) 32·94 (23·92 to 42·06) −0·04 (−0·09 to 0·05) 0·06 (0·02 to 0·11) 0·08 (0·02 to 0·15) Malaysia 2·04 (1·56 to 2·81) 39·53 (22·31 to 70·73) 2·29 (1·83 to 3·35) 29·22 (22·89 to 36·80) −0·07 (−0·08 to −0·05) −0·03 (−0·04 to −0·02) −0·01 (−0·03 to 0·01) Maldives 0·00 (0·00 to 0·00) 0·02 (0·01 to 0·03) 0·00 (0·00 to 0·00) 10·72 (7·60 to 15·10) −0·01 (−0·03 to 0·01) −0·02 (−0·04 to 0·00) −0·03 (−0·05 to −0·02) Mauritius 0·12 (0·08 to 0·19) 1·57 (0·94 to 2·56) 0·08 (0·07 to 0·09) 19·51 (15·45 to 23·79) −0·00 (−0·03 to 0·02) 0·04 (0·03 to 0·06) 0·12 (0·11 to 0·14) Myanmar 6·75 (1·35 to 18·92) 177·74 (39·79 to 645·94) 8·62 (1·68 to 36·64) 40·35 (30·88 to 51·89) −0·07 (−0·13 to 0·00) −0·01 (−0·07 to 0·05) −0·04 (−0·12 to 0·02) Philippines 33·31 (12·50 to 82·46) 273·65 (127·88 to 476·60) 3·55 (3·31 to 3·82) 32·57 (24·42 to 42·40) 0·09 (0·03 to 0·19) 0·09 (0·07 to 0·12) −0·05 (−0·06 to −0·04) Sri Lanka 0·21 (0·10 to 0·38) 2·21 (1·00 to 3·95) 0·05 (0·05 to 0·06) 25·91 (23·17 to 28·86) 0·07 (0·03 to 0·11) 0·03 (0·01 to 0·05) −0·02 (−0·03 to −0·01) Seychelles 0·01 (0·00 to 0·03) 0·14 (0·04 to 0·30) 0·01 (0·00 to 0·01) 29·89 (16·29 to 44·37) −0·02 (−0·40 to 0·15) 0·02 (−0·04 to 0·10) 0·01 (−0·03 to 0·05) Thailand 10·06 (2·85 to 21·66) 288·25 (139·33 to 514·46) 14·74 (10·34 to 21·49) 40·71 (34·54 to 51·96) −0·07 (−0·20 to 0·01) −0·01 (−0·03 to 0·00) −0·01 (−0·03 to 0·01) Timor-Les

(0·00 to 0·03) 0·14 (0·04 to 0·30) 0·01 (0·00 to 0·01) 29·89 (16·29 to 44·37) −0·02 (−0·40 to 0·15) 0·02 (−0·04 to 0·10) 0·01 (−0·03 to 0·05) Thailand 10·06 (2·85 to 21·66) 288·25 (139·33 to 514·46) 14·74 (10·34 to 21·49) 40·71 (34·54 to 51·96) −0·07 (−0·20 to 0·01) −0·01 (−0·03 to 0·00) −0·01 (−0·03 to 0·01) Timor-Les te 0·10 (0·00 to 0·44) 1·46 (0·03 to 6·86) 0·07 (0·00 to 0·37) 25·29 (9·12 to 49·66) 0·01 (−0·40 to 0·27) 0·05 (−0·07 to 0·20) 0·03 (−0·08 to 0·17) Vietnam 11·73 (2·48 to 33·85) 193·97 (38·99 to 718·95) 7·05 (1·39 to 29·18) 33·48 (23·62 to 44·85) −0·03 (−0·09 to 0·05) 0·03 (−0·01 to 0·08) 0·02 (−0·03 to 0·08) Oceania 1·62 (1·13 to 2·28) 28·85 (24·68 to 33·04) 0·99 (0·78 to 1·27) 52·65 (46·71 to 59·02) −0·05 (−0·08 to −0·01) 0·02 (−0·00 to 0·04) −0·04 (−0·06 to −0·02) American Samoa 0·00 (0·00 to 0·00) 0·02 (0·01 to 0·04) 0·00 (0·00 to 0·00) 28·49 (22·61 to 36·97) 0·04 (−0·06 to 0·09) 0·05 (0·01 to 0·08) 0·00 (−0·02 to 0·03) Federated States of Micronesia 0·01 (0·00 to 0·03) 0·11 (0·02 to 0·42) 0·00 (0·00 to 0·02) 24·74 (14·47 to 45·45) 0·07 (−0·10 to 0·20) 0·07 (−0·03 to 0·18) 0·02 (−0·08 to 0·13) Fiji 0·07 (0·04 to 0·13) 0·68 (0·32 to 1·24) 0·02 (0·02 to 0·03) 24·73 (21·67 to 28·32) 0·05 (0·03 to 0·07) 0·05 (0·04 to 0·06) 0·04 (0·02 to 0·06) Guam 0·01 (0·00 to 0·03) 0·14 (0·06 to 0·28) 0·00 (0·00 to 0·01) 30·04 (22·88 to 40·09) 0·04 (−0·05 to 0·09) 0·05 (0·01 to 0·08) −0·00 (−0·03 to 0·03) Kiribati 0·00 (0·00 to 0·00) 0·02 (0·01 to 0·04) 0·00 (0·00 to 0·00) 29·00 (21·80 to 38·81) 0·03 (−0·06 to 0·08) 0·03 (−0·01 to 0·06) −0·02 (−0·04 to 0·01) Marshall Islands 0·01 (0·00 to 0·02) 0·09 (0·01 to 0·39) 0·00 (0·00 to 0·02) 25·91 (15·20 to 47·44) 0·07 (−0·10 to 0·19) 0·07 (−0·05 to 0·19) 0·02 (−0·08 to 0·14) Northern Mariana Islands 0·01 (0·00 to 0·01) 0·05 (0·02 to 0·10) 0·00 (0·00 to 0·00) 24·00 (19·14 to 32·23) 0·04 (−0·06 to 0·08) 0·05 (0·01 to 0·08) 0·02 (−0·01 to 0·04) Papua New Guinea 1·31 (0·86 to 1·94) 24·57 (21·27 to 27·37) 0·84 (0·66 to 1·04) 54·46 (48·01 to 61·46) −0·06 (−0·10 to −0·02) 0·02 (−0·01 to 0·03) −0·05 (−0·07 to −0·02) Samoa 0·02 (0·00 to 0·05) 0·21 (0·03 to 0·92) 0·01 (0·00 to 0·04) 27·11 (16·02 to 47·08) 0·07 (−0·10 to 0·20) 0·07 (−0·04 to 0·18) 0·02 (−0·07 to 0·13) Solomon Islands 0·05 (0·01 to 0·16) 0·64 (0·08 to 2·87) 0·02 (0·00 to 0·15) 25·34 (15·33 to 44·47) 0·08 (−0·10 to 0·20) 0·07 (−0·04 to 0·18) 0·02 (−0·08 to 0·13) Tonga 0·01 (0·00 to 0·02) 0·06 (0·02 to 0·11) 0·00 (0·00 to 0·00) 22·65 (18·24 to 30·85) 0·08 (−0·02 to 0·14) 0·10 (0·0

7 (−0·04 to 0·18) 0·02 (−0·07 to 0·13) Solomon Islands 0·05 (0·01 to 0·16) 0·64 (0·08 to 2·87) 0·02 (0·00 to 0·15) 25·34 (15·33 to 44·47) 0·08 (−0·10 to 0·20) 0·07 (−0·04 to 0·18) 0·02 (−0·08 to 0·13) Tonga 0·01 (0·00 to 0·02) 0·06 (0·02 to 0·11) 0·00 (0·00 to 0·00) 22·65 (18·24 to 30·85) 0·08 (−0·02 to 0·14) 0·10 (0·0 5 to 0·14) 0·06 (0·02 to 0·09) Vanuatu 0·02 (0·00 to 0·08) 0·31 (0·04 to 1·28) 0·01 (0·00 to 0·07) 25·63 (15·26 to 45·39) 0·07 (−0·09 to 0·20) 0·07 (−0·05 to 0·19) 0·02 (−0·08 to 0·14) North Africa and Middle East 12·39 (8·53 to 17·51) 137·94 (113·08 to 172·80) 7·54 (6·25 to 9·30) 19·07 (15·88 to 22·65) −0·02 (−0·05 to 0·01) 0·02 (−0·01 to 0·03) 0·01 (−0·01 to 0·03) Afghanistan 0·70 (0·14 to 2·04) 4·50 (0·92 to 13·98) 0·21 (0·04 to 0·73) 4·17 (3·08 to 5·86) 0·09 (0·01 to 0·17) 0·06 (−0·05 to 0·14) 0·02 (−0·11 to 0·12) Algeria 0·31 (0·01 to 0·95) 6·47 (2·15 to 11·52) 0·23 (0·12 to 0·38) 61·55 (54·76 to 68·03) −0·13 (−0·48 to 0·04) 0·04 (−0·02 to 0·06) 0·01 (−0·04 to 0·05) Bahrain 0·06 (0·01 to 0·12) 0·53 (0·23 to 1·02) 0·02 (0·01 to 0·03) 17·69 (13·63 to 23·05) 0·04 (−0·08 to 0·11) 0·03 (−0·01 to 0·07) −0·01 (−0·04 to 0·02) Egypt 0·95 (0·48 to 1·65) 6·80 (3·37 to 11·62) 0·21 (0·17 to 0·26) 17·68 (15·09 to 20·70) 0·09 (0·06 to 0·13) 0·09 (0·07 to 0·11) 0·03 (0·02 to 0·05) Iran 1·13 (0·62 to 2·06) 11·49 (5·78 to 21·26) 0·55 (0·43 to 0·76) 15·31 (13·65 to 17·17) 0·03 (−0·02 to 0·07) 0·03 (0·01 to 0·05) 0·04 (0·01 to 0·08) Iraq 0·51 (0·13 to 1·05) 3·63 (1·66 to 6·87) 0·11 (0·08 to 0·15) 16·20 (12·70 to 21·54) 0·07 (−0·07 to 0·14) 0·09 (0·04 to 0·13) 0·06 (0·03 to 0·10) Jordan 0·03 (0·01 to 0·06) 0·28 (0·12 to 0·56) 0·01 (0·01 to 0·02) 20·24 (15·73 to 26·00) 0·02 (−0·12 to 0·09) 0·02 (−0·01 to 0·06) −0·00 (−0·03 to 0·03) Kuwait 0·01 (0·00 to 0·03) 0·14 (0·06 to 0·28) 0·01 (0·01 to 0·01) 19·37 (15·23 to 25·32) 0·03 (−0·11 to 0·09) 0·01 (−0·03 to 0·04) −0·10 (−0·11 to −0·08) Lebanon 0·13 (0·02 to 0·38) 1·94 (0·41 to 7·96) 0·09 (0·02 to 0·45) 35·17 (22·65 to 62·67) 0·02 (−0·04 to 0·09) 0·01 (−0·06 to 0·08) 0·01 (−0·07 to 0·06) Libya 0·23 (0·01 to 0·94) 2·43 (0·15 to 10·48) 0·11 (0·00 to 0·55) 19·73 (13·40 to 28·26) 0·04 (−0·14 to 0·16) 0·05 (−0·06 to 0·16) 0·02 (−0·09 to 0·13) Morocco 0·67 (0·39 to 1·06) 8·62 (4·20 to 15·19) 0·36 (0·27 to 0·47) 24·58 (22·74 to 26·56) −0·01 (−0·03 to 0·02) 0·04 (0·02 to 0·05) 0·08 (0·04 to 0·11) Oman 0·14 (0·09 to 0·19) 1·83 (0·97 to 2·95) 0·07 (0·05 to 0·09) 33·16 (27·89 to 39·27) −0·02 (−0·05 to 0·01) 0·01 (−0·01 to 0·02) 0·04 (−0·01 to

0·02 (−0·09 to 0·13) Morocco 0·67 (0·39 to 1·06) 8·62 (4·20 to 15·19) 0·36 (0·27 to 0·47) 24·58 (22·74 to 26·56) −0·01 (−0·03 to 0·02) 0·04 (0·02 to 0·05) 0·08 (0·04 to 0·11) Oman 0·14 (0·09 to 0·19) 1·83 (0·97 to 2·95) 0·07 (0·05 to 0·09) 33·16 (27·89 to 39·27) −0·02 (−0·05 to 0·01) 0·01 (−0·01 to 0·02) 0·04 (−0·01 to 0·07) Palestine 0·06 (0·02 to 0·12) 0·45 (0·21 to 0·90) 0·02 (0·01 to 0·02) 17·23 (13·38 to 22·07) 0·03 (−0·08 to 0·10) 0·06 (0·02 to 0·10) 0·04 (0·01 to 0·08) Qatar 0·01 (0·00 to 0·03) 0·14 (0·06 to 0·27) 0·01 (0·00 to 0·01) 17·24 (13·30 to 22·43) 0·02 (−0·11 to 0·08) −0·03 (−0·06 to 0·00) −0·07 (−0·09 to −0·03) Saudi Arabia 1·06 (0·50 to 2·04) 11·58 (5·82 to 25·18) 0·49 (0·25 to 1·34) 23·37 (19·48 to 28·20) 0·02 (−0·05 to 0·08) 0·03 (−0·03 to 0·07) 0·01 (−0·06 to 0·05) Sudan 4·31 (1·25 to 8·66) 55·38 (41·64 to 70·69) 4·32 (3·20 to 5·31) 10·02 (7·44 to 13·08) −0·09 (−0·20 to −0·02) −0·01 (−0·05 to 0·02) 0·01 (−0·02 to 0·04) Syria 0·04 (0·01 to 0·07) 0·66 (0·15 to 1·51) 0·03 (0·01 to 0·05) 18·22 (15·12 to 22·01) 0·10 (−0·02 to 0·14) 0·11 (0·01 to 0·17) 0·06 (−0·02 to 0·11) Tunisia 0·28 (0·13 to 0·53) 2·62 (1·16 to 4·69) 0·09 (0·07 to 0·12) 23·64 (20·75 to 27·28) 0·05 (0·00 to 0·09) 0·07 (0·05 to 0·09) 0·10 (0·07 to 0·13) Turkey 0·72 (0·28 to 1·28) 8·07 (3·64 to 13·75) 0·19 (0·14 to 0·24) 32·60 (26·92 to 39·90) 0·01 (−0·06 to 0·05) 0·08 (0·04 to 0·10) 0·02 (−0·01 to 0·05) United Arab Emirates 0·54 (0·03 to 2·02) 5·69 (0·34 to 25·28) 0·27 (0·01 to 1·35) 16·98 (12·17 to 23·71) 0·04 (−0·15 to 0·16) 0·05 (−0·06 to 0·16) 0·02 (−0·09 to 0·13) Yemen 0·51 (0·10 to 1·46) 4·52 (1·02 to 14·33) 0·16 (0·03 to 0·60) 35·75 (28·14 to 45·07) 0·04 (−0·04 to 0·12) 0·04 (−0·05 to 0·12) −0·02 (−0·15 to 0·09) South Asia 206·83 (171·79 to 249·70) 2966·01 (2767·85 to 3183·75) 135·26 (127·05 to 144·88) 25·55 (23·78 to 27·11) −0·01 (−0·02 to 0·01) −0·02 (−0·03 to −0·01) −0·06 (−0·07 to −0·05) Bangladesh 0·51 (0·11 to 1·54) 6·69 (1·44 to 22·77) 0·27 (0·05 to 1·11) 15·57 (12·24 to 19·89) −0·02 (−0·07 to 0·06) 0·09 (0·04 to 0·14) 0·13 (0·07 to 0·19) Bhutan 0·06 (0·01 to 0·18) 0·64 (0·14 to 2·16) 0·02 (0·00 to 0·06) 28·44 (18·86 to 44·60) 0·02 (−0·05 to 0·09) 0·04 (−0·04 to 0·12) −0·04 (−0·17 to 0·07) India 196·60 (164·77 to 237·89) 2881·13 (2702·40 to 3078·82) 131·56 (124·14 to 138·94) 25·82 (24·18 to 27·36) −0·01 (−0·03 to 0·01) −0·02 (−0·03 to −0·02) −0·06 (−0·07 to −0·05) Nepal 1·11 (0·23 to 3·34) 31·55 (6·90 to 118·68) 1·92 (0·37 to 8·04) 31·21 (24·54 to 39·54) −0·13 (−0·19 to −0·06) −0

2) −0·04 (−0·17 to 0·07) India 196·60 (164·77 to 237·89) 2881·13 (2702·40 to 3078·82) 131·56 (124·14 to 138·94) 25·82 (24·18 to 27·36) −0·01 (−0·03 to 0·01) −0·02 (−0·03 to −0·02) −0·06 (−0·07 to −0·05) Nepal 1·11 (0·23 to 3·34) 31·55 (6·90 to 118·68) 1·92 (0·37 to 8·04) 31·21 (24·54 to 39·54) −0·13 (−0·19 to −0·06) −0 ·01 (−0·05 to 0·03) 0·03 (−0·03 to 0·09) Pakistan 8·55 (1·66 to 25·89) 45·99 (9·66 to 139·53) 1·48 (0·25 to 5·20) 5·87 (4·00 to 8·40) 0·15 (0·07 to 0·24) 0·15 (0·06 to 0·24) 0·14 (0·01 to 0·24) Sub-Saharan Africa 1847·99 (1656·94 to 2051·52) 29 439·54 (28 678·60 to 30 195·25) 859·00 (804·61 to 912·93) 42·35 (41·05 to 43·58) −0·03 (−0·05 to −0·02) −0·00 (−0·00 to 0·00) −0·08 (−0·09 to −0·08) Southern sub–Saharan Africa 710·08 (604·65 to 836·17) 11 408·43 (10 926·35 to 11 882·56) 228·94 (211·90 to 250·82) 51·04 (48·91 to 53·20) −0·03 (−0·04 to −0·01) 0·01 (0·01 to 0·01) −0·09 (−0·09 to −0·08) Botswana 23·51 (14·49 to 33·86) 431·89 (394·30 to 473·07) 8·07 (5·00 to 10·58) 61·72 (54·17 to 69·76) −0·01 (−0·06 to 0·04) 0·01 (0·00 to 0·01) −0·08 (−0·11 to −0·04) Lesotho 24·75 (17·77 to 34·10) 354·36 (323·40 to 389·73) 12·57 (9·82 to 16·14) 36·43 (32·75 to 40·32) −0·02 (−0·05 to 0·01) 0·02 (0·01 to 0·02) −0·04 (−0·05 to −0·02) Namibia 14·05 (10·04 to 18·47) 253·41 (238·00 to 268·89) 5·09 (3·73 to 6·50) 51·55 (46·41 to 57·20) −0·03 (−0·06 to 0·00) 0·01 (0·01 to 0·01) −0·09 (−0·11 to −0·07) South Africa 529·67 (440·94 to 630·39) 8409·55 (7978·87 to 8850·42) 155·19 (140·96 to 172·68) 50·95 (48·45 to 53·56) −0·03 (−0·05 to −0·01) 0·02 (0·02 to 0·02) −0·08 (−0·08 to −0·06) Swaziland 13·91 (9·02 to 18·32) 263·04 (244·42 to 280·64) 5·89 (4·58 to 7·59) 52·50 (46·87 to 58·25) −0·06 (−0·10 to −0·03) 0·02 (0·01 to 0·02) −0·07 (−0·09 to −0·05) Zimbabwe 104·20 (53·05 to 173·34) 1696·17 (1543·57 to 1878·76) 42·12 (34·03 to 52·38) 51·64 (45·42 to 58·51) −0·01 (−0·08 to 0·06) −0·01 (−0·02 to −0·01) −0·13 (−0·15 to −0·10) Western sub-Saharan Africa 444·71 (334·29 to 571·29) 6417·10 (6036·15 to 6873·38) 249·30 (212·45 to 288·59) 29·09 (26·33 to 31·90) −0·03 (−0·06 to −0·00) −0·00 (−0·01 to 0·01) −0·05 (−0·07 to −0·04) Benin 5·40 (3·57 to 8·07) 83·05 (72·48 to 94·41) 2·36 (1·71 to 3·17) 43·94 (39·15 to 49·42) −0·02 (−0·07 to 0·03) −0·00 (−0·01 to 0·01) −0·10 (−0·13 to −0·08) Burkina Faso 6·02 (3·20 to 9·53) 101·91 (85·21 to 121·99) 3·19 (2·44 to 3·93) 51·64 (45·07 to 59·57) −0·01 (−0·08 to 0·05) −0·03 (−0·05 to −0·02) −0·16 (−0·18 to −0·14) Cameroon 48·59 (28·44 to 73·57) 659·83 (570·86 to

71 to 3·17) 43·94 (39·15 to 49·42) −0·02 (−0·07 to 0·03) −0·00 (−0·01 to 0·01) −0·10 (−0·13 to −0·08) Burkina Faso 6·02 (3·20 to 9·53) 101·91 (85·21 to 121·99) 3·19 (2·44 to 3·93) 51·64 (45·07 to 59·57) −0·01 (−0·08 to 0·05) −0·03 (−0·05 to −0·02) −0·16 (−0·18 to −0·14) Cameroon 48·59 (28·44 to 73·57) 659·83 (570·86 to 754·45) 33·19 (23·70 to 42·93) 21·55 (18·54 to 25·09) −0·03 (−0·08 to 0·01) −0·00 (−0·02 to 0·01) −0·02 (−0·04 to −0·01) Cape Verde 0·32 (0·14 to 0·84) 3·83 (2·92 to 5·18) 0·10 (0·07 to 0·14) 32·10 (22·93 to 41·72) −0·01 (−0·07 to 0·13) 0·03 (0·00 to 0·07) −0·04 (−0·07 to −0·01) Chad 9·24 (3·70 to 17·19) 166·86 (127·28 to 214·51) 8·92 (5·99 to 11·87) 33·57 (25·61 to 43·33) −0·09 (−0·19 to −0·02) −0·02 (−0·04 to −0·00) −0·06 (−0·08 to −0·03) Côte d'Ivoire 41·71 (24·65 to 64·08) 547·27 (464·67 to 633·88) 22·38 (17·99 to 28·08) 29·06 (23·81 to 34·68) −0·02 (−0·06 to 0·03) −0·01 (−0·03 to 0·00) −0·07 (−0·09 to −0·06) The Gambia 0·91 (0·35 to 1·68) 17·82 (14·35 to 22·44) 0·66 (0·43 to 0·95) 22·24 (17·69 to 27·59) −0·10 (−0·21 to −0·03) 0·01 (−0·01 to 0·03) −0·01 (−0·03 to 0·01) Ghana 17·30 (9·85 to 26·44) 282·24 (238·35 to 330·83) 11·25 (8·22 to 14·98) 37·00 (30·63 to 44·34) −0·04 (−0·09 to 0·01) −0·03 (−0·04 to −0·02) −0·09 (−0·12 to −0·07) Guinea 8·68 (3·79 to 14·06) 133·76 (109·08 to 158·21) 5·15 (4·06 to 6·69) 25·87 (21·95 to 30·18) −0·05 (−0·12 to 0·00) 0·01 (−0·02 to 0·03) −0·03 (−0·05 to −0·02) Guinea-Bissau 1·91 (0·69 to 3·61) 41·33 (34·66 to 48·64) 1·76 (0·98 to 2·56) 25·34 (22·37 to 28·67) −0·11 (−0·21 to −0·03) 0·02 (0·01 to 0·04) 0·02 (−0·02 to 0·07) Liberia 2·65 (1·23 to 4·58) 36·97 (31·57 to 44·16) 2·17 (1·74 to 2·64) 22·36 (18·09 to 26·50) −0·02 (−0·10 to 0·07) −0·04 (−0·06 to −0·02) −0·07 (−0·09 to −0·05) Mali 11·43 (5·99 to 17·86) 148·12 (111·67 to 187·99) 6·74 (4·96 to 9·08) 23·07 (19·78 to 26·82) −0·03 (−0·09 to 0·00) −0·00 (−0·02 to 0·02) −0·03 (−0·06 to −0·01) Mauritania 0·12 (0·01 to 0·42) 7·01 (1·37 to 29·75) 0·36 (0·06 to 1·70) 40·28 (26·39 to 55·32) −0·19 (−0·33 to −0·07) −0·04 (−0·07 to 0·00) −0·05 (−0·10 to −0·00) Niger 2·47 (0·79 to 4·96) 64·26 (53·92 to 76·75) 3·46 (2·94 to 4·06) 34·12 (28·38 to 40·03) −0·11 (−0·23 to −0·03) −0·06 (−0·08 to −0·04) −0·09 (−0·11 to −0·07) Nigeria 274·19 (168·30 to 396·73) 3874·25 (3517·41 to 4307·34) 136·42 (100·74 to 174·02) 28·98 (24·68 to 33·55) −0·02 (−0·07 to 0·01) 0·01 (−0·00 to 0·02) −0·04 (−0·07 to −0·01) São Tomé and Príncipe 0·00 (0·00 to 0·00) 0·03 (0·02 to 0·04) 0·00 (0·00 to 0·00) 54·78 (48·49

) −0·06 (−0·08 to −0·04) −0·09 (−0·11 to −0·07) Nigeria 274·19 (168·30 to 396·73) 3874·25 (3517·41 to 4307·34) 136·42 (100·74 to 174·02) 28·98 (24·68 to 33·55) −0·02 (−0·07 to 0·01) 0·01 (−0·00 to 0·02) −0·04 (−0·07 to −0·01) São Tomé and Príncipe 0·00 (0·00 to 0·00) 0·03 (0·02 to 0·04) 0·00 (0·00 to 0·00) 54·78 (48·49 to 61·57) −0·01 (−0·07 to 0·04) 0·03 (0·01 to 0·05) −0·06 (−0·09 to −0·02) Senegal 4·43 (1·36 to 7·44) 66·44 (52·59 to 81·28) 2·36 (1·70 to 3·01) 47·05 (40·37 to 55·06) −0·05 (−0·16 to −0·00) −0·00 (−0·02 to 0·01) −0·05 (−0·09 to −0·02) Sierra Leone 4·13 (1·23 to 7·09) 61·20 (50·35 to 72·68) 2·79 (2·11 to 3·76) 18·47 (14·21 to 23·83) −0·06 (−0·18 to 0·00) 0·01 (−0·02 to 0·03) −0·00 (−0·02 to 0·02) Togo 5·20 (2·59 to 9·20) 120·86 (105·83 to 138·66) 6·05 (4·77 to 7·59) 29·99 (25·43 to 34·83) −0·09 (−0·17 to −0·02) −0·04 (−0·05 to −0·02) −0·06 (−0·07 to −0·05) Eastern sub-Saharan Africa 618·52 (527·49 to 714·55) 10 437·57 (10 025·86 to 10 889·54) 318·68 (293·79 to 347·87) 42·82 (40·73 to 44·74) −0·04 (−0·05 to −0·02) −0·01 (−0·01 to −0·00) −0·10 (−0·11 to −0·09) Burundi 6·65 (3·45 to 11·37) 111·53 (94·00 to 130·90) 3·68 (2·79 to 4·79) 37·92 (31·85 to 44·95) −0·02 (−0·09 to 0·06) −0·03 (−0·04 to −0·01) −0·12 (−0·14 to −0·10) Comoros 0·06 (0·01 to 0·19) 0·41 (0·09 to 1·31) 0·02 (0·00 to 0·08) 12·81 (7·27 to 20·97) 0·02 (−0·09 to 0·13) 0·01 (−0·06 to 0·08) 0·03 (−0·05 to 0·11) Djibouti 0·51 (0·18 to 0·99) 7·37 (5·10 to 10·70) 0·38 (0·25 to 0·52) 24·81 (20·41 to 29·94) −0·02 (−0·10 to 0·05) −0·02 (−0·05 to −0·00) −0·06 (−0·10 to −0·02) Eritrea 1·49 (0·68 to 2·51) 21·22 (16·16 to 28·23) 0·80 (0·52 to 1·21) 36·89 (26·24 to 49·02) −0·00 (−0·08 to 0·07) −0·02 (−0·04 to 0·00) −0·09 (−0·13 to −0·05) Ethiopia 39·14 (19·59 to 62·13) 768·04 (651·03 to 904·91) 28·65 (22·04 to 34·85) 51·92 (45·82 to 59·09) 0·01 (−0·10 to 0·11) −0·05 (−0·06 to −0·04) −0·16 (−0·19 to −0·13) Kenya 137·20 (112·46 to 166·10) 1883·96 (1787·64 to 1987·73) 51·70 (48·19 to 55·64) 38·60 (36·42 to 40·72) 0·06 (0·04 to 0·08) −0·01 (−0·02 to −0·01) −0·13 (−0·13 to −0·12) Madagascar 2·00 (0·33 to 6·34) 42·51 (8·58 to 169·61) 4·42 (0·76 to 19·79) 1·35 (0·76 to 2·72) −0·13 (−0·22 to −0·03) −0·05 (−0·14 to 0·01) −0·03 (−0·11 to 0·04) Malawi 55·62 (25·29 to 80·90) 1126·77 (988·92 to 1234·61) 28·41 (22·67 to 35·54) 49·58 (42·68 to 58·00) −0·07 (−0·15 to −0·02) −0·01 (−0·03 to −0·00) −0·13 (−0·14 to −0·10) Mozambique 122·32 (74·39 to 176·33) 1833·02 (1593·99 to 2073·78) 70·06 (58·14 to 83·59) 30·66 (26·49 to 35·28)

−0·14 to 0·01) −0·03 (−0·11 to 0·04) Malawi 55·62 (25·29 to 80·90) 1126·77 (988·92 to 1234·61) 28·41 (22·67 to 35·54) 49·58 (42·68 to 58·00) −0·07 (−0·15 to −0·02) −0·01 (−0·03 to −0·00) −0·13 (−0·14 to −0·10) Mozambique 122·32 (74·39 to 176·33) 1833·02 (1593·99 to 2073·78) 70·06 (58·14 to 83·59) 30·66 (26·49 to 35·28) −0·06 (−0·10 to −0·02) 0·01 (0·00 to 0·02) −0·02 (−0·04 to −0·00) Rwanda 8·02 (4·46 to 12·17) 202·70 (180·73 to 224·59) 4·52 (3·29 to 5·88) 58·05 (51·98 to 64·37) −0·08 (−0·14 to −0·03) −0·01 (−0·02 to −0·00) −0·15 (−0·18 to −0·13) Somalia 1·69 (0·73 to 3·13) 24·24 (16·88 to 33·77) 1·58 (1·17 to 2·09) 8·65 (5·53 to 13·34) −0·08 (−0·18 to −0·01) −0·02 (−0·06 to 0·01) −0·01 (−0·04 to 0·02) South Sudan 9·76 (3·69 to 16·82) 122·52 (79·09 to 174·43) 8·93 (4·58 to 12·48) 8·59 (5·58 to 13·19) −0·05 (−0·14 to 0·01) −0·01 (−0·05 to 0·03) −0·01 (−0·05 to 0·05) Tanzania 86·66 (51·18 to 135·07) 1494·12 (1315·85 to 1672·30) 47·86 (34·24 to 61·09) 47·58 (41·18 to 54·48) −0·04 (−0·08 to 0·00) −0·01 (−0·03 to −0·00) −0·11 (−0·14 to −0·09) Uganda 77·87 (33·96 to 126·32) 1491·60 (1310·27 to 1708·04) 36·32 (27·59 to 51·43) 43·44 (38·37 to 48·94) −0·07 (−0·15 to −0·01) 0·02 (0·00 to 0·03) −0·09 (−0·11 to −0·06) Zambia 69·14 (50·55 to 88·86) 1300·28 (1215·65 to 1382·85) 31·12 (26·20 to 36·90) 52·70 (47·08 to 58·71) −0·06 (−0·10 to −0·03) 0·01 (0·00 to 0·01) −0·11 (−0·13 to −0·10) Central sub-Saharan Africa 74·67 (46·65 to 121·85) 1176·44 (1054·44 to 1312·53) 62·08 (54·72 to 70·25) 26·41 (23·29 to 29·65) −0·07 (−0·12 to −0·02) −0·03 (−0·04 to −0·02) −0·06 (−0·07 to −0·05) Angola 22·35 (11·92 to 36·24) 285·93 (229·98 to 350·10) 11·10 (6·36 to 16·37) 28·29 (23·41 to 33·79) −0·03 (−0·10 to 0·02) 0·03 (0·01 to 0·05) −0·01 (−0·05 to 0·02) Central African Republic 9·87 (4·92 to 17·25) 137·53 (115·36 to 162·34) 8·20 (6·84 to 9·68) 22·21 (18·48 to 26·35) −0·02 (−0·08 to 0·03) −0·04 (−0·05 to −0·02) −0·06 (−0·08 to −0·04) Congo 7·08 (3·54 to 10·54) 97·57 (73·72 to 119·26) 4·83 (3·57 to 6·10) 21·03 (18·08 to 26·01) −0·03 (−0·07 to 0·01) −0·01 (−0·03 to 0·01) −0·06 (−0·08 to −0·04) Democratic Republic of the Congo 32·55 (9·52 to 78·70) 588·53 (492·83 to 709·81) 35·90 (30·47 to 42·09) 24·84 (19·80 to 30·10) −0·12 (−0·26 to −0·01) −0·06 (−0·07 to −0·04) −0·07 (−0·09 to −0·05) Equatorial Guinea 0·63 (0·17 to 1·55) 24·45 (20·84 to 28·49) 0·81 (0·42 to 1·18) 31·82 (26·45 to 39·19) −0·18 (−0·33 to −0·06) 0·03 (0·01 to 0·04) −0·02 (−0·07 to 0·03) Gabon 2·19 (0·70 to 4·39) 42·43 (35·47 to 50·80) 1·24

·09) 24·84 (19·80 to 30·10) −0·12 (−0·26 to −0·01) −0·06 (−0·07 to −0·04) −0·07 (−0·09 to −0·05) Equatorial Guinea 0·63 (0·17 to 1·55) 24·45 (20·84 to 28·49) 0·81 (0·42 to 1·18) 31·82 (26·45 to 39·19) −0·18 (−0·33 to −0·06) 0·03 (0·01 to 0·04) −0·02 (−0·07 to 0·03) Gabon 2·19 (0·70 to 4·39) 42·43 (35·47 to 50·80) 1·24 (0·82 to 1·55) 60·25 (52·88 to 67·83) −0·07 (−0·18 to 0·02) −0·02 (−0·04 to 0·00) −0·07 (−0·11 to −0·04) Data in parentheses are 95% uncertainty intervals. New infections and HIV/AIDS deaths are cumulative for the whole of 2015. The number of people living with HIV is the point prevalence (as a count) at the end of 2015. The number of people living with HIV receiving ART and the total number of people living with HIV are year-end point prevalences. We calculated numerators for incidence, prevalence, and mortality rates with counts as previously described. The denominator for each rate was population at mid-year. We age-standardised rates with the WHO age standard. We calculated ARC as the slope from the log of the value in 2015, to the log of the value in 2005: (log[value 2015]–log[value 2005])/10. ART=antiretroviral therapy. ARC=annualised rate of change. SDI=sociodemographic index.

Introduction Antiretroviral therapy (ART) has substantially improved the prognosis for HIV-infected children, transforming HIV-1 infection from a life-threatening disease to a chronic infection. Furthermore, with new evidence,1 universal ART is now recommended2, 3 for all people living with HIV, including children and adolescents, even without major immunosuppression or HIV-related symptoms. Therefore, the population of children, adolescents, and young adults on life-long ART is growing.4 For this population, innovative treatment strategies are needed to address their lifestyle needs, to help maintain long-term retention-in-care, and to improve adherence to ART, which is particularly problematic during adolescence.4, 5, 6

opulation of children, adolescents, and young adults on life-long ART is growing.4 For this population, innovative treatment strategies are needed to address their lifestyle needs, to help maintain long-term retention-in-care, and to improve adherence to ART, which is particularly problematic during adolescence.4, 5, 6 Short cycle therapy aims to maintain suppression of HIV-1 RNA during planned short breaks from ART, thereby reducing ART intake, long-term toxic effects, and costs. First proof-of-concept studies suggested the feasibility of a 7 days on and 7 days off ART strategy;7, 8, 9 however, this approach proved inferior to continuous therapy in two randomised controlled trials in adults.10, 11 Single-arm studies with shorter breaks in ART (4 days on and 3 days off) reported inconsistent results.12, 13 However, two small randomised controlled trials confirmed that a short cycle therapy strategy of 5 days on and 2 days off ART is achievable: in the FOTO trial, including 60 US adults,14, 15 and in a larger randomised controlled trial in 103 Ugandan adults,10 short cycle therapy was non-inferior to continuous therapy in terms of maintained viral load suppression over 48 weeks with the added benefit of less toxicity. Most participants in both trials were on efavirenz, which has a long plasma half-life (40–91 h), and lamivudine, which has an intermediate long intracellular half-life (22 h).16 However, whereas participants in the US study received tenofovir disoproxil fumarate as the third drug (intracellular half-life 60–180 h),16 those in the Ugandan trial received shorter-acting stavudine or zidovudine.

a half-life (40–91 h), and lamivudine, which has an intermediate long intracellular half-life (22 h).16 However, whereas participants in the US study received tenofovir disoproxil fumarate as the third drug (intracellular half-life 60–180 h),16 those in the Ugandan trial received shorter-acting stavudine or zidovudine. Research in context Evidence before this study

a half-life (40–91 h), and lamivudine, which has an intermediate long intracellular half-life (22 h).16 However, whereas participants in the US study received tenofovir disoproxil fumarate as the third drug (intracellular half-life 60–180 h),16 those in the Ugandan trial received shorter-acting stavudine or zidovudine. Research in context Evidence before this study We searched PubMed up to March 1, 2016, with the search terms “HIV” AND (“short” AND “cycle” OR “short-cycle”) AND “therapy” and the references from the retrieved manuscripts. More than a decade ago, small proof-of-concept studies in adults suggested that structured treatment interruptions with 7 days on and 7 days off cycles of antiretroviral therapy (ART) could maintain virological suppression, particularly if drugs with long half-lives were used. However, this strategy proved inferior to continuous therapy in two randomised controlled trials in adults. Single-arm studies of a short-cycle therapy strategy with 4 days on and 3 days off showed inconsistent results: although there was no confirmed viral rebound in adults on different ART regimens in a French study, a study in highly treated adolescents and young adults on protease inhibitor-based therapy in the USA showed high rates of viral rebound. Adult studies of short cycle therapy with 2 days per week off efavirenz-based ART showed promising results: following a single arm study of 5 days on and 2 days off ART, which showed rates of virological suppression of about 90% over 48 weeks, two small randomised controlled trials in adults (one US, one Ugandan) confirmed non-inferiority of maintaining virological suppression with this short cycle therapy strategy compared with continuous therapy. No published trials have assessed 5 days on and 2 days off ART in children or adolescents.

t 90% over 48 weeks, two small randomised controlled trials in adults (one US, one Ugandan) confirmed non-inferiority of maintaining virological suppression with this short cycle therapy strategy compared with continuous therapy. No published trials have assessed 5 days on and 2 days off ART in children or adolescents. Added value of this study To our knowledge, this is the first randomised controlled trial to investigate the feasibility and acceptability of efavirenz-based short cycle therapy (5 days on and 2 days off) in a geographically diverse group of children, adolescents, and young adults with no previous treatment failure. The short cycle therapy was acceptable and non-inferior in terms of maintaining virological suppression (including to very low viral loads). No significant differences were noted in immune activation, total HIV-1 DNA, or development of resistance, and the short cycle therapy group had fewer ART-related adverse events than did the continuous therapy group. Additionally, participants expressed a strong preference for this short cycle therapy compared with continuous treatment, once they had adapted to the new routine. Implications of all the available evidence

To our knowledge, this is the first randomised controlled trial to investigate the feasibility and acceptability of efavirenz-based short cycle therapy (5 days on and 2 days off) in a geographically diverse group of children, adolescents, and young adults with no previous treatment failure. The short cycle therapy was acceptable and non-inferior in terms of maintaining virological suppression (including to very low viral loads). No significant differences were noted in immune activation, total HIV-1 DNA, or development of resistance, and the short cycle therapy group had fewer ART-related adverse events than did the continuous therapy group. Additionally, participants expressed a strong preference for this short cycle therapy compared with continuous treatment, once they had adapted to the new routine. Implications of all the available evidence The findings of this trial, supported by previous adult studies, show that a short cycle therapy strategy with 5 days on and 2 days off efavirenz -based ART with a standard dose of efavirenz (maximum 600 mg adult equivalent daily dose) is a viable option for virologically suppressed children, adolescents, and young adults with 29% reduction in the cost of drugs. 2 year extended follow-up of the trial is ongoing to address sustainability of this strategy over a longer duration and results will be available in 2017. Further studies are warranted to assess short cycle therapy with lower doses of efavirenz and other long-acting ART regimens in settings with less frequent viral load testing than the quarterly monitoring done in trials reporting to date.

ty of this strategy over a longer duration and results will be available in 2017. Further studies are warranted to assess short cycle therapy with lower doses of efavirenz and other long-acting ART regimens in settings with less frequent viral load testing than the quarterly monitoring done in trials reporting to date. No randomised trials of short cycle therapy have been done in children or adolescents, who face longer-term ART than adults. We aimed to assess whether short cycle therapy on first-line efavirenz-based ART in children, adolescents, and young adults was non-inferior to continous therapy in terms of maintaining virological suppression and adherence to ART, while improving quality of life.

follow-up visit. Adherence questionnaires were completed by carers and participants at weeks 0, 4, 12, 24, and 48. Acceptability questionnaires for those randomised to short cycle therapy were completed at randomisation and at final visit (or at time of change from short cycle therapy to continous therapy if earlier). The trial incorporated three substudies. The virology and immunology substudy assessed low level viraemia (viral load <20 copies per mL), total HIV-1 DNA, and 19 biomarkers of inflammation, vascular injury, and disordered thrombogenesis; all were measured retrospectively on stored plasma and cell samples. The ultrasensitive quantitative HIV-1 RNA and DNA assays used the Qiagen QIAsymphony SP (Hilden, Germany) for nucleic acid extraction. An ABI Prism 7500 real-time thermal cycler (Foster City, CA, USA) was used for amplification of HIV-1 RNA and DNA using Invitrogen RT-PCR (Waltham, MA, USA) and Qiagen Multiplex PCR (Hilden, Germany) reagents, respectively. An in-house standard curve calibrated against the WHO HIV International standard in IU per mL was used for HIV-1 RNA quantification (appendix). The quantitation of HIV-1 DNA was based on a standard curve using the 8E5 cell line, which carries one HIV proviral genome per cell; cell numbers were estimated with the single copy gene for pyruvate dehydrogenase; results were reported as copies of HIV-1 DNA per million cells.

n or adolescents, who face longer-term ART than adults. We aimed to assess whether short cycle therapy on first-line efavirenz-based ART in children, adolescents, and young adults was non-inferior to continous therapy in terms of maintaining virological suppression and adherence to ART, while improving quality of life. Methods Study design and participants In this open-label, randomised, parallel group non-inferiority phase 2/3 trial (BREATHER [BREaks in Adolescent and child THerapy using Efavirenz and two nRtis] PENTA 16), participants aged 8–24 years were eligible if they had a CD4 cell count 350 cells per μL or higher, suppressed viral load less than 50 copies per mL for at least 12 months on an efavirenz based regimen with two or three nucleoside or nucleotide reverse transcriptase inhibitors (NRTIs) and no previous treatment failure (first-line ART). Children on nevirapine or boosted protease inhibitor ART who had not had treatment failure and with undetectable viral load could be enrolled if they substituted efavirenz and viral load remained undetectable for 12 weeks or longer before enrolment. Previous two-drug ART, substitution of NRTIs, or both were allowed, provided these were not for regimen failure. Previous monotherapy was only allowed if taken perinatally for prevention of mother-to-child-transmission. Participants were not eligible if they were pregnant, on concomitant medications for acute illness, or if their creatinine or liver transaminases results were grade 3 or higher at screening. Parents or guardians and older participants provided written consent; young children gave assent appropriate for age and knowledge of HIV status, as per guidelines for each participating country.

nt medications for acute illness, or if their creatinine or liver transaminases results were grade 3 or higher at screening. Parents or guardians and older participants provided written consent; young children gave assent appropriate for age and knowledge of HIV status, as per guidelines for each participating country. The trial protocol was approved by the ethics committees in participating centres in Europe, Africa, and the Americas, and is available online. Randomisation and masking Patients were randomly assigned (1:1) to remain on continuous therapy or change to short cycle therapy and randomisation was done centrally by the MRC Clinical Trials Unit at UCL (London, UK), according to a computer-generated randomisation list, using permuted blocks of varying size, stratified by age (8–12 years, 13–17 years, 18–24 years) and site (African vs non-African). The randomisation list was prepared by the trial statistician and securely incorporated within the database. Randomisation of study participants was done via a web service accessed by site clinician or one of the three coordinating trials units.

(8–12 years, 13–17 years, 18–24 years) and site (African vs non-African). The randomisation list was prepared by the trial statistician and securely incorporated within the database. Randomisation of study participants was done via a web service accessed by site clinician or one of the three coordinating trials units. Procedures An initial 3 week randomised pilot safety phase in selected clinical centres was done in 32 participants (in which 15 participants were allocated to the short cycle therapy group) to ensure those in the short cycle therapy group maintained undetectable viral load (<50 copies per mL) after the 2 day break (Saturday and Sunday) and before resuming weekday ART on Monday. Recruitment to the main trial commenced after review of three consecutive Monday morning viral load results per participant by the Independent Data Monitoring Committee (IDMC).

p maintained undetectable viral load (<50 copies per mL) after the 2 day break (Saturday and Sunday) and before resuming weekday ART on Monday. Recruitment to the main trial commenced after review of three consecutive Monday morning viral load results per participant by the Independent Data Monitoring Committee (IDMC). In the main trial, participants randomly assigned to short cycle therapy chose 2 consecutive days off ART (Friday and Saturday or Saturday and Sunday; occasionally 2 other days: referred to as weekends off), and continued this cycle throughout. Participants on continuous therapy remained on continuous efavirenz-based ART. Substitutions for simplification (except efavirenz) or toxicity (all drugs) were allowed. Participants were randomised 2–4 weeks after screening and assessed clinically at weeks 4 and 12, then every 12 weeks until the last participant had completed 48 weeks' follow-up. Examination for lipodystrophy, Tanner stage, and a pregnancy test (in postmenarchal girls) were done at randomisation and repeated every 24 weeks.

ipants were randomised 2–4 weeks after screening and assessed clinically at weeks 4 and 12, then every 12 weeks until the last participant had completed 48 weeks' follow-up. Examination for lipodystrophy, Tanner stage, and a pregnancy test (in postmenarchal girls) were done at randomisation and repeated every 24 weeks. Viral load and T lymphocytes were measured at every visit; participants with viral load of 50 copies per mL or higher had a repeat test within 1 week; those on short cycle therapy with confirmed viral rebound (two test results with viral load >50 copies per mL) recommenced continuous ART. Additional assessment of treatment adherence and a stored sample for resistance testing were requested for all participants with viral rebound. Haematology and biochemistry tests were done at screening and randomisation; thereafter, haematology was done every 12 weeks and biochemistry as per local practice. Blood lipids, including total cholesterol, high density lipoprotein, low density lipoprotein, and very low density lipoprotein, were measured at weeks 0, 24, and 48. Plasma and cells were stored for additional immunology and virology tests (see below) at baseline and weeks 4, 8, and 12, and then every 12 weeks for plasma and 24 weeks for cells. Questions on compliance to the strategy were asked at every follow-up visit. Adherence questionnaires were completed by carers and participants at weeks 0, 4, 12, 24, and 48. Acceptability questionnaires for those randomised to short cycle therapy were completed at randomisation and at final visit (or at time of change from short cycle therapy to continous therapy if earlier).

for HIV-1 RNA quantification (appendix). The quantitation of HIV-1 DNA was based on a standard curve using the 8E5 cell line, which carries one HIV proviral genome per cell; cell numbers were estimated with the single copy gene for pyruvate dehydrogenase; results were reported as copies of HIV-1 DNA per million cells. 19 biomarkers (thrombomodulin, ICAM-1, ICAM-3, VCAM-1, CD62E, CD62P, VEGF, angiopoietin 1 and 2, serum amyloid, C-reactive protein, interleukin 1Ra, interleukin 6, interleukin 8, interleukin 10, TNFα, MCP-1, tissue factor, and D-dimers) were analysed with Meso Scale Discovery (Gaithersburg, MD, USA) or by ELISA kits (Quantikine ELISA Human Coagulation Factor III/Tissue Factor [R&D Systems, MN, USA] and TECHNOZYM D-dimer ELISA assay Technoclone [Vienna, Austria]). CD4 and CD8 lymphocyte subsets were quantified locally on fresh samples; CD45RA and CD45RO subpopulations of CD4 and CD8 cells were assessed on fresh or stored frozen cell samples at selected sites. The adherence substudy assessed adherence in participants from selected sites by recording bottle openings using a Medication Event Monitoring System (MEMS) capped container. MEMS caps were placed on the container with most frequently taken antiretrovirals. The longitudinal qualitative substudy focused on participants' experiences of the trial and acceptability of short cycle therapy.17

The adherence substudy assessed adherence in participants from selected sites by recording bottle openings using a Medication Event Monitoring System (MEMS) capped container. MEMS caps were placed on the container with most frequently taken antiretrovirals. The longitudinal qualitative substudy focused on participants' experiences of the trial and acceptability of short cycle therapy.17 Outcomes The primary endpoint was confirmed viral load of 50 copies per mL or higher by week 48. Secondary outcomes were: confirmed viral load of 400 copies per mL or higher by week 48; cumulative number and type of major HIV-1 RNA resistance mutations in those with viral rebound; change in CD4% and CD4 cell count, glucose, blood lipids from baseline to week 48; changes in ART regimen; change back to continuous therapy (short cycle therapy only); adherence; acceptability (short cycle therapy); division of AIDS grade 3 or 4 clinical or laboratory adverse events,18 and treatment-modifying adverse events of any grade; and new US Centers for Disease Control (CDC) stage B or C diagnoses or death.

regimen; change back to continuous therapy (short cycle therapy only); adherence; acceptability (short cycle therapy); division of AIDS grade 3 or 4 clinical or laboratory adverse events,18 and treatment-modifying adverse events of any grade; and new US Centers for Disease Control (CDC) stage B or C diagnoses or death. Statistical analysis 160 participants (80 per group) provided 80% power to exclude a non-inferiority margin of 12% for the difference in proportion of participants reaching the primary endpoint, assuming 10% of participants have confirmed viral load 50 copies per mL or higher in the continuous therapy group and a one-sided α of 0·05 (two-sided α=0·1). The Trial Steering Committee decided to continue recruitment until the end of the planned randomisation period to allow sites to recruit patients already invited for screening and to avoid the study being underpowered if the proportion of participants reaching the primary endpoint in the continuous therapy group was lower than expected (the Trial Steering Committee did not have access to event rates during the trial).

on period to allow sites to recruit patients already invited for screening and to avoid the study being underpowered if the proportion of participants reaching the primary endpoint in the continuous therapy group was lower than expected (the Trial Steering Committee did not have access to event rates during the trial). In the primary, intent-to-treat analysis, the proportion of participants who had viral rebound (≥50 copies per mL) was estimated with Kaplan-Meier methods, with adjustment for baseline stratification factors, censoring at week 54 (upper band of week 48 assessment window) or last follow-up date if not seen at week 48. The difference in proportion (between the short cycle therapy group and continuous therapy group) of participants who had viral rebound was estimated and two-sided 90% CIs of the difference was obtained with bootstrap SE (1000 replicates).19 In a prespecified sensitivity analysis on the per-protocol population, individuals were censored if they had a break in treatment for longer than 7 days, discontinued efavirenz for longer than 7 days, or changed strategy to continuous therapy for reasons other than viral rebound. The intent-to-treat analysis was also repeated without adjustment for stratification factors. Confirmed viral load of 400 copies per mL or higher was estimated by the same approach. Major resistance mutations were summarised.

r than 7 days, or changed strategy to continuous therapy for reasons other than viral rebound. The intent-to-treat analysis was also repeated without adjustment for stratification factors. Confirmed viral load of 400 copies per mL or higher was estimated by the same approach. Major resistance mutations were summarised. Immunology (including substudy biomarkers), HIV-DNA, haematology, biochemistry, and lipids were assessed at week 48 by fitting normal regression models with adjustment for randomised group and baseline values. Natural log transformations were applied as appropriate (for HIV-DNA, biomarkers and ratios of CD45RA [naive]:CD45RO [memory] cells and CD8RA [naive]:CD8RO [memory] cells). Change from baseline is presented as change from mean at baseline in all participants. Categorical variables were compared with Fisher's exact tests, or McNemar's tests for paired data; rates used Poisson regression (including a random effect for participant where appropriate). Generalised estimating equations (independent correlation structure) were used to compare self-reported adherence across randomised groups over time. Stata version 13.1 was used for all analyses (StataCorp 2013, College Station, TX, USA: StataCorp LP). To assess adherence to allocated strategy, the number of days that MEMS cap was opened at least once divided by number of days that MEMS cap was in use during the trial was calculated for each day of the week.

r time. Stata version 13.1 was used for all analyses (StataCorp 2013, College Station, TX, USA: StataCorp LP). To assess adherence to allocated strategy, the number of days that MEMS cap was opened at least once divided by number of days that MEMS cap was in use during the trial was calculated for each day of the week. Pilot phase data were included in the analysis. The IDMC reviewed full interim data on three occasions, viral load and enrolment data at a fourth meeting, and analyses of viral load results alone on six further occasions during the trial. The trial was registered with EudraCT, number 2009-012947-40), ISRCTN, number 97755073, and CTA, number 27505/0005/001-0001. Role of the funding source The funders had no direct role in the study design, data collection, data analysis, data interpretation, report writing, or decision to submit the report for publication. The corresponding author had access to all data and responsibility for submission for publication. Results Between April 1, 2011, and June 28, 2013, 227 participants were screened (figure 1), of whom 199 from 24 sites were randomly assigned (99 to short cycle therapy and 100 to continuous therapy). One participant in the continuous therapy group moved location and withdrew consent at week 24; the remaining 198 were followed up to at least week 48.

nd June 28, 2013, 227 participants were screened (figure 1), of whom 199 from 24 sites were randomly assigned (99 to short cycle therapy and 100 to continuous therapy). One participant in the continuous therapy group moved location and withdrew consent at week 24; the remaining 198 were followed up to at least week 48. Of those patients randomly assigned, 70 (35%) were recruited from Uganda (35 in the short cycle therapy group and 35 in the continuous therapy group), 48 (24%) from western Europe, 36 (18%) from Thailand, 20 (10%) from Ukraine, 14 (13%) from the USA, and 11 (6%) from Argentina (appendix p 5). Baseline characteristics were similar between the groups (table 1). Although CD4% and count were high and well matched between groups, fewer participants had CDC stage C disease in the short cycle therapy group than in the continuous therapy group. Pre-trial ART exposure was comparable between groups: median time on ART at randomisation was 6·1 years (IQR 3·8–8·4), 82 (41%) were on their initial ART regimen at baseline, 29 (15%) had previously substituted a protease inhibitor, but following the exclusion criteria, none had switched ART for failure.

therapy group. Pre-trial ART exposure was comparable between groups: median time on ART at randomisation was 6·1 years (IQR 3·8–8·4), 82 (41%) were on their initial ART regimen at baseline, 29 (15%) had previously substituted a protease inhibitor, but following the exclusion criteria, none had switched ART for failure. 13 participants had a confirmed viral load 50 of copies per mL or higher at any time up to 48 weeks (table 2), an estimated probability of viral rebound of 6·1% in short cycle therapy versus 7·3% in continuous therapy (difference −1·2%, 90% CI −7·3 to 4·9, test for difference, bootstrap p=0·75; figure 2A). Thus, the 4·9% upper band of the two-sided 90% confidence limit was well within the 12% non-inferiority margin. The per-protocol analysis gave a similar estimated difference of −1·1% (90% CI −6·8 to 4·6), as did analysis without adjustment for stratification factors (figure 2B).

ifference, bootstrap p=0·75; figure 2A). Thus, the 4·9% upper band of the two-sided 90% confidence limit was well within the 12% non-inferiority margin. The per-protocol analysis gave a similar estimated difference of −1·1% (90% CI −6·8 to 4·6), as did analysis without adjustment for stratification factors (figure 2B). After viral rebound, five (83%) of six participants in the short cycle therapy group resuppressed viral load (three on the same regimen but changed to continuous daily ART, two following regimen change) compared with only three (43%) of seven participants in the continuous therapy group (two resuppressed while continuing the same regimen and one after regimen change). The remaining five participants (one in the short cycle therapy group and four in the continuous therapy group) remained non-suppressed; three on first-line ART (one in the short cycle therapy group and two in the continuous therapy group) and two (in the continuous therapy group) after switching to second-line ART. Results repeating the primary analysis, adjusted for CDC stage at baseline were qualitatively unchanged: −1·3% difference between groups, in favour of short cycle therapy (90% CI −7·4 to 4·7, test for difference, bootstrap p=0·72).

apy group) and two (in the continuous therapy group) after switching to second-line ART. Results repeating the primary analysis, adjusted for CDC stage at baseline were qualitatively unchanged: −1·3% difference between groups, in favour of short cycle therapy (90% CI −7·4 to 4·7, test for difference, bootstrap p=0·72). To determine whether the risk of reaching the primary endpoint was related to type of NRTI (short-acting or long-acting), a Cox model adjusted for randomised group and NRTI received (zidovudine vs abacavir or tenofovir disoproxil fumarate) was fitted (exploratory analysis); results showed no significant differences between continuous therapy and short cycle therapy (p=0·81; data not shown). Six participants (two [2%] in the short cycle therapy group and four [4%] in the continuous therapy group) had confirmed viral load of 400 copies per mL or higher by week 48; estimated probability 2·1% in the short cycle therapy group versus 4·2% in the continuous therapy group (difference −2·1%, 90% CI −6·2 to 1·9, p=0·38). 12 participants changed ART regimen during the first 48 weeks (three in the short cycle therapy group and nine in the continuous therapy group, Fisher's exact p=0·13), five because of toxic effects (one in the short cycle therapy group and four in the continuous therapy group; Table 2, Table 3).

nd, Kaplan-Meier methods, with adjustment for study stratification factors. †Difference in estimated probability of viral rebound, Kaplan-Meier methods. ‡With exact confidence intervals. §Kaplan-Meier methods, censoring individuals who violated the profile at that time, with adjustment for study stratification factors. Figure 3 Proportion of days MEMS caps were opened Data for 31 participants in the short cycle therapy group and 30 participants in the continuous treatment group (including 23 in each group with data to 48 weeks). MEMS=Medication Event Monitoring System. Table 1 Baseline characteristics

Six participants (two [2%] in the short cycle therapy group and four [4%] in the continuous therapy group) had confirmed viral load of 400 copies per mL or higher by week 48; estimated probability 2·1% in the short cycle therapy group versus 4·2% in the continuous therapy group (difference −2·1%, 90% CI −6·2 to 1·9, p=0·38). 12 participants changed ART regimen during the first 48 weeks (three in the short cycle therapy group and nine in the continuous therapy group, Fisher's exact p=0·13), five because of toxic effects (one in the short cycle therapy group and four in the continuous therapy group; Table 2, Table 3). Of 13 participants reaching the primary endpoint, resistance results were available for nine (three in the short cycle therapy group and six in the continuous therapy group); the remaining four patients had samples with low viral load, insufficient to obtain a result (three in the short cycle therapy group: 56 copies per mL, 62 copies per mL, and 126 copies per mL; one in the continuous therapy group: 231 copies per mL). All four participants suppressed again after these blips, suggesting drug resistance was unlikely. Seven of nine participants with available results had resistance mutations: all seven had NNRTI mutations and two had Met184Val (table 2).

L, and 126 copies per mL; one in the continuous therapy group: 231 copies per mL). All four participants suppressed again after these blips, suggesting drug resistance was unlikely. Seven of nine participants with available results had resistance mutations: all seven had NNRTI mutations and two had Met184Val (table 2). No new CDC stage C and two CDC stage B events were recorded (bronchopneumonia in the short cycle therapy group and bronchitis in the continuous therapy group) and no significant differences were noted between groups in CD4% or CD4 cell count (table 2). With the exception of lower mean corpuscular volume in those on zidovudine and lower platelet levels in the short cycle therapy group compared with the continuous therapy group; haematological variables did not differ (appendix p 6). Concentration of low density lipoproteins was higher at week 24 in the short cycle therapy group than in the continuous therapy group, but we observed no difference at week 48 (appendix p 6). By week 48, eight participants in the short cycle therapy group had reverted to continuous therapy: six participants reached the primary endpoint, one developed gynaecomastia leading to efavirenz discontinuation and resumption of daily ART, and one had ART changed for poor adherence (without reaching the primary endpoint).

eek 48, eight participants in the short cycle therapy group had reverted to continuous therapy: six participants reached the primary endpoint, one developed gynaecomastia leading to efavirenz discontinuation and resumption of daily ART, and one had ART changed for poor adherence (without reaching the primary endpoint). By 48 weeks, 20 participants had 27 grade 3 or 4 adverse events, with decreased neutrophil count being the most common (two participants in the short cycle therapy group vs six participants in the continuous therapy group; Fisher's exact p=0·28; table 3). Two ART-related adverse events were reported in two participants in the short cycle therapy group compared with 14 events in ten participants in the continuous therapy group (Poisson p=0·02 for number of events; Fisher's exact p=0·03 for number of participants); this was the only significant difference in adverse events between groups). Lipodystrophy and gynaecomastia were the most common ART-related events. 13 serious adverse events were reported in nine participants (six in the short cycle therapy group and three in the continuous therapy group; table 3). There were five pregnancies (one in the short cycle therapy group and four in the continuous therapy group).

and gynaecomastia were the most common ART-related events. 13 serious adverse events were reported in nine participants (six in the short cycle therapy group and three in the continuous therapy group; table 3). There were five pregnancies (one in the short cycle therapy group and four in the continuous therapy group). Among 192 children (98 in the short cycle therapy group and 94 in the continuous therapy group) in the immunology and virology substudy, values for viral load less than 20 copies per mL, total HIV-1 DNA, and inflammatory markers, including interleukin 6 and D-dimer, were similar between randomised groups at baseline (table 1; appendix p 7). At week 48, 13 (13%) children in the short cycle therapy group and 14 (15%) in the continuous therapy group had viral load 20 copies per mL or higher (Fisher's exact p=0·84) and there were no significant differences between groups in total HIV-1 DNA (table 2; p=0·13), including after adjustment for differences at baseline or after exclusion of participants with evidence of viral rebound (data not shown). No differences between groups were noted at week 48 in the 19 biomarkers of inflammation, vascular injury, and disordered thrombogenesis, with the exception of D-dimer, which was lower in the short cycle therapy group than in the continuous therapy group by log 0·5 (p=0·05; table 2; appendix p 7). No differences were identified in CD8 cells, ratios of CD45RA (naive):CD45RO (memory) cells, and CD8RA (naive):CD8RO (memory) cells between groups at week 48 (data not shown).

on of D-dimer, which was lower in the short cycle therapy group than in the continuous therapy group by log 0·5 (p=0·05; table 2; appendix p 7). No differences were identified in CD8 cells, ratios of CD45RA (naive):CD45RO (memory) cells, and CD8RA (naive):CD8RO (memory) cells between groups at week 48 (data not shown). In the short cycle therapy group, 95% of weekend breaks were reported as taken (99% excluding time after return to continuous therapy). The MEMS cap substudy data supported these results. Among 61 participants enrolled in the substudy (31 in the short cycle therapy group and 30 in the continuous therapy group), 56 (28 in each group) continued to use MEMS caps until 36 weeks and 46 (23 in each group) were still using MEMS caps at week 48. The median number of cap openings per week was five (IQR 4–5) in the short cycle therapy group and seven (6–7) in the continuous therapy group. MEMS caps were opened at least once daily from Monday to Friday more than 80% of the time in both groups, with the percentage of bottle openings remaining high in the continuous therapy group at weekends, but dropping to less than 20% for those on short cycle therapy (figure 3).

nd, Kaplan-Meier methods, with adjustment for study stratification factors. †Difference in estimated probability of viral rebound, Kaplan-Meier methods. ‡With exact confidence intervals. §Kaplan-Meier methods, censoring individuals who violated the profile at that time, with adjustment for study stratification factors. Figure 3 Proportion of days MEMS caps were opened Data for 31 participants in the short cycle therapy group and 30 participants in the continuous treatment group (including 23 in each group with data to 48 weeks). MEMS=Medication Event Monitoring System. Table 1 Baseline characteristics Short cycle therapy (n=99) Continuous therapy (n=100) Total (n=199) Male 57 (58%) 48 (48%) 105 (53%) Age (years) 13·7 (11·7–17·7) 14·4 (12·0–17·5) 14·1 (11·9–17·6) 8–12 38 (38%) 39 (39%) 77 (39%) 13–17 39 (39%) 41 (41%) 80 (40%) 18–24 22 (22%) 20 (20%) 42 (21%) Ethnic origin Black (African or other) 58 (59%) 54 (54%) 112 (56%) White 24 (24%) 17 (17%) 41 (21%) Asian 15 (15%) 22 (22%) 37 (19%) Other 2 (2%) 7 (7%) 9 (5%) Route of infection Vertical 90 (91%) 90 (90%) 180 (90%) Sexual contact 7 (7%) 7 (7%) 14 (7%) Unknown/other* 2 (2%) 3 (3%) 5 (3%) CDC stage† N 16 (16%) 10 (10%) 26 (13%) A 25 (25%) 25 (25%) 50 (25%) B 45 (45%) 43 (43%) 88 (44%) C 13 (13%) 21 (21%) 34 (17%) Cumulative ART exposure before baseline (years) 6·2 (3·8–7·9) 5·9 (4·0–8·4) 6·1 (3·8–8·4) Baseline regimen is the initial ART regimen 40 (40%) 42 (42%) 82 (41%) Efavirenz plus Zidovudine plus lamivudine 52 (53%) 53 (53%) 105 (53%) Tenofovir plus lamivudine or emtricitabine 25 (25%) 27 (27%) 52 (26%) Abacavir plus lamivudine or emticitabine 22 (22%) 18 (18%) 40 (20%) Other‡ 0 (0%) 2 (2%) 2 (1%) CD4 percentage 34·5 (29·3–39·0) 34·0 (29·5–38·1) 34·0 (29·5–38·5) <25% 5 (5%) 6 (6%) 11 (6%) ≥25% to <40% 73 (74%) 76 (76%) 149 (75%) ≥40% 21 (21%) 18 (18%) 39 (20%) CD4 cell count (cells per μL) 722·5 (581·0–965·0) 747·3 (575·3–972·8) 735·0 (575·5–967·5) ≥350–500 16 (16%) 12 (12%) 28 (14%) >500 83 (84%) 88 (88%) 171 (86%) Viral load (copies per mL) <20§ 91 (93%) 86 (91%) 177 (92%) ≥20 7 (7%) 8 (9%) 15 (8%) Total HIV-1 DNA (copies per million cells) 420 (159–871) 309 (136–926) 347 (145–894) Interleukin 6 (pg/mL) 0·6 (0·4–0·9) 0·6 (0·4–0·9) 0·6 (0·4–0·9) D-dimers (ng/mL) 69·1 (3·13–135·4) 65·7 (4·8–80·3) 67·5 (3·1–152·2) CRP (pg/mL) 631·2 (303·8–2676·1) 621·6 (260·8–2164·1) 626·8 (288·8–2311·0) Data are median (IQR) or n (%). CDC=US Centers for Disease Control and Prevention. ART= antiretroviral therapy. CRP=C-reactive protein.

n (6–7) in the continuous therapy group. MEMS caps were opened at least once daily from Monday to Friday more than 80% of the time in both groups, with the percentage of bottle openings remaining high in the continuous therapy group at weekends, but dropping to less than 20% for those on short cycle therapy (figure 3). Based on ART logs, updated at each visit, one participant in the short cycle therapy group and seven participants in the continuous therapy group had a treatment interruption of 3 days or more (excluding weekend breaks in the short cycle therapy group; Fisher's exact test p=0·07). Adherence questionnaires were completed by 91 participants in the short cycle therapy group and 93 participants in the continuous therapy group at one or more visit (80 in both groups at four or more visits) to 48 weeks. Adherence was similar in both groups with 7% (29 of 414) of reports in the short cycle therapy group versus 10% of (40 of 409) reports in the continuous therapy group of missing ART in the week prior to the assessment visit (excluding weekend breaks in the short cycle therapy group; p=0·42). Adherence based on carers' questionnaires was also similar between the two groups (data not shown).

he short cycle therapy group versus 10% of (40 of 409) reports in the continuous therapy group of missing ART in the week prior to the assessment visit (excluding weekend breaks in the short cycle therapy group; p=0·42). Adherence based on carers' questionnaires was also similar between the two groups (data not shown). In acceptability questionnaires completed at baseline, 70 (88%) of 80 participants in the short cycle therapy group thought the approach would be easier than staying on continuous therapy. At end of follow-up 81 (90%) of 90 participants in the short cycle therapy group reported that weekend breaks made life easier than daily ART, mainly because going out with friends was easier: 15 (20%) of 76 participants who completed both questionnaires reported this was difficult pre-trial compared with only two of 76 during the trial (McNemar p=0·001; appendix p 8). The acceptability of short cycle therapy as further explored in the qualitative substudy will be reported elsewhere.17

th friends was easier: 15 (20%) of 76 participants who completed both questionnaires reported this was difficult pre-trial compared with only two of 76 during the trial (McNemar p=0·001; appendix p 8). The acceptability of short cycle therapy as further explored in the qualitative substudy will be reported elsewhere.17 Discussion We found no evidence that short cycle therapy was inferior to continuous therapy in maintaining viral load suppression with a very small non-significant difference between the groups favouring short cycle therapy. Further, five of six participants on short cycle therapy who had low level viraemia resuppressed on returning to daily ART. Results were essentially unchanged in further analyses that adjusted for small differences in CDC stage at baseline, and were done per protocol. Our results have broad generalisability because we recruited participants from diverse geographical, ethnic, and sociocultural backgrounds in 11 countries, including 21% who were young adults in their early twenties. There were fewer major resistance mutations among children failing on short cycle therapy than in those on continuous therapy, although this was not statistically significant. In both groups and similarly to the PENPACT-1 trial, which assessed timing of switch to second-line ART, NNRTI and Met184Val mutations emerged rapidly even at low level viraemia.20

tations among children failing on short cycle therapy than in those on continuous therapy, although this was not statistically significant. In both groups and similarly to the PENPACT-1 trial, which assessed timing of switch to second-line ART, NNRTI and Met184Val mutations emerged rapidly even at low level viraemia.20 Although virological suppression to less than 50 copies per mL was the primary endpoint, we further investigated the safety of short cycle therapy by assessing its effect on very low level viraemia and HIV reservoir and showed no differences between the short cycle therapy and continuous therapy groups. Methods of varying technical difficulty and biological meaning have been suggested to quantify the HIV reservoir, which is responsible for viral rebound following treatment interruption.21 We measured HIV-1 DNA because it is a surrogate for reservoir size in acute and chronic HIV infection.22, 23

inuous therapy groups. Methods of varying technical difficulty and biological meaning have been suggested to quantify the HIV reservoir, which is responsible for viral rebound following treatment interruption.21 We measured HIV-1 DNA because it is a surrogate for reservoir size in acute and chronic HIV infection.22, 23 Increases in chronic immune activation and inflammation have been reported in adult interruption trials designed to allow rebounds in viral load, and have been associated with adverse HIV-related outcomes.24 Immune activation with raised concentrations of biomarkers of inflammation and coagulation has also been reported in patients with virological suppression,25 both among elite controllers not on ART and ART recipients with supressed viral load, albeit at low levels.26 Therefore, we also measured the effect of the short cycle therapy strategy on 19 potentially important biomarkers and found no evidence of any differences between groups, with the exception of D-dimer which, by contrast with expectation, was lower in short cycle therapy than in continuous therapy, which could be a chance finding. The absence of a signal suggestive of any increased immune activation and inflammation adds further confidence that the short cycle therapy strategy was not causing subclinical injury. Furthermore, we recorded no differences in cellular markers previously shown to be rapidly deranged during treatment interruption.27

g. The absence of a signal suggestive of any increased immune activation and inflammation adds further confidence that the short cycle therapy strategy was not causing subclinical injury. Furthermore, we recorded no differences in cellular markers previously shown to be rapidly deranged during treatment interruption.27 Most safety profiles were similar between randomised groups, and there were more ART-related adverse events reported in the continuous therapy group. However, in an open-label trial, potential for reporting bias exists. Assuring adherence to the randomised strategy is crucial to the integrity of trial results. If participants randomly assigned to the continuous therapy group elected, of their own accord, to take breaks in therapy, non-inferiority of short cycle therapy and continuous therapy might be shown, because both groups could be taking similar breaks off-ART. Three independent indicators of adherence to assigned strategy (self-reported adherence, MEMS caps substudy, and differences between groups in mean corpuscular volume among zidovudine recipients) all showed that participants on short cycle therapy had appropriately less ART exposure than those on continuous therapy.

T. Three independent indicators of adherence to assigned strategy (self-reported adherence, MEMS caps substudy, and differences between groups in mean corpuscular volume among zidovudine recipients) all showed that participants on short cycle therapy had appropriately less ART exposure than those on continuous therapy. As well as being the first randomised trial in children, our results build on those from two adult trials with similar design, showing non-inferiority of short cycle therapy versus continuous therapy on efavirenz-based ART.10, 15 Only one non-randomised study of short cycle therapy in US adolescents and young adults has been reported in heavily ART-experienced participants taking a 3 day weekend break from protease inhibitor-based ART regimens.12 This study differed substantially from our study and the adult short cycle therapy trials in both design and ART used. More than a third of participants had viral rebound and more than half changed to continuous treatment for other reasons; with no control group and multiple previous ART regimens, viral load, and resistance test results are hard to interpret or compare with our trial. Protease inhibitor ART might not be ideal for short cycle therapy because half-lives are shorter than NNRTIs and might not protect against viral replication during days off. Furthermore, participants in the US study had breaks of 3 days, whereas those in our trial had breaks of only 2 days.

t or compare with our trial. Protease inhibitor ART might not be ideal for short cycle therapy because half-lives are shorter than NNRTIs and might not protect against viral replication during days off. Furthermore, participants in the US study had breaks of 3 days, whereas those in our trial had breaks of only 2 days. Acceptability of the short cycle therapy strategy was shown among participants from all backgrounds; in particular, it was valued because it allowed for more socialising with friends at weekends. Similar results were reported in the associated qualitative substudy, during which participants also discussed liking short cycle therapy because of perceived reduction of previously unreported and unrecognised ART side-effects, such as dizziness and reduced energy. The qualitative substudy17 provided insights into the complexities of physician–patient interactions, particularly relating to non-adherence. In particular, participants who are virologically suppressed might elect not to disclose adherence lapses because of a desire not to fail or disappoint their physician. Overall, the qualitative findings endorsed participants' enthusiasm for short cycle therapy, but also highlighted the need for support with early adaptation to weekend breaks off-ART.17

e virologically suppressed might elect not to disclose adherence lapses because of a desire not to fail or disappoint their physician. Overall, the qualitative findings endorsed participants' enthusiasm for short cycle therapy, but also highlighted the need for support with early adaptation to weekend breaks off-ART.17 The overall reduction in drug exposure could reduce long-term toxicity for individuals and, at a population level, result in cost savings, enabling more participants to receive treatment. The ENCORE1 trial28 showed that daily 400 mg efavirenz was non-inferior to 600 mg, with less toxicity; in both groups, efavirenz was given with daily tenofovir and emtricitabine. Efavirenz 400 mg daily is included as an alternative option to 600 mg efavirenz-based ART in revised WHO 2016 adult guidelines.29 Of note, the weekly cumulative dose of daily 400 mg efavirenz is almost the same as in our trial. Both strategies seem to be more acceptable to patients than 600 mg daily efavirenz and provide the possibility of individualisation of ART regimens to suit life situations.

based ART in revised WHO 2016 adult guidelines.29 Of note, the weekly cumulative dose of daily 400 mg efavirenz is almost the same as in our trial. Both strategies seem to be more acceptable to patients than 600 mg daily efavirenz and provide the possibility of individualisation of ART regimens to suit life situations. The results from this study show that short cycle therapy might be a promising strategy for adherent children and adolescents well established on ART. However, follow-up is relatively short. A 2-year trial extension is ongoing, which will provide further data on longer-term sustainability. More than 90% of participants have reconsented to stay on their randomised strategy and we expect results in 2017. Of note, this short cycle therapy strategy can be generalised only to children and young people taking efavirenz-based ART who have not had treatment failure, and where there is availability of viral load monitoring. Appropriate counselling and support is needed to explain that there should be a maximum of 2 days per week breaks in therapy. Furthermore, results presented here cannot necessarily be extrapolated to ART containing the reduced dose of efavirenz (equivalent to 400 mg for adults) or other ART regimens, or to settings where viral load monitoring is unavailable or infrequent. Further research is needed to address this, and could also assess short cycle therapy with other suitable long-acting drugs or drugs with a higher barrier to resistance such as tenofovir alafenamide and dolutegravir .30

lts) or other ART regimens, or to settings where viral load monitoring is unavailable or infrequent. Further research is needed to address this, and could also assess short cycle therapy with other suitable long-acting drugs or drugs with a higher barrier to resistance such as tenofovir alafenamide and dolutegravir .30 In conclusion, in an adherent and geographically diverse population of HIV-infected 8–24 year-olds on 600 mg efavirenz-based ART, a short cycle therapy strategy with 2-day weekend breaks was non-inferior to continuous therapy in terms of virological, immunological, inflammatory effects, and resulted in fewer adverse events. Treatment with ART 5 days per week instead of 7 provides potential for cost savings. Short cycle therapy was liked by participants; in particular, it improved their social lives. This short cycle therapy strategy is a viable option for adherent HIV-infected young people who are stable on efavirenz-based ART. Ongoing longer-term follow-up will further inform sustainability and further research is required for settings where viral load monitoring is less accessible. Correspondence to: Dr Anna Turkova, MRC Clinical Trials Unit at UCL, London, WC2B 6NH, UK a.turkova@ucl.ac.uk Supplementary Material Supplementary appendix

In conclusion, in an adherent and geographically diverse population of HIV-infected 8–24 year-olds on 600 mg efavirenz-based ART, a short cycle therapy strategy with 2-day weekend breaks was non-inferior to continuous therapy in terms of virological, immunological, inflammatory effects, and resulted in fewer adverse events. Treatment with ART 5 days per week instead of 7 provides potential for cost savings. Short cycle therapy was liked by participants; in particular, it improved their social lives. This short cycle therapy strategy is a viable option for adherent HIV-infected young people who are stable on efavirenz-based ART. Ongoing longer-term follow-up will further inform sustainability and further research is required for settings where viral load monitoring is less accessible. Correspondence to: Dr Anna Turkova, MRC Clinical Trials Unit at UCL, London, WC2B 6NH, UK a.turkova@ucl.ac.uk Supplementary Material Supplementary appendix Acknowledgments This trial is funded by the UK National Institute for Health Research Health Technology Assessment (projects: 08/53/25; 11/136/108); European Commission through EuroCoord (FP7/2007/2015); PENTA Foundation; UK Medical Research Council; and INSERM SC10-US19, France. The qualitative substudy was additionally funded by the Economic and Social Research Council (RES-062-23-2308). This trial was sponsored by the Paediatric European Network for Treatment of AIDS (PENTA) Foundation. We thank all of the young people, their families, and staff from the centres participating in the BREATHER trial.

litative substudy was additionally funded by the Economic and Social Research Council (RES-062-23-2308). This trial was sponsored by the Paediatric European Network for Treatment of AIDS (PENTA) Foundation. We thank all of the young people, their families, and staff from the centres participating in the BREATHER trial. The BREATHER (PENTA 16) Writing group Karina Butler, Anna Turkova*, Jamie Inshaw*, Alexandra Compagnucci, Julia Kenny, Yacine Saidi, Victor Musiime, Annett Nanduudu, Tim R Cressey, Suwalai Chalermpantmetagul, Karen Scott, Lynda Harper, Sam Montero, Yoann Riault, Torsak Bunupuradah, Alla Volokha, Patricia M Flynn, Rosa Bologna, Hilda Kizito, Jose T Ramos, Eleni Nastouli, Nigel Klein, Carlo Giaquinto, Deborah Ford*, Abdel Babiker*, Diana M Gibb, on behalf of the BREATHER (PENTA 16) trial team. *Equal contribution. A full list of contributors is given in the appendix.

ann Riault, Torsak Bunupuradah, Alla Volokha, Patricia M Flynn, Rosa Bologna, Hilda Kizito, Jose T Ramos, Eleni Nastouli, Nigel Klein, Carlo Giaquinto, Deborah Ford*, Abdel Babiker*, Diana M Gibb, on behalf of the BREATHER (PENTA 16) trial team. *Equal contribution. A full list of contributors is given in the appendix. Contributors The BREATHER (PENTA 16) trial was designed by KB, DMG, AB, AC, YS, and CG; JK, VM, TB, and JTR had input in the design and conduct of the pilot study. The trial was coordinated by LH, KS, SM, AT (MRC CTU at UCL, London, UK), AC, YS, YR (INSERM, France), TRC, and SC (PHPT, Thailand). HK, VM, TB, AN, AV, PMF, and RB recruited and collected data for substantial numbers of patients in the trial. AB and JI wrote the trial analysis plan, which was reviewed by KB, DMG, YS, and AC. JI and DF undertook the analysis. All authors contributed to interpretation of the data. DMG, JI, AT, JK, KS, KB, and DF wrote drafts of the report. NK, EN, and JK coordinated and implemented laboratory aspects of the study and contributed to manuscript preparation. All authors revised the report critically and approved the final version. Declaration of interests We declare no competing interests. Figure 1 Trial profile *One participant was unable to attend the randomisation visit due to a traffic accident and another participant was excluded because of unreliable attendance. Figure 2 Time to viral rebound

Contributors The BREATHER (PENTA 16) trial was designed by KB, DMG, AB, AC, YS, and CG; JK, VM, TB, and JTR had input in the design and conduct of the pilot study. The trial was coordinated by LH, KS, SM, AT (MRC CTU at UCL, London, UK), AC, YS, YR (INSERM, France), TRC, and SC (PHPT, Thailand). HK, VM, TB, AN, AV, PMF, and RB recruited and collected data for substantial numbers of patients in the trial. AB and JI wrote the trial analysis plan, which was reviewed by KB, DMG, YS, and AC. JI and DF undertook the analysis. All authors contributed to interpretation of the data. DMG, JI, AT, JK, KS, KB, and DF wrote drafts of the report. NK, EN, and JK coordinated and implemented laboratory aspects of the study and contributed to manuscript preparation. All authors revised the report critically and approved the final version. Declaration of interests We declare no competing interests. Figure 1 Trial profile *One participant was unable to attend the randomisation visit due to a traffic accident and another participant was excluded because of unreliable attendance. Figure 2 Time to viral rebound (A) Kaplan-Meier graph adjusted for stratification factors—time from randomisation to viral rebound (confirmed viral load ≥50 copies per mL). (B) Estimated difference in proportion of participants with viral rebound (two-sided 90% CI) between short cycle therapy and continuous therapy at week 48 for different analyses. *Difference in estimated probability of viral rebound, Kaplan-Meier methods, with adjustment for study stratification factors. †Difference in estimated probability of viral rebound, Kaplan-Meier methods. ‡With exact confidence intervals. §Kaplan-Meier methods, censoring individuals who violated the profile at that time, with adjustment for study stratification factors.

0·6 (0·4–0·9) 0·6 (0·4–0·9) 0·6 (0·4–0·9) D-dimers (ng/mL) 69·1 (3·13–135·4) 65·7 (4·8–80·3) 67·5 (3·1–152·2) CRP (pg/mL) 631·2 (303·8–2676·1) 621·6 (260·8–2164·1) 626·8 (288·8–2311·0) Data are median (IQR) or n (%). CDC=US Centers for Disease Control and Prevention. ART= antiretroviral therapy. CRP=C-reactive protein. * Three young people acquired HIV through blood products (one in the short cycle therapy group and two in the continuous therapy group), two had uncertain methods of transmission (one in the continuous therapy group and one in the short cycle therapy group). † One young person in the continuous therapy group with unknown US CDC stage at randomisation. ‡ The remaining nucleotide reverse transcriptase inhibitor backbones in two patients were (1) zidovudine plus lamivudine plus tenofovir and (2) didanosine plus abacavir. § Seven samples (one in the short cycle therapy group and six in the continuous therapy group) were not available for testing with an ultra-sensitive assay. Table 2 Trial efficacy from randomisation to week 48 assessment

‡ The remaining nucleotide reverse transcriptase inhibitor backbones in two patients were (1) zidovudine plus lamivudine plus tenofovir and (2) didanosine plus abacavir. § Seven samples (one in the short cycle therapy group and six in the continuous therapy group) were not available for testing with an ultra-sensitive assay. Table 2 Trial efficacy from randomisation to week 48 assessment Short cycle therapy (n=99) Continuous therapy (n=100) p value Primary endpoint Participants with confirmed viral load ≥50 copies/mL 6 (6%) 7 (7%) 0·75 Secondary endpoints Participants with confirmed viral load ≥400 copies/mL 2 (2%) 4 (4%) 0·38 Participants with change in ART regimen 3 (3%) 9 (9%) 0·13 Viral rebound 0 1 .. Toxicity* 1 4 .. Adherence problems 1 1 .. Simplification 1 3 .. Participants with mutations present at viral rebound† (participants with resistance test result available) 2 (3) 5 (6) 1·00 Number of NNRTI mutations None 1 1 .. 1–2 1 5 .. 3 or more 1 0 .. Number of NRTI mutations None 2 5 .. 1 1 1 .. Mean change in CD4 percentage (%) 0·2% (0·4) 0·1% (0·4) 0·76 Mean change in absolute CD4 count (cells per μL) −34·2 (20·9) −21·6 (21·1) 0·67 Substudy results n=98 n=94 Viral load ≥20 copies/mL at week 48 13 (13%) 14 (15%) 0·84 <20 copies/mL at week 48 85 (87%) 80 (85%) Mean change in total HIV-1 DNA (Ln copies per million cells) 0·1 (0·1) −0·2 (0·1) 0·13 Mean change in interleukin 6 (Ln pg/mL) 0·0 (0·1) 0·1 (0·1) 0·64 Mean change in D-dimers (Ln ng/mL) −0·5 (0·2) −0·0 (0·2) 0·05 Data are n (%) or mean change from randomisation (SE), unless otherwise stated. ART=antiretroviral therapy. NRTI=nucleoside/nucleotide reverse transcriptase inhibitors. NNRTI=non-nucleoside reverse transcriptase inhibitor. Ln=natural logarithm.

·1 (0·1) 0·64 Mean change in D-dimers (Ln ng/mL) −0·5 (0·2) −0·0 (0·2) 0·05 Data are n (%) or mean change from randomisation (SE), unless otherwise stated. ART=antiretroviral therapy. NRTI=nucleoside/nucleotide reverse transcriptase inhibitors. NNRTI=non-nucleoside reverse transcriptase inhibitor. Ln=natural logarithm. * One gynaecomastia in the short cycle therapy group, three lipodystrophy events with onset before baseline in the continuous therapy group, and one raised transaminases in the continuous therapy group. † Two participants on short cycle therapy: (1) Leu100Ile, Lys103Asn, Tyr188Cys and (2) Lys103Asn, Met184Val; five participants on continuous therapy: (1) Val106, Glu138Ala, (2) Lys103Asn, Val106Met, (3) Met230Leu, (4) Lys103Asn, Val106Met, and (5) Met184Val, Glys190Ser; samples from four additional patients (three in the short cycle therapy group and one in the continuous therapy group) with low level viraemia failed to amplify. Table 3 Adverse events from randomisation to week 48 assessment

† Two participants on short cycle therapy: (1) Leu100Ile, Lys103Asn, Tyr188Cys and (2) Lys103Asn, Met184Val; five participants on continuous therapy: (1) Val106, Glu138Ala, (2) Lys103Asn, Val106Met, (3) Met230Leu, (4) Lys103Asn, Val106Met, and (5) Met184Val, Glys190Ser; samples from four additional patients (three in the short cycle therapy group and one in the continuous therapy group) with low level viraemia failed to amplify. Table 3 Adverse events from randomisation to week 48 assessment Short cycle therapy (n=99) Continuous therapy (n=100) Total (n=199) Grade 3 and 4 adverse events 13 (8) 14 (12) 27 (20) Clinical Infections and infestations 3 1 4 Nervous system disorders* 2 1 3 Skin and subcutaneous tissue disorders 0 1 1 Surgical and medical procedures 0 2 2 Kaposi's sarcoma (AIDS related) 1 0 1 Suicidal ideation 0 1 1 Gynaecomastia 1 0 1 Laboratory Neutropenia 2 6 8 Low density lipoprotein cholesterol increased 1 1 2 Bilirubin increased 1 0 1 Calcium decreased 1 0 1 Glucose decreased 1 0 1 Alkaline phosphatase increased 0 1 1 ART-related adverse events (all grades) 2 (2) 14 (10) 16 (12) Lipodystrophy† 0 5 5 Gynaecomastia 1 2 3 Suicidal ideation 0 1 1 Dizziness 0 1 1 Headache and syncope 0 1 1 Spontaneous abortion 1 1 2 Neutropenia 0 1 1 Raised transaminases 0 2 2 Treatment-modifying adverse events (all grades)‡ 1 (1) 4 (4) 5 (5) Serious adverse events§ 7 (6) 6 (3) 13 (9) Serious adverse event rate per 100 person-years (95% CI) 6·9 (3·3–14·4) 5·9 (2·6–13·1) 6·4 (3·7–10·9) Data are number of episodes (number of participants), unless otherwise stated. ART=antiretroviral therapy. The only significant difference in numbers of adverse events or number of patients with adverse events were in ART-related adverse events (Poisson p=0·02, Fisher's exact p=0·03, respectively).

13·1) 6·4 (3·7–10·9) Data are number of episodes (number of participants), unless otherwise stated. ART=antiretroviral therapy. The only significant difference in numbers of adverse events or number of patients with adverse events were in ART-related adverse events (Poisson p=0·02, Fisher's exact p=0·03, respectively). * One headache (short cycle therapy group), one hemiparesis (short cycle therapy group), and one collapse or suspected seizure (continuous therapy group). † Lipodystrophy events (continuous therapy group): two new and three worsening events with onset before baseline. ‡ One gynaecomastia (short cycle therapy group), three worsening lipodystrophy events with onset before baseline (continuous therapy group), and one raised concentration of transaminases (continuous therapy group). § Serious adverse events in the short cycle therapy group: five admissions to hospital (one headache, one exacerbation of bronchiectasis, one Kaposi's sarcoma, one measles, and one epistaxis) and two other important medical conditions (one spontaneous abortion and one transient hemiparesis); in the continuous therapy group: one life threatening (suicidal ideation), four admissions to hospital (one contusion of chest, one collapse or suspected seizure, one spontaneous abortion, and one appendicitis) and one other important medical condition (neurosyphilis).

Introduction The AIDS response has had outstanding success in the development and global scale-up of antiretroviral therapy (ART).1 With a few exceptions, there has been less success in reducing the spread of the HIV epidemic.2, 3, 4 However, the sustainability of the response rests upon there being large reductions in new HIV infections in the coming years.5 To meet this challenge, prevention interventions have been the focus of much research and development investment; programmes are now in the fortunate position of having a range of interventions to consider, including male condoms, voluntary medical male circumcision (VMMC) services, increased HIV testing and initiation of ART for all diagnosed with HIV (early ART), and oral pre-exposure prophylaxis (PrEP). These interventions need to be prioritised for populations and groups at the greatest risk and who stand to benefit most.6, 7, 8 There is also a growing understanding of the barriers to impact that are faced. These barriers have been articulated as the prevention cascade (Hargreaves and colleagues,9 this issue) and include limitations in the supply of, demand for, and adherence to an intervention arising from logistic, structural, and behavioural factors. Arguably, innovative approaches might be able to allow these barriers to increasingly be overcome.

culated as the prevention cascade (Hargreaves and colleagues,9 this issue) and include limitations in the supply of, demand for, and adherence to an intervention arising from logistic, structural, and behavioural factors. Arguably, innovative approaches might be able to allow these barriers to increasingly be overcome. At the same time, there are many advances in the development of new technologies that offer new possible interventions to prevent HIV infections. In the medium term, intravaginal rings and long-acting injectable antiretroviral drugs might be efficacious and acceptable to potential users.10, 11, 12, 13, 14 In the longer term, effective transfusions of broadly neutralising antibodies (bNAbs), or a vaccine, might become available.15, 16 These new technological developments offer promise, but it should be remembered that these too will have to overcome the same prevention cascade barriers as other interventions.

12, 13, 14 In the longer term, effective transfusions of broadly neutralising antibodies (bNAbs), or a vaccine, might become available.15, 16 These new technological developments offer promise, but it should be remembered that these too will have to overcome the same prevention cascade barriers as other interventions. For programme planning and for determining investments in research and development, it is crucial to be able to establish priorities among this large portfolio. Usually, the value and probable effect of each intervention is considered separately. However, because they overlap in their functionality and the effect of one intervention affects that of others, a holistic perspective is needed to establish a coherent strategy for HIV prevention that maximises the reduction in HIV incidence and is aligned with information about the epidemic, and the potential risks, costs, and benefits of new developments. To do this, researchers should investigate which configuration of present and future interventions will maximise prevention with available resources, assess the contribution of each component to overall success to inform research and development priorities, and monitor how new information on the success of interventions (in scale-up or further development) informs the balance of the portfolio and the overall effect. Research in context Evidence before this study

For programme planning and for determining investments in research and development, it is crucial to be able to establish priorities among this large portfolio. Usually, the value and probable effect of each intervention is considered separately. However, because they overlap in their functionality and the effect of one intervention affects that of others, a holistic perspective is needed to establish a coherent strategy for HIV prevention that maximises the reduction in HIV incidence and is aligned with information about the epidemic, and the potential risks, costs, and benefits of new developments. To do this, researchers should investigate which configuration of present and future interventions will maximise prevention with available resources, assess the contribution of each component to overall success to inform research and development priorities, and monitor how new information on the success of interventions (in scale-up or further development) informs the balance of the portfolio and the overall effect. Research in context Evidence before this study Many empirical and modelling studies have investigated the effect and cost-effectiveness of different HIV prevention interventions, singly and in combination, and across different settings. More recently, modelling studies have done these analyses under the constraint of resource allocation—whereby multiple combinations of interventions are examined to understand the best way of spending a fixed budget. We searched PubMed for HIV prevention studies published between Jan 1, 2000, and Dec 31, 2015, with the terms “HIV prevention” AND (“budget allocation” OR “resource allocation”) AND (“model” OR “modeling” OR “modelling”). Our results included a 2013 systematic review of cost-effectiveness modelling studies of PrEP, which found that PrEP has the potential to be a cost-effective addition to HIV prevention programmes in specific settings, particularly when delivered to key populations at highest risk of HIV exposure. We identified 22 abstracts, of which 17 studies met our inclusion criteria of mathematical modelling studies. Most resource allocation models are from a health economics perspective with a short-term outlook, with only a few including a dynamic HIV transmission model (which account for future trends in the epidemic). Of two recent analyses including detailed economic and epidemiological information from Kenya and south India respectively, one shows the additional gains of geographical prioritisation of interventions by local epidemiology, and one shows that as budget levels increase, the optimum intervention strategy is to first increase intervention intensity before scaling up coverage. Other recent studies on new technologies highlight that long-acting PrEP for high-risk women could be cost-effective and that even a medium-efficacy vaccine could have a substantial impact on the epidemic. Resource allocation models are increasingly being developed and used at international and national levels to help guide HIV prevention policy.

technologies highlight that long-acting PrEP for high-risk women could be cost-effective and that even a medium-efficacy vaccine could have a substantial impact on the epidemic. Resource allocation models are increasingly being developed and used at international and national levels to help guide HIV prevention policy. Added value of this study We analyse present and future interventions within the framework of optimum resource allocation using a synthesis of all available data. We find that scaling up existing interventions is the most cost-effective way to stem new HIV infections in South Africa, with new technologies able to plug the gaps in cases where the predicted scale-ups are not possible. Implications of all the available evidence Our findings add to an increasing body of evidence that frontloading HIV prevention investments to maximise the use of interventions which are available now leads to the largest health gains in the long term. To achieve these aims, we developed a model of the HIV epidemic in South Africa. We incorporated the extent of current intervention scale-up and simulated the cost and effect of a wide portfolio of options for HIV prevention, encompassing the further scale-up of existing interventions and the introduction of interventions focused on new technological developments, to identify optimum allocations of resources.

ated the extent of current intervention scale-up and simulated the cost and effect of a wide portfolio of options for HIV prevention, encompassing the further scale-up of existing interventions and the introduction of interventions focused on new technological developments, to identify optimum allocations of resources. Methods Overview We briefly describe the analysis framework, model design, and assumptions about interventions' effects and costs (further details are provided in the appendix). We began by enumerating all the potential interventions (existing and future technologies) that could be scaled up and formed assumptions about the nature of their use (coverage, timing, priority groups) in a range of scenarios (table 1). We defined three sets of scale-up assumptions that span the range of possible scale-up scenarios for each intervention (except for vaccination, which has only two): the constant scenario assumed no introduction or additional scale-up of prevention interventions; the medium scenario characterised the degree of long-term scale-up that might be anticipated on the basis of current planning; and the maximum scenario determined the fullest possible extent to which those interventions could be deployed. The medium scenario is credible and is based on consultation with a working group of Bill & Melinda Gates Foundation staff for the development of each of these interventions (appendix p 13).

current planning; and the maximum scenario determined the fullest possible extent to which those interventions could be deployed. The medium scenario is credible and is based on consultation with a working group of Bill & Melinda Gates Foundation staff for the development of each of these interventions (appendix p 13). We characterised the optimum set of interventions across a wide range of assumptions for the budget available for HIV prevention. Next, we simulated the effect and evaluated the total cost of implementing the medium scenario from 2016 to 2050. We then explored all possible permutations of the scale-up of the interventions (n=4374, from seven interventions with three possible coverage levels and one intervention with two levels) to determine if other configurations of interventions could, with the same total cost as the medium scenario, achieve a greater impact (defined as more infections averted over the period 2016–50). We repeated the analysis under alternative sets of assumptions whereby the development of new technologies was not realised, and whereby further scale-up of existing interventions was not possible (table 2).

We characterised the optimum set of interventions across a wide range of assumptions for the budget available for HIV prevention. Next, we simulated the effect and evaluated the total cost of implementing the medium scenario from 2016 to 2050. We then explored all possible permutations of the scale-up of the interventions (n=4374, from seven interventions with three possible coverage levels and one intervention with two levels) to determine if other configurations of interventions could, with the same total cost as the medium scenario, achieve a greater impact (defined as more infections averted over the period 2016–50). We repeated the analysis under alternative sets of assumptions whereby the development of new technologies was not realised, and whereby further scale-up of existing interventions was not possible (table 2). Model design We developed a deterministic compartmental model of heterosexual HIV transmission in South Africa (full description and parameter values are given in Cremin and colleagues23 and the appendix). The model represents age, sex, behavioural risk, HIV infection, declining CD4 cell count, ART scale-up, and the eight HIV prevention interventions, and is calibrated to South African demography, age-specific and sex-specific HIV incidence and prevalence, the historical scale-up of ART, circumcision, and patterns of condom use (appendix pp 1–12).24, 25 We assume that almost all HIV-positive individuals with CD4 counts below 200 cells per μL present for care and receive ART (late ART).

South African demography, age-specific and sex-specific HIV incidence and prevalence, the historical scale-up of ART, circumcision, and patterns of condom use (appendix pp 1–12).24, 25 We assume that almost all HIV-positive individuals with CD4 counts below 200 cells per μL present for care and receive ART (late ART). Eight prevention methods were included in the analysis (table 1). Efficacy assumptions for all interventions represent the underlying biological efficacy among people who adhere to the method in question (appendix pp 19–22). For existing methods (male condoms, VMMC, early ART [ie, outreach testing and offering ART to all diagnosed with HIV], and oral PrEP), efficacy estimates were based on current data.17, 18, 19, 20, 21, 22, 26, 27, 28 For new PrEP products (intravaginal rings, long-acting antiretroviral drugs, and bNAbs), the effectiveness of intravaginal rings was based on that observed among the women who were most highly adherent in the Ring Study,11 and long-acting antiretroviral drugs and bNAbs were assumed to have a similar effectiveness to oral PrEP but potentially reached a wider part of the population. We included two vaccine formulations: the P5-like vaccine is similar to the pox-protein vaccines in development with 50% efficacy and the idealised vaccine is assumed to have a higher efficacy (70%) and lower attrition rate (5% compared with 10% per year) due to the need for less frequent booster administrations.16 In scenarios with vaccination available, we assumed that the P5-like vaccine will be available from 2024, initially for adults and switching to teenagers aged 14 years from 2026. The idealised vaccine will replace the P5-like vaccine from 2030 onwards.

year) due to the need for less frequent booster administrations.16 In scenarios with vaccination available, we assumed that the P5-like vaccine will be available from 2024, initially for adults and switching to teenagers aged 14 years from 2026. The idealised vaccine will replace the P5-like vaccine from 2030 onwards. For each intervention, we defined coverage levels for seven population subgroups (female sex workers aged 15–49 years, high-risk women aged 15–29 years, low-risk women aged 15–29 years, high-risk women aged 30–49 years, low-risk women aged 30–49 years, high-risk men aged 15–49 years, low-risk men aged 15–49 years) under each set of scale-up assumptions (appendix pp 15–24). For the PrEP products (oral PrEP, intravaginal rings, long-acting antiretroviral drugs, and bNAbs), the coverage level was the proportion of persons taking PrEP with sufficient adherence such that they benefit fully from the assumed efficacy. A product cannibalism assumption is also incorporated into the PrEP products' coverage levels, such that the introduction of a new product takes some users from existing products but always increases total PrEP coverage (appendix p 14). The relative coverage of each product when implemented in combination is proportional to its coverage in that subgroup when no other PrEP product is available. When any vaccine is introduced, we assumed that oral PrEP, intravaginal rings, long-acting antiretroviral drugs, and bNAbs are scaled back to zero over 5 years from 2035.

ve coverage of each product when implemented in combination is proportional to its coverage in that subgroup when no other PrEP product is available. When any vaccine is introduced, we assumed that oral PrEP, intravaginal rings, long-acting antiretroviral drugs, and bNAbs are scaled back to zero over 5 years from 2035. Most interventions (condoms, oral PrEP, long-acting antiretroviral drugs, bNAbs) prioritise female sex workers and young women, intravaginal rings prioritise female sex workers, VMMC prioritises young men, and early ART and vaccination are assumed to have uniform coverage by risk, with vaccination primarily targeted to teenagers aged 14 years. All interventions are scaled up linearly over a given time period.

female sex workers and young women, intravaginal rings prioritise female sex workers, VMMC prioritises young men, and early ART and vaccination are assumed to have uniform coverage by risk, with vaccination primarily targeted to teenagers aged 14 years. All interventions are scaled up linearly over a given time period. Costs are intended to represent the cost of fully delivering the intervention with the stated scale and efficacy. The rationale for assumptions around fixed and variable costs is given in the appendix. Fixed costs (one-time or recurring annually, but not related to scale) were derived from information about existing product launches. Variable costs (dependent on scale) were based on the population group in question and include the commodity plus a combination of service delivery, testing, laboratory costs, outreach and demand incentives, as appropriate. For all products, there is a step change in the fixed and variable cost when coverage rises above the medium scenario, such that the marginal cost is greater to deliver the intervention at maximum coverage. The total costs for all interventions were scaled up by a factor of 1·43 to represent the 30% of the HIV budget allocated to indirect costs that are not explicitly included here. Throughout, we refer to each of these interventions with a short-hand label (eg, condoms), but the meaning remains that a full set of activities consistent with increasing the use of condoms as specified in our assumptions (eg, distribution, promotion, supply chain management, social marketing).

Costs are intended to represent the cost of fully delivering the intervention with the stated scale and efficacy. The rationale for assumptions around fixed and variable costs is given in the appendix. Fixed costs (one-time or recurring annually, but not related to scale) were derived from information about existing product launches. Variable costs (dependent on scale) were based on the population group in question and include the commodity plus a combination of service delivery, testing, laboratory costs, outreach and demand incentives, as appropriate. For all products, there is a step change in the fixed and variable cost when coverage rises above the medium scenario, such that the marginal cost is greater to deliver the intervention at maximum coverage. The total costs for all interventions were scaled up by a factor of 1·43 to represent the 30% of the HIV budget allocated to indirect costs that are not explicitly included here. Throughout, we refer to each of these interventions with a short-hand label (eg, condoms), but the meaning remains that a full set of activities consistent with increasing the use of condoms as specified in our assumptions (eg, distribution, promotion, supply chain management, social marketing). Role of the funding source A working group of Bill & Melinda Gates Foundation staff (appendix p 13) together with the coauthors contributed to the study design and data collection. With the exception of coauthors, the funder had no further role in data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

ributed to the study design and data collection. With the exception of coauthors, the funder had no further role in data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results The effect of each intervention used in isolation was assessed (appendix p 48). All interventions have the potential to reduce HIV incidence substantially over the period 2016–50. Vaccination has the largest potential impact when scaled up to maximum coverage, followed by long-acting antiretroviral drugs, oral PrEP, bNAbs, and condoms.

h intervention used in isolation was assessed (appendix p 48). All interventions have the potential to reduce HIV incidence substantially over the period 2016–50. Vaccination has the largest potential impact when scaled up to maximum coverage, followed by long-acting antiretroviral drugs, oral PrEP, bNAbs, and condoms. The cost and effect of all permutations of the scale-up of interventions were examined in the model (figure 1). A frontier can be constructed across these permutations that shows the maximum effect (infections averted with respect to the projected epidemic under the constant assumptions for all interventions) that can be achieved for a given cost. At the low end of the frontier, at which point interventions are scaled up minimally, 1·9 million infections would be averted and the cost of the programme would be US$44 billion ($1·3 billion per year). Most of these costs are attached to spending on late ART for those entering care (84%, data not shown) whereas the effect is produced by increased use of condoms and VMMC (appendix p 45), together with the effect of late ART in reducing transmission. At the high end, at which point all interventions are scaled up to the maximum possible extent, an additional 4·7 million infections are averted at a marginal cost of $26 billion. With increasing available budget, additional interventions are incorporated incrementally (appendix p 45). Condoms, VMMC, early ART, oral PrEP, and a vaccine are implemented at the low end of the frontier, with the inclusion of the intravaginal rings, long-acting antiretroviral drugs, and finally bNAbs, as resources increase.

of risk, actual risk, duration of risk, and duration of protection provided by an intervention need to be considered when thinking about how to generate a cascade where perception of risk is included. More consideration of such issues will be required when prospectively collecting data to populate prevention cascades. The data analysed here were not collected with prevention cascade analyses in mind. Other studies could more explicitly include questions relating to each step on the cascades. Prevention cascades are most useful when applied to the underlying population at risk (ie, using population based surveys) rather than those adopting services, since the latter only capture those already taking up services. For planners looking to organise effective HIV prevention this population-level view can come from surveys such as the current HIV Impact Assessments.23 However, implementers can still consider in their monitoring and assessment plans the steps of the cascade that they need to improve for their interventions to have effect.

creasing available budget, additional interventions are incorporated incrementally (appendix p 45). Condoms, VMMC, early ART, oral PrEP, and a vaccine are implemented at the low end of the frontier, with the inclusion of the intravaginal rings, long-acting antiretroviral drugs, and finally bNAbs, as resources increase. The medium scenario, where medium coverage is implemented for all interventions, implies a cost of $50 billion over 34 years. The configuration of interventions represented in this scenario is almost on the frontier, signifying that it is approaching the optimum use of resources (figure 1). For the same cost, an alternative configuration of interventions would reduce HIV incidence more rapidly at first and generate 110 000 extra infections averted between 2016 and 2050 (figure 1). Comparison between the optimum allocation configuration of interventions and the medium coverage level throughout (representing current programmatic aims) shows that increasing coverage of existing interventions (VMMC and early ART) to a greater extent than current projections would be prioritised in the optimum allocation scenario, together with implementing intravaginal rings (figure 1). The scale-up of condoms, oral PrEP, and vaccines in the optimum allocation is consistent with the medium scenario. However, increasing coverage of long-acting antiretroviral drugs is only partly selected and bNAbs are not selected in the optimum configuration of interventions.

mplementing intravaginal rings (figure 1). The scale-up of condoms, oral PrEP, and vaccines in the optimum allocation is consistent with the medium scenario. However, increasing coverage of long-acting antiretroviral drugs is only partly selected and bNAbs are not selected in the optimum configuration of interventions. In the optimum allocation configuration, the largest amount of resources for prevention interventions is used for VMMC programming (28%; figure 1). If the variable cost of bNAbs was reduced by just 7% it would enter the optimum allocation at low levels and it would require more than 25% reduction in variable cost to be included at higher coverage (appendix p 43). An additional cost of at least $4·5 billion would be required to push vaccination to below medium or maximum coverage in the optimum allocation (appendix p 44).

uld enter the optimum allocation at low levels and it would require more than 25% reduction in variable cost to be included at higher coverage (appendix p 43). An additional cost of at least $4·5 billion would be required to push vaccination to below medium or maximum coverage in the optimum allocation (appendix p 44). This analysis was repeated under different assumptions about the availability of interventions (figure 2). The failure for long-term future products to be developed (scenarios B and D in table 2), or the failure for a vaccine to be developed (scenario C and D in table 2), does not substantially affect the overall configuration of the programme in the respective optimum allocation for each. Future products do not attract substantial resources in the optimum allocation scenarios so their loss has little effect (figure 2). However, the financial resources released by the loss of a vaccine are outweighed by those needed to fund the additional ART used if there were no vaccine (given the higher rate of new infections), which also requires the removal of the intravaginal rings from the optimum allocation (figure 2).

ittle effect (figure 2). However, the financial resources released by the loss of a vaccine are outweighed by those needed to fund the additional ART used if there were no vaccine (given the higher rate of new infections), which also requires the removal of the intravaginal rings from the optimum allocation (figure 2). The optimum allocation of resources does change if a greater coverage of existing interventions cannot be achieved. If condom programmes are unable to achieve higher coverage than the current level (constant scenario), then the optimum allocation configuration still includes a strong emphasis on VMMC, early ART, intravaginal rings, and vaccination, and maintains a moderate emphasis on oral PrEP (figure 2). If existing interventions (condoms, VMMC, early ART) are all unable to expand further then the optimum configuration of resources includes high levels of oral PrEP and intravaginal rings, plus a greater emphasis on long-acting antiretroviral drugs (figure 2). The biggest single determinant of overall effect is the development of a vaccine (figure 3); the expected number of infections averted is 35% lower without a vaccine. Overall effect is also reduced by up to 23% if the scale-up of condom and other programmes is limited. By contrast, the failure of bNAbs or long-acting antiretroviral drugs to develop has little influence on the overall maximum effect, provided that a vaccine is developed.

ctions averted is 35% lower without a vaccine. Overall effect is also reduced by up to 23% if the scale-up of condom and other programmes is limited. By contrast, the failure of bNAbs or long-acting antiretroviral drugs to develop has little influence on the overall maximum effect, provided that a vaccine is developed. A sensitivity analysis where the model was recalibrated for Cross River State, Nigeria, shows a different optimum allocation (for details of calibration see appendix pp 4–6, 46). Cross River State has a concentrated HIV epidemic with lower overall HIV prevalence than South Africa, a smaller group at highest risk driving the epidemic, and very high VMMC coverage.29 In this setting, the budget available to allocate across interventions is much lower ($5·6 billion), meaning that fewer interventions are affordable and interventions that target those at highest risk of infection are preferred. In the optimum allocation, condoms, intravaginal rings, and vaccination are all implemented at maximum levels, with all other interventions remaining constant.

ns is much lower ($5·6 billion), meaning that fewer interventions are affordable and interventions that target those at highest risk of infection are preferred. In the optimum allocation, condoms, intravaginal rings, and vaccination are all implemented at maximum levels, with all other interventions remaining constant. Discussion Overall, this analysis highlights the need to exploit fully the prevention interventions available today. We find that with a budget consistent with a reasonable set of aims (ie, the budget of the medium scenario), the greatest effect is generated if prevention efforts focus on high coverage of VMMC and early ART programming, use of intravaginal rings particularly by sex workers and, later, achieving high coverage of a vaccine. The single most crucial intervention is the development of a vaccine. However, if it is not possible to scale up existing interventions to the extent envisioned here, the use of PrEP products would become more important to generate the most impact possible. In this light, a focus on removing the barriers that exist to greater uptake of VMMC, together with early ART, especially, is the priority. An exclusive focus on the development and scaling up of new products, which themselves will have to confront similar barriers, would not generate the greatest effect. However, the widening range of interventions provides opportunities to maintain effect by compensating for failures in other potential interventions.

ty. An exclusive focus on the development and scaling up of new products, which themselves will have to confront similar barriers, would not generate the greatest effect. However, the widening range of interventions provides opportunities to maintain effect by compensating for failures in other potential interventions. There are two major limitations to this analysis. First, the representation of product coverage in the model includes three discrete patterns over the different population sub-groups rather than an exploration of all the possible combinations. Although these represent expectations among the authors and the technical group (appendix p 13), this could affect the cost–benefit profile of interventions. For example, products assumed to be targeted to high-risk groups (eg, condom use, PrEP products) might seem more attractive than those that have uniform coverage by risk or age (eg, early ART), and if interventions could be scaled-up differentially by age or risk group, a different combination might have been identified in this analysis as a priority. Furthermore, some interventions could have already reached saturation—for example, condom prevention programmes in hyperendemic settings have been aiming to increase condom use for many years with mixed success.

lly by age or risk group, a different combination might have been identified in this analysis as a priority. Furthermore, some interventions could have already reached saturation—for example, condom prevention programmes in hyperendemic settings have been aiming to increase condom use for many years with mixed success. Second, cost assumptions in the model are not perfectly informed by data. This limitation is inevitable when making projections for scale-up of future interventions. However, we aimed to make consistent assumptions across the interventions for the analysis to be directionally correct; and the total spending in the model for recent years is similar to South Africa HIV programme costs.30, 31 Estimated condom and VMMC costs are higher in this cost model than the recent South African HIV and TB Investment Case, but the results are broadly consistent with both identified as important priorities.31 We do not include research and development costs but these might also be substantial; neither do we include the potential development of drug resistance or the extra treatment costs these could incur.

ican HIV and TB Investment Case, but the results are broadly consistent with both identified as important priorities.31 We do not include research and development costs but these might also be substantial; neither do we include the potential development of drug resistance or the extra treatment costs these could incur. The analysis focuses on South Africa, but the sensitivity analysis for a lower prevalence setting shows that the results are highly dependent on the epidemic setting. The setting will determine budget level and relative effectiveness of the different interventions according to the size and level of risk across the different population groups. This study complements substantial other modelling work about the allocation of resources in South Africa by including future interventions and taking a long-term perspective.31, 32 It adds to a growing body of literature on resource allocation for HIV prevention, which highlights the need to account for local epidemiology to maximise the efficiency of HIV prevention planning.6, 33, 34, 35, 36, 37

resources in South Africa by including future interventions and taking a long-term perspective.31, 32 It adds to a growing body of literature on resource allocation for HIV prevention, which highlights the need to account for local epidemiology to maximise the efficiency of HIV prevention planning.6, 33, 34, 35, 36, 37 This analysis represents an interrogation of the information available today, with a view on the evidence of interventions and risk and benefits that is to an extent dependent upon the perspective of the authors. The intention would be that such an analysis would be updated as new information becomes available. A strategic approach in which limited resources are used to maximise prevention effect would focus on strengthening the scale-up of existing interventions, while urgently pursuing a workable vaccine and developing other approaches that can be used if increasing use of existing interventions is limited. Supplementary Material Supplementary appendix Acknowledgments The authors thank James Hargreaves and Gina Dallabetta for their helpful comments. Contributors The ideas behind the analysis were developed by all authors. Model development and analysis was done by JAS with assistance from S-JA and JBM. KLH led the costing estimates. All authors contributed to the interpretation of findings. The manuscript was drafted by JAS and TBH. All authors approved the final version of the article for submission.

ysis were developed by all authors. Model development and analysis was done by JAS with assistance from S-JA and JBM. KLH led the costing estimates. All authors contributed to the interpretation of findings. The manuscript was drafted by JAS and TBH. All authors approved the final version of the article for submission. Declaration of interests JAS reports personal fees from Anansi Health, during the conduct of the study, and grants from the Bill & Melinda Gates Foundation, outside the submitted work. S-JA and JBM report personal fees from the Bill & Melinda Gates Foundation, Avenir Health, and Anansi Health, outside the submitted work. EL declares no competing interests. KLH and GPG are employees of the Bill & Melinda Gates Foundation. TBH reports grants from the Bill & Melinda Gates Foundation, World Bank, UNAIDS, Rush Foundation, Wellcome Trust, and personal fees from the Bill & Melinda Gates Foundation, during the conduct of the study. Figure 1 Effect achievable with the full range of interventions

Declaration of interests JAS reports personal fees from Anansi Health, during the conduct of the study, and grants from the Bill & Melinda Gates Foundation, outside the submitted work. S-JA and JBM report personal fees from the Bill & Melinda Gates Foundation, Avenir Health, and Anansi Health, outside the submitted work. EL declares no competing interests. KLH and GPG are employees of the Bill & Melinda Gates Foundation. TBH reports grants from the Bill & Melinda Gates Foundation, World Bank, UNAIDS, Rush Foundation, Wellcome Trust, and personal fees from the Bill & Melinda Gates Foundation, during the conduct of the study. Figure 1 Effect achievable with the full range of interventions (A) The impact and cost of all possible interventions (grey dots), with the frontier highlighted (blue line). The medium scale-up scenario (purple dot), the optimum allocation point on the frontier with the same cost (green dot), and the maximum scale-up scenario (red dot) are highlighted. (B) Trajectory of HIV incidence in 15–49-year-olds, 2016–50, for the constant scale-up, medium scale-up, optimum allocation, and maximum scale-up scenarios. (C) The level of scale-up for each intervention in the optimum allocation scenario. (D) The distribution of costs by intervention type, 2016–50, in the optimum allocation scenario. VMMC=voluntary medical male circumcision. ART=antiretroviral therapy. PrEP=pre-exposure prophylaxis. IVR=intravaginal ring. LA-ARVs=long-acting antiretrovirals. bNAbs=broadly neutralising antibodies.

um allocation scenario. (D) The distribution of costs by intervention type, 2016–50, in the optimum allocation scenario. VMMC=voluntary medical male circumcision. ART=antiretroviral therapy. PrEP=pre-exposure prophylaxis. IVR=intravaginal ring. LA-ARVs=long-acting antiretrovirals. bNAbs=broadly neutralising antibodies. Figure 2 The level of scale-up for each intervention in the optimum allocation scenario, under different assumptions for the availability of interventions (see table 2) Scenario A, all interventions available; scenario B, as A without LA-ARVs or bNAbs; scenario C, as A without vaccine; scenario D, as A without vaccine, LA-ARVs, or bNAbs; scenario E, all interventions available, condom use limited to constant level; scenario F, all interventions available, condom use, VMMC, and early ART limited to constant levels. VMMC=voluntary medical male circumcision. ART=antiretroviral therapy. PrEP=pre-exposure prophylaxis. IVR=intravaginal ring. LA-ARVs=long-acting antiretrovirals. bNAbs=broadly neutralising antibodies. Figure 3 Maximum impact (HIV infections averted 2016–50) achievable with the resources implied in the medium scenario, under different assumptions for the availability of interventions (see table 2)

Scenario A, all interventions available; scenario B, as A without LA-ARVs or bNAbs; scenario C, as A without vaccine; scenario D, as A without vaccine, LA-ARVs, or bNAbs; scenario E, all interventions available, condom use limited to constant level; scenario F, all interventions available, condom use, VMMC, and early ART limited to constant levels. VMMC=voluntary medical male circumcision. ART=antiretroviral therapy. PrEP=pre-exposure prophylaxis. IVR=intravaginal ring. LA-ARVs=long-acting antiretrovirals. bNAbs=broadly neutralising antibodies. Figure 3 Maximum impact (HIV infections averted 2016–50) achievable with the resources implied in the medium scenario, under different assumptions for the availability of interventions (see table 2) Scenario A, all interventions available; scenario B, as A without long-acting antiretrovirals or broadly neutralising antibodies; scenario C, as A without vaccine; scenario D, as A without vaccine, long-acting antiretrovirals, or broadly neutralising antibodies; scenario E, all interventions available, condom use limited to constant level; scenario F, all interventions available, condom use, voluntary medical male circumcision, and early antiretroviral therapy limited to constant levels. Table 1 Summary of intervention assumptions

Scenario A, all interventions available; scenario B, as A without long-acting antiretrovirals or broadly neutralising antibodies; scenario C, as A without vaccine; scenario D, as A without vaccine, long-acting antiretrovirals, or broadly neutralising antibodies; scenario E, all interventions available, condom use limited to constant level; scenario F, all interventions available, condom use, voluntary medical male circumcision, and early antiretroviral therapy limited to constant levels. Table 1 Summary of intervention assumptions Efficacy Available from (medium and maximum scenarios) Main target group Effective coverage in main target group Fixed cost (US$) Variable cost (US$) Condoms 90%17 Now Female sex workers Constant: 29%; medium: 60%; maximum: 80% $3·43–8·23 million per partnership type per year $0·31–0·37 per condom VMMC 60%18, 19, 20 Now Young men Constant: 43%; medium: 80%; maximum: 80% $5 million launch (high coverage only) plus $27·5–36·3 million per population group per year $42 per person Early ART 85%21 Now All Constant: 0%; medium: 40%; maximum: 60% $10–11·6 million per year $275–295 per person per year Oral PrEP 90%22 Now Female sex workers, high-risk young women Constant: 0%; medium: 45%; maximum: 80% $5–15 million launch plus $3·3–11·4 million per year $170–190 per person per year IVR 65%11 2017 Female sex workers Constant: 0%; medium: 30%; maximum: 80% $10 million launch plus $5 million per year $107–115 per person per year LA-ARVs 90% 2020 Female sex workers, high-risk young women Constant: 0%; medium: 50%; maximum: 80% $10 million launch plus $5 million per year $180–200 per person per year bNAbs 90% 2028 Female sex workers, high-risk young women Constant: 0%; medium: 50%; maximum: 80% $10 million launch plus $5 million per year $190–210 per person per year P5-like vaccine 50% 2024 Teenagers aged 14 years Constant: 0%; medium and maximum: 70% $65 million launch plus $5–15 million per year $40–60 per person per year in first year, $3·5–4·5 per person per year thereafter Idealised vaccine 70% 2030 Teenagers aged 14 years Constant: 0%; medium and maximum: 80% $5 million per year throughout $50–60 per person per year in first year, $3·5–4·5 per person per year thereafter Efficacy refers to the protection afforded by perfect use of a product. Effective coverage is the proportion of people who fully adhere to a product such that they benefit from its protection. Full details on coverage, costing, and implementation across all population sub-groups are provided in the appendix. In the Effective coverage in main target group column, the three levels given indicate the assumptions under the constant, medium, and maximum scenarios. VMMC=voluntary medical male circumcision.

ection. Full details on coverage, costing, and implementation across all population sub-groups are provided in the appendix. In the Effective coverage in main target group column, the three levels given indicate the assumptions under the constant, medium, and maximum scenarios. VMMC=voluntary medical male circumcision. ART=antiretroviral therapy. PrEP=pre-exposure prophylaxis. IVR=intravaginal ring. LA-ARVs=long-acting antiretrovirals. bNAbs=broadly neutralising antibodies. Table 2 Primary (A) and alternative (B–F) scenarios, where some interventions are removed from consideration or restricted in use Condoms VMMC Early ART Oral PrEP IVR LA-ARVs bNAbs Vaccine All interventions available A ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ High-risk technologies not brought to market B ✓ ✓ ✓ ✓ ✓ × × ✓ No vaccine developed C ✓ ✓ ✓ ✓ ✓ ✓ ✓ × D ✓ ✓ ✓ ✓ ✓ × × × Condom interventions cannot be scaled up E * ✓ ✓ ✓ ✓ ✓ ✓ ✓ F * ✓ ✓ * * ✓ ✓ ✓ Scenario A, all interventions available; scenario B, as A without LA-ARVs or bNAbs; scenario C, as A without vaccine; scenario D, as A without vaccine, LA-ARVs, or bNAbs; scenario E, all interventions available, condom use limited to constant level; scenario F, all interventions available, condom use, VMMC, and early ART limited to constant levels. VMMC=voluntary medical male circumcision. ART=antiretroviral therapy. PrEP=pre-exposure prophylaxis. IVR=intravaginal ring. LA-ARVs=long-acting antiretrovirals. bNAbs=broadly neutralising antibodies. ✓=intervention is available at all coverage levels. X=intervention is never available. * Intervention is available only at constant scenario levels.

Introduction To halt the spread of HIV requires that efficacious treatment and preventive products or behaviours are used by those at risk. These products or behaviours are the direct mechanisms through which HIV prevention programmes exert their effect. Analyses of whether programmes are effectively preventing HIV transmission from mothers to children, and whether programmes are effectively treating patients with HIV, have used the concept of a cascade to explore the steps required to prevent infection or death.1, 2, 3, 4 This approach has powerfully illustrated what needs to be done to suppress HIV viral loads in many populations and is increasingly being used to compare the performance of prevention of mother to child transmission (PMTCT) and treatment programmes.5, 6 An HIV prevention cascade could be a similarly powerful approach to complement the treatment cascade, by defining the different steps in successful implementation of prevention interventions, providing estimates of the proportions of populations lost at each step in implementation, demonstrating points at which inefficiencies occur, and providing a framework for planning further actions.

mplement the treatment cascade, by defining the different steps in successful implementation of prevention interventions, providing estimates of the proportions of populations lost at each step in implementation, demonstrating points at which inefficiencies occur, and providing a framework for planning further actions. The idea of using models to describe the steps in controlling infectious diseases was initially developed to explore the effectiveness of tuberculosis control in the 1960s.7 Illustrating the coverage of diagnosis and treatment was extended to HIV, and has become common, although different approaches have been applied.6 The characteristics of the HIV treatment cascade depend upon whether data comes from health facilities tracking patients or from population based studies where all those infected with HIV can be included.6 The cascade also differs according to whether it provides a cross-sectional view of the population at a particular point in time or follows cohorts longitudinally describing their progression over time through the cascade.8 The description of care9 and other HIV services,10 in addition to treatment in cascades, can change the steps included and has culminated in the HIV prevention, treatment, and care cascade in the WHO Consolidated Strategic Information Guidelines for HIV in the Health Sector11 and the HIV Cascade Framework for Key Populations.12 These cascades attempt to include HIV prevention, but emphasise the steps of diagnosis, linkage to treatment, treatment initiation, and viral suppression. Often HIV testing is seen as an essential gateway to both HIV treatment and prevention, and is pivotal in the HIV prevention, treatment, and care cascade.11 The ambition of capturing all HIV prevention interventions in one cascade is challenging because various approaches exist that focus on either HIV infected or HIV uninfected people and use alternative products and approaches such as treatment as prevention, condoms, pre-exposure prophylaxis, voluntary medical male circumcision (VMMC), and reduction in sexual partner numbers.11 To explore the steps required for a particular approach to prevent HIV, each intervention can be separated out and described in terms of how it can succeed or fail in the epidemiological context in which it is used.

sure prophylaxis, voluntary medical male circumcision (VMMC), and reduction in sexual partner numbers.11 To explore the steps required for a particular approach to prevent HIV, each intervention can be separated out and described in terms of how it can succeed or fail in the epidemiological context in which it is used. This does not imply that each intervention should be used in isolation or that each provider should focus on a single intervention, rather it describes the steps that need attention for an intervention to be a worthwhile part of HIV prevention. Research in context Evidence before this study The HIV or cascade treatment cascade has become a tool to evaluate programme performance describing the steps required for those HIV infected to be successfully virally suppressed. In addition to its advocacy role the HIV treatment and care cascade can act as a guide to improving programme performance. The HIV treatment, prevention, and care cascade has started to be discussed as a comprehensive HIV cascade. However, to date such a cascade has yet to be fully populated with data. To explore the use of cascades in the HIV field PubMed was searched using the terms “HIV treatment cascade”, “HIV prevention cascade”, “HIV care cascade”, with last search on Feb 23, 2016. Added value of this study

The HIV or cascade treatment cascade has become a tool to evaluate programme performance describing the steps required for those HIV infected to be successfully virally suppressed. In addition to its advocacy role the HIV treatment and care cascade can act as a guide to improving programme performance. The HIV treatment, prevention, and care cascade has started to be discussed as a comprehensive HIV cascade. However, to date such a cascade has yet to be fully populated with data. To explore the use of cascades in the HIV field PubMed was searched using the terms “HIV treatment cascade”, “HIV prevention cascade”, “HIV care cascade”, with last search on Feb 23, 2016. Added value of this study This study describes how HIV prevention cascades might be considered in theory and then proceeds to use field data from rural Zimbabwe to generate HIV prevention cascades for voluntary male medical circumcision and for condom use and partner reduction after testing and counselling. The study illustrates the use of prevention cascades and provides a template for prospective studies to develop the practical use of the concept further in planning, implementing, and evaluating HIV prevention interventions. Implications of the available evidence

This study describes how HIV prevention cascades might be considered in theory and then proceeds to use field data from rural Zimbabwe to generate HIV prevention cascades for voluntary male medical circumcision and for condom use and partner reduction after testing and counselling. The study illustrates the use of prevention cascades and provides a template for prospective studies to develop the practical use of the concept further in planning, implementing, and evaluating HIV prevention interventions. Implications of the available evidence In this population in rural Zimbabwe HIV prevention interventions have improved but are having little impact. VMMC was limited by supply in 2009–11, but is increasingly limited by demand. HIV testing and counselling has little effect because men and particularly women do not change their behaviour. Perception of risk is a barrier to uptake on interventions particularly among men. For cascades to be useful, in addition to a theoretically justified description of HIV treatment and prevention, practical sources of data need to be available with which to populate the cascades. In this paper, we propose a new conceptual framework for HIV prevention cascades and illustrate its application with data collected in a large-scale general-population cohort in eastern Zimbabwe.

ription of HIV treatment and prevention, practical sources of data need to be available with which to populate the cascades. In this paper, we propose a new conceptual framework for HIV prevention cascades and illustrate its application with data collected in a large-scale general-population cohort in eastern Zimbabwe. Methods Conceptual framework To be effective, prevention products or protective behaviours have to be used by those who can benefit from them, and that use depends upon delivery, acceptance, and adoption by the appropriate population as well as adherence. The treatment cascade looks at the HIV infected population and the steps as they progress to viral suppression. An analogous client-centric prevention cascade could explore how those at risk of acquiring HIV (as a starting point) could avoid infection by perceiving that risk and acting on that risk by both adopting (ie, taking up) and adhering to an efficacious prevention approach (figure 1). An alternative would be an intervention-centric approach, in which programme staff identify a priority population, make an intervention available, see the uptake of that product, and see that it is used appropriately and that it is efficacious (figure 1). These prevention cascades progress from a population at risk of acquiring infection over a given period of time. The steps show the reduction in effect, and, as a corollary, the effect that improving each step can achieve. Efficacy, defined as the according-to-protocol estimate of effect of an intervention in an individual randomised controlled trial, is the last step in the cascade and is of limited importance if an intervention is not delivered, adopted, and adhered to. An intervention needs to be made available through supply of products, information, or procedures to the population at risk. Then the priority population has to take up the products, information, or procedures. This requirement for supply and then demand could be illustrated the other way round with demand being necessary and then limited by supply. Supply and demand are inextricably linked, with both required for effective prevention programmes.

the priority population has to take up the products, information, or procedures. This requirement for supply and then demand could be illustrated the other way round with demand being necessary and then limited by supply. Supply and demand are inextricably linked, with both required for effective prevention programmes. To describe the prevention cascades defining the population at risk of acquiring infection is an important first step. Theoretically, if we could accurately identify those that would acquire the infection without the intervention, then those lost at each step would acquire infection and the population left after all the steps of the cascade would be those staying uninfected. However, we can only identify those with a risk of infection at the start, so losses would be those remaining at risk and the end would represent those no longer at risk. Better focus on priority populations would imply a smaller population that could benefit from prevention to be included in programmes and a smaller number in the prevention cascade. The same proportional reductions in effect size through lack of supply, take-up, adherence, or efficacy apply to those who have a risk of acquiring infection or those who would actually acquire infection over a period. Thus, the starting for the prevention cascade could be the predicted incidence of infection without the intervention, and the probable number of infections occurring with the intervention would be determined by the number dropping out along the cascade. Risk behaviour is not constant so one could see the denominator at risk population changing over time. Acknowledging the not-at-risk population and allowing movement from not at risk to at risk may be possible (figure 1). This implies that individuals are removed entirely from potential risk of exposure; alternatively, their risk could have been reduced through successful use of a different prevention approach, which would mean that the population at risk of acquiring HIV at the start of one prevention cascade is a function of other prevention cascades (figure 1).

s are removed entirely from potential risk of exposure; alternatively, their risk could have been reduced through successful use of a different prevention approach, which would mean that the population at risk of acquiring HIV at the start of one prevention cascade is a function of other prevention cascades (figure 1). Data sources To explore the potential of cascades to track the implementation and impact of HIV prevention interventions, we used data from a general population open cohort HIV serosurvey done in Manicaland, eastern Zimbabwe, the Manicaland HIV/STD Prevention Project. Details of the cohort have been published previously.1, 2, 13, 14 In brief, a baseline survey and five rounds of follow-up have been done in a random sample of between 8000 and 15 000 adults recruited in a prospective census in 12 sites (eight sites in round six) in small towns, agricultural estates, roadside settlements, and rural villages spread across three districts in Zimbabwe's Manicaland province. Numbers varied between rounds because of demographic changes in the underlying population and changes in the sampling fraction following fluctuations in funding. The data collected in each round included information on HIV infection status, sexual behaviour (collected with a method to reduce social desirability bias),15 and availability, uptake, and adherence to HIV services such as VMMC, HIV testing and counselling (HTC), and so on. HIV prevalence in the study areas declined from 24·2% in 1998–2000 to 14·5% in 2012–13; HIV incidence was 0·80% in 2006–11 and 0·67% in 2009–13.

lected with a method to reduce social desirability bias),15 and availability, uptake, and adherence to HIV services such as VMMC, HIV testing and counselling (HTC), and so on. HIV prevalence in the study areas declined from 24·2% in 1998–2000 to 14·5% in 2012–13; HIV incidence was 0·80% in 2006–11 and 0·67% in 2009–13. Here, we used data from the eight sites covered in both of the two most recent rounds of follow-up of the Manicaland general population cohort study (2009–11 and 2012–13). Eligibility was restricted to adults aged 15–54 years who were resident in and had stayed in a household in the study areas for at least four nights in the last month. Descriptive statistics were calculated for uptake of each HIV prevention approach by the study population at each round. Intervention-centric HIV prevention cascades (based on local availability of services) and client-centric HIV prevention cascades (based on respondents' perceived personal risk of acquiring HIV infection) were constructed for all men and women, separately, who reported ever having had sex. Further intervention and client centred HIV prevention cascades were constructed for those reporting casual sexual partners in the 3 years before each survey to investigate the effect of HTC in reducing risk through sexual partner reduction and increased condom use. Our aim was to see how much impact HTC was having on risk behaviour and HIV incidence and test whether in its current form in this community it should be classed as an HIV prevention intervention.

re each survey to investigate the effect of HTC in reducing risk through sexual partner reduction and increased condom use. Our aim was to see how much impact HTC was having on risk behaviour and HIV incidence and test whether in its current form in this community it should be classed as an HIV prevention intervention. Availability, risk perception, uptake, and adherence for each HIV prevention approach within the study population were measured with existing questions rather than questions specifically designed to explore prevention cascades. Respondents were taken to have a service available locally when they knew that the service was being provided at a facility within 20 km of their homestead. Uptake of VMMC was measured with a variable derived from the question “Have you ever been circumcised yourself?” (coded “Yes–full”, “Yes–partial”, and “No”) and a question on the person performing the circumcision (doctor, nurse, traditional healer, tribal elder). Men with partial circumcisions or circumcisions done by traditional practitioners or tribal elders were treated as uncircumcised in the HIV prevention cascades for VMMC. Uptake of HTC was measured using the question “On how many different occasions have you had an HIV test in your lifetime?” and a question on collection of most recent test results. Variables for VMMC and HTC uptake in the last 3 years were also created with responses to questions on age at circumcision and number of HIV tests taken in the previous 3 years. For HTC, variables on adherence to sexual partner reduction and increased condom use were constructed from responses to the questions “After the HIV test, did you start having more or fewer sexual partners? / use condoms more or less than before?” We have estimates of the efficacy of VMMC of 60%,16 but for partner reduction and condom use we have only reports about whether or not behaviour changed so have to assume by how much. For illustrative purposes, we assume a 50% reduction in risk associated with reducing partner numbers, which is commensurate with the change in risk observed in earlier studies of risk reduction in this population13, 17 and 80% reduction associated with condom use reported in systematic reviews.18 For the client-centric HIV prevention cascades, perceived risk among individuals currently not infected was measured using the question “If you are not infected (currently), do you think you are in danger of getting infected now or in the future?”

% reduction associated with condom use reported in systematic reviews.18 For the client-centric HIV prevention cascades, perceived risk among individuals currently not infected was measured using the question “If you are not infected (currently), do you think you are in danger of getting infected now or in the future?” Data analysis To construct intervention-centric cascades, data were available on the number of people reporting that a service was available in their area, the number who had taken up the service. The efficacy of the intervention was estimated as described above. For circumcision adherence is not relevant, but for HIV testing and counselling adherence is represented by the proportion reporting reduced partner numbers or increased condom use. In the client-centric cascade the first reduction was based on the number reporting that they did not perceive that they were at risk of HIV and then the next step was based on whether they had taken up the intervention. To normalise the cascades and make them comparable the cascades were created for 1000 individuals with proportionate reductions at each step based on the numbers in the corresponding round of the Manicaland GPCS cohort. So for each 1000 population the number protected was first reduced by not having the service available or not perceiving risk, then reduced by the number not taking up the service, then by the number not adhering (zero in the case of circumcision), and then by the number where there would not be efficacy leaving the number who would be protected from those thousand individuals. To explore in more detail perceptions of risk we calculated age adjusted odds ratios comparing perception in risk between men and women and uptake in HTC as a function of perception of risk.

d then by the number where there would not be efficacy leaving the number who would be protected from those thousand individuals. To explore in more detail perceptions of risk we calculated age adjusted odds ratios comparing perception in risk between men and women and uptake in HTC as a function of perception of risk. For both types of cascade, estimates of potential HIV incidence in the absence of a given intervention Ip were calculated with the formula Ip = [N / (N – e × Nc)] × Io, where Io and N are the observed (actual) HIV incidence and the sample size for the specific populations of 15–49 year olds in the relevant round of the Manicaland GPCS, Nc is the number of individuals in the population who take up and adhere to the service, and e is the efficacy of the service, which can be measured in the according-to-protocol analysis in randomised controlled trials. This estimate is based on the reduction in incidence from what would have been the case without an intervention being the product of effective coverage and efficacy for that intervention. For example if observed incidence in a population is 1% and 400 of 1000 people have adhered to an intervention with 50% efficacy then the incidence without the intervention would have been 1·25%. Details of the HIV incidence estimates observed in the Manicaland GPCS are given in the appendix. Confidence intervals were calculated for binomial outcomes using Stata SE12.

is 1% and 400 of 1000 people have adhered to an intervention with 50% efficacy then the incidence without the intervention would have been 1·25%. Details of the HIV incidence estimates observed in the Manicaland GPCS are given in the appendix. Confidence intervals were calculated for binomial outcomes using Stata SE12. Role of the funding source GPG is an employee of the Bill & Melinda Gates Foundation, other than this the funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results In the study sites, 3·1% (95% CI 2·5–3·7) of the 3131 men interviewed and not infected with HIV in 2009–11 were circumcised: 1·8% (1·4–2·3) had received VMMC and 1·3% (0·9–1·7) had received traditional circumcision (table). VMMC coverage was somewhat higher (2·7%; 2·0–3·6) among the 1873 men who reported having passed their sexual debut. 3 years later, 6·9% (5·9–8·1) of the 2245 men interviewed in 2012–13 had been circumcised: 4·1% (3·4–5·1) for VMMC and 2·8% (2·2–3·6) for traditional circumcision.

ceived traditional circumcision (table). VMMC coverage was somewhat higher (2·7%; 2·0–3·6) among the 1873 men who reported having passed their sexual debut. 3 years later, 6·9% (5·9–8·1) of the 2245 men interviewed in 2012–13 had been circumcised: 4·1% (3·4–5·1) for VMMC and 2·8% (2·2–3·6) for traditional circumcision. HIV infections occurring because of unavailability of local VMMC services were reduced from 96·4% to 68·4% over a period of about 3 years between surveys (figure 2). However, the number of infections prevented increased only modestly from 16·3 per 1000 to 22·8 per 1000 (assuming an efficacy of 60%) because of limited uptake (12·0%) among men who knew that VMMC services were locally available. In the absence of the VMMC programme, HIV incidence would have been 0·69% per person-year of exposure compared to 0·67% rate observed between 2009 and 2013 in the GPCS, a reduction of just 0·02%.

assuming an efficacy of 60%) because of limited uptake (12·0%) among men who knew that VMMC services were locally available. In the absence of the VMMC programme, HIV incidence would have been 0·69% per person-year of exposure compared to 0·67% rate observed between 2009 and 2013 in the GPCS, a reduction of just 0·02%. The client-centric HIV prevention cascade for 2012–13 (figure 2) shows that 85·3% (1876 of 2199) of uninfected men in Manicaland did not perceive a personal risk of HIV infection. However, even in those who did believe themselves to be at risk, uptake of VMMC services was just 4·0% (13 of 323). In 2009–11, 27% of men (867 of 3154) and 57% of women (1284 of 2270) reported ever having had an HIV test, these proportions increased to 61% (2872 of 4703) and 80% (2625 of 3297) in 2012–13 (table). In 2009–11, lack of known availability and, particularly, lack of uptake when services were known to be available, were the principal obstacles to HTC having a major effect in preventing HIV infections (figure 3). Both of these obstacles were reduced substantially by 2012–13, at which point, the main obstacle was lack of partner reduction among men who took up HTC (table). In 2009–11, HTC, by bringing about reductions in numbers of sexual partners, prevented 49 per 1000 infections and reduced HIV incidence from 0·99% (in the absence of the programme) to 0·94% per person-year. In 2012–13, the HIV prevention effect of the HTC programme almost doubled to 90 per 1000 infections, and HIV incidence was reduced from 0·74% to 0·67%.

uctions in numbers of sexual partners, prevented 49 per 1000 infections and reduced HIV incidence from 0·99% (in the absence of the programme) to 0·94% per person-year. In 2012–13, the HIV prevention effect of the HTC programme almost doubled to 90 per 1000 infections, and HIV incidence was reduced from 0·74% to 0·67%. Much of the failure to prevent HIV infections resulting from lack of partner reduction stems from the presence of men who report abstinence or monogamy before taking up HTC in the population at risk. Restricting the analysis for 2009–11 to men with one or more casual sexual partners in the previous 3 years, reported availability and uptake of HTC were unchanged, but fewer men reported lack of partner reduction (145 per 1000 compared with 200 per 1000; appendix p 7). Given their higher rate of new infections, this yielded a somewhat greater absolute reduction in HIV incidence in this group from 1·58% to 1·45%.

ous 3 years, reported availability and uptake of HTC were unchanged, but fewer men reported lack of partner reduction (145 per 1000 compared with 200 per 1000; appendix p 7). Given their higher rate of new infections, this yielded a somewhat greater absolute reduction in HIV incidence in this group from 1·58% to 1·45%. For women, availability of HTC services and uptake when services were known were higher than for men, but availability and particularly uptake were still problematic in 2009–11 (figure 3). Lack of uptake remained a limiting factor in 2012–13 (figure 3) but the main obstacle in HTC contributing to reductions in HIV risk for women in both rounds was lack of partner reduction (table). Even more so than for men, this reflects large numbers of sexually experienced women who reported abstinence or monogamy prior to taking up HTC and conceal the greater prevention effect in women with casual partners (appendix p 8). However, overall, the effect of HTC in reducing numbers of sexual partners only reduced HIV incidence from 0·80% to 0·79% in the general population in the 3 year period to 2012–13.

orted abstinence or monogamy prior to taking up HTC and conceal the greater prevention effect in women with casual partners (appendix p 8). However, overall, the effect of HTC in reducing numbers of sexual partners only reduced HIV incidence from 0·80% to 0·79% in the general population in the 3 year period to 2012–13. The contributions of lack of availability and lack of uptake of HTC for condom use (figure 4) are the same as for reducing numbers of sexual partners. Fewer men report increasing condom use after HTC than report reduced sexual partners (appendix p 4; table 1), which offsets the smaller (assumed) lack of efficacy of condom use (20% vs 50%; figure 4). For example, in 2012–13, the effect of HTC in reducing HIV risk in sexually experienced men by increasing condom use was 61 per 1000 (appendix p 13) compared to 90 per 1000 (appendix p 9) achieved by reducing sexual partners. For women, lack of condom use after HTC, once again, was the main factor contributing to the very small effect of HTC in preventing HIV infections in those otherwise at risk (figure 4). However, for women, increased condom use after HTC had a slightly larger effect in prevention of HIV infections (24 per 1000 in round six) than in reductions in numbers of sexual partners (13 per 1000).

ctor contributing to the very small effect of HTC in preventing HIV infections in those otherwise at risk (figure 4). However, for women, increased condom use after HTC had a slightly larger effect in prevention of HIV infections (24 per 1000 in round six) than in reductions in numbers of sexual partners (13 per 1000). In the client-centric HIV prevention cascades for 2012–13, 15·3%, of sexually active men (223 of 1457) and 36·8% (1000 of 2717) women perceived a risk of HIV infection (p<0·0001; figure 3). 124 (55·6%) of 223 men who felt at risk had taken up HTC compared with 805 (65·2%) of 1234 men who did not (age-adjusted odds ratio 0·66; p=0·005). By contrast, 846 (84·6%) of 1000 women who felt at risk used HTC compared with 1359 (79·1%) of 1717 women who did not feel at risk (1·34; p=0·007). Among those who felt at risk and had taken up HTC, 43·7% of men (49 of 112) and 2·6% of women (21 of 814), had subsequently reduced their number of sexual partners; and 23·2% (26 of 112) and 4·4% (36 of 815) reported an increase in condom use (figure 3).

9 (79·1%) of 1717 women who did not feel at risk (1·34; p=0·007). Among those who felt at risk and had taken up HTC, 43·7% of men (49 of 112) and 2·6% of women (21 of 814), had subsequently reduced their number of sexual partners; and 23·2% (26 of 112) and 4·4% (36 of 815) reported an increase in condom use (figure 3). Discussion In this population before 2011, circumcision was not widely available. By 2013, availability had improved somewhat but uptake was still low. In addition to low availability, men at risk did not perceive that risk; even among those who acknowledged a risk, uptake of circumcision was low. Over the past 2 years, circumcision has become more readily available in Zimbabwe and barriers to uptake are being addressed.19 However, in the rural populations studied here, acknowledgment and perception of risk of HIV is a substantial challenge, which has many potential causes.20 Analysis of HIV prevention cascades can help identify the barriers to effective HIV prevention in populations, which is illustrated by applying different formulations of prevention cascades to data from a population at high risk of HIV acquisition in rural Zimbabwe. This application of the concept for prevention interventions used by susceptible individuals highlights the lack of prevention service delivery, low perceptions of risk, and poor uptake of HIV prevention in rural Zimbabwe. VMMC, condom use, and reduced partner numbers after HTC would be efficacious ways of reducing HIV risk, but they are not widely used in Manicaland.

erventions used by susceptible individuals highlights the lack of prevention service delivery, low perceptions of risk, and poor uptake of HIV prevention in rural Zimbabwe. VMMC, condom use, and reduced partner numbers after HTC would be efficacious ways of reducing HIV risk, but they are not widely used in Manicaland. HIV testing and counselling is often described as a gateway to both HIV treatment and HIV prevention and included in HIV prevention budgets, but without associated changes in behaviour it will not reduce HIV risk. Our analysis shows that, in rural eastern Zimbabwe, HIV testing has little effect as HIV prevention among susceptible people. This is in line with some previous studies of the impact of HIV testing and counselling, which does not change the risk behaviour of HIV-negative people.21, 22 Over time, the supply and uptake of HTC has improved in both men and women, but the lack of change in partner numbers and condom use associated with HTC undermines its prevention effect. This is particularly notable for women. In such circumstances, inclusion of HTC as an HIV prevention intervention may not be justifiable.

time, the supply and uptake of HTC has improved in both men and women, but the lack of change in partner numbers and condom use associated with HTC undermines its prevention effect. This is particularly notable for women. In such circumstances, inclusion of HTC as an HIV prevention intervention may not be justifiable. In applying prevention cascades to data, limitations arose from relying on self-reported data on availability including distances to services, uptake of services, and sexual behaviour. Cross-sectional approaches to measuring steps in HIV prevention cascades, as used here, can also be problematic. For example, cross-sectional measurements of perceived risk can be unreliable since perceived risk can lead to VMMC uptake which, in turn, reduces perceived risk, generating higher VMMC in those with no perceived risk. Furthermore the perception of risk, actual risk, duration of risk, and duration of protection provided by an intervention need to be considered when thinking about how to generate a cascade where perception of risk is included. More consideration of such issues will be required when prospectively collecting data to populate prevention cascades.

lanners looking to organise effective HIV prevention this population-level view can come from surveys such as the current HIV Impact Assessments.23 However, implementers can still consider in their monitoring and assessment plans the steps of the cascade that they need to improve for their interventions to have effect. There is no one correct approach to studying HIV prevention cascades, although, for comparability, some agreement on approach would be worthwhile. However, to our knowledge this is the first time that HIV prevention cascades have included data. Other approaches need to be similarly tested with data to assess their utility. In our opinion, it is only by breaking out separate prevention interventions that the prevention cascade can be practically applied, as done here. There are five major limitations to the cascades. First, they only explore reductions in risk of acquiring HIV, not in transmitting HIV, so do not illustrate the importance and role of reduced transmissibility among people with HIV, the impact of which can be assessed with the treatment cascade. Second, they quantify the direct protection provided to susceptible individuals but not the onward protection provided when that individual does not acquire infection; thus they do not quantify the full effect of an intervention but only a minimum level of protection. Third, they do not show the combined protection provided by two or more protection measures. Fourth, they are based on understanding who is at risk of acquiring HIV infection, which can only be inferred from epidemiological data. Generating useful HIV prevention cascades requires a good understanding of patterns of risk of HIV within populations. Fifth, the casades explore the effect of the interventions only for those at risk at the time of the study not for those who could enter the at-risk population in the future and the potential of interventions to change their future behaviours. Potentially, combining cascades in transmission dynamic models could extend our understanding of the impact of our interventions.

tions only for those at risk at the time of the study not for those who could enter the at-risk population in the future and the potential of interventions to change their future behaviours. Potentially, combining cascades in transmission dynamic models could extend our understanding of the impact of our interventions. The cascades shown here average risk over the adult population and look at the effects of prevention interventions in the time period when the data are collected. Additional insights could be gained by disaggregating the analysis of risk by age (see appendices) or cohort and looking at the age and cohort specific effects of interventions. In considering the error in measuring cascades, confidence intervals for each step should be based on the binomial distribution as people either do or do not continue to the next step in the cascade. The cascade is a measure of the potential that interventions can have to reduce HIV acquisition and can act as priors in a Bayesian model of the effects of interventions on a population.

ls for each step should be based on the binomial distribution as people either do or do not continue to the next step in the cascade. The cascade is a measure of the potential that interventions can have to reduce HIV acquisition and can act as priors in a Bayesian model of the effects of interventions on a population. The importance of advocacy and monitoring of an HIV prevention cascade has been recognised for some time, but previous attempts have struggled to generate a single prevention cascade covering HIV positive and negative people with multiple prevention approaches. Our concept has been to break down HIV prevention according to both population and prevention approach, so that the important steps can be highlighted and where there are gaps they can be recognised. We can now start to develop an understanding of the more specific problems and what approaches might enable better supply, demand, adherence, or efficacy. This is not to say that delivery of HIV prevention interventions should be siloed as multiple approaches may be necessary and should be combined where this is most efficient. However, it does emphasis the requirements for each approach to have an impact. To illustrate this we have been able to generate HIV prevention cascades for populations in rural Zimbabwe, which illustrate just how far we are from effective HIV prevention targeted at susceptible populations and the scale of the challenges to overcome. Supplementary Material Supplementary appendix

The importance of advocacy and monitoring of an HIV prevention cascade has been recognised for some time, but previous attempts have struggled to generate a single prevention cascade covering HIV positive and negative people with multiple prevention approaches. Our concept has been to break down HIV prevention according to both population and prevention approach, so that the important steps can be highlighted and where there are gaps they can be recognised. We can now start to develop an understanding of the more specific problems and what approaches might enable better supply, demand, adherence, or efficacy. This is not to say that delivery of HIV prevention interventions should be siloed as multiple approaches may be necessary and should be combined where this is most efficient. However, it does emphasis the requirements for each approach to have an impact. To illustrate this we have been able to generate HIV prevention cascades for populations in rural Zimbabwe, which illustrate just how far we are from effective HIV prevention targeted at susceptible populations and the scale of the challenges to overcome. Supplementary Material Supplementary appendix Acknowledgments Thanks to Jen Kates, Nancy Padian, Gina Dallabetta, Paulin Basinga, Mike Isbell, and Judith Auerbach for initial discussions and about prevention cascades. Contributions The ideas behind the analysis were developed by all the authors. The data analysis was carried out by SG with assistance from AT and RR. The paper was drafted by GPG and SG.

Acknowledgments Thanks to Jen Kates, Nancy Padian, Gina Dallabetta, Paulin Basinga, Mike Isbell, and Judith Auerbach for initial discussions and about prevention cascades. Contributions The ideas behind the analysis were developed by all the authors. The data analysis was carried out by SG with assistance from AT and RR. The paper was drafted by GPG and SG. Declaration of interests GPG is an employee of the Bill & Melinda Gates Foundation. TBH, JH, MW, SG report grants from Bill & Melinda Gates Foundation. TBH reports grants from World Bank, UNAIDS, Rush Foundation, Wellcome Trust, and personal fees form Bill & Melinda Gates Foundation, New York University, WHO Global Fund for AIDS Tuberculosis and Malaria outside the submitted work. SG reports grants from Wellcome Trust, during the conduct of the study; grants from UNAIDS, grants from UNFPA, grants from European Research Council, grants from RIATT, and grants from ESRC-DFID, outside the submitted work; and owns some ordinary shares in Astra Zeneca and GlaxoSmithKline. Figure 1 Generic conceptual HIV prevention cascades applied to the population who would otherwise acquire HIV or are at risk of acquiring HIV

Declaration of interests GPG is an employee of the Bill & Melinda Gates Foundation. TBH, JH, MW, SG report grants from Bill & Melinda Gates Foundation. TBH reports grants from World Bank, UNAIDS, Rush Foundation, Wellcome Trust, and personal fees form Bill & Melinda Gates Foundation, New York University, WHO Global Fund for AIDS Tuberculosis and Malaria outside the submitted work. SG reports grants from Wellcome Trust, during the conduct of the study; grants from UNAIDS, grants from UNFPA, grants from European Research Council, grants from RIATT, and grants from ESRC-DFID, outside the submitted work; and owns some ordinary shares in Astra Zeneca and GlaxoSmithKline. Figure 1 Generic conceptual HIV prevention cascades applied to the population who would otherwise acquire HIV or are at risk of acquiring HIV (A) Steps to prevention taking the clients' perspective and their perception of risk into account. (B) Steps to prevention taking the intervention perspective and whether there is supply of the product available. (C) Including the possibility that those at risk can move into a not at risk population and vice versa. (D) Combining two prevention cascades where the population at risk for the start of the second intervention is the population that remain at risk after the application of the first intervention. Figure 2 HIV prevention cascades for voluntary medical male circumcision in sexually experienced uninfected men aged 15–54 years in Manicaland, Zimbabwe

(A) Steps to prevention taking the clients' perspective and their perception of risk into account. (B) Steps to prevention taking the intervention perspective and whether there is supply of the product available. (C) Including the possibility that those at risk can move into a not at risk population and vice versa. (D) Combining two prevention cascades where the population at risk for the start of the second intervention is the population that remain at risk after the application of the first intervention. Figure 2 HIV prevention cascades for voluntary medical male circumcision in sexually experienced uninfected men aged 15–54 years in Manicaland, Zimbabwe (A) Cascade based on service availability for 1888 individuals in 2009–11. (B) Cascade based on service availability for 1476 individuals in 2012–13. (C) Cascade based on participant's perception of personal risk of acquiring HIV infection for 1476 individuals in 2012–13. Figure 3 HIV prevention cascades for HIV testing and counselling and sexual partner reduction in sexually experienced uninfected adults aged 15–54 years in Manicaland, Zimbabwe

(A) Cascade based on service availability for 1888 individuals in 2009–11. (B) Cascade based on service availability for 1476 individuals in 2012–13. (C) Cascade based on participant's perception of personal risk of acquiring HIV infection for 1476 individuals in 2012–13. Figure 3 HIV prevention cascades for HIV testing and counselling and sexual partner reduction in sexually experienced uninfected adults aged 15–54 years in Manicaland, Zimbabwe (A) Cascade for 1888 men based on service availability in 2009–11. (B) Cascade for 1468 men based on service availability in 2012–13. (C) Cascade for 3793 women based on service availability in 2009–11. (D) Cascade for 2743 women based on service availability in 2012–13. (E) Cascade for 1468 men based on participant's perception of personal risk of acquiring HIV infection in 2012–13. (F) Cascade for 2743 women based on participant's perception of personal risk of acquiring HIV infection in 2012–13. Figure 4 HIV prevention cascades for HIV testing and counselling and increased condom use in sexually experienced uninfected adults aged 15–54 years in Manicaland, Zimbabwe

(A) Cascade for 1888 men based on service availability in 2009–11. (B) Cascade for 1468 men based on service availability in 2012–13. (C) Cascade for 3793 women based on service availability in 2009–11. (D) Cascade for 2743 women based on service availability in 2012–13. (E) Cascade for 1468 men based on participant's perception of personal risk of acquiring HIV infection in 2012–13. (F) Cascade for 2743 women based on participant's perception of personal risk of acquiring HIV infection in 2012–13. Figure 4 HIV prevention cascades for HIV testing and counselling and increased condom use in sexually experienced uninfected adults aged 15–54 years in Manicaland, Zimbabwe (A) Cascade for 1888 men based on service availability in 2009–11. (B) Cascade for 1468 men based on service availability in 2012–13. (C) Cascade for 3793 women based on service availability in 2009–11. (D) Cascade for 2743 women based on service availability in 2012–13. (E) Cascade for 1468 men based on participant's perception of personal risk of acquiring HIV infection in 2012–13. (F) Cascade for 2743 women based on participant's perception of personal risk of acquiring HIV infection in 2012–13. Table Uptake of voluntary medical male circumcision and HIV testing and counselling in HIV-uninfected adults aged 15–54 years in Manicaland, Zimbabwe

(A) Cascade for 1888 men based on service availability in 2009–11. (B) Cascade for 1468 men based on service availability in 2012–13. (C) Cascade for 3793 women based on service availability in 2009–11. (D) Cascade for 2743 women based on service availability in 2012–13. (E) Cascade for 1468 men based on participant's perception of personal risk of acquiring HIV infection in 2012–13. (F) Cascade for 2743 women based on participant's perception of personal risk of acquiring HIV infection in 2012–13. Table Uptake of voluntary medical male circumcision and HIV testing and counselling in HIV-uninfected adults aged 15–54 years in Manicaland, Zimbabwe Men Women 2009–11 2012–13 2009–11 2012–13 n/N % 95% CI n/N % 95% CI n/N % 95% CI n/N % 95% CI Male circumcision All men Traditional 40/3131 1·28 (0·91–1·74) 63/2245 2·81 (2·16–3·58) .. .. .. .. .. .. VMMC 56/3131 1·79 (1·35–2·32) 93/2245 4·14 (3·36–5·05) .. .. .. .. .. .. Sexually active men VMMC Lifetime 51/1873 2·72 (2·03–3·56) 56/1476 3·79 (2·88–4·90) .. .. .. .. .. .. Past 3 years 2/1873 0·11 (0·01–0·39) 24/1476 1·63 (1·04–2·41) .. .. .. .. .. .. HIV testing and counselling All respondents Lifetime uptake 867/3154 27·5 (25·9–29·1) 1284/2270 56·6 (54·5–58·6) 2872/4703 61·1 (59·7–62·5) 2625/3297 79·6 (78·2–81·0) Past 3 years 0 2411/3154 76·4 .. 1114/2270 49·1 .. 2177/4703 46·3 .. 921/3297 27·9 .. 1 505/3154 16·0 .. 594/2270 26·2 .. 1359/4703 28·9 .. 858/3297 26·0 .. ≥2 238/3154 7·6 .. 562/2270 24·8 .. 1167/4703 24·8 .. 1518/3297 46·0 .. Sexually active respondents Lifetime 742/1888 39·3 (37·1–41·6) 1046/1474 71·0 (68·6–73·3) 2742/3793 72·3 (70·8–73·7) 2475/2773 89·3 (88·0–90·4) Past 3 years 632/1888 33·5 (31·4–35·7) 941/1474 63·8 (61·3–66·3) 2404/3793 63·4 (61·8–64·9) 2243/2773 80·9 (79·4–82·3) Behaviour change after HTC* Number of sexual partners More 6/563 1·1 .. 15/875 1·7 .. 6/2198 0·3 .. 9/2141 0·4 .. Same 372/563 6·1 .. 594/875 67·9 .. 2080/2198 94·6 .. 2057/2141 96·1 .. Fewer 185/563 32·9 .. 266/875 30·4 .. 112/2198 5·1 .. 75/2141 3·5 .. Condom use More 79/563 14·0 .. 112/874 12·8 .. 90/2199 4·1 .. 86/2143 4·0 .. Same 420/563 74·6 .. 637/874 72·9 .. 2011/2199 91·5 .. 1965/2143 91·7 .. Fewer 64/563 11·4 .. 125/874 14·3 .. 98/2199 4·5 .. 92/2143 4·3 .. Data given for those who provided responses on the questionnaire, so differ from total sample size owing to missing data.

/563 14·0 .. 112/874 12·8 .. 90/2199 4·1 .. 86/2143 4·0 .. Same 420/563 74·6 .. 637/874 72·9 .. 2011/2199 91·5 .. 1965/2143 91·7 .. Fewer 64/563 11·4 .. 125/874 14·3 .. 98/2199 4·5 .. 92/2143 4·3 .. Data given for those who provided responses on the questionnaire, so differ from total sample size owing to missing data. * HIV testing and counselling in the last 3 years. VMMC=voluntary medical male circumcision. HTC=HIV testing and counselling.

Introduction The provision of antiretroviral therapy (ART) has substantially reduced HIV mortality.1 With timely diagnosis, treatment initiation, and good adherence, life-expectancy for people with HIV can approach that of uninfected people.2 However, life-years are still lost to AIDS, and in sub-Saharan Africa, hundreds of thousands of AIDS-related deaths occur each year.3

T) has substantially reduced HIV mortality.1 With timely diagnosis, treatment initiation, and good adherence, life-expectancy for people with HIV can approach that of uninfected people.2 However, life-years are still lost to AIDS, and in sub-Saharan Africa, hundreds of thousands of AIDS-related deaths occur each year.3 Reasons for continued health losses to HIV when ART is widely available are poorly understood. The care cascade describes the series of engagements with the health system through which people with HIV must pass to benefit fully from ART, beginning with HIV testing, and ending with regular monitoring of patients in a state of sustained viral suppression. In 2011, Rosen and Fox4 showed that fewer patients reach each successive stage of HIV care in sub-Saharan Africa, and representations of the care cascade in different settings followed.5 However, many attempts to quantify the care cascade have been limited by not being able to follow a cohort of patients through all stages of care. Data are typically available only for people with HIV who present to clinics and therefore exclude those who never engage in care and who are likely to have the greatest health losses.6, 7 Differences in care-seeking behaviour between patients who actively seek care (through clinics) and those who are actively sought (through outreach programmes) lead to fundamental uncertainties in the operation of the care cascade, in particular, the extent to which health-care-seeking behaviour enables patients to present for care and initiate treatment on becoming ill, bypassing the typical stages of pre-ART care, and monitoring until reaching ART eligibility (in a process that we have termed previously as reaching ART via the side-door).6

ade, in particular, the extent to which health-care-seeking behaviour enables patients to present for care and initiate treatment on becoming ill, bypassing the typical stages of pre-ART care, and monitoring until reaching ART eligibility (in a process that we have termed previously as reaching ART via the side-door).6 Research in context Evidence before this study

ade, in particular, the extent to which health-care-seeking behaviour enables patients to present for care and initiate treatment on becoming ill, bypassing the typical stages of pre-ART care, and monitoring until reaching ART eligibility (in a process that we have termed previously as reaching ART via the side-door).6 Research in context Evidence before this study With global aspirations to eliminate AIDS as a cause of population morbidity and death, attention has turned to identification and reduction of weaknesses in HIV care programmes. As the breadth of interventions targeting aspects of the cascade of care expands, care programmes must identify which strategies will bring about the largest impact for the lowest cost. Mathematical models are well placed to answer these questions. We searched PubMed for HIV modelling studies published between Jan 1, 2000, and Aug 2, 2016, with the terms “HIV” AND (“ART” OR “antiretroviral therapy”) AND (“cascade” OR “continuum”) AND (“modelling” OR “modeling” OR “model”) without any language restrictions. The search yielded 65 abstracts and 12 met our inclusion criteria of being mathematical modelling studies. Previous studies have relied on aggregated routine data from disparate analyses and populations to estimate the cascade, and consequently most have used deterministic compartmental models to do investigations on individual interventions. Only four studies have assessed the cost-effectiveness of interventions targeting improvements in care. Among these, analysis of the effect of early ART initiation in a study in India showed such an intervention would be highly cost-effective, but the researchers stressed the importance of the modulatory effects of retention in care. Findings from analyses of the cascade in the USA corroborate these results. The only study from our search to use an individual-based model described HIV transmission and cascade progression in South Africa. Results indicated that among the interventions implemented, returning and reinitiating patients onto ART is highly cost-effective along with improving retention in care.

ate these results. The only study from our search to use an individual-based model described HIV transmission and cascade progression in South Africa. Results indicated that among the interventions implemented, returning and reinitiating patients onto ART is highly cost-effective along with improving retention in care. Added value of this study Our model shows that single interventions have a modest effect on improving present care programmes. A combination of interventions concurrently strengthening various aspects of care is the most cost-effective strategy to improve outcomes for patients. This strategy is likely to be more cost-effective and generate greater impacts than immediate ART that is not accompanied by improvements to the care cascade. Our model represents HIV care and treatment in western Kenya, but our overall conclusions will have the same broad relevance to the many other settings with large generalised epidemics and established ART programmes. Implications of all the available evidence Many health systems do not capture a large proportion of HIV-related deaths occurring outside of the clinic. There is considerable scope for care programmes to improve care throughout the cascade. Detailed longitudinal data about ART health-care programmes could improve future projections, characterise country-specific gaps in care, and identify cost-effective strategies to achieve future treatment targets.

ide of the clinic. There is considerable scope for care programmes to improve care throughout the cascade. Detailed longitudinal data about ART health-care programmes could improve future projections, characterise country-specific gaps in care, and identify cost-effective strategies to achieve future treatment targets. Many strategies to improve the care cascade have been proposed and tested.8, 9, 10 These variously aim to improve testing, linkage to care, retention in care before starting treatment, retention on ART, and rates of viral suppression. However, evidence about the effectiveness of individual interventions is only partly informative about the best strategy to improve HIV care, and most studies have not been able to measure the eventual population-health benefits resulting from improved provision of care. The expansion of ART eligibility to potentially all people with HIV as recommended in the latest WHO guidelines (ie, immediate ART initiation)11 and increased outreach to populations for testing (eg, through a universal test-and-treat strategy) are proposals for transformational expansions to treatment.12, 13 In hypothetical idealised programmes, the persuasive argument in favour of such approaches is that the additional initial costs would later become offset by savings resulting from reductions in HIV transmission and need for ART.12 However, whether these savings will be realised with more realistic implementation assumptions is questionable.14

d programmes, the persuasive argument in favour of such approaches is that the additional initial costs would later become offset by savings resulting from reductions in HIV transmission and need for ART.12 However, whether these savings will be realised with more realistic implementation assumptions is questionable.14 Definitions for optimal strategies to improve HIV care most efficiently in health systems require insight into the sources of HIV mortality and morbidity for patients at each stage of care and a comparison of the population-level health impact of a range of candidate interventions that act at different stages of care. We created a mathematical model of the HIV epidemic and care cascade in Kenya. The model is parameterised with data from an HIV care programme in western Kenya supported by the Academic Model Providing Access To Healthcare (AMPATH) collaboration, which uniquely includes data about people before testing and disengagement from care, through an integrated household-based testing intervention. We used the model to quantify the previous care experience of those dying from HIV in a setting with a mature (>8 years since established) ART programme, and to simulate the cost and effect of HIV care interventions, in isolation and in combinations. From this information, we aimed to generate recommendations about optimal strategies for HIV care programmes to maximise health costs effectively.

m HIV in a setting with a mature (>8 years since established) ART programme, and to simulate the cost and effect of HIV care interventions, in isolation and in combinations. From this information, we aimed to generate recommendations about optimal strategies for HIV care programmes to maximise health costs effectively. Methods Data sources We constructed an individual-based microsimulation model representing the HIV epidemic in Kenya and capturing the care experience of individuals as they progress through an ART programme. The model represents births, ageing, deaths, and HIV transmission in the Kenyan population (appendix pp 5–8). After infection, disease progression is modelled by an individual progressing to a lower CD4 count category (>500 cells per μL, 350–500 cells per μL, 200–350 cells per μL, and <200 cells per μL) and of greater disease severity (WHO defined stages 1, 2, 3, and 4; appendix p 12). Both CD4 cell count and disease state affect the risk of mortality. Additionally, WHO disease stage affects the propensity to seek care (appendix pp 12–18). When treatment is initiated, patients can transition to a higher CD4 cell count and a lower (healthier) WHO clinical stage (appendix p 12). Model design The model describes the pathway through care for each individual infected with HIV (figure 1, appendix p 19). The baseline scenario represents no additional intervention.

The model represents births, ageing, deaths, and HIV transmission in the Kenyan population (appendix pp 5–8). After infection, disease progression is modelled by an individual progressing to a lower CD4 count category (>500 cells per μL, 350–500 cells per μL, 200–350 cells per μL, and <200 cells per μL) and of greater disease severity (WHO defined stages 1, 2, 3, and 4; appendix p 12). Both CD4 cell count and disease state affect the risk of mortality. Additionally, WHO disease stage affects the propensity to seek care (appendix pp 12–18). When treatment is initiated, patients can transition to a higher CD4 cell count and a lower (healthier) WHO clinical stage (appendix p 12). Model design The model describes the pathway through care for each individual infected with HIV (figure 1, appendix p 19). The baseline scenario represents no additional intervention. The HIV care and treatment components of the model were parameterised with observed patients' data from Bunyala, western Kenya (appendix pp 19–44). HIV testing, HIV care, and treatment services were established in 2006 in district hospital and health centres, supported by the AMPATH collaboration. AMPATH is a partnership established in 2001 between Moi Teaching and Referral Hospital, Moi University School of Medicine, and a consortium of universities from North America in response to the HIV epidemic in Kenya (appendix pp 19–20). All patients' visits from 2004 have been recorded electronically in the AMPATH Medical Record System, furnishing information about retention and outcomes. For patients to seek care in non-AMPATH clinics they must leave the AMPATH catchment area (appendix pp 19–20). Patients lost from care are traced and actively followed up to ascertain their outcome, either through direct contact or discussions with family members if patients cannot be found. In the area of Bunyala, there have been two rounds of home-based counselling and testing campaigns since 2010, the first of which achieved more than 85% coverage of the community.15 These early rounds of home-based counselling and testing involved passive referral of patients infected with HIV, with no active follow-up to facilitate linkage to health care (appendix p 24). Present AMPATH home-based counselling and testing campaigns now include active follow-up, and we expect linkage rates to be substantially higher.

of home-based counselling and testing involved passive referral of patients infected with HIV, with no active follow-up to facilitate linkage to health care (appendix p 24). Present AMPATH home-based counselling and testing campaigns now include active follow-up, and we expect linkage rates to be substantially higher. We characterised the main parameters of the care cascade through the analysis of the linked individual clinical and home-based counselling and testing records after extensive removal of incorrectly entered data and removal of duplication (appendix p 21).15 For model parameters that did not correspond to quantities that can be directly observed, values were inferred through fitting the model output to observed data (table 1, appendix pp 22, 23, 25–28). 62% of people in both AMPATH and model data were diagnosed in mid-2010: 41% through voluntary testing and counselling in AMPATH and 42% in the model; 21% provide initiated counselling and testing in each. Of those diagnosed, AMPATH data showed 34% were on ART in 2010 and the model estimated 33%; by 2014 an estimated 91% (AMPATH) and 82% (model) of diagnosed people were on ART.

in mid-2010: 41% through voluntary testing and counselling in AMPATH and 42% in the model; 21% provide initiated counselling and testing in each. Of those diagnosed, AMPATH data showed 34% were on ART in 2010 and the model estimated 33%; by 2014 an estimated 91% (AMPATH) and 82% (model) of diagnosed people were on ART. Calibration yielded several sets of parameters for the model of the care cascade, which are variously in better agreement with different indicators (appendix pp 25–44). The ART programme costs at baseline were estimated from the perspective of a health-care provider. Unit cost estimates were based on the CHAI MATCH study17 of ART facilities, and comprised the cost of ART care, the cost of pre-ART clinic visits, and the cost of CD4 laboratory-based tests (appendix pp 45, 46). Sensitivity analysis was done with consideration to the variations in the unit cost of ART programme components (appendix pp 53–55). Analysis of the care cascade Interventions on the care cascade can be divided into those that aim to increase testing, linkage, and retention in pre-ART health care, or retention and suppression for patients on ART. We reviewed the medical literature to identify realistic and empirically based assumptions for the efficacy and cost of representative interventions in each of these categories (table 2, appendix p 48).

to increase testing, linkage, and retention in pre-ART health care, or retention and suppression for patients on ART. We reviewed the medical literature to identify realistic and empirically based assumptions for the efficacy and cost of representative interventions in each of these categories (table 2, appendix p 48). To assess the impact of individual interventions, each was simulated for the duration 2010–30, and the effect on patients' outcomes compared with the baseline scenario in the absence of any interventions. The effect of the programme was quantified as disability-adjusted life-years (DALYs) averted (appendix p 47), additional cost of care (appendix pp 45, 46), and HIV-related deaths averted, compared with a baseline programme without any intervention (taken to be similar in structure to AMPATH before the launch of household-based testing). Because of the stochastic nature of the model, we present results as the mean of ten repeat simulations. An optimal combination of individual cascade interventions (excluding universal test and treat, which is a composite of home-based counselling and testing and immediate ART) was identified by simulation of all possible combinations and selection of those that provided the greatest increase in health for a range of budget levels. We imposed the additional constraint that once an intervention had been included in the combination at one budget level, it cannot be removed at higher budget levels.

identified by simulation of all possible combinations and selection of those that provided the greatest increase in health for a range of budget levels. We imposed the additional constraint that once an intervention had been included in the combination at one budget level, it cannot be removed at higher budget levels. To assess the cost-effectiveness of interventions, the cost per DALY averted was compared with the gross domestic product (GDP) per capita for Kenya in 2013 (US$1242).24 We assume that an intervention is likely to be cost-effective if the cost per DALY is less than 50% GDP per capita.25, 26 Costs and DALYs were both discounted at 6% per annum from 2010. Role of the funding source The funders had no role in model construction, data collection, data analysis, data interpretation, or writing of the report. The corresponding author worked with co-authors to analyse the data in the study and had final responsibility for the decision to submit for publication.

To assess the cost-effectiveness of interventions, the cost per DALY averted was compared with the gross domestic product (GDP) per capita for Kenya in 2013 (US$1242).24 We assume that an intervention is likely to be cost-effective if the cost per DALY is less than 50% GDP per capita.25, 26 Costs and DALYs were both discounted at 6% per annum from 2010. Role of the funding source The funders had no role in model construction, data collection, data analysis, data interpretation, or writing of the report. The corresponding author worked with co-authors to analyse the data in the study and had final responsibility for the decision to submit for publication. Results We projected the baseline model in the absence of any interventions between 2010 and 2030 and analysed the status of care of those dying from HIV-related causes in two timeframes: 2010–2015 and 2025–2030. Between 2010 and 2030, most people will have initiated treatment (61%), but many will never have been diagnosed (25%) or will have been diagnosed but never started ART (14%). Among all AIDS-related deaths between 2010 and 2015, most occurred in individuals who had initiated treatment (figure 2). Among these, most died because they had initiated treatment late (CD4 count <200 cells per μL). The largest proportion of deaths was in people who were never diagnosed, with the remainder those who were diagnosed but did not initiate ART. By contrast, between 2025 and 2030, the distribution of mortality shifted to the latter stages of care. Most deaths still occurred in individuals who had initiated treatment, with the largest single proportion being in patients who had disengaged from ART care.

mainder those who were diagnosed but did not initiate ART. By contrast, between 2025 and 2030, the distribution of mortality shifted to the latter stages of care. Most deaths still occurred in individuals who had initiated treatment, with the largest single proportion being in patients who had disengaged from ART care. We applied each of the 12 interventions in isolation and calculated the DALYs averted and additional costs between 2010 and 2030 compared with the baseline scenario (figure 3 and table 3). Costs and effect are generally closely related, with low-cost interventions having a low impact. The effects of most single interventions affecting engagement in pre-ART health care cluster together with relatively low impact and low cost (figure 3).

and 2030 compared with the baseline scenario (figure 3 and table 3). Costs and effect are generally closely related, with low-cost interventions having a low impact. The effects of most single interventions affecting engagement in pre-ART health care cluster together with relatively low impact and low cost (figure 3). One exception is home-based counselling and testing (with passive referral), which has a high cost per DALY averted (figure 3). We assumed that only 30% of people diagnosed for the first time at home-based counselling and testing will link to care without further intervention. This assumption is based on observations in the earliest home-based counselling and testing campaign at AMPATH, which used passive referral to care.15 AMPATH's present home-based counselling and testing programme incorporates active referral and, based on preliminary data, yields considerably higher linkage rates (a scenario with 90% linkage, consistent with WHO targets, is included for comparison in the appendix p 56). Home-based counselling and testing (with point-of-care CD4) averts more DALYs than any other single intervention because of the reduced time to confirm eligibility and increased probability of linkage. However, the cost per DALY averted of $1617 is 130% GDP per capita of Kenya (table 3).

mparison in the appendix p 56). Home-based counselling and testing (with point-of-care CD4) averts more DALYs than any other single intervention because of the reduced time to confirm eligibility and increased probability of linkage. However, the cost per DALY averted of $1617 is 130% GDP per capita of Kenya (table 3). By contrast, the on-ART outreach intervention, which seeks and returns 40% of people disengaged from health care after ART initiation, has the third lowest cost per DALY averted, and a larger impact than almost all interventions (table 3). These data are consistent with the large proportion of deaths and high mortality in people disengaged from care (figure 2). The two interventions that represent large changes to ART eligibility, immediate ART and universal test and treat, differ substantially in their impact and cost. Immediate ART provides treatment to those who present for care, and is effective because it eliminates the potential for patients to be lost from care before they are confirmed to be eligible for ART. The cost per DALY averted is $895 (72% GDP per capita), which is not considered to be cost-effective at the threshold of 50% GDP per capita (which is becoming widely used), but would be considered cost-effective with previous thresholds.27 The universal test-and-treat intervention has a much larger impact than any other intervention because the home-based testing results in a higher population ART coverage, but is much more costly because of high HIV testing costs. This approach is also not cost-effective ($1760 per DALY averted; 142% GDP per capita).

The two interventions that represent large changes to ART eligibility, immediate ART and universal test and treat, differ substantially in their impact and cost. Immediate ART provides treatment to those who present for care, and is effective because it eliminates the potential for patients to be lost from care before they are confirmed to be eligible for ART. The cost per DALY averted is $895 (72% GDP per capita), which is not considered to be cost-effective at the threshold of 50% GDP per capita (which is becoming widely used), but would be considered cost-effective with previous thresholds.27 The universal test-and-treat intervention has a much larger impact than any other intervention because the home-based testing results in a higher population ART coverage, but is much more costly because of high HIV testing costs. This approach is also not cost-effective ($1760 per DALY averted; 142% GDP per capita). Most of the projected benefits of single interventions are because of the direct therapeutic effect of ART averting morbidity and mortality, rather than the secondary effects of fewer HIV transmissions (appendix p 51). Furthermore, with adjusted calibrations, if health-care seeking behaviour is greater and is not strongly related to symptoms, pretreatment interventions generate less impact as patients enter the health-care system earlier without any additional intervention (appendix p 49). By contrast, if care is only sought when a patient is symptomatic, the outreach intervention before ART initiation has twice the impact, as patients do not return to care faster when lost (appendix p 50).

e less impact as patients enter the health-care system earlier without any additional intervention (appendix p 49). By contrast, if care is only sought when a patient is symptomatic, the outreach intervention before ART initiation has twice the impact, as patients do not return to care faster when lost (appendix p 50). Finally, sensitivity analysis varying the unit costs of different aspects of care showed that the absolute cost of interventions is most sensitive to the cost of ART, but the rank order of cost per DALY averted for interventions is preserved when unit costs are varied over reasonable (±50%) ranges (appendix pp 53–55). We identified a combination of five interventions that maximise the health gained from a budget of $700 million (table 4). This budget was chosen because it is about equal to the cost of implementing immediate ART, but is less than the cost of implementing home-based counselling and testing in isolation. The five selected interventions, in the order in which they were added with increasing budget, were: pre-ART outreach, facilitated linkage, voluntary counselling and testing point-of-care CD4, point-of-care CD4, and on-ART outreach. The next intervention to be added would be enhanced counselling and testing, comprising a set of interventions that act on each part of the cascade at a total budget 67% lower than that of the home-based counselling and testing intervention in isolation (table 4).

nt-of-care CD4, point-of-care CD4, and on-ART outreach. The next intervention to be added would be enhanced counselling and testing, comprising a set of interventions that act on each part of the cascade at a total budget 67% lower than that of the home-based counselling and testing intervention in isolation (table 4). Collectively, this combination of interventions reduces AIDS deaths by 19% relative to baseline, the same reduction as the universal test-and-treat intervention, and averts 69% as many DALYs. It averts 77% more DALYs than the immediate ART intervention. However, the combination of cascade interventions costed 22% as much as universal test and treat and only 14% more than immediate ART. This combination approach is estimated to be cost-effective, at a cost of $571 per DALY averted (46% GDP per capita; figure 3). The comparatively low cost and high impact of the combination cascade intervention with this budget is the result of a collection of interventions operating synergistically, whereas the universal test and treat and immediate ART interventions incur inefficiencies because of the remaining weaknesses in the cascade. These synergies among interventions are exemplified by the point-of-care CD4 and pre-ART outreach interventions, for which the incremental cost-effectiveness ratio of both interventions together is lower than for either intervention alone (table 4). A combination of both strengthening of the cascade and changing eligibility to immediate ART averted 31% of AIDS-related deaths relative to baseline, at a cost of 54% GDP per capita (table 4).

the incremental cost-effectiveness ratio of both interventions together is lower than for either intervention alone (table 4). A combination of both strengthening of the cascade and changing eligibility to immediate ART averted 31% of AIDS-related deaths relative to baseline, at a cost of 54% GDP per capita (table 4). Discussion Our results suggest that ART programmes can be substantially and cost-effectively improved by strengthening each part of the care cascade through a combination of interventions. By contrast, simply moving to immediate ART would have less impact for the same cost as a combination of interventions. Although a universal test-and-treat strategy would generate greater health benefits than a combination of interventions, it would not be cost-effective if the weaknesses in the cascade were not resolved. In the coming years, one of the most cost-efficient interventions would be to find people who have disengaged from ART care (on-ART outreach). However, we reported that no individual pre-ART intervention has had a large effect on patient outcomes except annual testing interventions with substantial costs (figure 3). This finding is a result of the multifaceted nature of weakness in the present pre-ART care cascade. We do find that combinations of interventions at all parts of the cascade can have a large effect and be cost-effective. The reason that interventions affecting a single care stage have a modest effect is because there are weaknesses throughout the care cascade, so any potential impact is attenuated by remaining weaknesses elsewhere.

combinations of interventions at all parts of the cascade can have a large effect and be cost-effective. The reason that interventions affecting a single care stage have a modest effect is because there are weaknesses throughout the care cascade, so any potential impact is attenuated by remaining weaknesses elsewhere. Published trials have tended to examine the effect of single interventions on the cascade,8 because measurements of the effect of a combination of changes simultaneously poses challenges for robust experimental designs. Our model allows the effect of interventions, applied individually or in combination, to be assessed across the entire cascade. However, implementation studies investigating the effect and feasibility of complete combinations of interventions are needed to validate our findings. If greater synergy among interventions can be achieved than our model simulations, greater benefits might be realised at lower costs. Alternatively, the increased complexity of operations could lead to higher costs and fewer benefits. Therefore, trials of each individual and combination of interventions are needed to further confirm our results in western Kenya.

e achieved than our model simulations, greater benefits might be realised at lower costs. Alternatively, the increased complexity of operations could lead to higher costs and fewer benefits. Therefore, trials of each individual and combination of interventions are needed to further confirm our results in western Kenya. Attention has focused on the marginal therapeutic benefits to a patient who has had early initiation of ART. A potentially large benefit of earlier ART initiation is to reduce the risk of losing a patient from the pretreatment monitoring phase, so lessening the chance of further transmission and the risk of death from AIDS before such a time as the patient might return to care. Furthermore, earlier initiation in the form of immediate ART would continue to prioritise treatment for patients with low CD4 counts as per WHO guidelines.11 Our model did not consider the potential effect of interventions that reposition ART initiation (eg, home initiation), which have been shown in recent trials.28

. Furthermore, earlier initiation in the form of immediate ART would continue to prioritise treatment for patients with low CD4 counts as per WHO guidelines.11 Our model did not consider the potential effect of interventions that reposition ART initiation (eg, home initiation), which have been shown in recent trials.28 Neither of the strategies (combinations of cascade interventions or immediate ART for those presenting) would reduce the risk of deaths from AIDS among those not already diagnosed (appendix p 52). This issue would potentially exacerbate disparities in overall health between those who are better able to seek care. The effect of immediate ART is enhanced by further strengthening of care through a combination of interventions (figure 3). This result would lead to reductions in mortality before and after starting ART by outreach strategies returning more patients to treatment than in the absence of immediate ART.

able to seek care. The effect of immediate ART is enhanced by further strengthening of care through a combination of interventions (figure 3). This result would lead to reductions in mortality before and after starting ART by outreach strategies returning more patients to treatment than in the absence of immediate ART. Our model suggests that a large proportion of HIV-related deaths occur in individuals never diagnosed and those who were diagnosed but never initiated treatment. These results are in agreement with data from general population cohort studies in Rakai (Uganda) and uMkhanyakude (South Africa).29 Many health data systems account only for individuals who have registered with the clinic, and therefore will not capture this source of AIDS mortality (figure 2). Monitoring and evaluation frameworks for the cascade should therefore seek to quantify the extent to which deaths from AIDS are among those undiagnosed and treatment guidelines should recognise testing as an integral part of treatment programmes.

and therefore will not capture this source of AIDS mortality (figure 2). Monitoring and evaluation frameworks for the cascade should therefore seek to quantify the extent to which deaths from AIDS are among those undiagnosed and treatment guidelines should recognise testing as an integral part of treatment programmes. Therefore, to achieve the vision of eliminating deaths from AIDS, substantial active outreach is required to identify all individuals infected with HIV before needing ART. This goal is likely to incur substantial costs, but the exact cost is unknown and in well documented existing care programmes, costs vary between studies.17 For instance, an independent modelling analysis of a home-based counselling and testing intervention with active follow-up of patients piloted in Kwazulu-Natal (South Africa) reported that the intervention was cost effective,9 whereas home-based counselling and testing with passive referral was not cost-effective in our analysis. This discrepancy is likely because the KwaZulu-Natal study with active follow-up achieved a much higher rate of linkage to care (90%) than the 30% we assumed for our simulated home-based counselling and testing intervention with passive referral. Ongoing analysis of AMPATH home-based counselling and testing rounds will provide an opportunity to examine both efficacy and cost-effectiveness of this intervention with active referral in a different setting, which might show important ways in which its efficacy can be maximised.

g intervention with passive referral. Ongoing analysis of AMPATH home-based counselling and testing rounds will provide an opportunity to examine both efficacy and cost-effectiveness of this intervention with active referral in a different setting, which might show important ways in which its efficacy can be maximised. Comparison of our alternative model parameterisations (appendix pp 49, 50) shows how patterns of health-care seeking behaviour can modify the effect of interventions on population health, which has not been readily apparent from empirical observation.30 Health-care seeking behaviours have an important effect on the value of outreach interventions. Such behaviours are hard to measure empirically. Additionally, intrinsic care seeking behaviour and the functionality of provider-initiated counselling and testing are hard to distinguish from each other in many contexts. For these reasons, extrapolation of the findings of different studies from different populations into a common framework, as we had to do in this model, is hard. As a result, although each assumption about the interventions is based on a real study, our findings can only be directionally informative. Furthermore, uncertainty increases over the 20 year simulation period, particularly as major changes to health care and treatment are difficult to predict. For example, if a functional cure for HIV were developed or other important changes to treatment made, our results could become obsolete. Although the results presented rely largely on HIV health-care data from western Kenya served by AMPATH, we believe that our results will have the same broad relevance to other settings with large generalised epidemics in rural areas with an established ART programme.

treatment made, our results could become obsolete. Although the results presented rely largely on HIV health-care data from western Kenya served by AMPATH, we believe that our results will have the same broad relevance to other settings with large generalised epidemics in rural areas with an established ART programme. Findings from other modelling studies that have relied on aggregated routine data to provide insight into care are in broad agreement with our results. For example, a similar modelling study focusing on South Africa has shown the potential impact and cost-effectiveness of a combination of interventions strengthening the cascade.31 Additionally, other modelling studies have described the cost-effectiveness of immediate ART (Rwanda)32 and treatment re-initiation interventions (South Africa).33 These differences in intervention cost-effectiveness emphasise variations in the state of care by location and the risk of directly comparing model outputs resulting from potential inconsistencies in approaches and assumptions.

ness of immediate ART (Rwanda)32 and treatment re-initiation interventions (South Africa).33 These differences in intervention cost-effectiveness emphasise variations in the state of care by location and the risk of directly comparing model outputs resulting from potential inconsistencies in approaches and assumptions. Aspirations for HIV care and treatment have increased rapidly in recent years. The UNAIDS 90-90-90 strategy set out three ambitious targets to be achieved by 2020: 90% of people with HIV diagnosed, 90% of those on treatment, and 90% of them virally suppressed.13 Some countries are already moving ahead to universal eligibility for ART;1, 34 new data are emerging on clinical benefits of ART,35 many studies on cascade interventions are being reported,8 and WHO has recently released new guidance for ART programmes encouraging immediate initiation of treatment for all individuals positive for HIV.11 As countries move towards these targets and consider moves to new guidelines, integrating all available care cascade data with the perspective of improving health for the population is going to be especially important. Our results suggest that there is substantial scope for programmes to improve population health and that alternative sets of strategies are available that will be consistent with their particular aims and budget. For more information on the HIV Modelling Consortium see http://www.hivmodelling.org Supplementary Material Supplementary appendix

Aspirations for HIV care and treatment have increased rapidly in recent years. The UNAIDS 90-90-90 strategy set out three ambitious targets to be achieved by 2020: 90% of people with HIV diagnosed, 90% of those on treatment, and 90% of them virally suppressed.13 Some countries are already moving ahead to universal eligibility for ART;1, 34 new data are emerging on clinical benefits of ART,35 many studies on cascade interventions are being reported,8 and WHO has recently released new guidance for ART programmes encouraging immediate initiation of treatment for all individuals positive for HIV.11 As countries move towards these targets and consider moves to new guidelines, integrating all available care cascade data with the perspective of improving health for the population is going to be especially important. Our results suggest that there is substantial scope for programmes to improve population health and that alternative sets of strategies are available that will be consistent with their particular aims and budget. For more information on the HIV Modelling Consortium see http://www.hivmodelling.org Supplementary Material Supplementary appendix Acknowledgments We thank the Bill & Melinda Gates Foundation for funding support to the HIV Modelling Consortium) and their support to AMPATH for the merging of data from home-based testing with electronic medical records. This study was made possible through joint support from AMPATH and the United States Agency for International Development (USAID). The contents are the sole responsibility of the authors and do not necessarily reflect the views of USAID or the United States Government. Funded in part through a supplement to the National Institutes of Allergy and Infectious Diseases award 2U01AI069911-06 and a supplement to the East Africa International Epidemiologic Databases to Evaluate AIDS (IeDEA) Consortium (NIH Award U01 AI069911). Work by JWH was partially supported by NIH Award R01 AI 108441 and Lifespan/Tufts/Brown Center for AIDS Research (NIH Award P30 AI 42853).

and Infectious Diseases award 2U01AI069911-06 and a supplement to the East Africa International Epidemiologic Databases to Evaluate AIDS (IeDEA) Consortium (NIH Award U01 AI069911). Work by JWH was partially supported by NIH Award R01 AI 108441 and Lifespan/Tufts/Brown Center for AIDS Research (NIH Award P30 AI 42853). Contributors JJO, JWE, and TBH developed the mathematical model, did all simulations, and wrote the first draft of the paper. PB and JWH oversaw the creation and analysis of the data, and provided substantial intellectual contributions on model development and interpretation of data. ES and MN cleaned and analysed the AMPATH dataset. All authors were involved in manuscript revisions and approved the final version of the article for submission. Declaration of interests JJO, EM, and JWE received grants from the Bill & Melinda Gates Foundation during the conduct of the study. TBH received grants from the Bill & Melinda Gates Foundation, World Bank, UNAIDS, Rush Foundation, Wellcome Trust, and personal fees from the Bill & Melinda Gates Foundation and WHO during the conduct of the study. JWH, PB, MN, SK, and ES declare no competing interests. Figure 1 Operational steps involved in navigating an ART programme Blue arrows show linkage step in which patients were seen by a clinician and had blood taken for a CD4 test. Grey arrows show the shortcut to immediate ART initiation taken by individuals presenting with WHO stage 3 or 4 symptoms. Dashed arrow shows ART reinitiation after loss from ART care (does not occur in the baseline programme). ART=antiretroviral therapy.

ents were seen by a clinician and had blood taken for a CD4 test. Grey arrows show the shortcut to immediate ART initiation taken by individuals presenting with WHO stage 3 or 4 symptoms. Dashed arrow shows ART reinitiation after loss from ART care (does not occur in the baseline programme). ART=antiretroviral therapy. Figure 2 Distribution of care experience of patients who died from HIV ART=antiretroviral therapy. Late initiation is defined as a person with a CD4 count of less than 200 cells per μL. Figure 3 Disability-adjusted life-years averted and additional cost of care (based on 2013 US$) for interventions acting on the cascade between 2010 and 2030 Optimal combination of interventions includes facilitated linkage, on-ART outreach, VCT point-of-care CD4, pre-ART outreach, and point-of-care CD4. ART=antiretroviral therapy. HBCT=home-based counselling and testing. VCT=voluntary counselling and testing. Table 1 Summary of agreement between AMPATH data and the model

Figure 3 Disability-adjusted life-years averted and additional cost of care (based on 2013 US$) for interventions acting on the cascade between 2010 and 2030 Optimal combination of interventions includes facilitated linkage, on-ART outreach, VCT point-of-care CD4, pre-ART outreach, and point-of-care CD4. ART=antiretroviral therapy. HBCT=home-based counselling and testing. VCT=voluntary counselling and testing. Table 1 Summary of agreement between AMPATH data and the model AMPATH data Model 2007–10 2010–11 2011–14 2007–10 2010–11 2011–14 People diagnosed with HIV who entered care by route into care HBCT .. 7% 60% .. 13% 64% VCT 66% 47% 20% 65% 48% 20% PICT 34% 46% 20% 35% 39% 16% Proportion of individuals in CD4 strata at ART initiation >500 cells per μL 9% 14% 19% 10% 12% 9% 350–500 cells per μL 7% 8% 18% 11% 12% 9% 200–350 cells per μL 18% 21% 41% 14% 12% 46% <200 cells per μL 66% 57% 22% 65% 64% 36% AMPATH data were analysed in three discrete time periods: 2007–10 marking the period of time before household-based testing where individuals could only seek care through VCT or PICT, 2010–11 in which HBCT was rolled out in Bunyala (Kenya), and 2011–14 when HBCT was fully implemented and treatment eligibility guidelines had been updated to less than 350 cells per μL or WHO stage 3–4. See appendix pp 25–28 for corresponding figures. AMPATH=Academic Model Providing Access To Healthcare. HBCT=home-based counselling and testing. VCT=voluntary counselling and testing. PICT=provider-initiated counselling and testing. ART=antiretroviral therapy.

been updated to less than 350 cells per μL or WHO stage 3–4. See appendix pp 25–28 for corresponding figures. AMPATH=Academic Model Providing Access To Healthcare. HBCT=home-based counselling and testing. VCT=voluntary counselling and testing. PICT=provider-initiated counselling and testing. ART=antiretroviral therapy. Table 2 Summary of individual interventions designed to target various aspects of care

been updated to less than 350 cells per μL or WHO stage 3–4. See appendix pp 25–28 for corresponding figures. AMPATH=Academic Model Providing Access To Healthcare. HBCT=home-based counselling and testing. VCT=voluntary counselling and testing. PICT=provider-initiated counselling and testing. ART=antiretroviral therapy. Table 2 Summary of individual interventions designed to target various aspects of care Assumptions Cost (2013 US$) HBCT (passive referral) Every 4 years, 90% testing coverage; 30% linked to care immediately if not previously diagnosed; 40% if previously diagnosed $18·00 per HBCT person tested ($8·00 home visit9* and $10·00 rapid HIV test16) Enhanced counselling and testing The rate of HIV testing is 125% that of baseline $50·00 per person tested ($28·00 clinic visit,17 $10·00 rapid HIV test,16 and $12·00 CD4 laboratory test17) HBCT (with point-of-care CD4) Every 4 years, 90% testing coverage of population; 65% linked to care if not previously diagnosed, 70% if previously diagnosed (point-of-care CD4 reduces non-linkage by 50%) $60·00 per HBCT person tested ($8·00 home-visit,9* $10·00 rapid HIV-test,16 and $42·00 point-of-care CD4 test18) Facilitated linkage The risk of failure-to-link is reduced by 50% $2·61 per diagnosed but not linked patient per year19 VCT point-of-care CD4 At VCT, a point-of-care CD4 test is given to patients reducing the risk of not linking to 0% $80·00 per point-of-care CD4 test ($28·00 clinic visit,17 $10·00 rapid HIV test,16 and $42·00 point-of-care CD4 test18 Pre-ART outreach In the middle of each year, 20% of tested individuals lost from pre-ART care are sought and returned $19·55 per patient sought20 Improved care The risk of a patient missing an appointment is reduced by 50% $7·05 per patient per clinic visit21, 22 Point-of-care CD4 A point-of-care CD4 test removes the 10% disengagement from care between CD4 test and receiving result $70·00 per point-of-care CD4 test ($28·00 clinic visit17 and $42·00 point-of-care CD4 test18) On-ART outreach In the middle of each year, 40% of patients who had initiated ART and were lost from care are sought and returned $19·55 per patient sought.20 Adherence At ART initiation, adherence to ART increases by 50% $33·54 per person on ART per year23 Immediate ART No pre-ART care, all individuals who enter care are initiated onto ART immediately Only additional costs due to increased usage of ART (appendix p 45) Universal test and treat Immediate ART and HBCT (every 4 years, 90% testing coverage.

adherence to ART increases by 50% $33·54 per person on ART per year23 Immediate ART No pre-ART care, all individuals who enter care are initiated onto ART immediately Only additional costs due to increased usage of ART (appendix p 45) Universal test and treat Immediate ART and HBCT (every 4 years, 90% testing coverage. 30% linked if not previously diagnosed, 40% if previously diagnosed) $18·00 per HBCT person tested ($8·00 home visit9* and $10·00 rapid HIV-test16) When HBCT is applied in isolation in the model this incorporates only a passive referral of patients. Universal test and treat is a combination of HBCT and immediate ART. All interventions except immediate ART and universal test and treat, were considered when identifying the optimal combination of interventions acting on the cascade by selecting interventions with the lowest cost per disability-adjusted life-year averted. HBCT=home-based counselling and testing. VCT=voluntary counselling and testing. ART=antiretroviral therapy. * Secondary analysis of data from van Rooyen and colleagues;9 see appendix p 48 for further details. Table 3 DALYs averted and additional cost of care for individual interventions between 2010 and 2030

30% linked if not previously diagnosed, 40% if previously diagnosed) $18·00 per HBCT person tested ($8·00 home visit9* and $10·00 rapid HIV-test16) When HBCT is applied in isolation in the model this incorporates only a passive referral of patients. Universal test and treat is a combination of HBCT and immediate ART. All interventions except immediate ART and universal test and treat, were considered when identifying the optimal combination of interventions acting on the cascade by selecting interventions with the lowest cost per disability-adjusted life-year averted. HBCT=home-based counselling and testing. VCT=voluntary counselling and testing. ART=antiretroviral therapy. * Secondary analysis of data from van Rooyen and colleagues;9 see appendix p 48 for further details. Table 3 DALYs averted and additional cost of care for individual interventions between 2010 and 2030 DALYS averted between 2010 and 2030 (million) Additional cost between 2010 and 2030 (2013; million US$) Cost per DALY averted compared with baseline (ACER)* AIDS deaths averted (%) HBCT 0·96 $2241·11 $2324·76 11·56 Enhanced counselling and testing 0·24 $253·76 $1062·06 2·94 HBCT (with point-of-care CD4) 1·61 $2600·75 $1616·71 19·12 Facilitated linkage 0·10 $39·75 $383·97 1·02 VCT point-of-care CD4 0·13 $63·93 $474·19 1·46 Pre-ART outreach 0·26 $217·74 $822·99 3·81 Improved care 0·16 $156·74 $1008·24 2·01 Point-of-care CD4 0·11 $107·40 $953·35 1·66 On-ART outreach 0·71 $355·92 $499·41 13·85 Adherence 0·46 $364·41 $787·45 5·56 Immediate ART 0·62 $552·02 $895·12 8·32 Universal test and treat 1·60 $2813·84 $1760·10 19·11 DALY=disability-adjusted life-year. ACER=average cost-effectiveness ratio. HBCT=home-based counselling and testing. VCT=voluntary counselling and testing. ART=antiretroviral therapy.

1 13·85 Adherence 0·46 $364·41 $787·45 5·56 Immediate ART 0·62 $552·02 $895·12 8·32 Universal test and treat 1·60 $2813·84 $1760·10 19·11 DALY=disability-adjusted life-year. ACER=average cost-effectiveness ratio. HBCT=home-based counselling and testing. VCT=voluntary counselling and testing. ART=antiretroviral therapy. * Calculations had an SE of US$150. Table 4 DALYs averted and additional cost of implementing a combination of interventions between 2010 and 2030 DALYs averted (million) Additional cost relative to baseline (2013; million US$) ICER* ACER† Facilitated linkage 0·10 $39·75 $383·97 $383·97 Facilitated linkage and on-ART outreach 0·81 $406·78 $518·30 $501·17 Facilitated linkage, on-ART outreach, and VCT point-of-care CD4 0·88 $457·03 $783·02 $521·82 Facilitated linkage, on-ART outreach, VCT point-of-care CD4, pre-ART outreach 1·09 $623·33 $774·53 $571·57 Facilitated linkage, on-ART outreach, VCT point-of-care CD4, pre-ART outreach, and point-of-care CD4 1·10 $630·99 $543·24 $571·21 Facilitated linkage, on-ART outreach, VCT point-of-care CD4, pre-ART outreach, point-of-care CD4, and immediate ART 1·68 $1123·06 $852·10 $667·64 Intervention results in table 3 cannot be combined additively to arrive at those listed above because these results are generated with a dynamic model. Interventions were considered cost-effective if ACER was less than 50% of GDP per capita for Kenya in 2013 (US$1242).24 DALY=disability-adjusted life-year. ICER=incremental cost-effectiveness ratio. ACER=average cost-effectiveness ratio. ART=antiretroviral therapy. VCT=voluntary counselling and testing.

dynamic model. Interventions were considered cost-effective if ACER was less than 50% of GDP per capita for Kenya in 2013 (US$1242).24 DALY=disability-adjusted life-year. ICER=incremental cost-effectiveness ratio. ACER=average cost-effectiveness ratio. ART=antiretroviral therapy. VCT=voluntary counselling and testing. * Cost per DALY averted compared with previous increment. † Cost per DALY averted compared with baseline.

ing school.12, 23 Diagnostic HIV testing was not done during the household visit because it could have affected participation in the prevalence survey, but dedicated research staff were available at the primary health-care clinics so that those attending for HIV testing did not have to wait in the routine clinic queue. The use of incentives to increase HIV testing is grounded in two economic concepts related to decision making. First, an economic incentive might mitigate indirect costs of HIV testing incurred by clients, such as loss of income through time taken off work and transport costs. These could be an even larger cost consideration for a child who is likely to be economically dependent. Second, some individuals might display what is termed present-biased preferences of a behaviour. They place disproportionate emphasis on the immediate costs and benefits, such as economic burden or fear of a positive result compared with future costs and benefits.13 Incentives might bring forward in time the benefits and sway the decision of the child, the caregiver, or both.

ent-biased preferences of a behaviour. They place disproportionate emphasis on the immediate costs and benefits, such as economic burden or fear of a positive result compared with future costs and benefits.13 Incentives might bring forward in time the benefits and sway the decision of the child, the caregiver, or both. Incentives have been used for the completion of goal-directed activities such as hepatitis B vaccination, tuberculosis screening, and testing for sexually transmitted infections.11, 24 Several studies19, 20 have shown improved uptake of HIV testing by young people and first-time testers in sub-Saharan Africa. However, incentivised HIV testing in children and adolescents has never been investigated. Findings from a recent study25 in Tanzania showed that incentivising universal HIV testing in adults with $1·30–6·40 was highly cost-effective. The costs per quality-adjusted life-year gained was $70 for prevalent and $620 for incident HIV infections. However, HIV prevalence is generally lower in children and adolescents than in adults and therefore cannot be generalised to this age group. This might be off set partly by the fact that children and adolescents have more unlived life-years and are not at ongoing risk of being HIV infected until they become sexually active. HIV testing is therefore a one-off activity in childhood, which is particularly important because the sustainability of incentivisation strategies is of concern, particularly for enforcing long-term changes in health behaviours, such as adherence to ART.26, 27

Introduction For 20 years, combination antiretroviral therapy (ART) has been the standard approach to treating HIV-1 infection in Europe and North America. The first ART regimens were inferior to those currently available, which better suppress HIV replication, are less toxic, and have higher genetic barriers to resistance, reduced pill burden (often one a day), and fewer side-effects.1, 2 Other improvements in health care since 1996 for people living with HIV include treatment and prophylaxis for opportunistic infections and management of comorbidities.3 Improvements in intensive care management, disease screening, and health promotion might also have improved prognosis. Therefore, people living with HIV who started ART more recently might have improved survival compared with those treated earlier in the ART era.

portunistic infections and management of comorbidities.3 Improvements in intensive care management, disease screening, and health promotion might also have improved prognosis. Therefore, people living with HIV who started ART more recently might have improved survival compared with those treated earlier in the ART era. The Antiretroviral Therapy Cohort Collaboration (ART-CC) previously reported that despite improvements in virological response to ART, mortality 1 year after initiation of ART did not decrease between 1998 and 2003.4 This absence of improvement in survival might have been related to changes in patients' characteristics: increasing numbers were from areas with a high prevalence of tuberculosis infection.4 Some studies have reported improvements in overall survival and changing causes of death, with proportionately fewer AIDS-related deaths in more recent years,5, 6, 7, 8, 9 but none has investigated trends in prognosis after starting ART by calendar period. We examined changes in all-cause and cause-specific mortality in the first 3 years of ART during 1996–2013, and investigated trends in life expectancy.

, with proportionately fewer AIDS-related deaths in more recent years,5, 6, 7, 8, 9 but none has investigated trends in prognosis after starting ART by calendar period. We examined changes in all-cause and cause-specific mortality in the first 3 years of ART during 1996–2013, and investigated trends in life expectancy. Methods Participants We combined data from 18 European and North American HIV cohorts participating in ART-CC, which includes ART-naive people living with HIV aged 16 years or older who started treatment with three or more antiretroviral drugs between 1996 and 2010.10 Cohorts were approved by ethics committees or institutional review boards, used standardised data collection methods, and scheduled follow-up visits at least every 6 months. Cohorts included in this paper were the French Hospital Database on HIV (FHDH); the Italian Cohort of Antiretroviral-naive patients (ICONA); the Swiss HIV Cohort Study (SHCS); the AIDS Therapy Evaluation project, Netherlands (ATHENA); the Multicenter Study Group on EuroSIDA; the Aquitaine Cohort, France; the Royal Free Hospital Cohort, UK; the South Alberta Clinic Cohort, Canada; the Danish HIV Cohort Study, Denmark; HAART Observational Medical Evaluation and Research (HOMER) Cohort, Canada; HIV Atlanta Veterans Affairs Cohort Study (HAVACS), USA; Osterreichische HIV-Kohortenstudie (OEHIVKOS), Austria; Proyecto para la Informatizacion del Seguimiento Clinico-epidemiologico de la Infeccion por HIV y SIDA (PISCIS), Spain; University of Washington HIV Cohort, USA; VACH, Spain; Veterans Aging Cohort Study (VACS), USA; Vanderbilt, USA; and the Koln/Bonn Cohort, Germany.

erreichische HIV-Kohortenstudie (OEHIVKOS), Austria; Proyecto para la Informatizacion del Seguimiento Clinico-epidemiologico de la Infeccion por HIV y SIDA (PISCIS), Spain; University of Washington HIV Cohort, USA; VACH, Spain; Veterans Aging Cohort Study (VACS), USA; Vanderbilt, USA; and the Koln/Bonn Cohort, Germany. Research in context Evidence before this study We ran three PubMed searches for articles published between Jan 1, 2001, and June 1, 2016, with the terms (1) “HIV”, “calendar year”, and “mortality”; (2) “HIV”, “life expectancy”, and “mortality”; (3) “HIV”, “causes of death”, and “mortality”. A study of 2675 patients in the Australian HIV cohort found lower mortality in individuals who started antiretroviral therapy (ART) from 2004 onwards compared with earlier years, although that study focused on the effect of duration of treatment rather than trends in early mortality on ART. A large study by D:A:D found lower all-cause, cardiovascular, and liver disease mortality in patients followed up in 2009–11 compared with 1999–2000, but did not analyse deaths by period of starting ART. We previously reported that between 1995 and 2003, virological response to ART improved but early mortality did not decrease. Surveillance data from the USA showed increasing life expectancy with year of HIV diagnosis between 1996 and 2005. We also reported that life expectancy in individuals starting ART increased between 1996 and 2005; studies in Europe and the USA have also shown increases in life expectancy over time. Added value of this study

We ran three PubMed searches for articles published between Jan 1, 2001, and June 1, 2016, with the terms (1) “HIV”, “calendar year”, and “mortality”; (2) “HIV”, “life expectancy”, and “mortality”; (3) “HIV”, “causes of death”, and “mortality”. A study of 2675 patients in the Australian HIV cohort found lower mortality in individuals who started antiretroviral therapy (ART) from 2004 onwards compared with earlier years, although that study focused on the effect of duration of treatment rather than trends in early mortality on ART. A large study by D:A:D found lower all-cause, cardiovascular, and liver disease mortality in patients followed up in 2009–11 compared with 1999–2000, but did not analyse deaths by period of starting ART. We previously reported that between 1995 and 2003, virological response to ART improved but early mortality did not decrease. Surveillance data from the USA showed increasing life expectancy with year of HIV diagnosis between 1996 and 2005. We also reported that life expectancy in individuals starting ART increased between 1996 and 2005; studies in Europe and the USA have also shown increases in life expectancy over time. Added value of this study Our study, based on a large collaboration of cohorts in Europe and North America, found that substantial declines in mortality for individuals starting ART in 2008–10, compared with earlier years, has resulted in increased life expectancy. However, this life expectancy remains lower than that of the general population. Declines in mortality were greater for the second and third years after starting ART than for the first year after starting ART. Rates of non-AIDS-related deaths, particularly deaths from cardiovascular disease, were substantially lower in 2008–10 than previously. There was little evidence that mortality has declined in people who inject drugs.

greater for the second and third years after starting ART than for the first year after starting ART. Rates of non-AIDS-related deaths, particularly deaths from cardiovascular disease, were substantially lower in 2008–10 than previously. There was little evidence that mortality has declined in people who inject drugs. Implications of all the available evidence Improvements in the care of people living with HIV since the introduction of ART 20 years ago have led to improved survival and increased life expectancy in those starting ART. These improvements probably reflect the availability of superior antiretroviral agents, more options for the management of patients developing resistance, fewer drug interactions, better management of opportunistic infections and chronic diseases, and introduction of screening and prevention programmes. Prognostic models and estimates of life expectancy should be updated to account for these improvements. Eligible patients started ART at least 3 years before the cohort-specific database close date, which varied from May 31, 2012, to July 31, 2013, and had a baseline CD4 cell count measured within a window 3 months before until 2 weeks after starting ART. We defined CD4 cell count and viral load 1 year after the start of ART as the closest measurement before 1 year within a 9–12 month window.

database close date, which varied from May 31, 2012, to July 31, 2013, and had a baseline CD4 cell count measured within a window 3 months before until 2 weeks after starting ART. We defined CD4 cell count and viral load 1 year after the start of ART as the closest measurement before 1 year within a 9–12 month window. Patients were followed up for all-cause and cause-specific mortality from the time of starting ART, considered lost to follow-up if there was a gap of more than 1 year between the dates they were last known to be alive and the database close date, and censored 6 months after the last recorded measurement. Mortality information was obtained through linkage with Vital Statistics agencies and hospitals or physician report and active follow-up of participants. Methods for classifying causes of death, with an adaptation of the CoDe project protocol, are described elsewhere.11 Deaths were coded as AIDS-related if there was a serious AIDS defining condition close to death or a low CD4 cell count (<100 cells per μL) before death, and a diagnosis compatible with AIDS as cause of death.

s for classifying causes of death, with an adaptation of the CoDe project protocol, are described elsewhere.11 Deaths were coded as AIDS-related if there was a serious AIDS defining condition close to death or a low CD4 cell count (<100 cells per μL) before death, and a diagnosis compatible with AIDS as cause of death. Statistical analysis We compared characteristics of patients by calendar period of initiation of ART (1996–99, 2000–03, 2004–07, 2008–10). We used Cox models stratified by cohort to estimate unadjusted and adjusted mortality hazard ratios (HRs) by period of initiation of ART. Models were adjusted for sex, injecting drug use, AIDS at baseline, age (16–29, 30–39, 40–49, 50–59, ≥60 years), CD4 cell count (0–24, 25–49, 50–99, 100–199, 200–349, 350–499, ≥500 cells per μL), and HIV-1 RNA viral load (0 to <10 000, ≥10 000 to <100 000, ≥100 000 copies per mL) at the start of ART. Because mortality is higher in the first year of ART, we fitted separate models for the first year after starting ART and the second and third years after starting ART. To investigate the mediating effect of response to therapy, we additionally adjusted the second and third year analysis for CD4 cell count and viral load measured 1 year after initiation of ART. We compared HR for regimen failure, defined as switching regimen within 6 months of a viral load measurement of more than 1000 copies per mL after first achieving viral suppression, by period of ART initiation. In patients with regimen failure, we estimated mortality HR by period of ART initiation.

r initiation of ART. We compared HR for regimen failure, defined as switching regimen within 6 months of a viral load measurement of more than 1000 copies per mL after first achieving viral suppression, by period of ART initiation. In patients with regimen failure, we estimated mortality HR by period of ART initiation. We used Cox models to investigate the consistency of mortality trends across subgroups of patients defined by sex and transmission risk (men who have sex with men [MSM], male heterosexual, female heterosexual, men who inject drugs, women who inject drugs) in European patients (because transmission risk was missing for large numbers of North American patients); CD4 cell count categories (<100, 100–199, 200–349, ≥350 cells per μL); age; and region (Europe, North America). In a sensitivity analysis, patients were considered lost to follow-up at 6 months rather than 12 months. We repeated analyses with the following causes of death as the outcome: AIDS-related, non-AIDS-related (ie, non-AIDS infection, malignancies not caused by AIDS or hepatitis, liver disease, cardiovascular disease, and other [causes with ≤20 cases]), and unnatural causes (suicide, accident or other violent death, euthanasia, and substance abuse), and those missing or unknown. We estimated sex-specific life expectancy by period of ART initiation, overall and by region (North America, Europe). We used a Poisson model to estimate mortality in age bands of 5 years, which were used to construct life tables and estimate average age at death for those aged 20 years at initiation of ART. We compared these estimates with those from the French and US general populations. For comparability across periods and to investigate the effect of higher mortality in the first year of ART, we estimated life expectancy on the basis of mortality in the first 3 years of follow-up, and then with mortality in the second and third years of follow-up. Because there were few patients aged 70 years or older, for the oldest open-ended age group we used the French general population mortality multiplied by the mean rate ratio in ART-CC compared with the French general population (chosen because France contributed the most patients) for ages 60–64 and 65–69 years. Per-period changes in life expectancy were estimated with metaregression. We used Stata (version 14) for analyses.

ch general population mortality multiplied by the mean rate ratio in ART-CC compared with the French general population (chosen because France contributed the most patients) for ages 60–64 and 65–69 years. Per-period changes in life expectancy were estimated with metaregression. We used Stata (version 14) for analyses. Role of the funding source The funders of the study, the UK Medical Research Council, UK Department for International Development, and the European Union, had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results 88 504 patients were eligible for our analyses. During 84 621 person-years, 2106 (2%) patients died in the first year after starting ART (24·9 per 1000 person-years). 81 608 (92%) individuals remained in the study for more than 1 year, of whom 2302 (3%) died during 153 813 person-years (15·0 per 1000 person-years). 4594 (5%) patients were lost to follow-up during the first year after starting ART and 6674 patients (8%) were lost-to follow-up during the second and third years. Patients who were lost to follow-up had lower viral loads and higher CD4 cell counts than were those not lost to follow-up, and they were also more likely to be female, younger, and people who inject drugs (data not shown).

er starting ART and 6674 patients (8%) were lost-to follow-up during the second and third years. Patients who were lost to follow-up had lower viral loads and higher CD4 cell counts than were those not lost to follow-up, and they were also more likely to be female, younger, and people who inject drugs (data not shown). The proportion of women increased from 20% in 1996–99 to 28% in 2004–07, then decreased to 21% in 2008–10 (table 1). Median age increased between 1996–99 and 2008–10, whereas the proportion of people who inject drugs decreased from 17% to 7%. Median CD4 count 1 year after ART initiation increased substantially, from 370 cells per μL (IQR 211–572) in 1996–99 to 430 cells per μL (295–570) in 2008–10, and the proportion of patients with HIV-1 RNA viral load of 500 copies per mL or lower increased from 71% in 1996–99 to 93% in 2008–10.Table 1 Characteristics of patients at the time of starting antiretroviral therapy with number of deaths, by period of initiation

2) in 1996–99 to 430 cells per μL (295–570) in 2008–10, and the proportion of patients with HIV-1 RNA viral load of 500 copies per mL or lower increased from 71% in 1996–99 to 93% in 2008–10.Table 1 Characteristics of patients at the time of starting antiretroviral therapy with number of deaths, by period of initiation 1996–99 2000–03 2004–07 2008–10 Total deaths 1431/24 445 (6%) 1452/25 683 (6%) 1084/24 462 (4%) 441/13 914 (3%) Men 1240/19 529 (6%) 1231/18 653 (7%) 882/17 570 (5%) 377/10 995 (3%) Women 191/4916 (4%) 221/7030 (3%) 202/6892 (3%) 64/2919 (2%) No injecting drug use 1161/20 186 (6%) 1187/22 182 (5%) 877/22 020 (4%) 385/12 885 (3%) Injecting drug use 270/4259 (6%) 265/3501 (8%) 207/2442 (8%) 56/1029 (5%) No AIDS 810/19 118 (4%) 799/19 589 (4%) 595/19 262 (3%) 239/11 570 (2%) AIDS 621/5327 (12%) 653/6094 (11%) 489/5200 (9%) 202/2344 (9%) Age (years) 36 (31–43) 37 (31–45) 39 (32–46) 40 (32–47) 16–29 117/4457 (3%) 95/4620 (2%) 55/4118 (1%) 23/2208 (1%) 30–39 508/11 176 (5%) 399/10 320 (4%) 256/8846 (3%) 74/4651 (2%) 40–49 433/5729 (8%) 485/6714 (7%) 352/7294 (5%) 142/4311 (3%) 50–59 257/2270 (11%) 313/2998 (10%) 261/3003 (9%) 115/1919 (6%) ≥60 116/813 (14%) 160/1031 (16%) 160/1201 (13%) 87/825 (11%) CD4 count (cells per μL) 238 (93–394) 200 (81–326) 219 (115–310) 265 (157–351) 0–24 356/2511 (14%) 355/2789 (13%) 232/1921 (12%) 97/849 (11%) 25–49 173/1508 (11%) 179/1753 (10%) 102/1291 (8%) 37/559 (7%) 50–99 197/2332 (8%) 244/2771 (9%) 173/2203 (8%) 57/891 (6%) 100–199 288/4212 (7%) 331/5461 (6%) 246/5370 (5%) 69/2272 (3%) 200–349 237/6171 (4%) 217/7354 (3%) 240/9090 (3%) 115/5764 (2%) 350–499 109/4258 (3%) 76/3171 (2%) 58/2673 (2%) 44/2384 (2%) ≥500 71/3453 (2%) 50/2384 (2%) 33/1914 (2%) 22/1195 (2%) HIV-1 RNA (log copies per mL) 4·9 (4·2–5·4) 4·9 (4·3–5·4) 4·8 (4·1–5·3) 4·7 (4·1–5·2) 0–3·99 177/4439 (4%) 179/4693 (4%) 132/5317 (2%) 70/3224 (2%) 4–4·99 396/9397 (4%) 426/9355 (5%) 332/8937 (4%) 158/5675 (3%) ≥5 858/10 609 (8%) 847/11 635 (7%) 620/10 208 (6%) 213/5015 (4%) Regimen NNRTI-based 210/4178 (5%) 609/11 391 (5%) 460/12 126 (4%) 215/7902 (3%) Protease inhibitor-based 1159/19 184 (6%) 602/9520 (6%) 555/10 496 (5%) 197/5398 (4%) NRTI/abacavir* 18/364 (5%) 196/4040 (5%) 47/1312 (4%) 4/117 (3%) NRTI/non-abacavir† 28/542 (5%) 29/564 (5%) 6/240 (3%) 0/29 (0%) Other 16/177 (9%) 16/168 (10%) 16/288 (6%) 25/468 (5%) Data are number of deaths/number of patients (%) or median (IQR). NNRTI=non-nucleoside reverse transcriptase inhibitor.

%) 197/5398 (4%) NRTI/abacavir* 18/364 (5%) 196/4040 (5%) 47/1312 (4%) 4/117 (3%) NRTI/non-abacavir† 28/542 (5%) 29/564 (5%) 6/240 (3%) 0/29 (0%) Other 16/177 (9%) 16/168 (10%) 16/288 (6%) 25/468 (5%) Data are number of deaths/number of patients (%) or median (IQR). NNRTI=non-nucleoside reverse transcriptase inhibitor. NRTI=nucleoside reverse transcriptase inhibitor. * Triple NRTI including abacavir. † Triple NRTI not including abacavir. During 1996–99, most patients started a protease inhibitor-based regimen, whereas after 2000 non-nucleoside reverse transcriptase inhibitor (NNRTI)-based regimens were most common (table 2). The protease inhibitors indinavir, nelfinavir, and saquinavir were replaced by atazanavir, darunavir, and lopinavir. Of the NNRTIs, efavirenz was the most commonly used third regimen drug from 2000 onwards. The NRTIs didanosine, stavudine, and zidovudine were replaced by abacavir and tenofovir.Table 2 Proportion of patients prescribed specific antiretroviral drugs as their first regimen, by period of initiation

unavir, and lopinavir. Of the NNRTIs, efavirenz was the most commonly used third regimen drug from 2000 onwards. The NRTIs didanosine, stavudine, and zidovudine were replaced by abacavir and tenofovir.Table 2 Proportion of patients prescribed specific antiretroviral drugs as their first regimen, by period of initiation 1996–99 (n=24 445) 2000–03 (n=25 683) 2004–07 (n=24 462) 2008–10 (n=13 914) Protease inhibitors Amprenavir 0 1% 0 0 Atazanavir 0 0 13% 19% Darunavir 0 0 0 3% Fosamprenavir 0 0 6% 2% Indinavir 37% 9% 1% 0 Lopinavir 0 8% 16% 13% Nelfinavir 25% 16% 3% 0 Ritonavir 7% 1% 1% 2% Saquinavir 10% 3% 3% 1% Non-nucleoside reverse transcriptase inhibitors Efavirenz 6% 30% 42% 50% Nevirapine 12% 15% 8% 7% Entry inhibitors Enfuvirtide 0 0 1% 0 Integrase inhibitors Raltegravir 0 0 0 3% Nucleoside reverse transcriptase inhibitors Abacavir 2% 18% 16% 11% Didanosine 17% 14% 7% 1% Emtricitabine 0 0 37% 80% Lamivudine 80% 90% 59% 19% Stavudine 40% 20% 2% 0 Tenofovir 0 5% 49% 79% Zalcitabine 3% 0 0 0 Zidovudine 59% 68% 34% 8%

% Entry inhibitors Enfuvirtide 0 0 1% 0 Integrase inhibitors Raltegravir 0 0 0 3% Nucleoside reverse transcriptase inhibitors Abacavir 2% 18% 16% 11% Didanosine 17% 14% 7% 1% Emtricitabine 0 0 37% 80% Lamivudine 80% 90% 59% 19% Stavudine 40% 20% 2% 0 Tenofovir 0 5% 49% 79% Zalcitabine 3% 0 0 0 Zidovudine 59% 68% 34% 8% Compared with patients who started ART in 2000–03, all-cause mortality during the first year of ART was similar in patients who started ART between 1996 and 2007, but substantially lower for those who started ART in 2008–10 (adjusted HR 0·71, 95% CI 0·61–0·83; figure 1). The adjusted HR per calendar period was 0·90 (0·87–0·95). Declines in 1 year mortality over calendar time were consistent across subgroups of patients defined by their characteristics at the start of ART, apart from individuals who inject drugs and those starting ART with CD4 counts less than 100 cells per μL or CD4 count greater than or equal to 350 cells per μL (appendix).Figure 1 Unadjusted and adjusted all-cause mortality hazard ratios for the first year after starting antiretroviral therapy (ART), by period of initiation *Adjusted for age, sex, AIDS, risk group, CD4 cell count, and HIV-1 RNA at the time of starting ART.

Compared with patients who started ART in 2000–03, all-cause mortality during the first year of ART was similar in patients who started ART between 1996 and 2007, but substantially lower for those who started ART in 2008–10 (adjusted HR 0·71, 95% CI 0·61–0·83; figure 1). The adjusted HR per calendar period was 0·90 (0·87–0·95). Declines in 1 year mortality over calendar time were consistent across subgroups of patients defined by their characteristics at the start of ART, apart from individuals who inject drugs and those starting ART with CD4 counts less than 100 cells per μL or CD4 count greater than or equal to 350 cells per μL (appendix).Figure 1 Unadjusted and adjusted all-cause mortality hazard ratios for the first year after starting antiretroviral therapy (ART), by period of initiation *Adjusted for age, sex, AIDS, risk group, CD4 cell count, and HIV-1 RNA at the time of starting ART. All-cause mortality in the second and third years after starting ART declined substantially over calendar time (adjusted HR per calendar period 0·78, 95% CI 0·75–0·82; figure 2). Declines were consistent across Europe and North America, age groups, and CD4 cell count at ART initiation. The decline in mortality was less in people who inject drugs (adjusted HR per calendar period 0·90 [0·80–1·02] for men and 0·95 [0·76–1·20] for women; appendix p 1) than in other groups. We examined the mediating effects of CD4 cell count and viral load measured 1 year after starting ART on mortality trends during the second and third years of ART in 53 244 (65%) eligible patients with available measurements (figure 2). Additionally adjusting for 1 year CD4 cell count and viral load attenuated the adjusted HR per calendar period to 0·90 (0·85–0·96). The proportion of patients with regimen failure declined over time (adjusted HR per calendar period 0·73, 0·72–0·75), but among those with regimen failure there was no evidence of an improvement in survival (adjusted HR per calendar period 0·98, 0·87–1·09; appendix p 2). Results of sensitivity analyses in which patients were considered lost to follow-up at 6 months rather than 12 months were similar to the main analyses (data not shown).Figure 2 All-cause mortality hazard ratios for the second and third years after starting antiretroviral therapy (ART), by period of initiation

ix p 2). Results of sensitivity analyses in which patients were considered lost to follow-up at 6 months rather than 12 months were similar to the main analyses (data not shown).Figure 2 All-cause mortality hazard ratios for the second and third years after starting antiretroviral therapy (ART), by period of initiation *Adjusted for age, sex, AIDS, risk group, CD4 cell count, and HIV-1 RNA at the time of starting ART. Causes were classified for 3126 (71%) of 4408 deaths. Rates of deaths from AIDS during the first year of ART declined over calendar time (adjusted HR per calendar period 0·93, 95% CI 0·86–0·99), with greater declines for the second and third year of ART (0·69, 0·64–0·76). Rates of non-AIDS-related death during the first year of follow-up declined over calendar time (0·87, 0·80–0·95), as did rates during the second and third years of follow-up (0·75, 0·69–0·81). Declines in mortality were consistent across causes of mortality (table 3): the greatest decline was in liver-related deaths during the second and third years of ART.Table 3 Adjusted hazard ratios for specific causes of death by period of antiretroviral therapy (ART) initiation for first year of ART and second and third years of ART

rtality were consistent across causes of mortality (table 3): the greatest decline was in liver-related deaths during the second and third years of ART.Table 3 Adjusted hazard ratios for specific causes of death by period of antiretroviral therapy (ART) initiation for first year of ART and second and third years of ART Number of deaths Period of ART initiation Per period 1996–99 2000–03 2004–07 2008–10 First year of follow-up AIDS 902 0·98 (0·83–1·16) 1 0·94 (0·79–1·11) 0·71 (0·56–0·90) 0·93 (0·86–0·99) Non-AIDs 525 1·04 (0·83–1·30) 1 1·09 (0·88–1·35) 0·48 (0·34–0·67) 0·87 (0·80–0·95) Non-AIDS infection 117 0·89 (0·55–1·43) 1 1·11 (0·71–1·74) 0·45 (0·22–0·94) 0·91 (0·76–1·10) Non-AIDS, non-hepatitis malignancies 122 1·01 (0·61–1·68) 1 1·53 (0·97–2·39) 0·65 (0·34–1·24) 0·99 (0·83–1·18) Liver-related 76 1·07 (0·60–1·88) 1 0·93 (0·52–1·67) 0·36 (0·14–0·96) 0·80 (0·63–1·01) Cardiovascular 64 0·95 (0·52–1·73) 1 0·72 (0·39–1·35) 0·19 (0·06–0·62) 0·71 (0·55–0·92) Other 146 1·18 (0·78–1·78) 1 1·03 (0·67–1·59) 0·64 (0·34–1·19) 0·87 (0·73–1·02) Unnatural* 107 1·37 (0·82–2·29) 1 1·49 (0·90–2·48) 0·62 (0·29–1·34) 0·89 (0·73–1·08) Missing/unknown 572 1·00 (0·81–1·24) 1 0·86 (0·69–1·08) 0·98 (0·76–1·23) 0·97 (0·89–1·05) Second and third years of follow-up AIDS 646 1·34 (1·12–1·60) 1 0·74 (0·59–0·92) 0·35 (0·24–0·51) 0·69 (0·64–0·76) Non-AIDS 770 1·12 (0·94–1·34) 1 0·86 (0·71–1·03) 0·29 (0·21–0·40) 0·75 (0·69–0·81) Non-AIDS infection 132 0·79 (0·52–1·19) 1 0·66 (0·43–1·04) 0·27 (0·12–0·59) 0·79 (0·66–0·95) Non-AIDS, non-hepatitis malignancies 206 1·48 (1·03–2·13) 1 1·40 (0·97–2·00) 0·50 (0·28–0·87) 0·82 (0·71–0·94) Liver-related 127 0·94 (0·63–1·40) 1 0·49 (0·30–0·79) 0·15 (0·05–0·42) 0·66 (0·54–0·80) Cardiovascular 100 0·82 (0·50–1·34) 1 0·79 (0·49–1·29) 0·21 (0·08–0·53) 0·78 (0·64–0·95) Other 205 1·47 (1·05–2·05) 1 0·93 (0·64–1·34) 0·29 (0·14–0·59) 0·69 (0·60–0·80) Unnatural* 176 1·06 (0·74–1·53) 1 0·91 (0·62–1·35) 0·32 (0·16–0·63) 0·79 (0·68–0·92) Missing/unknown 710 1·23 (1·02–1·48) 1 0·73 (0·58–0·91) 1·19 (0·96–1·49) 0·93 (0·87–1·00) Data are hazard ratio (95% CI), mutually adjusted for age, sex, AIDS, risk group, CD4 cell count, HIV-1 RNA, and stratified by cohort, with 2000–03 as comparator.

06 (0·74–1·53) 1 0·91 (0·62–1·35) 0·32 (0·16–0·63) 0·79 (0·68–0·92) Missing/unknown 710 1·23 (1·02–1·48) 1 0·73 (0·58–0·91) 1·19 (0·96–1·49) 0·93 (0·87–1·00) Data are hazard ratio (95% CI), mutually adjusted for age, sex, AIDS, risk group, CD4 cell count, HIV-1 RNA, and stratified by cohort, with 2000–03 as comparator. * Unnatural deaths include suicide, accident or other violent death, euthanasia, and substance abuse.

06 (0·74–1·53) 1 0·91 (0·62–1·35) 0·32 (0·16–0·63) 0·79 (0·68–0·92) Missing/unknown 710 1·23 (1·02–1·48) 1 0·73 (0·58–0·91) 1·19 (0·96–1·49) 0·93 (0·87–1·00) Data are hazard ratio (95% CI), mutually adjusted for age, sex, AIDS, risk group, CD4 cell count, HIV-1 RNA, and stratified by cohort, with 2000–03 as comparator. * Unnatural deaths include suicide, accident or other violent death, euthanasia, and substance abuse. Life expectancy increased with calendar period of initiation of ART, for both men and women (figure 3 and appendix p 2). Expected average ages at death for Europeans aged 20 years starting ART in 2008–10, on the basis of mortality during the first 3 years of ART, were 67·6 years (95% CI 66·7–68·5) for men and 67·9 years (67·2–68·7) for women, lower than in the French general population (79 years in men and 85 years in women). Life expectancy was lower in North America (expected age at death 65·9 years [65·0–66·8] in men and 63·2 years [62·2–64·3] in women for patients aged 20 years starting ART in 2008–10) than in Europe and was lower than in the US general population (78 years in men and 82 years in women). When estimates of life expectancy were based on mortality during the second and third years of ART, the average ages at death were around 10 years higher. Increases in life expectancy over calendar time were similar when estimated with data for the first 3 years and for the second and third years of follow-up. The expected age at death of a 20-year-old patient starting ART during 2008–10, who had a CD4 count of more than 350 cells per μL 1 year after starting ART, was 78·0 years (77·7–78·3).Figure 3 Expected age at death of men and women living with HIV starting antiretroviral therapy (ART) aged 20 years, by period of initiation

expected age at death of a 20-year-old patient starting ART during 2008–10, who had a CD4 count of more than 350 cells per μL 1 year after starting ART, was 78·0 years (77·7–78·3).Figure 3 Expected age at death of men and women living with HIV starting antiretroviral therapy (ART) aged 20 years, by period of initiation Estimates of life expectancy were based on mortality during the first 3 years of follow-up and the second and third years of follow-up. Data are for all regions. Discussion Between 1996 and 2013, survival of people living with HIV in the first 3 years since ART initiation improved substantially. During the first year of ART, mortality was similar in patients who started ART between 1996 and 2007, but lower during 2008–10. Survival over calendar time improved consistently during the second and third years after initiation of ART. Declines in mortality were lower in people who inject drugs than in other groups. Response to ART, measured by CD4 cell count and viral load 1 year after starting ART, only partly explained improvements in survival during the second and third years of ART. AIDS-related and non-AIDS-related mortality declined over calendar time, during the first year and second–third years after ART initiation. Life expectancy in patients starting ART has increased by about 10 years during the ART era, but remains lower than in the general population. Patients who started ART during 2008–10 whose CD4 counts exceeded 350 cells per μL 1 year after ART initiation have estimated life expectancy approaching that of the general population.

ectancy in patients starting ART has increased by about 10 years during the ART era, but remains lower than in the general population. Patients who started ART during 2008–10 whose CD4 counts exceeded 350 cells per μL 1 year after ART initiation have estimated life expectancy approaching that of the general population. Reduced mortality during the first year of ART is likely to be explained by better initial regimens with greater effectiveness and improved tolerability with fewer side-effects, because rates of regimen modification are highest soon after starting ART.12 Improvements in survival during the second and third years of ART are probably caused by increased viral suppression, declining rates of viral failure, and increasing treatment options.13, 14 Simpler regimens might have contributed to improvements in both short-term and long-term adherence to ART. Drug pharmacokinetics have improved and since 2006, single daily pill formulations with fewer drug interactions have been available.2 Mortality soon after starting ART is strongly influenced by the proportion of patients who start ART with severe immunodeficiency (late presentation), but reductions in the proportion of such patients starting ART do not explain our findings, because we controlled for previous AIDS and for CD4 cell count at ART initiation. Better management of patients with late presentation of HIV infection and more general improvements in health care for people living with HIV could have contributed to improved survival. With the perception that HIV-positive people will live into old age, clinicians are screening for and treating comorbidities more aggressively, including common disorders such as cardiovascular disease, hepatitis C, and cancer. Increasing life expectancy might encourage patients to engage in risk reduction programmes, to cease smoking, and to increase adherence to ART.15 In the USA, improved survival after 2010 could in part be a result of the introduction of the National HIV/AIDS Strategy, which aimed to increase access to care, improve health outcomes, and address health inequities among people living with HIV.

duction programmes, to cease smoking, and to increase adherence to ART.15 In the USA, improved survival after 2010 could in part be a result of the introduction of the National HIV/AIDS Strategy, which aimed to increase access to care, improve health outcomes, and address health inequities among people living with HIV. Rates of AIDS deaths during the first year of ART were substantially lower in the most recent period, which is probably caused by declining rates of more serious AIDS events, such as AIDS-defining malignancies.6 As we reported previously,16 the proportion of deaths from AIDS has declined over time. The substantial decline in cardiovascular mortality could be a result of more aggressive screening and treatment of cardiovascular risk factors, decreasing contraindications for lipid-lowering medications, especially statins (because of drug interactions or poor overall condition), and reduced use of abacavir in individuals who have high viral loads, are HLA B5701 positive or are at high risk of myocardial infarction.17 Moreover, the incidence of cardiovascular disease has decreased in the general population over time, and therapeutic interventions have improved.18

ons or poor overall condition), and reduced use of abacavir in individuals who have high viral loads, are HLA B5701 positive or are at high risk of myocardial infarction.17 Moreover, the incidence of cardiovascular disease has decreased in the general population over time, and therapeutic interventions have improved.18 Between 1996–99 and 2008–10, life expectancy in people living with HIV starting ART increased by around 10 years for both sexes, in Europe and North America. However, the 12-year improvements that we found were less than the 15–24 year increases over similar time periods reported by other studies in the UK and North America.19, 20, 21, 22 Two of these studies examined trends by period of follow-up21, 22 and two19, 20 did not control for duration of ART, which tends to increase estimated life expectancy in earlier periods relative to later periods23 because mortality is higher soon after the start of ART than after successful treatment for a number of years. Estimates of life expectancy in patients who survive the first year of ART are much higher than at ART initiation, reflecting the importance of starting therapy early in the course of HIV infection.20, 23

ds23 because mortality is higher soon after the start of ART than after successful treatment for a number of years. Estimates of life expectancy in patients who survive the first year of ART are much higher than at ART initiation, reflecting the importance of starting therapy early in the course of HIV infection.20, 23 We analysed data for many people living with HIV who were receiving routine clinical care in western Europe, the USA, and Canada. Some cohorts, such as the Danish HIV cohort, FHDH (France), and ATHENA (Netherlands) cover most of their countries. Others are regional but are representative of public care for the areas in which they operate. Therefore, our findings should be generalisable to treated people living with HIV in high-income settings. We compared patients with the same potential years of follow-up between calendar periods and accounted for heterogeneity in death rates between cohorts.24 Our results might be affected by confounding: patient characteristics have changed during the 20 years that ART has been available, with a smaller proportion of infections in people who inject drugs in more recent years and changing patterns of migration from sub-Saharan Africa. Transmission group is sometimes misclassified, and transmission for people who inject drugs does not necessarily imply continuing drug use. Outcomes in patients lost to follow-up within 3 years of starting ART are uncertain, but most cohorts link to death registries. CD4 cell count and viral load 1 year after starting ART were missing for some patients who might have been less engaged in care and likely to have worse prognosis. Mortality was estimated on pooled data and therefore the estimated life expectancy reflects patients' average experience. Estimates of life expectancy are sensitive to mortality in the oldest age groups, for which data are sparse. Our study only includes patients who started ART, whereas most deaths in people with HIV infection occur in the untreated population.

herefore the estimated life expectancy reflects patients' average experience. Estimates of life expectancy are sensitive to mortality in the oldest age groups, for which data are sparse. Our study only includes patients who started ART, whereas most deaths in people with HIV infection occur in the untreated population. Our study tracks the progress made in treating people living with HIV between 1996 and 2013. Monitoring survival can clarify when and how improvements were achieved, and provides a benchmark against which current or future interventions, such as treatment with integrase inhibitors, guidelines recommending earlier treatment, or limiting CD4 cell count and viral load monitoring in stable patients, can be measured. Prognostic information is important to patients, their relatives, and clinicians, and can be used to inform health-care planning at the individual, clinic, and government levels. Improvements in survival with better ART regimens could provide evidence to policy makers that modern palatable and effective treatments should continue to be used in preference to older antiretroviral drugs that are becoming available as cheaper generics. Information about life expectancy in people living with HIV and the knowledge that it could be approaching that of the general population is important to motivate at-risk individuals to test for HIV and convince those infected to start ART immediately, and might decrease stigmatisation of people living with HIV and help them to obtain insurance or employment.

le living with HIV and the knowledge that it could be approaching that of the general population is important to motivate at-risk individuals to test for HIV and convince those infected to start ART immediately, and might decrease stigmatisation of people living with HIV and help them to obtain insurance or employment. Since modern ART is highly effective and has low toxicity, the excess mortality in people living with HIV is unlikely to be addressed by further development of antiretroviral drugs. Instead, lifestyle issues that affect adherence to ART and non-AIDS mortality, and diagnosis and treatment of comorbidities in people living with HIV will need to be addressed. Interventions are needed to promote modern therapy to vulnerable populations, such as people who inject drugs, who currently do not fully benefit from ART. Improved access to opioid substitution treatment programmes and direct acting antiviral drugs for hepatitis C virus co-infection should be a priority for this group.25 Continued efforts are required to address late diagnosis and presentation to care to decrease mortality soon after starting ART, and support lifelong adherence to ART. Treatment guidelines changed in 2015 after results of the START trial showed clear benefits of immediate versus deferred treatment.26 Most future patients diagnosed with HIV are likely to start ART immediately (both for their own health and to prevent transmission to others), but this will only result in improved survival if the problems of late HIV diagnosis and access to care are addressed.

d clear benefits of immediate versus deferred treatment.26 Most future patients diagnosed with HIV are likely to start ART immediately (both for their own health and to prevent transmission to others), but this will only result in improved survival if the problems of late HIV diagnosis and access to care are addressed. Correspondence to: Mr Adam Trickey, School of Social and Community Medicine, University of Bristol, Bristol BS8 2PS, UK adam.trickey@bristol.ac.uk Supplementary Material Supplementary appendix

d clear benefits of immediate versus deferred treatment.26 Most future patients diagnosed with HIV are likely to start ART immediately (both for their own health and to prevent transmission to others), but this will only result in improved survival if the problems of late HIV diagnosis and access to care are addressed. Correspondence to: Mr Adam Trickey, School of Social and Community Medicine, University of Bristol, Bristol BS8 2PS, UK adam.trickey@bristol.ac.uk Supplementary Material Supplementary appendix Acknowledgments We thank all patients, doctors, and study nurses associated with the participating cohort studies. This work was supported by the UK Medical Research Council (MRC; grant number MR/J002380/1) and the UK Department for International Development (DFID) under the MRC/DFID Concordat agreement and is also part of the EDCTP2 programme supported by the European Union. JACS is funded by National Institute for Health Research Senior Investigator award NF-SI-0611-10168. Data from 11 European cohorts were pooled in June, 2014, within COHERE in EuroCoord. COHERE receives funding from the European Union Seventh Framework Programme (FP7/2007-2013) under EuroCoord grant agreement number 260694. Sources of funding of individual cohorts include the Agence Nationale de Recherche sur le SIDA et les hépatites virales (ANRS), the Institut National de la Santé et de la Recherche Médicale (INSERM), the French, Italian, and Spanish Ministries of Health, the Swiss National Science Foundation (grant 33CS30_134277), the Ministry of Science and Innovation and the Spanish Network for AIDS Research (RIS; ISCIII-RETIC RD06/006), the Stichting HIV Monitoring, the European Commission (EuroCoord grant 260694), the British Columbia and Alberta Governments, the National Institutes of Health (NIH), UW Center for AIDS Research (CFAR; NIH grant P30 AI027757), UAB CFAR (NIH grant P30-AI027767), the Vanderbilt-Meharry CFAR (NIH grant P30 AI54999), National Institute on Alcohol Abuse and Alcoholism (U10-AA13566, U24-AA020794), the US Department of Veterans Affairs, the Michael Smith Foundation for Health Research, the Canadian Institutes of Health Research, the VHA Office of Research and Development and unrestricted grants from Abbott, Gilead, Tibotec-Upjohn, ViiV Healthcare, MSD, GlaxoSmithKline, Pfizer, Bristol-Myers Squibb, Roche, and Boehringer Ingelheim. The Danish HIV Cohort Study is funded by the Preben and Anne Simonsens Foundation. Instituto de Salud Carlos III, Ministerio de Economía y Competitividad, Madrid (Spain), provided funding to JMM under a personal intensification research grant (INT15/00168) during 2016.

tol-Myers Squibb, Roche, and Boehringer Ingelheim. The Danish HIV Cohort Study is funded by the Preben and Anne Simonsens Foundation. Instituto de Salud Carlos III, Ministerio de Economía y Competitividad, Madrid (Spain), provided funding to JMM under a personal intensification research grant (INT15/00168) during 2016. Contributors AT did the statistical analyses and wrote the first draft of the paper. The writing committee contributed to study design, data collection, data interpretation, writing the report, and approved the final version. AT had full access to the data and acts as guarantor for the report. MTM and JACS had the original concept for the study and designed and supervised the analyses.

ft of the paper. The writing committee contributed to study design, data collection, data interpretation, writing the report, and approved the final version. AT had full access to the data and acts as guarantor for the report. MTM and JACS had the original concept for the study and designed and supervised the analyses. The Antiretroviral Therapy Cohort Collaboration Writing committee—Adam Trickey (School of Social and Community Medicine, University of Bristol, Bristol, UK); Margaret T May (School of Social and Community Medicine, University of Bristol, Bristol, UK); Jorg-Janne Vehreschild (German Centre for Infection Research, partner site Bonn-Cologne, Cologne, Germany; Department I for Internal Medicine, University Hospital of Cologne, Cologne, Germany); Niels Obel (Department of Infectious Diseases, Copenhagen University Hospital, Copenhagen, Denmark); M John Gill (Division of Infectious Diseases, University of Calgary, Calgary, Canada); Heidi M Crane (Center for AIDS Research, University of Washington, Seattle, WA, USA); Christoph Boesecke (Department of Internal Medicine, University of Bonn, Bonn, Germany); Sophie Patterson (Epidemiology and Population Health Program, British Columbia Centre for Excellence in HIV/AIDS, Vancouver, Canada; and Faculty of Health Sciences, Simon Fraser University, Burnaby, Canada); Sophie Grabar (Sorbonne Universités, UPMC University Paris 06, UMR_S 1136, Institut Pierre Louis d'Epidémiologie et de Santé Publique, Paris, France; INSERM, UMR_S 1136, Institut Pierre Louis d'Epidémiologie et de Santé Publique, Paris, France); Charles Cazanave (Centre Hospitalier Universitaire de Bordeaux, Hôpital Pellegrin, Bordeaux, France); Matthias Cavassini (Service of Infectious Diseases, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland); Leah Shepherd (Research Department of Infection and Population Health, UCL Medical School, London, UK); Antonella d'Arminio Monforte (Clinic of Infectious Diseases and Tropical Medicine, San Paolo Hospital, University of Milan, Italy); Ard van Sighem (Stichting HIV Monitoring, Amsterdam, Netherlands); Michael Saag (Division of Infectious Disease, Department of Medicine, University of Alabama, Birmingham, USA); Fiona Lampe (Research Department of Infection and Population Health, UCL Medical School, London, UK); Vicky Hernando (Red de Investigación en Sida, Centro Nacional de Epidemiología, Instituto de Salud Carlos III, Madrid, Spain); Marta Montero (La Fe Hospital, Valencia, Spain); Robert Zan

y of Alabama, Birmingham, USA); Fiona Lampe (Research Department of Infection and Population Health, UCL Medical School, London, UK); Vicky Hernando (Red de Investigación en Sida, Centro Nacional de Epidemiología, Instituto de Salud Carlos III, Madrid, Spain); Marta Montero (La Fe Hospital, Valencia, Spain); Robert Zan gerle (Innsbruck Medical University, Innsbruck, Austria); Amy C Justice (Yale University School of Medicine, New Haven, CT, USA; VA Connecticut Healthcare System, West Haven, CT, USA); Timothy Sterling (Vanderbilt University School of Medicine, Nashville, TN, USA); Jose M Miro (Hospital Clínic-IDIBAPS, University of Barcelona, Barcelona, Spain); Suzanne M Ingle (School of Social and Community Medicine, University of Bristol, Bristol, UK); Jonathan A C Sterne (School of Social and Community Medicine, University of Bristol, Bristol, UK).

ity School of Medicine, Nashville, TN, USA); Jose M Miro (Hospital Clínic-IDIBAPS, University of Barcelona, Barcelona, Spain); Suzanne M Ingle (School of Social and Community Medicine, University of Bristol, Bristol, UK); Jonathan A C Sterne (School of Social and Community Medicine, University of Bristol, Bristol, UK). Steering group—Andrew Boulle (IeDEA Southern Africa), Christoph Stephan (Frankfurt), Jose M Miro (PISCIS), Matthias Cavassini (SHCS), Geneviève Chêne (Aquitaine), Dominique Costagliola (FHDH), François Dabis (Aquitaine), Antonella D'Arminio Monforte (ICONA), Julia del Amo (CoRIS-MD), Ard Van Sighem (ATHENA), Jorg-Janne Vehreschild (Koln/Bonn), M John Gill (South Alberta Clinic), Jodie Guest (HAVACS), David Hans-Ulrich Haerry (EATG), Robert Hogg (HOMER), Amy C Justice (VACS), Leah Shepherd (EuroSIDA), Niels Obel (Denmark), Heidi M Crane (Washington), Colette Smith (Royal Free), Peter Reiss (ATHENA), Michael Saag (Alabama), Timothy Sterling (Vanderbilt-Meherry), Ramon Teira (VACH), Matthew Williams (UK-CAB), Robert Zangerle (Austria). Coordinating team—Jonathan A C Sterne, Margaret T May (principal investigators); Suzanne M Ingle, Adam Trickey (statisticians).

ton), Colette Smith (Royal Free), Peter Reiss (ATHENA), Michael Saag (Alabama), Timothy Sterling (Vanderbilt-Meherry), Ramon Teira (VACH), Matthew Williams (UK-CAB), Robert Zangerle (Austria). Coordinating team—Jonathan A C Sterne, Margaret T May (principal investigators); Suzanne M Ingle, Adam Trickey (statisticians). Declaration of interests The following members of the writing committee, or their institution, received fees from the following entities for work unrelated to this paper. JMM was provided funding by Instituto de Salud Carlos III, Ministerio de Economia y Competitividad, Madrid (Spain), under a personal intensification research grant (INT15/00168) during 2016. JMM has also received research and academic medical grants, payments for lectures and advisory board participation from Abbvie, Genentech, Gilead, Medtronic, Merck, Novartis, Pfizer, and ViiV Healthcare. CB has received personal fees from Abbvie, Gilead, Merck, and ViiV Healthcare. MM has received personal fees and grants for advisory board participation from Bristol-Myers Squibb, ViiV Healthcare, Merck, Abbvie, Gilead Sciences, and Janssen Cilag. JJV has received research grants, travel grants, non-financial support, and personal fees from Merck, Gilead, Astellas, Pfizer, Basilea, and Infectopharm. MJG has received personal fees as an ad-hoc member of the Canadian HIV advisory boards of Gilead, Merck, and ViiV. MS has received personal fees for consultancy and grants paid to the University of Alabama from Merck, Gilead, ViiV, and Bristol-Myers Squibb. The European Centre for Disease Prevention and Control, ViiV, Gilead, and Janssen Cilag funded Stichting HIV Monitoring for work by AVS. All other members of the writing committee declare no competing interests.

Introduction Cervical cancer is the most common cancer affecting women in low-income and middle-income countries,1 and one of the most common cancers in women living with HIV.2 Women living with HIV have higher prevalence of genital high-risk oncogenic human papillomavirus (HPV) infection than do the general population,3 they are also more likely to have persistent infection4 and progression of cervical intraepithelial neoplasia (CIN) lesions.5 As combined antiretroviral therapy (ART) is scaled up, the effect on cervical cancer due to longer survival is unknown. The interactions of ART and the natural history of high-risk HPV and cervical lesions in women living with HIV are poorly understood. Observational studies differ with respect to study design, outcomes, timing of ART initiation and effectiveness of ART use, making it difficult to estimate the true effect of ART. Previous systematic reviews have explored the association of ART and high-risk HPV and cervical lesions,5, 6, 7 but to our knowledge no meta-analysis has quantified the risk of high-risk HPV infection and cervical lesions among ART users compared with ART-naive women. In view of the large and increasing number of women on ART, improved understanding of the interplay of ART, immune recovery, and virological control on the natural history of high-risk HPV infection and CIN progression is needed to guide screening programmes. Research in context Evidence before this study

The interactions of ART and the natural history of high-risk HPV and cervical lesions in women living with HIV are poorly understood. Observational studies differ with respect to study design, outcomes, timing of ART initiation and effectiveness of ART use, making it difficult to estimate the true effect of ART. Previous systematic reviews have explored the association of ART and high-risk HPV and cervical lesions,5, 6, 7 but to our knowledge no meta-analysis has quantified the risk of high-risk HPV infection and cervical lesions among ART users compared with ART-naive women. In view of the large and increasing number of women on ART, improved understanding of the interplay of ART, immune recovery, and virological control on the natural history of high-risk HPV infection and CIN progression is needed to guide screening programmes. Research in context Evidence before this study Women living with HIV have higher prevalence of genital high-risk oncogenic human papillomavirus (HPV) infection than the general population and are more likely to have persistent infection and progression of cervical intraepithelial neoplasia (CIN) lesions. Increased access to antiretroviral therapy (ART) has increased the life expectancy of women living with HIV, but many remain susceptible to high-risk HPV incidence and persistence and cervical lesion incidence and progression. The precise effect of ART on the natural history of high-risk HPV infection and cervical lesion progression is not well established, and studies evaluating this association have reported conflicting results. We searched all available publications in English in the MEDLINE and Embase databases from Jan 1, 1996, to May 6, 2017, which reported the association of ART with prevalence of high-risk HPV or prevalence, incidence, progression, or regression of histological (CIN) or cytological (squamous intraepithelial lesions [SIL]) cervical abnormalities, or incidence of invasive cervical cancer. We found 31 studies of the association of ART with prevalence of high-risk HPV (6537 women living with HIV), and CIN of grade 2 or higher (CIN2+) diagnosed by histology or high-grade SIL (HSIL+) diagnosed by cytology only (9288 women living with HIV). Furthermore, 17 studies reported the association of ART with longitudinal cervical lesion outcomes (any CIN or SIL), providing data for 6864 women living with HIV, and three studies reported the association of ART with incidence of invasive cervical cancer among 15 826 women living with HIV.

288 women living with HIV). Furthermore, 17 studies reported the association of ART with longitudinal cervical lesion outcomes (any CIN or SIL), providing data for 6864 women living with HIV, and three studies reported the association of ART with incidence of invasive cervical cancer among 15 826 women living with HIV. Added value of this study

288 women living with HIV). Furthermore, 17 studies reported the association of ART with longitudinal cervical lesion outcomes (any CIN or SIL), providing data for 6864 women living with HIV, and three studies reported the association of ART with incidence of invasive cervical cancer among 15 826 women living with HIV. Added value of this study We found that prevalence of high-risk HPV and histology diagnosed HSIL-CIN2+ was lower among ART users compared with those not on treatment. ART was associated with a decreased risk of histology diagnosed HSIL-CIN2+ incidence, cytology diagnosed SIL incidence, and SIL progression. Women living with HIV on ART had an increased likelihood of histology diagnosed CIN or cytology diagnosed SIL regression and a decreased risk of invasive cervical cancer incidence. To our knowledge, this is the first study to quantify the effect of ART on prevalent high-risk HPV, high-grade cervical lesion outcomes, and invasive cervical cancer in a meta-analysis. Studies that adjusted for either nadir or current CD4 cell count and time-varying effects of ART were more likely to show a protective effect of ART on these outcomes. Studies from Africa and Europe or North America provide indication that ART was associated with lower prevalence of high-risk HPV and cervical lesions, and over prolonged duration, ART can prevent cervical lesion incidence and progression, promote regression, and prevent incidence of invasive cervical cancer. Fewer studies exist from Asia and Latin America with the majority being cross-sectional in design, and these studies were less likely to report any protective association of ART. Because some studies from Latin America have reported an increased risk of high-risk HPV and CIN2+ among women with a lower nadir CD4 cell count, the lack of association might reflect the timing of ART in relation to HPV infection and cervical lesion development in these populations. Our findings highlight the importance of early ART initiation (before reaching a low nadir CD4 cell count) and sustained effectiveness, as evidenced by duration, high adherence, virological control, and CD4 cell recovery, in controlling HPV infection and cervical disease progression.

n development in these populations. Our findings highlight the importance of early ART initiation (before reaching a low nadir CD4 cell count) and sustained effectiveness, as evidenced by duration, high adherence, virological control, and CD4 cell recovery, in controlling HPV infection and cervical disease progression. Implications of all the available evidence The current recommendation of encouraging earlier ART initiation, coupled with rapid virological control, and sustained adherence is likely to lead to an earlier and possibly more functionally complete mucosal immune reconstitution. ART users with low or unknown nadir CD4 cell count should be screened frequently because their risk of high-risk HPV infection and cervical lesion progression remains high. Longitudinal studies in the era of immediate unconditional ART initiation should capture the greater benefit of ART treatment on cervical disease and cancer. We aimed to review and to summarise the literature about the association of ART with high-risk HPV prevalence, and with cervical lesion prevalence, incidence, progression and regression, and invasive cervical cancer incidence. We also aimed to investigate the role of HIV-related cofactors that might modify these associations, such as ART duration, timing of treatment initiation, immune suppression, and recovery.

HPV prevalence, and with cervical lesion prevalence, incidence, progression and regression, and invasive cervical cancer incidence. We also aimed to investigate the role of HIV-related cofactors that might modify these associations, such as ART duration, timing of treatment initiation, immune suppression, and recovery. Methods Search strategy and selection criteria We searched MEDLINE and Embase databases for publications in English with search terms for human papillomavirus, CIN, SIL, invasive cervical cancer, and ART (appendix p 1). Reference lists of review articles and all articles identified in the systematic search were checked. We did the search from Jan 1, 1996 (when highly active ART came into use), up to May 6, 2017. One author (HK) screened all abstracts. Two authors (HK and PM) obtained full-text copies of relevant publications, assessed them for eligibility, and reached consensus on potential relevance. Studies were eligible if they reported the association of combination ART or highly active ART use (referred to as ART from now on) with the following outcomes: prevalence of high-risk HPV; prevalence, incidence, progression, or regression of SIL diagnosed with cytology or CIN diagnosed with histology; and incidence of invasive cervical cancer among women living with HIV. We also considered studies eligible if they provided raw data to calculate an unadjusted effect estimate.

prevalence of high-risk HPV; prevalence, incidence, progression, or regression of SIL diagnosed with cytology or CIN diagnosed with histology; and incidence of invasive cervical cancer among women living with HIV. We also considered studies eligible if they provided raw data to calculate an unadjusted effect estimate. For high-risk HPV outcomes, we included studies reporting genital high-risk HPV. There were no exclusions on HPV test methods. For the prevalent lesion outcomes, studies reporting cervical lesions using visual inspection with acetic acid or Lugol's iodine but without high-resolution colposcopy were excluded because of the poor sensitivity and specificity of visual inspection alone in detecting high-grade lesions. For prevalent outcomes, cross-sectional studies were included if they reported the association of ART use with high-risk HPV or any grade of histological or cytological cervical lesion. Cohort studies were included if participants initiated ART at enrolment, were followed up, and had measures of high-risk HPV at baseline and in the follow-up visit.

sectional studies were included if they reported the association of ART use with high-risk HPV or any grade of histological or cytological cervical lesion. Cohort studies were included if participants initiated ART at enrolment, were followed up, and had measures of high-risk HPV at baseline and in the follow-up visit. For the longitudinal outcomes, we included cohort studies reporting the association of ART with the incidence, progression, and regression of any CIN grade diagnosed by histology or any SIL grade diagnosed by cytology (which could include atypical squamous cells of undetermined significance as well as low-grade and high-grade lesions) because SIL represent various incremental degrees of high-risk HPV persistence and subsequent lesion development. Only cohort studies examining invasive cervical cancer incidence among ART users and treatment-naive women in the ART era were included because they provide the most robust direct comparison of the effect of therapy on invasive cervical cancer. For publications that reported results from the same cohort, but at different follow-up visits, the publication that gave the most relevant description of the cohort and study design and the most complete set of results was included. There was no restriction on age or geographical location.

For the longitudinal outcomes, we included cohort studies reporting the association of ART with the incidence, progression, and regression of any CIN grade diagnosed by histology or any SIL grade diagnosed by cytology (which could include atypical squamous cells of undetermined significance as well as low-grade and high-grade lesions) because SIL represent various incremental degrees of high-risk HPV persistence and subsequent lesion development. Only cohort studies examining invasive cervical cancer incidence among ART users and treatment-naive women in the ART era were included because they provide the most robust direct comparison of the effect of therapy on invasive cervical cancer. For publications that reported results from the same cohort, but at different follow-up visits, the publication that gave the most relevant description of the cohort and study design and the most complete set of results was included. There was no restriction on age or geographical location. Data extraction From the consensus list, one author (HK) extracted the data and a second author (HAW) checked a random sample of 25%. For studies reporting prevalence of high-risk HPV or cervical lesions, odds ratios (ORs) were extracted. For studies reporting cervical lesion incidence, progression or regression, hazard ratios (HRs) or ORs were extracted.

st, one author (HK) extracted the data and a second author (HAW) checked a random sample of 25%. For studies reporting prevalence of high-risk HPV or cervical lesions, odds ratios (ORs) were extracted. For studies reporting cervical lesion incidence, progression or regression, hazard ratios (HRs) or ORs were extracted. Methodological quality assessment We assessed studies primarily on adjustment for HIV-related factors (current and nadir CD4 cell count and ART duration). We considered cross-sectional studies that adjusted for either current or nadir CD4 cell count or ART duration separately in sensitivity analyses, as were cohort studies that adjusted for time on ART during follow-up. We also assessed study quality by participant selection, statistical method, HPV test used, and cervical lesion (cytological or histological) classification (appendix pp 6–18). Statistical analysis We did meta-analyses for the discrete outcomes of high-risk HPV prevalence, high-grade lesion (high-grade squamous intraepithelial lesion or cervical intraepithelial neoplasia grade 2 or higher, diagnosed by cytology or histology [HSIL-CIN2+]) prevalence, incidence, progression and regression of any histology diagnosed CIN or cytology diagnosed SIL, and incidence of invasive cervical cancer.

, high-grade lesion (high-grade squamous intraepithelial lesion or cervical intraepithelial neoplasia grade 2 or higher, diagnosed by cytology or histology [HSIL-CIN2+]) prevalence, incidence, progression and regression of any histology diagnosed CIN or cytology diagnosed SIL, and incidence of invasive cervical cancer. We report adjusted effect estimates when available. For the cross-sectional studies in which adjusted effect estimates were not reported but raw data were provided, we calculated crude ORs (HK) and independently verified them (HAW and PM). We contacted authors when the paper suggested that relevant data were collected but not reported. We used random-effects meta-analysis to estimate pooled effects to account for between-study heterogeneity.8 We examined heterogeneity using the I2 statistic and publication bias using funnel plots and Begg's test for correlation between the effect estimate and their variances.9, 10 We did an influence analysis to assess the robustness of the pooled summary effects by excluding each of the studies from the pooled estimate. We did subgroup analyses by geographical region to compare pooled effects and heterogeneity. We did sensitivity analyses excluding studies unadjusted for HIV-related factors. We analysed data using Stata version 14. This review is reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)11 and the Meta-analysis Of Observational Studies in Epidemiology (MOOSE) guidelines.12 The review protocol and the dataset are available online.

We used random-effects meta-analysis to estimate pooled effects to account for between-study heterogeneity.8 We examined heterogeneity using the I2 statistic and publication bias using funnel plots and Begg's test for correlation between the effect estimate and their variances.9, 10 We did an influence analysis to assess the robustness of the pooled summary effects by excluding each of the studies from the pooled estimate. We did subgroup analyses by geographical region to compare pooled effects and heterogeneity. We did sensitivity analyses excluding studies unadjusted for HIV-related factors. We analysed data using Stata version 14. This review is reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)11 and the Meta-analysis Of Observational Studies in Epidemiology (MOOSE) guidelines.12 The review protocol and the dataset are available online. Role of the funding source There was no funding source for this study. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

This review is reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)11 and the Meta-analysis Of Observational Studies in Epidemiology (MOOSE) guidelines.12 The review protocol and the dataset are available online. Role of the funding source There was no funding source for this study. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results We identified 605 publications for the association of ART and high-risk HPV prevalence through MEDLINE and Embase searches, 198 of which were duplicates and removed; and we excluded 343 after abstract review, leaving 64 articles for full-text review. Finally, 16 articles matched inclusion criteria and we identified three additional publications through cross-referencing (figure 1). Data were extracted from 19 publications (12 cross-sectional; seven cohort) representing 20 discrete populations and providing data from 6537 women living with HIV, of whom 3677 (56%) were taking ART (range 19–85% in cross-sectional studies), 2032 (31%) were ART-naive, and 828 (13%) were ART initiators. Four studies13, 14, 15, 16 compared high-risk HPV before and after ART initiation (ie, women acted as their own controls; table 1; appendix p 2). One publication provided data from two countries,17 and was considered as two individual studies in the analysis, resulting in 20 included studies.Figure 1 Study selection for outcomes of high-risk HPV (A) and cervical lesions (B)

and after ART initiation (ie, women acted as their own controls; table 1; appendix p 2). One publication provided data from two countries,17 and was considered as two individual studies in the analysis, resulting in 20 included studies.Figure 1 Study selection for outcomes of high-risk HPV (A) and cervical lesions (B) HPV=human papillomavirus. HAART=highly active antiretroviral therapy. CIN=cervical intraepithelial neoplasia. SIL=squamous intraepithelial lesions. HSIL=high-grade SIL. LSIL=low-grade SIL. ASCUS= atypical squamous cells of undetermined significance. *Some studies contributed to more than one outcome (ie, incidence and progression, or progression and regression). Individual studies are summarised in table 1. Table 1 Summary of studies reporting the association of ART use with high-risk HPV, cervical lesion outcomes and invasive cervical cancer incidence

HPV=human papillomavirus. HAART=highly active antiretroviral therapy. CIN=cervical intraepithelial neoplasia. SIL=squamous intraepithelial lesions. HSIL=high-grade SIL. LSIL=low-grade SIL. ASCUS= atypical squamous cells of undetermined significance. *Some studies contributed to more than one outcome (ie, incidence and progression, or progression and regression). Individual studies are summarised in table 1. Table 1 Summary of studies reporting the association of ART use with high-risk HPV, cervical lesion outcomes and invasive cervical cancer incidence Location Study period Total sample Mean or median age (IQR), years ART users (%) Cervical lesions Definition Diagnostic method High risk HPV prevalence Zeier et al (2015)13 Western Cape, South Africa 2009–11 300 36 (ART); 31 (ART-naive) 68% initiated during follow-up* ·· ·· Rositch et al (2013)14 Rakai, Uganda 2007–10 96 35 (31–44) 0%* ·· ·· Minkoff et al (2010)15 5 cities, USA 1994–2002 286 NR 0%* ·· ·· Fife et al (2009)16 Puerto Rico/USA 2001–05 146 35 0%* ·· ·· Kelly et al (2017)17 Ouagadougou, Burkina Faso 2011–12 570 36 (31–41) 67% ·· ·· Kelly et al (2017)17 Johannesburg, South Africa 2011–12 613 34 (30–40) 65% ·· ·· Ezechi et al (2014)18 Ogun and Lagos, Nigeria NR 220 37 (31–45) 72% ·· ·· Reddy et al (2014)19 Lilongwe, Malawi 2011–12 294 36 (30–43) 85% ·· ·· De Vuyst et al (2012)20 Nairobi, Kenya 2009 497 38 75% ·· ·· Jaquet et al (2012)21 Abidjan, Côte d'Ivoire Jun to Oct, 2010 254 36 (32–42) 75% ·· ·· Veldhuijzen et al (2011)22 Kigali, Rwanda 2006–09 124 27 (23–32) 40% ·· ·· Menezes et al (2016)23 Chennai, India July to Aug, 2011 50 33 48% ·· ·· Zhang et al (2014)24 Yunnan, China NR 301 34 64% ·· ·· Mane et al (2012)25 Pune, India NR 277 33 56% ·· ·· Aggarwal et al (2012)26 Chandigarh, India NR 130 34 75% ·· ·· Rocha-Brischiliari et al (2014)27 Maringa city, Brazil Apr to Oct, 2011 178 Range: 18–66 years 79% ·· ·· Dames et al (2014)28 Nassau, Bahamas Feb to Sep, 2008 165 40 81% ·· ·· Grinsztejn et al (2009)29 Rio de Janeiro, Brazil 1996–2006 634 36 (29–43) 68% ·· ·· Konopnicki et al (2013)30 Brussels, Belgium 2002–11 652 38 (31–45) 79% ·· ·· Blitz et al (2013)31 11 cities, Canada 1993–2002 750 33 (28–38) 19% ·· ·· HSIL-CIN2+ prevalence Kelly et al (2017)17 Ouagadougou, Burkina Faso 2011–12 530 36 (31–41) 73% HSIL-CIN2+ Histology Kelly et al (2017)17 Johannesburg, South Africa 2011–12 566 34 (30–40) 65% HSIL-CIN2+ Histology De Vuyst et al (2012)20 Nairobi, Kenya 2009 470 38 75% HSIL-CIN2+ Histology Memiah et al (2015)32 Kiambu, Kenya 2009–10 686 52% <40 years 16% HSIL-CIN2+ Histology Huchko et al (2014)33 Kisumu, Kenya 2007–10 3185 33 (29–39) 50% HSIL-CIN2+ Histology Mabeya et al (2012)34 Eldoret, Kenya NR 149 34 67% HSIL-CIN2+ Histology Ezechi et al (201

012)20 Nairobi, Kenya 2009 470 38 75% HSIL-CIN2+ Histology Memiah et al (2015)32 Kiambu, Kenya 2009–10 686 52% <40 years 16% HSIL-CIN2+ Histology Huchko et al (2014)33 Kisumu, Kenya 2007–10 3185 33 (29–39) 50% HSIL-CIN2+ Histology Mabeya et al (2012)34 Eldoret, Kenya NR 149 34 67% HSIL-CIN2+ Histology Ezechi et al (201 4)35 Ogun and Lagos, Nigeria NR 490 37 (31–45) 76% HSIL-CIN2+ Cytology Firnhaber et al (2010)36 Johannesburg, South Africa NR 1010 34 (18–65) 65% HSIL-CIN2+ Cytology Mogtomo et al (2009)37 Douala, Cameroon NR 70 35 50% HSIL-CIN2+ Cytology Feng et al (2017)† Yunnan, China 2009 301 34 64% HSIL-CIN2+ Histology Sahasrabuddhe et al (2010)38 Pune, India 2006–07 271 30 (27–34) 26% HSIL-CIN2+ Histology De Andrade et al (2011)39 Rio de Janeiro, Brazil 1996–2007 340 34 (28–41) 26% HSIL-CIN2+ Histology Patrelli et al (2013)40 Parma, Italy 1993–2010 194 41 66% HSIL-CIN2+ Cytology Kitchener et al (2007)41 6 cities, Europe 2000–04 1026 33 56–79% HSIL-CIN2+ Cytology SIL-CIN incidence Minkoff et al (2010)15 5 cities, USA 1994–2002 286 NR All ART initiators Normal to ASCUS+ Cytology Kelly et al (2017)17 Johannesburg, South Africa 2011–12 379 34 (30–40) 71% at end of follow-up <CIN2 to CIN2/3 Histology Adler et al (2012)54 Soweto, South Africa 2003–10 767 33 2% at baseline; 17% initiation during follow-up Normal to ASCUS Cytology Firnhaber et al (2012)55 Johannesburg, South Africa NR 326 35 (31–41) 71% at baseline Normal to ASCUS+ Cytology Kreitchmann et al (2013)56 Porto Alegre, Brazil 1997–2007 349 32 38% <LSIL to LSIL+, Cytology Sirera et al (2008)57 Barcelona, Spain 1997–2006 127 35 71% at baseline Normal to LSIL+ Cytology Soncini et al (2007)58 Parma, Italy 1993–2003 101 NR 43% through follow-up Normal to LSIL+ Cytology Lehtovirta et al (2006)59 Helsinki, Finland 1989–2003 55 30–36 48% at baseline; 64% at follow-up Normal to LSIL+ Cytology Heard et al (2006)60 Paris, France 1993–2005 298 33 (29–38) 49% through follow-up Normal to ASCUS+ Cytology Schuman et al (2003)61 4 cities, USA 1993–95 629 35 33% at baseline Normal to LSIL+ Cytology Ellerbrock et al (2000)62 New York, USA 1991–96 328 47% <35 years 54% on ≥1 ARV during study period Normal to ASCUS+ Cytology Clifford et al (2016)63 5 cities, Switzerland 1995–2013 1451 NR 54% <CIN2 to CIN2/3 Histology SIL progression Blitz et al (2013)31 11 cities, Canada 1993–2002 326 33 (28–38) 19% at baseline; 64% by study end ASCUS to any grade higher Cytology Adler et al (2012)54 Soweto, South Africa 2003–10 1123 33 2% at baseline; 17%

Clifford et al (2016)63 5 cities, Switzerland 1995–2013 1451 NR 54% <CIN2 to CIN2/3 Histology SIL progression Blitz et al (2013)31 11 cities, Canada 1993–2002 326 33 (28–38) 19% at baseline; 64% by study end ASCUS to any grade higher Cytology Adler et al (2012)54 Soweto, South Africa 2003–10 1123 33 2% at baseline; 17% initiation during follow-up Subsequent smear with worsening dysplasia Cytology Firnhaber et al (2012)55 Johannesburg, South Africa NR 326 35 (31–41) 71% at baseline Normal to LSIL+; LSIL to HSIL+ Cytology Schuman et al (2003)61 4 cities, USA 1993–95 629 35 33% at baseline Normal/ASCUS to LSIL+; LSIL to HSIL Cytology Zeier et al (2012)64 Western Cape, South Africa 2004–09 1048 33 18% LSIL to HSIL+ Cytology Omar et al (2011)65 Soweto, South Africa 2003–10 1074 32 (28–37) 6% at baseline; 20% initiated during follow-up Normal to LSIL+; LSIL to HSIL+/ASC-H Cytology Kim et al (2013)66 New York, USA 1991–2011 245 37 NR Normal to ASCUS+; ASCUS to LSIL+ Cytology Paramsothy et al (2009)67 4 cities, USA 1996–2000 537 34 47% during follow-up Normal to ASCUS; ASCUS to LSIL; LSIL to HSIL Cytology Minkoff et al (2001)68 6 cities, USA 1994–95 741 37 1% at baseline Subsequent smear any grade higher than baseline Cytology Lillo et al (2001)69 Milan, Italy 1995–97 163 34 46% through follow-up Normal to LSIL+; LSIL to HSIL Cytology SIL or CIN regression Minkoff et al (2010)15 5 cities, USA 1994–2002 286 NR All ART initiators SIL to lower grade Cytology Blitz et al (2013)31 11 cities, Canada 1993–2002 326 33 (28–38) 19% at baseline; 64% by study end ≥ASCUS to <ASCUS Cytology Adler et al (2012)54 Soweto, South Africa 2003–10 1123 33 2% at baseline; 17% initiation during follow-up Subsequent improvement in cytological results Cytology Schuman et al (2003)61 4 cities, USA 1993–95 629 35 33% at baseline LSIL or HSIL to <LSIL Cytology Zeier et al (2012)64 Western Cape, South Africa 2004–09 1048 33 18% ≥LSIL to <LSIL Cytology Paramsothy et al (2009)67 4 cities, USA 1996–2000 537 34 47% during follow-up HSIL to LSIL; LSIL to ASCUS; ASCUS to normal Cytology Minkoff et al (2001)68 6 cities, USA 1994–95 741 37 1% at baseline Lower grade abnormality than baseline Cytology Massad et al (2004)70 6 cities, USA 1994–2002 202 38 22% CIN1 to normal Histology Heard et al (2002)71 Paris, France 1993–99 168 33 56% through follow-up Reversion to normal or from high to low grade Cytology Del Mistro et al (2004)72 Vicenza and Padova, Italy 1994–2002 201 33 37% Normal or lower SIL grade at subsequent exam

et al (2004)70 6 cities, USA 1994–2002 202 38 22% CIN1 to normal Histology Heard et al (2002)71 Paris, France 1993–99 168 33 56% through follow-up Reversion to normal or from high to low grade Cytology Del Mistro et al (2004)72 Vicenza and Padova, Italy 1994–2002 201 33 37% Normal or lower SIL grade at subsequent exam Cytology Invasive cervical cancer incidence Clifford et al (2016)63 5 cities, Switzerland 1995–2013 80 NR 54% <CIN2 to ICC Unclear Chen et al (2014)73 Taiwan 2000–08 1360 32 28% Incidence of CIS or ICC Unclear Guiguet et al (2009)74 62 French university hospitals, France 1998–2006 14 406 39 (35–44) 17% Incidence of ICC ICD10 SIL diagnosed by cytology or CIN diagnosed by histology. Detailed description of studies in appendix (pp 2–5). HPV=human papillomavirus. HSIL=high-grade squamous intraepithelial lesion. CIN=cervical intraepithelial neoplasia. ASCUS=atypical squamous cells of undetermined significance. LSIL=low-grade squamous intraepithelial lesion. ARV=antiretroviral. ART=antiretroviral therapy. ASC-H=atypical squamous cells-cannot exclude HSIL. CIS=carcinoma in situ. NR=not reported. ICD10=International Classification of Diseases version 10. ICC=invasive cervical cancer. * Studies that included women who initiated ART at enrolment. † Personal communication.

Detailed description of studies in appendix (pp 2–5). HPV=human papillomavirus. HSIL=high-grade squamous intraepithelial lesion. CIN=cervical intraepithelial neoplasia. ASCUS=atypical squamous cells of undetermined significance. LSIL=low-grade squamous intraepithelial lesion. ARV=antiretroviral. ART=antiretroviral therapy. ASC-H=atypical squamous cells-cannot exclude HSIL. CIS=carcinoma in situ. NR=not reported. ICD10=International Classification of Diseases version 10. ICC=invasive cervical cancer. * Studies that included women who initiated ART at enrolment. † Personal communication. The pooled OR among 20 studies13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31 indicates that women living with HIV on ART had a lower risk of high-risk HPV prevalence compared with women who were ART-naive (crude OR 0·82, 95% CI 0·68–0·98); but there was a high degree of heterogeneity between studies (I2=71%, p value for heterogeneity<0·0001; table 2, figure 2). Restricting the analysis to the 12 studies that adjusted for either current or nadir CD4 cell count, or ART duration,13, 15, 16, 17, 19, 20, 21, 24, 25, 29, 30 the OR was similar but with a moderate degree of heterogeneity (adjusted [a] OR 0·85, 95% CI 0·73–1·00, adjusted for nadir or current CD4 cell count; aOR 0·83, 95% CI 0·70–0·99, I2=51%, p value for heterogeneity=0·02, with additional adjustment for duration on ART). The reduction in heterogeneity on adjustment for confounding was most noticeable among the studies from Africa; among six studies13, 17, 19, 20, 21 the aOR was 0·70 (95% CI 0·56–0·88) with no evidence of heterogeneity (I2=0·0%, p=0·97). Similarly, among studies from Europe or North America, three studies15, 16, 30] showed a similar reduction in high-risk HPV (aOR 0·74, 95% CI 0·59–0·93; I2=48%, p=0·14). This was by contrast with the two studies from Asia24, 25 (1·72, 1·10–2·68; I2=0%, p=0·34) and three from Latin America27, 28, 29 (crude OR 1·08, 95% CI 0·84–1·39; I2=0%, p =0·99).Figure 2 Meta-analysis of the prevalence of high-risk HPV and HSIL-CIN2+ among ART users compared with ART-naive

0·59–0·93; I2=48%, p=0·14). This was by contrast with the two studies from Asia24, 25 (1·72, 1·10–2·68; I2=0%, p=0·34) and three from Latin America27, 28, 29 (crude OR 1·08, 95% CI 0·84–1·39; I2=0%, p =0·99).Figure 2 Meta-analysis of the prevalence of high-risk HPV and HSIL-CIN2+ among ART users compared with ART-naive Weights are from random-effects analysis. HPV=human papillomavirus. HSIL=high-grade squamous intraepithelial lesion. CIN2+=cervical intraepithelial lesion, grade 2 or higher. ART=antiretroviral therapy. NR=not reported. *Studies that adjusted for any of ART duration, current or nadir CD4 cell count. †Personal communication. ‡Includes France, Ireland, Italy, Poland, and the UK ((authors report rate ratio of cytology-diagnosed HSIL+ among ART users over follow-up as opposed to odds ratio). Table 2 Meta-analysis of the association of ART with the prevalence of high-risk HPV and HSIL-CIN2+ among women living with HIV

Weights are from random-effects analysis. HPV=human papillomavirus. HSIL=high-grade squamous intraepithelial lesion. CIN2+=cervical intraepithelial lesion, grade 2 or higher. ART=antiretroviral therapy. NR=not reported. *Studies that adjusted for any of ART duration, current or nadir CD4 cell count. †Personal communication. ‡Includes France, Ireland, Italy, Poland, and the UK ((authors report rate ratio of cytology-diagnosed HSIL+ among ART users over follow-up as opposed to odds ratio). Table 2 Meta-analysis of the association of ART with the prevalence of high-risk HPV and HSIL-CIN2+ among women living with HIV Crude analysis* Adjusted analysis† n studies OR (95%CI) I2 p value for heterogeneity n studies OR (95%CI) I2 p value for heterogeneity High-risk HPV prevalence All 20 0·82 (0·68–0·98) 71·0% <0·0001 12 0·83 (0·70–0·99) 51·0% 0·02 Africa 9 0·67 (0·52–0·88) 58·8% 0·01 6 0·70 (0·56–0·88) 0% 0·97 Asia 4 1·60 (0·93–2·75) 38·6% 0·18 2 1·72 (1·10–2·68) 0% 0·34 Latin America 3 1·08 (0·84–1·39) 0% 0·99 ·· ·· ·· ·· Europe or North America 4 0·75 (0·63–0·88) 29·9% 0·23 3 0·74 (0·59–0·93) 48·4% 0·14 HSIL-CIN2+ prevalence All 14 0·92 (0·70–1·20) 56·6% 0·01 4 0·65 (0·40–1·06) 29·5% 0·25 Africa 9 0·84 (0·64–1·10) 45·5% 0·07 3 0·70 (0·48–1·01) 0% 0·40 Asia 2 0·66 (0·05–9·37) 83·7% 0·01 ·· ·· ·· ·· Latin America 1 2·31 (1·02–5·23) ·· ·· ·· ·· ·· ·· Europe or North America 2 0·83 (0·43–1·57) 32·2% 0·23 ·· ·· ·· ·· HPV=human papillomavirus. OR=odds ratio. HSIL-CIN2+=high-grade squamous intraepithelial lesions or cervical intraepithelial neoplasia, grade 2 or higher. ART=antiretroviral therapy.

3·7% 0·01 ·· ·· ·· ·· Latin America 1 2·31 (1·02–5·23) ·· ·· ·· ·· ·· ·· Europe or North America 2 0·83 (0·43–1·57) 32·2% 0·23 ·· ·· ·· ·· HPV=human papillomavirus. OR=odds ratio. HSIL-CIN2+=high-grade squamous intraepithelial lesions or cervical intraepithelial neoplasia, grade 2 or higher. ART=antiretroviral therapy. * Includes studies with no adjustment and studies that adjust for sociodemographic factors only but no adjustment for HIV-related factors. † Adjusted for at least one of the following: current CD4 cell count, nadir CD4 cell count, and ART duration. The pooled estimate from four cohort studies that followed women before and after ART initiation13, 14, 15, 16 provides strong evidence of a reduced prevalence of high-risk HPV after ART compared with before ART initiation (crude OR 0·80, 95% CI 0·72–0·89; aOR 0·79, 95% CI 0·71–0·88; I2=48%, p=0·15; data not shown). Nine studies reported the association of ART duration with high-risk HPV prevalence.17, 19, 20, 21, 23, 24, 28, 30 Although high-risk HPV prevalence was similar among the ART-naive and short-duration users (<2 years), the pooled OR suggests that prevalence of high-risk HPV was lower among prolonged ART users (≥2 years) than in short-duration users and ART-naive combined (crude OR 0·65, 95% CI 0·55–0·77; I2=0%, p=0·92; appendix p 19). Among the seven studies adjusted for current and nadir CD4 cell count,17, 19, 20, 21, 24, 30 the association was similar (aOR 0·65, 95% CI 0·55–0·78; I2=0%, p=0·91, data not shown).

ART users (≥2 years) than in short-duration users and ART-naive combined (crude OR 0·65, 95% CI 0·55–0·77; I2=0%, p=0·92; appendix p 19). Among the seven studies adjusted for current and nadir CD4 cell count,17, 19, 20, 21, 24, 30 the association was similar (aOR 0·65, 95% CI 0·55–0·78; I2=0%, p=0·91, data not shown). There was no evidence to suggest publication bias (ie, smaller studies were not more likely to report a positive association; Beggs rank correlation test p=0·12 for the crude analysis, p=0·34 for adjusted analysis). We identified 1158 publications for the association of ART and any cervical lesion outcome, of which 127 duplicates were removed and 889 excluded after abstract review, leaving 142 articles for full review. Finally, we identified 38 articles that matched the inclusion criteria and ten additional publications through cross-referencing (figure 1). Data from an ongoing but unpublished study on association of ART with the prevalence of histology diagnosed HSIL-CIN2+ (Feng et al, 2017) was also included (data provided by Y-L Qiao, personal communication, appendix p 3).

ched the inclusion criteria and ten additional publications through cross-referencing (figure 1). Data from an ongoing but unpublished study on association of ART with the prevalence of histology diagnosed HSIL-CIN2+ (Feng et al, 2017) was also included (data provided by Y-L Qiao, personal communication, appendix p 3). 13 studies17, 20, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41 reported the association of ART with the prevalence of cytology or histology diagnosed HSIL-CIN2+ among 9288 women living with HIV, of whom 5161 (56%) were taking ART (range across studies 16% to 79%) and 4127 (44%) were ART-naive (table 1). One publication provided data from two countries,17 and was considered as two individual studies in the analysis. 12 further studies reported the association of ART with the prevalence of combined cytology diagnosed outcomes of atypical squamous cells of undetermined significance (or higher),42, 43, 44, 45, 46, 47, 48 and low-grade SIL (or higher),49 histology diagnosed CIN (grade 1 or higher),50, 51 and abnormalities on visual inspection with colposcopy (appendix p 20).52, 53

ART with the prevalence of combined cytology diagnosed outcomes of atypical squamous cells of undetermined significance (or higher),42, 43, 44, 45, 46, 47, 48 and low-grade SIL (or higher),49 histology diagnosed CIN (grade 1 or higher),50, 51 and abnormalities on visual inspection with colposcopy (appendix p 20).52, 53 Ten studies reported the association of ART with cytology diagnosed SIL incidence,15, 54, 55, 56, 57, 58, 59, 60, 61, 62 and two studies with histology diagnosed HSIL-CIN2+ incidence17, 63 from a combined total of 5096 women (table 1). We included ten studies31, 54, 55, 61, 64, 65, 66, 67, 68, 69 for cytology diagnosed SIL progression from a combined total of 6212 women, and ten studies15, 31, 54, 61, 64, 67, 68, 70, 71, 72 for regression of histology diagnosed CIN or cytology diagnosed SIL from a combined total of 5261 women (table 1). Only one study reported the regression from histological CIN grade 1 to normal.70 Three studies63, 73, 74 reported the association of ART with invasive cervical cancer incidence among 15 846 women. Studies reporting the association of ART with cervical lesion incidence, progression and regression, and invasive cervical cancer incidence are summarised in figure 3.Figure 3 Meta-analysis of cervical lesion incidence, progression and regression, and invasive cervical cancer incidence among ART users compared with ART-naive

s reporting the association of ART with cervical lesion incidence, progression and regression, and invasive cervical cancer incidence are summarised in figure 3.Figure 3 Meta-analysis of cervical lesion incidence, progression and regression, and invasive cervical cancer incidence among ART users compared with ART-naive Weights are from random effects analysis. Only studies that reported HR from time-to-event analysis included in the meta-analysis (table 3). HR=hazard ratio. OR=odds ratio. SIL=squamous intraepithelial lesion. CIN=cervical intraepithelial neoplasia.*Adjusted for the time-varying effects of ART or CD4 cell count.

s reporting the association of ART with cervical lesion incidence, progression and regression, and invasive cervical cancer incidence are summarised in figure 3.Figure 3 Meta-analysis of cervical lesion incidence, progression and regression, and invasive cervical cancer incidence among ART users compared with ART-naive Weights are from random effects analysis. Only studies that reported HR from time-to-event analysis included in the meta-analysis (table 3). HR=hazard ratio. OR=odds ratio. SIL=squamous intraepithelial lesion. CIN=cervical intraepithelial neoplasia.*Adjusted for the time-varying effects of ART or CD4 cell count. The pooled OR among 14 studies17, 20, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41 reporting the association of ART and cervical lesion prevalence suggests no evidence of an association of ART with the prevalence of HSIL-CIN2+ diagnosed by either cytology or histology (crude OR 0·92, 95% CI 0·70–1·20; I2=56·6%, p=0·01; table 2, figure 2). Restricting the analysis to those studies17, 20, 32, 33, 34, 38, 39 with histological confirmation found no evidence of an association (crude OR 0·99, 95% CI 0·69–1·41; I2=58·7%, p=0·01; data not shown) but when analyses were restricted to studies17, 20 that adjusted for both current CD4 cell count and ART duration, there was some evidence that ART users had decreased prevalence of HSIL-CIN2+ compared with ART-naive women (aOR 0·85, 95% CI 0·62–1·18; I2=0%, p=0·56, adjusted for current CD4 cell count alone; aOR 0·65, 95% CI 0·40–1·06; I2=29·5%, p=0·25, with additional adjustment for duration on ART). Three studies,17, 20 all from the African region, reported the association of ART duration with prevalent HSIL-CIN2+ diagnosed by histology. The pooled OR suggests that CIN2+ was lower among prolonged ART users (≥2 years) than in short-duration users (<2 years) and ART-naive combined (aOR 0·68, 0·49–0·94; I2=2.5%, p=0·36, adjusted for age and current CD4 cell count; appendix p 19).

ssociation of ART duration with prevalent HSIL-CIN2+ diagnosed by histology. The pooled OR suggests that CIN2+ was lower among prolonged ART users (≥2 years) than in short-duration users (<2 years) and ART-naive combined (aOR 0·68, 0·49–0·94; I2=2.5%, p=0·36, adjusted for age and current CD4 cell count; appendix p 19). Study size varied widely (range 70–3185 women living with HIV). The largest study33 enrolled 3185 women (34% of participants included in the meta-analysis). However, excluding this study did not change the overall results. We found no evidence to suggest publication bias among studies reporting cervical lesion prevalence (Begg's rank correlation test; crude analysis p=0·48, adjusted analysis p=0·50). An additional sensitivity analysis including low-grade lesion outcomes (atypical squamous cells of undetermined significance or low-grade SIL diagnosed by cytology, CIN [grade 1 or higher] diagnosed by histology, and abnormality on visual inspection with colposcopy) suggests that ART is associated with a reduction in these outcomes, although these associations were not significant (appendix pp 20–22).

amous cells of undetermined significance or low-grade SIL diagnosed by cytology, CIN [grade 1 or higher] diagnosed by histology, and abnormality on visual inspection with colposcopy) suggests that ART is associated with a reduction in these outcomes, although these associations were not significant (appendix pp 20–22). The pooled HR among ten studies15, 54, 55, 56, 57, 58, 59, 60, 61, 62 reporting the association of ART and cervical lesion incidence provides weak evidence of an association of ART with cytology diagnosed SIL incidence (crude HR 0·75, 95% CI 0·56–1·00; I2=41%, p =0·09; table 3). Among five studies that adjusted for the time-varying effects of ART,15, 54, 58, 60, 62 we found evidence of a reduction in SIL incidence among ART users (aHR 0·64, 95% CI 0·47–0·86; I2=19·4%, p=0·29). There was no evidence to suggest publication bias for these studies (Beggs rank correlation test; crude analysis p=0·42, adjusted analysis p=1·000).Table 3 Meta-analysis of the association of ART with cervical lesion incidence, progression and regression, and invasive cervical cancer incidence among women living with HIV

sk of being HIV infected until they become sexually active. HIV testing is therefore a one-off activity in childhood, which is particularly important because the sustainability of incentivisation strategies is of concern, particularly for enforcing long-term changes in health behaviours, such as adherence to ART.26, 27 In low-income settings, lotteries might be a more affordable strategy than fixed incentives. In our study, the proportion of participants in the lottery group who underwent HIV testing was almost three times the proportion of participants in the control group who had an HIV test, and the effect was similar to that of a fixed incentive. These findings are in contrast with results from studies investigating the effect of fixed financial incentives or lottery, or both, to enhance uptake of circumcision.28, 29 Fixed incentives increased uptake of circumcision, but lotteries had no or a non-significant effect.28, 29 Contextual factors need to be taken into account when designing an incentivisation strategy. Careful consideration is needed to determine the amount, type, and frequency of incentives and the probability of receiving an incentive.17 These factors affect both the likelihood of affecting the desired behaviour and enable autonomic decision making by the client.

ere was no evidence to suggest publication bias for these studies (Beggs rank correlation test; crude analysis p=0·42, adjusted analysis p=1·000).Table 3 Meta-analysis of the association of ART with cervical lesion incidence, progression and regression, and invasive cervical cancer incidence among women living with HIV Crude analysis* Adjusted analysis† n studies HR (95%CI)‡ I2 p value for heterogeneity n studies HR (95%CI)‡ I2 p value for heterogeneity SIL incidence All 10 0·75 (0·56–1·00) 40·9% 0·09 5 0·64 (0·47–0·86) 19·4% 0·29 Africa 2 0·59 (0·44–0·80) 0% 0·71 1 0·62 (0·42–0·91) ·· ·· Latin America 1 1·90 (0·90–4·01) ·· ·· ·· ·· ·· ·· Europe or North America 7 0·73 (0·52–1·03) 14·0% 0·32 4 0·64 (0·40–1·02) 39·0% 0·18 SIL progression All 6 0·64 (0·56–0·74) 0% 0·42 4 0·64 (0·54–0·75) 17·8% 0·30 Africa 3 0·67 (0·56–0·79) 0% 0·68 2 0·68 (0·57–0·80) 0% 0·65 Europe or North America 3 0·62 (0·43–0·90) 46·4% 0·16 2 0·57 (0·39–0·85) 58·0% 0·12 SIL-CIN regression All 6 1·61 (1·31–1·97) 18·3% 0·30 5 1·54 (1·30–1·82) 0% 0·42 Africa ·· ·· ·· ·· 1 1·71 (1·29–2·27) ·· ·· Europe or North America 5 1·62 (1·21–2·16) 28·4% 0·23 4 1·45 (1·17–1·81) 1·8% 0·38 Invasive cervical cancer incidence All 2 0·40 (0·18–0·87) 32·7% 0·22 1 0·50 (0·29–0·87) ·· ·· * Includes studies with no adjustment potential confounders and studies that adjust for sociodemographic factors only but no adjustment for HIV related factors. † Includes studies that adjusted for time-varying ART or time-varying CD4 cell count.

Crude analysis* Adjusted analysis† n studies HR (95%CI)‡ I2 p value for heterogeneity n studies HR (95%CI)‡ I2 p value for heterogeneity SIL incidence All 10 0·75 (0·56–1·00) 40·9% 0·09 5 0·64 (0·47–0·86) 19·4% 0·29 Africa 2 0·59 (0·44–0·80) 0% 0·71 1 0·62 (0·42–0·91) ·· ·· Latin America 1 1·90 (0·90–4·01) ·· ·· ·· ·· ·· ·· Europe or North America 7 0·73 (0·52–1·03) 14·0% 0·32 4 0·64 (0·40–1·02) 39·0% 0·18 SIL progression All 6 0·64 (0·56–0·74) 0% 0·42 4 0·64 (0·54–0·75) 17·8% 0·30 Africa 3 0·67 (0·56–0·79) 0% 0·68 2 0·68 (0·57–0·80) 0% 0·65 Europe or North America 3 0·62 (0·43–0·90) 46·4% 0·16 2 0·57 (0·39–0·85) 58·0% 0·12 SIL-CIN regression All 6 1·61 (1·31–1·97) 18·3% 0·30 5 1·54 (1·30–1·82) 0% 0·42 Africa ·· ·· ·· ·· 1 1·71 (1·29–2·27) ·· ·· Europe or North America 5 1·62 (1·21–2·16) 28·4% 0·23 4 1·45 (1·17–1·81) 1·8% 0·38 Invasive cervical cancer incidence All 2 0·40 (0·18–0·87) 32·7% 0·22 1 0·50 (0·29–0·87) ·· ·· * Includes studies with no adjustment potential confounders and studies that adjust for sociodemographic factors only but no adjustment for HIV related factors. † Includes studies that adjusted for time-varying ART or time-varying CD4 cell count. ‡ Only studies that reported HR from time-to-event analysis included in the meta-analysis. HR=hazard ratio. SIL=squamous intraepithelial lesions. CIN=cervical intraepithelial neoplasia.

Crude analysis* Adjusted analysis† n studies HR (95%CI)‡ I2 p value for heterogeneity n studies HR (95%CI)‡ I2 p value for heterogeneity SIL incidence All 10 0·75 (0·56–1·00) 40·9% 0·09 5 0·64 (0·47–0·86) 19·4% 0·29 Africa 2 0·59 (0·44–0·80) 0% 0·71 1 0·62 (0·42–0·91) ·· ·· Latin America 1 1·90 (0·90–4·01) ·· ·· ·· ·· ·· ·· Europe or North America 7 0·73 (0·52–1·03) 14·0% 0·32 4 0·64 (0·40–1·02) 39·0% 0·18 SIL progression All 6 0·64 (0·56–0·74) 0% 0·42 4 0·64 (0·54–0·75) 17·8% 0·30 Africa 3 0·67 (0·56–0·79) 0% 0·68 2 0·68 (0·57–0·80) 0% 0·65 Europe or North America 3 0·62 (0·43–0·90) 46·4% 0·16 2 0·57 (0·39–0·85) 58·0% 0·12 SIL-CIN regression All 6 1·61 (1·31–1·97) 18·3% 0·30 5 1·54 (1·30–1·82) 0% 0·42 Africa ·· ·· ·· ·· 1 1·71 (1·29–2·27) ·· ·· Europe or North America 5 1·62 (1·21–2·16) 28·4% 0·23 4 1·45 (1·17–1·81) 1·8% 0·38 Invasive cervical cancer incidence All 2 0·40 (0·18–0·87) 32·7% 0·22 1 0·50 (0·29–0·87) ·· ·· * Includes studies with no adjustment potential confounders and studies that adjust for sociodemographic factors only but no adjustment for HIV related factors. † Includes studies that adjusted for time-varying ART or time-varying CD4 cell count. ‡ Only studies that reported HR from time-to-event analysis included in the meta-analysis. HR=hazard ratio. SIL=squamous intraepithelial lesions. CIN=cervical intraepithelial neoplasia. When analyses were restricted to two studies17, 63 that reported incidence of HSIL-CIN2+ determined by histology, there was strong evidence that prolonged duration ART users had reduced incidence compared with ART-naive women (aOR 0·59, 95% CI 0·40–0·87 [adjusted for nadir CD4 cell count]; I2=0%, p=0·35, data not shown).

When analyses were restricted to two studies17, 63 that reported incidence of HSIL-CIN2+ determined by histology, there was strong evidence that prolonged duration ART users had reduced incidence compared with ART-naive women (aOR 0·59, 95% CI 0·40–0·87 [adjusted for nadir CD4 cell count]; I2=0%, p=0·35, data not shown). The pooled HR among six studies31, 55, 64, 65, 66, 67 suggests a reduced hazard of cytology diagnosed SIL progression among ART users (crude HR 0·64, 95% CI 0·56–0·74; I2=0%, p=0·42; table 3). Restricting the analysis to four studies64, 65, 66, 67 that adjusted for time-varying ART did not alter the estimate (aHR 0·64, 95% CI 0·54–0·75; I2=17·8%, p=0·30). Similarly, there was no variation in HR by region. The pooled HR among six studies15, 31, 64, 67, 70, 71 suggests an increased likelihood of regression of cytology diagnosed SIL or histology diagnosed CIN among ART users (crude HR 1·61, 95% CI 1·31–1·97; I2=18·3%, p=0·30; table 3). Restricting the analysis to five studies15, 64, 67, 70, 71 that adjusted for time-varying ART during follow-up did not alter the estimate (aHR 1·54, 95% CI 1·30–1·82; I2=0%, p=0·42).

account when designing an incentivisation strategy. Careful consideration is needed to determine the amount, type, and frequency of incentives and the probability of receiving an incentive.17 These factors affect both the likelihood of affecting the desired behaviour and enable autonomic decision making by the client. Ethicists have raised concerns regarding coercion and equity when using incentives to promote healthy behaviour.30 In particular, when considering incentivisation of caregivers for health-related activities targeting their children, the potential of coercion of children from their caregivers should be considered. Lottery systems might be ethically less problematic because receipt of the incentive does not rely exclusively on displaying the desired behaviour but includes an element of chance. The use of lottery incentive systems to encourage HIV testing in the general population has been discussed in the national HIV testing campaign South African Right To Know.31

ession of cytology diagnosed SIL or histology diagnosed CIN among ART users (crude HR 1·61, 95% CI 1·31–1·97; I2=18·3%, p=0·30; table 3). Restricting the analysis to five studies15, 64, 67, 70, 71 that adjusted for time-varying ART during follow-up did not alter the estimate (aHR 1·54, 95% CI 1·30–1·82; I2=0%, p=0·42). Although most studies reported progression or regression of any cytology diagnosed SIL grade, one study reported progression of low-grade SIL to a higher grade, and regression from high-grade to low-grade SIL, diagnosed by cytology.64 No change in the estimate was observed when excluding that study for either the progression or regression outcomes. No evidence suggests publication bias for the progression studies (Beggs rank correlation test, p=0·85), but there is some evidence for bias in the regression studies (p=0·04) because more of the small studies report a positive association of ART with regression. However, the largest study (enrolling 1048 women living with HIV followed up over a median 18 months64) finds a significant increased likelihood of regression among ART users compared with ART-naive participants (aHR 1·71, 95% CI 1·29–2·27, adjusted for ART duration, age, and excision treatment), suggesting a real beneficial effect of ART. The pooled HR among two studies73, 74 suggests a decreased risk of invasive cervical cancer incidence among ART users (crude HR 0·40, 95% CI 0·18–0·87; I2=32·7%, p=0·22; table 3). There is no evidence to suggest publication bias for these studies (Beggs rank correlation test; p=0·32, data not shown).

Although most studies reported progression or regression of any cytology diagnosed SIL grade, one study reported progression of low-grade SIL to a higher grade, and regression from high-grade to low-grade SIL, diagnosed by cytology.64 No change in the estimate was observed when excluding that study for either the progression or regression outcomes. No evidence suggests publication bias for the progression studies (Beggs rank correlation test, p=0·85), but there is some evidence for bias in the regression studies (p=0·04) because more of the small studies report a positive association of ART with regression. However, the largest study (enrolling 1048 women living with HIV followed up over a median 18 months64) finds a significant increased likelihood of regression among ART users compared with ART-naive participants (aHR 1·71, 95% CI 1·29–2·27, adjusted for ART duration, age, and excision treatment), suggesting a real beneficial effect of ART. The pooled HR among two studies73, 74 suggests a decreased risk of invasive cervical cancer incidence among ART users (crude HR 0·40, 95% CI 0·18–0·87; I2=32·7%, p=0·22; table 3). There is no evidence to suggest publication bias for these studies (Beggs rank correlation test; p=0·32, data not shown). Discussion Our results indicate that women on ART had a lower prevalence of high-risk HPV and a reduction in the incidence of histology diagnosed HSIL-CIN2+ and invasive cervical cancer, after adjustment for CD4 cell count and treatment duration.

The pooled HR among two studies73, 74 suggests a decreased risk of invasive cervical cancer incidence among ART users (crude HR 0·40, 95% CI 0·18–0·87; I2=32·7%, p=0·22; table 3). There is no evidence to suggest publication bias for these studies (Beggs rank correlation test; p=0·32, data not shown). Discussion Our results indicate that women on ART had a lower prevalence of high-risk HPV and a reduction in the incidence of histology diagnosed HSIL-CIN2+ and invasive cervical cancer, after adjustment for CD4 cell count and treatment duration. To our knowledge, this is the first meta-analysis to investigate the associations between ART and high-risk HPV and cytology and histology diagnosed cervical lesion and invasive cervical cancer outcomes. Doing a meta-analysis of observational studies for high-risk HPV and cervical lesion outcomes has difficulties because of inherent differences in study populations, definitions of exposure and timescale of outcomes used, and the varying approaches to adjustment of effect estimates. The particular challenge with cross-sectional studies concerns the timing of HPV infection and development of cervical lesions, which might take several years, in relation to ART initiation and immune restoration that can happen more rapidly but is dependent on nadir CD4 cell count. The discordances in natural histories of HPV, CIN, and HIV disease might explain the observed lack of effect of ART on prevalent high-grade cervical lesions in this analysis.

t take several years, in relation to ART initiation and immune restoration that can happen more rapidly but is dependent on nadir CD4 cell count. The discordances in natural histories of HPV, CIN, and HIV disease might explain the observed lack of effect of ART on prevalent high-grade cervical lesions in this analysis. Restricting analyses to those studies that adjusted for nadir or current CD4 cell count or ART duration suggests that ART is associated with a reduction in high-risk HPV or cervical lesion outcomes, with less between-study heterogeneity. In studies that report limited or no association, immune reconstitution by ART might not have been established early enough after HPV infection to prevent or to reverse the development of high-risk HPV persistence or CIN2+. However, prospective studies that adjusted for the time-varying effects of ART use and CD4 cell count suggested a reduction in the incidence of CIN2+ and incidence and progression of SIL.

ess problematic because receipt of the incentive does not rely exclusively on displaying the desired behaviour but includes an element of chance. The use of lottery incentive systems to encourage HIV testing in the general population has been discussed in the national HIV testing campaign South African Right To Know.31 The strengths of this study include an incentivisation strategy directed at caregivers who are the gatekeepers to children accessing health care, clear denominators, and a large sample size. We acknowledge several limitations. First, the trial was nested in a prevalence survey involving household visits. Whether the interaction between fieldworkers and household members and the information provided during these visits had any effect on the uptake of testing is unknown. Second, the number of households randomised to each trial arm were relatively balanced, but the number of households eligible to participate were not, which might have introduced selection bias. However, adjustments were made for the number of children per household to account for imbalance. Households randomised to not receive an incentive were more likely to indicate that they did not have a child in the target age group and therefore were ineligible. These households might have silently refused to participate but felt uncomfortable refusing openly. Thus the incentives might have increased the participation in the trial and uptake of testing. Household characteristics of the participating household were similar between the three groups except for the number of children in each household. This did not affect the effect estimate, as the outcome was measured on household level and adjusted for the number of children in a household. Third, children in the non-incentivised group might have tested without identifying themselves as trial participants, resulting in differential outcome misclassification and possibly overestimation of the effect. Fourth, as previously discussed, the effect of incentives is context-specific. Although the broad principle might be generalisable to other settings, the size of the effect is less likely to be.

ve been established early enough after HPV infection to prevent or to reverse the development of high-risk HPV persistence or CIN2+. However, prospective studies that adjusted for the time-varying effects of ART use and CD4 cell count suggested a reduction in the incidence of CIN2+ and incidence and progression of SIL. Several studies reported that a high nadir CD4 cell count was associated with a 36–70% reduced risk of high-risk HPV29, 30 and a 36–80% reduced risk of CIN2+33, 39, 63 compared with those with low nadir CD4 cell count. Other studies17, 30 have shown that, once on ART, effective therapy (ie, patients with prolonged duration, sustained HIV-1 viral suppression and stable high CD4 cell count) was associated with a reduction in high-risk HPV persistence and histology diagnosed CIN2+. Further evidence suggests that high-risk HPV prevalence and incidence decreased and cytology diagnosed SIL regression increased in women who were highly adherent to ART.15 Of crucial importance, ART is associated with a reduction in incidence of invasive cervical cancer, especially if started at higher nadir CD4 cell count,63 and used over longer durations by adherent patients.73 This encouraging finding contrasts with previous studies that had shown a paradoxical increase in invasive cervical cancer incidence after the introduction of highly active ART.75 This could be because, in the early ART era, therapy was initiated at a lower nadir CD4 cell count, at which full restoration of cervical mucosal immunity was not obtained while life expectancy of patients and their likelihood to develop cancers were higher.

ncer incidence after the introduction of highly active ART.75 This could be because, in the early ART era, therapy was initiated at a lower nadir CD4 cell count, at which full restoration of cervical mucosal immunity was not obtained while life expectancy of patients and their likelihood to develop cancers were higher. The representation of studies from African settings has been steadily increasing; many of the earlier studies were done in the USA or Europe, leading to a geographical and period heterogeneity. The African studies17, 54, 55, 64, 65 provide encouraging indication that earlier initiation and effective ART over a prolonged duration can prevent cervical lesion incidence and progression and promote regression. Conversely, we found fewer studies from Latin America and Asia and most were cross-sectional in design. These studies24, 25, 26, 27, 28, 29, 38, 39 reported an opposite increased risk of high-risk HPV and high-grade cervical lesions among ART users. The lack of prospective studies in these regions prohibits a more direct assessment of the role of ART on longitudinal outcomes. An increased frequency of cervical cancer screening visits remains important especially among women on ART if they have started at a low nadir CD4 cell count. This concerns a generation of women who might have started ART with older guidelines at specific lower CD4 cell count thresholds and who might never have fully recovered their HPV-specific mucosal immune response.

visits remains important especially among women on ART if they have started at a low nadir CD4 cell count. This concerns a generation of women who might have started ART with older guidelines at specific lower CD4 cell count thresholds and who might never have fully recovered their HPV-specific mucosal immune response. We encountered several limitations in this review. Firstly, most cross-sectional studies used a binary category of ART users and treatment-naive. A more informative analysis would be to measure the effect of ART duration because there is a non-comparability among women initiating ART with decreasing CD4 cell count compared with those with higher CD4 cell count not yet needing treatment. Women who initiate ART are more likely to have advanced HIV disease, lower nadir CD4 cell counts, and higher HIV-1 viral loads than are those who have not yet started ART. The definition of ART-naive participants also varied across studies, which in some cases included women on monotherapy or dual-therapy regimens, and we cannot rule out the possibility that these women could have had lower or less stable CD4 cell counts to justify ART initiation.

are those who have not yet started ART. The definition of ART-naive participants also varied across studies, which in some cases included women on monotherapy or dual-therapy regimens, and we cannot rule out the possibility that these women could have had lower or less stable CD4 cell counts to justify ART initiation. The outcome definitions for cervical lesions varied between studies, in particular the use of cytological and histological measurement and definition of progression and regression between grades. Most prospective studies used cytological outcomes instead of the more desirable histological endpoint and grouping of cytology diagnosed grades of SIL varied; this, coupled with the variation in ART exposure between populations (eg, varying regimens and duration), makes interpretation of pooled data less clear. The possibility of unmeasured confounding also exists. Additionally, many studies did not report on likely predictors or effect modifiers of progression or regression of cervical lesions, which include nadir CD4 cell count, ART adherence, and HIV virological control. When available, we did sensitivity analysis that adjusted for time-varying effects of ART. Finally, individual patient-level data meta-analysis would allow for better harmonisation of these definitions and adjustments, which would provide a more precise and robust estimate of the association of ART and high-risk HPV and cervical lesion outcomes.

itivity analysis that adjusted for time-varying effects of ART. Finally, individual patient-level data meta-analysis would allow for better harmonisation of these definitions and adjustments, which would provide a more precise and robust estimate of the association of ART and high-risk HPV and cervical lesion outcomes. Our review has practical implications for the management of HIV patients and cervical cancer control worldwide. The current recommendation of encouraging earlier ART initiation, coupled with rapid virological control, and sustained adherence is likely to lead to an earlier and possibly more functionally complete mucosal immune reconstitution. We expect that this should in turn lead to a more rapid clearance of high-risk HPV, thus reducing cytology diagnosed SIL and histology diagnosed CIN incidence or progression and ultimately reducing cervical cancer incidence in this high-risk population. ART users with low or unknown nadir CD4 cell count remain at significant high risk despite ART initiation and should be screened frequently. Supplementary Material Supplementary appendix

Our review has practical implications for the management of HIV patients and cervical cancer control worldwide. The current recommendation of encouraging earlier ART initiation, coupled with rapid virological control, and sustained adherence is likely to lead to an earlier and possibly more functionally complete mucosal immune reconstitution. We expect that this should in turn lead to a more rapid clearance of high-risk HPV, thus reducing cytology diagnosed SIL and histology diagnosed CIN incidence or progression and ultimately reducing cervical cancer incidence in this high-risk population. ART users with low or unknown nadir CD4 cell count remain at significant high risk despite ART initiation and should be screened frequently. Supplementary Material Supplementary appendix Acknowledgments We wish to thank the ART and HPV Review Group who were principal authors of some of the studies included in this review and provided clarifications and re-analysis of their published data. We did this analysis without a dedicated funding source; however, SDS and YB are supported by public grants from the Instituto de Salud Carlos III CIBERESP and the Agència de Gestió d'Ajuts Universitaris i de Recerca (2014 SGR 756) and PM and HK by UK Medical Research Council (MRC) PHINDS scheme (PH01/14–39).

hed data. We did this analysis without a dedicated funding source; however, SDS and YB are supported by public grants from the Instituto de Salud Carlos III CIBERESP and the Agència de Gestió d'Ajuts Universitaris i de Recerca (2014 SGR 756) and PM and HK by UK Medical Research Council (MRC) PHINDS scheme (PH01/14–39). Contributors HK, SDS, and PM conceptualised the study, and developed the research protocol. HK and PM identified articles for full-text review. HK and HAW extracted data from studies that matched inclusion criteria. HK did the statistical analyses with input from HAW and YB. All authors contributed to the writing of the manuscript. Declaration of interests We declare no competing interests. ART and HPV Review Group Y-L Qiao and R-M Feng (Cancer Institute & Hospital, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, China), H De Vuyst and V Tenet (International Agency for Research on Cancer, Lyon, France), A Jaquet (Université Bordeaux, Bordeaux, France), D Konopnicki (Centre Hospitalier Universitaire Saint-Pierre, Brussels, Belgium), T Omar (University of the Witwatersrand and National Health Laboratory Service, Johannesburg, South Africa), L Menezes (University of South Florida, Tampa, FL, USA), and C Moucheraud and R Hoffman (University of California, Los Angeles, CA, USA).

Introduction Antiretroviral therapy (ART) effectively prevents progression to AIDS and death in people with HIV and decreases the likelihood of onward transmission. The number of HIV-related deaths in adolescents, however, has more than tripled in the past decade. Adolescents are the only age group in which HIV-associated mortality is increasing, despite the global scale-up of ART programmes.1 Delayed diagnosis of young people living with HIV increases the risk of immunosuppression resulting in increased mortality.2 Additionally, initiation of ART at advanced stages of disease is associated with much poorer outcomes than if initiated at early stages.3, 4 The prevalence of undiagnosed HIV is particularly high in older children and adolescents.5, 6 Findings from a recent meta-analysis from South Africa estimated that only 14% of children and adolescents aged 15–24 years who live with HIV were accessing ART.7

ith much poorer outcomes than if initiated at early stages.3, 4 The prevalence of undiagnosed HIV is particularly high in older children and adolescents.5, 6 Findings from a recent meta-analysis from South Africa estimated that only 14% of children and adolescents aged 15–24 years who live with HIV were accessing ART.7 HIV testing is the essential entry point for both treatment and prevention efforts. Conventional HIV testing strategies such as facility-based, provider-initiated HIV testing and counselling, recommended by WHO since 2007 in high HIV prevalence settings, have not been sufficient to reduce the burden of undiagnosed HIV in this age group.8 Community-based strategies, such as mobile testing units and door-to-door testing, and one-stop campaigns have been effective in adults, but tend to either exclude adolescents or be less effective in increasing uptake of HIV testing in this age group.9, 10 This might partly be due to issues of consent to HIV testing. Novel approaches are therefore needed to improve coverage of HIV diagnosis and treatment in this age group. Research in context Evidence before this study

HIV testing is the essential entry point for both treatment and prevention efforts. Conventional HIV testing strategies such as facility-based, provider-initiated HIV testing and counselling, recommended by WHO since 2007 in high HIV prevalence settings, have not been sufficient to reduce the burden of undiagnosed HIV in this age group.8 Community-based strategies, such as mobile testing units and door-to-door testing, and one-stop campaigns have been effective in adults, but tend to either exclude adolescents or be less effective in increasing uptake of HIV testing in this age group.9, 10 This might partly be due to issues of consent to HIV testing. Novel approaches are therefore needed to improve coverage of HIV diagnosis and treatment in this age group. Research in context Evidence before this study Survival has substantially improved since the advent of antiretroviral therapy (ART). The crucial step to accessing HIV treatment is HIV testing and counselling. The prevalence of undiagnosed HIV is particularly high in older children and adolescents, and coverage of ART is therefore much lower than in adults. Existing HIV testing and counselling strategies either exclude or are insufficient to meet the needs of this age group. Novel strategies will be required if we are to meet the ambitious UNAIDS 90-90-90 targets in this age group, which stipulate that 90% of HIV-infected individuals should be diagnosed. Incentivisation is a strategy that has been used in various public health programmes to influence health-related behaviour or to achieve specific targets. The effect of incentives on uptake of HIV testing has been investigated in two recent systematic reviews. We searched the Cochrane Review database, ClinicalTrials.gov, the WHO International Clinical Trial Registry, MEDLINE, Embase, and Web of Science with the terms “HIV”, “incentives”, “voucher”, “lottery”, “conditional cash transfer”, and “prize draw” for papers not included in the systematic reviews. We identified one randomised controlled trial from the USA, one randomised controlled trial from Malawi, and two observational studies in high-risk groups (unemployed men and adolescents) in South Africa. In all the studies, uptake of HIV testing was higher in the incentivised groups than in the non-incentivised groups, but none of the studies used a lottery approach. The only randomised controlled trial investigating the effect of incentives in sub-Saharan Africa was focused on adults and was done in 2004, before ART became widely available.

ake of HIV testing was higher in the incentivised groups than in the non-incentivised groups, but none of the studies used a lottery approach. The only randomised controlled trial investigating the effect of incentives in sub-Saharan Africa was focused on adults and was done in 2004, before ART became widely available. Added value of this study Our study is the first randomised controlled trial to test incentives to improve uptake of HIV testing by older children and adolescents in sub-Saharan Africa. Notably, use of the household as the unit of randomisation acknowledges the central role of the family and caregiver in making the decision about whether the child or adolescent is tested or not. We used two different incentivisation strategies, namely a fixed incentive of US$2 or a lottery with a one in eight chance to receive $5 or $10. Although uptake of HIV testing was higher in households randomised to fixed incentives than in households receiving no incentives, participation in the lottery tripled uptake. Lottery might be a more cost-effective strategy in resource-constrained settings and potentially less coercive because the participant is aware that an incentive might not be forthcoming. The strategy has potential for scalability and sustainability for identifying children with HIV acquired perinatally because there is no ongoing risk until sexual debut. Implications of all the available evidence

Our study is the first randomised controlled trial to test incentives to improve uptake of HIV testing by older children and adolescents in sub-Saharan Africa. Notably, use of the household as the unit of randomisation acknowledges the central role of the family and caregiver in making the decision about whether the child or adolescent is tested or not. We used two different incentivisation strategies, namely a fixed incentive of US$2 or a lottery with a one in eight chance to receive $5 or $10. Although uptake of HIV testing was higher in households randomised to fixed incentives than in households receiving no incentives, participation in the lottery tripled uptake. Lottery might be a more cost-effective strategy in resource-constrained settings and potentially less coercive because the participant is aware that an incentive might not be forthcoming. The strategy has potential for scalability and sustainability for identifying children with HIV acquired perinatally because there is no ongoing risk until sexual debut. Implications of all the available evidence Financial incentives show promise for improving engagement in HIV testing, especially in high-risk groups. A better understanding of durability, scalability, ease of implementation, sustainability, and cost-effectiveness of these different approaches is needed to maximise the effect of incentives in reaching the ambitious UNAIDS 90-90-90 targets.

omise for improving engagement in HIV testing, especially in high-risk groups. A better understanding of durability, scalability, ease of implementation, sustainability, and cost-effectiveness of these different approaches is needed to maximise the effect of incentives in reaching the ambitious UNAIDS 90-90-90 targets. Incentivisation is a strategy that has been used with varying success in health programmes to influence behaviours, including smoking, illicit substance use, and poor diet, and to achieve specific targets such as completion of vaccination.11, 12 The principle underlying use of incentives is the psychological theory of contingency management, whereby stimulus control and positive reinforcement are used to change behaviour.13 Conditional and unconditional incentives reduce pregnancy rates and sexual risk behaviour for HIV acquisition in adolescents and young adults in Kenya, Malawi, and South Africa.14, 15, 16, 17 Economic incentives have also been applied to encourage testing for sexually transmitted infections including HIV.18 The provision of financial incentives increased uptake of HIV testing in adults in Malawi19 and unemployed men in South Africa.20

scents and young adults in Kenya, Malawi, and South Africa.14, 15, 16, 17 Economic incentives have also been applied to encourage testing for sexually transmitted infections including HIV.18 The provision of financial incentives increased uptake of HIV testing in adults in Malawi19 and unemployed men in South Africa.20 In sub-Saharan Africa, where 90% of the world's children with HIV live, testing of minors requires consent from caregivers with the exception of emancipated minors. The age of ability to give independent consent varies between countries but is 18 years in most sub-Saharan African countries.21 For minors to access testing requires the willingness and engagement of caregivers. The aim of this study was to assess the effect of financial incentives provided to caregivers on uptake of HIV testing and counselling in older children and adolescents aged 8–17 years in Harare, Zimbabwe. Methods Study design and participants In this three-arm household-randomised controlled trial, we compared the effect on HIV test uptake at primary health-care clinics by children aged 8–17 years of provision of no incentives (control) versus either a fixed incentive of US$2 or participation in a lottery (interventions). The trial was done and analysed according to the CONSORT guidelines, and ethical approval was obtained from the Medical Research Council of Zimbabwe, the London School of Hygiene & Tropical Medicine Ethics Committee, and the Institutional Review Board of the Biomedical Research and Training Institute, Harare, Zimbabwe.

The trial was done and analysed according to the CONSORT guidelines, and ethical approval was obtained from the Medical Research Council of Zimbabwe, the London School of Hygiene & Tropical Medicine Ethics Committee, and the Institutional Review Board of the Biomedical Research and Training Institute, Harare, Zimbabwe. The trial was nested within a household survey to estimate the prevalence of undiagnosed HIV in children aged 8–17 years in seven communities in Harare. As part of the prevalence survey, participants were anonymously tested for HIV by providing oral fluid samples. Participants and caregivers did not receive these results. Each community is served by primary health-care clinics that provide acute and antenatal care services. The survey took place between Jan 1, and Dec 18, 2015. Results of the prevalence survey have been reported.8 In brief, a sample of census enumeration areas, defined as the smallest delimited census area in the study communities, was selected from the 2012 National Census sampling frame using simple random sampling. All households in the selected census enumeration areas were enumerated, and any household with one or more residents aged 8–17 years was eligible to participate in the prevalence survey. Households were eligible for the trial if they included at least one prevalence survey participant whose HIV status was unknown.

ng. All households in the selected census enumeration areas were enumerated, and any household with one or more residents aged 8–17 years was eligible to participate in the prevalence survey. Households were eligible for the trial if they included at least one prevalence survey participant whose HIV status was unknown. Written informed consent in Shona was sought from the caregiver and assent from the participants. Consent to participate in the trial was sought separately from consent to participate in the prevalence survey. Households with one or more survey participants could therefore decline to participate in the trial.

ng. All households in the selected census enumeration areas were enumerated, and any household with one or more residents aged 8–17 years was eligible to participate in the prevalence survey. Households were eligible for the trial if they included at least one prevalence survey participant whose HIV status was unknown. Written informed consent in Shona was sought from the caregiver and assent from the participants. Consent to participate in the trial was sought separately from consent to participate in the prevalence survey. Households with one or more survey participants could therefore decline to participate in the trial. Randomisation and masking After enumeration, eligible households were randomly assigned (1:1:1) to one of three groups that would either receive no incentive, receive a $2 incentive, or participate in a lottery to win a cash prize if a survey participant in the household presented to the primary health-care clinic in the study community for HIV testing. US$ has been the official currency in Zimbabwe since 2009. The gross domestic product in Zimbabwe was $1008·6 per capita.22 $2 would pay for a return journey for two individuals from the outskirts of Harare to the city centre. Random allocation was built into the tablet used for data collection. Participants who were randomly assigned to the lottery had a one in eight chance of winning $5 or $10. There was no separate draw for $5 and $10 because both were in the same box at each clinic. Randomisation was done at the household level because it was not feasible to allocate participants in one household to different trial arms. An independent statistician used Stata version 14.0 to randomly allocate households. Randomisation was done on the basis of the list of households enumerated before the prevalence survey. This included households that were subsequently deemed ineligible because they did not have a child in the target age group. However, we randomised the enumerated households rather than those eligible for the survey to prevent fieldworkers from influencing allocation. If more than one survey participant from a household who was randomised to the intervention groups attended testing, each would be given the incentive.

in the target age group. However, we randomised the enumerated households rather than those eligible for the survey to prevent fieldworkers from influencing allocation. If more than one survey participant from a household who was randomised to the intervention groups attended testing, each would be given the incentive. Because the trial was embedded in the prevalence survey, the survey fieldworkers enrolled children into both the survey and the trial, and recruitment into the trial occurred on the same visit as that for enrolment into the prevalence survey. Procedures Fieldworkers visited eligible households and, after obtaining informed consent, collected data on household sociodemographic characteristics. If a child from an eligible household was absent at the first visit, two additional visits were made within 2 weeks unless the household head reported the child was expected to be absent for more than 2 weeks (in which case the child was coded as unavailable). History of previous HIV testing, including the date and location of the test (or tests) and whether participants were taking ART or co-trimoxazole prophylaxis, was recorded for each participant with a questionnaire administered to the participant's caregiver. Participants were asked to provide documentary evidence of previous HIV testing, and all participants underwent anonymised HIV testing.8 All households participating in the prevalence survey were provided with written information about the benefits of HIV testing.

stionnaire administered to the participant's caregiver. Participants were asked to provide documentary evidence of previous HIV testing, and all participants underwent anonymised HIV testing.8 All households participating in the prevalence survey were provided with written information about the benefits of HIV testing. Households with at least one survey participant who had either no documented evidence of a positive HIV test, had a negative HIV test result more than 6 months ago, or had never tested for HIV were invited to participate in the trial. Participants were given vouchers stating their survey study number and the trial arm to which their household had been assigned. Free HIV testing at primary health-care clinics was available for all trial participants and other members of the household at any time, but incentives were only provided for those with a trial voucher. Research assistants were available at the clinics for HIV testing and counselling. HIV testing was done according to national guidelines, and those who tested HIV positive were referred for HIV care at the same clinic. As per national guidelines, HIV testing required both caregiver consent and child assent. Staff at the clinics and the research assistant had repeated training to provide age-appropriate information, testing, and counselling to prevent coercion. A research assistant based at the clinics reported any adverse events and ensured appropriate follow-up and linkage to care for any child diagnosed with HIV.

ssent. Staff at the clinics and the research assistant had repeated training to provide age-appropriate information, testing, and counselling to prevent coercion. A research assistant based at the clinics reported any adverse events and ensured appropriate follow-up and linkage to care for any child diagnosed with HIV. Outcomes The primary outcome was proportion of households with at least one child taking an HIV test within 4 weeks of enrolment. A household was categorised as having tested for HIV if at least one child in the participating household presented for HIV testing at the primary health-care clinic. Statistical analysis Sample size calculations were based on the assumption that if 20% of households in the control group sent a child for HIV testing at the clinic, 392 participating households per trial arm would provide 90% power to detect a 50% increase in uptake of testing in an intervention arm versus the control arm. Allowing for 25% refusal, we aimed to recruit 1568 households.

mption that if 20% of households in the control group sent a child for HIV testing at the clinic, 392 participating households per trial arm would provide 90% power to detect a 50% increase in uptake of testing in an intervention arm versus the control arm. Allowing for 25% refusal, we aimed to recruit 1568 households. Data were collected by fieldworkers on Nexus 7 2013 tablets running Open Data Kit software and transferred to Stata version 14.0 for data analysis. Descriptive statistics were done on the sociodemographic characteristics of the eligible households and the participants. We calculated median and IQR for continuous and non-parametric variables, and we estimated frequencies and percentages for categorical variables. Odds ratios were estimated with logistic regression to compare household HIV testing uptake (ie, at least one child testing for HIV) between the intervention arms and the control group, adjusting for community and number of children in the household as fixed effects and research assistant as a random effect. Adjustment for community and research assistant were made a priori. Adjustment for number of children was done to account for imbalance in different trial arms. Logistic regression was chosen as the method for analysis to account for the effect of clustering within communities and by research assistant. Research assistant was included as a random effect to allow for the possibility that some research assistants were better than others at explaining the study or convincing caregivers to take children for testing. All analyses were by intention to treat.

effect of clustering within communities and by research assistant. Research assistant was included as a random effect to allow for the possibility that some research assistants were better than others at explaining the study or convincing caregivers to take children for testing. All analyses were by intention to treat. We did a sensitivity analysis to investigate individual HIV test uptake by trial arm, adjusting for community and number of children in the household as fixed effects and household and research assistant as a random effect. Odds ratios were estimated for factors that predict individual HIV test uptake with logistic regression for children in the control group, adjusted for household as a fixed effect and research assistant as a random effect. Children's schooling was recoded into two categories on the basis of the recommended school grade for their age (appropriate grade for their age, any higher grade, or one grade below vs more than one grade below their age-appropriate grade or never in school). Reported general health was recorded as excellent or good or as fair or poor. The trial is registered with the Pan African Clinical Trials Registry, number PACTR201605001615280. Data sharing The prevalence survey dataset is stored in the DataCompass secure online repository, curated by the London School of Hygiene & Tropical Medicine (http://dx.doi.org/10.17037/DATA.174).

We did a sensitivity analysis to investigate individual HIV test uptake by trial arm, adjusting for community and number of children in the household as fixed effects and household and research assistant as a random effect. Odds ratios were estimated for factors that predict individual HIV test uptake with logistic regression for children in the control group, adjusted for household as a fixed effect and research assistant as a random effect. Children's schooling was recoded into two categories on the basis of the recommended school grade for their age (appropriate grade for their age, any higher grade, or one grade below vs more than one grade below their age-appropriate grade or never in school). Reported general health was recorded as excellent or good or as fair or poor. The trial is registered with the Pan African Clinical Trials Registry, number PACTR201605001615280. Data sharing The prevalence survey dataset is stored in the DataCompass secure online repository, curated by the London School of Hygiene & Tropical Medicine (http://dx.doi.org/10.17037/DATA.174). Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Data sharing The prevalence survey dataset is stored in the DataCompass secure online repository, curated by the London School of Hygiene & Tropical Medicine (http://dx.doi.org/10.17037/DATA.174). Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Between Aug 4, and Dec 18, 2015, 2050 households were eligible to participate in the prevalence survey on the basis of the randomly selected census enumeration areas (figure). 649 (32%) households were randomly assigned to receive no incentive, 740 (34%) households to receive $2, and 661 (32%) households to participate in the lottery. 1703 households participated in the prevalence survey. Of the participating households, 942 (55%) had one child, 496 (29%) had two children, 188 (11%) had three children, 55 (3%) had four children, and 22 (1%) had more than four children. The $2 incentive group was larger than the other intervention groups partly because of chance imbalance at randomisation. Households in the control group were more likely to have an absent child at the time of the survey visit. These households were therefore not eligible to participate in the trial. Children were unavailable in 148 households in the no-incentive group, 63 households in the $2 incentive group, and 81 households in the lottery group. 1688 households had at least one child with unknown HIV status and were enrolled into the trial. 22 households had no undiagnosed child, and one household refused consent. The primary outcome of HIV testing was assessed in 472 (28%) households in the no-incentive group, 654 (39%) households in the $2 incentive group, and 562 (33%) households in the lottery group.Figure Study profile

d were enrolled into the trial. 22 households had no undiagnosed child, and one household refused consent. The primary outcome of HIV testing was assessed in 472 (28%) households in the no-incentive group, 654 (39%) households in the $2 incentive group, and 562 (33%) households in the lottery group.Figure Study profile Child unavailable refers to a child that was absent at initial household visits and absent at two further visits or household head reporting that the child was expected to be absent for more than 2 weeks. *The households remained in the analysis as other children in the household participated. Socioeconomic characteristics were balanced between the three trial arms (table 1). The characteristics of individual trial participants by trial arm are shown in the appendix. Most household heads had at least secondary education, and almost half of the households owned their dwelling. Half of the households did not have a regular income or had a monthly income of less than $100. Most caregivers felt comfortable with the idea of an HIV-infected child visiting the household or for their child to share food and play with an HIV-infected child.Table 1 Baseline characteristics

holds owned their dwelling. Half of the households did not have a regular income or had a monthly income of less than $100. Most caregivers felt comfortable with the idea of an HIV-infected child visiting the household or for their child to share food and play with an HIV-infected child.Table 1 Baseline characteristics No incentive (N=472) US$2 (N=654) Lottery (N=562) Household size 5 (4–6) 5 (4–6) 5 (4–6) Eligible children in household 1 (1–2) 2 (1–2) 2 (1–2) Age of household head* 41 (35–49) 42 (36–51) 41 (35–49) Education of household head* None or primary 14 (3%) 34 (5%) 28 (5%) Secondary 397 (84%) 521 (80%) 468 (83%) Higher 60 (13%) 99 (15%) 66 (12%) Ownership of dwelling* Own dwelling 199 (42%) 314 (48%) 234 (42%) Rent 249 (53%) 307 (47%) 298 (53%) Use dwelling without rent 23 (5%) 33 (5%) 30 (5%) Household owns fridge* 429 (91%) 614 (94%) 518 (92%) Household owns car or truck* 71 (15%) 112 (17%) 85 (15%) Household owns television* 460 (98%) 650 (99%) 549 (98%) Number of household members earning regular salary* None 188 (40%) 265 (41%) 255 (45%) One 249 (53%) 322 (49%) 257 (46%) More than one 34 (7%) 67 (10%) 50 (9%) Regular household income per month* No regular income or <US$200 274 (58%) 355 (54%) 338 (60%) $200–500 128 (27%) 161 (25%) 140 (25%) >$500 69 (15%) 138 (21%) 84 (15%) Caregiver very comfortable with child playing with HIV-positive child* 447 (95%) 627 (96%) 544 (97%) Caregiver very comfortable with HIV-positive child visiting household* 442 (94%) 618 (95%) 529 (94%) Caregiver very comfortable with child sharing food with HIV-positive child* 430 (91%) 606 (93%) 523 (93%) Children aged 8–17 years in the household diagnosed with HIV 6 (1%) 19 (3%) 12 (2%) Children aged 8–17 years in the household living with HIV 8 (2%) 30 (5%) 24 (4%) Data are n (%) or median (IQR).

8 (95%) 529 (94%) Caregiver very comfortable with child sharing food with HIV-positive child* 430 (91%) 606 (93%) 523 (93%) Children aged 8–17 years in the household diagnosed with HIV 6 (1%) 19 (3%) 12 (2%) Children aged 8–17 years in the household living with HIV 8 (2%) 30 (5%) 24 (4%) Data are n (%) or median (IQR). * Data are missing for one patient in the no-incentive group. 93 (20%) of 472 households in the control group had at least one child tested for HIV within 4 weeks of enrolment, whereas at least one child was tested in 316 (48%) of 654 households in the $2 incentive group (adjusted odds ratio [OR] 3·67, 95% CI 2·77–4·85) and in 223 (40%) of 562 households in the lottery group (2·66, 2·00–3·55; table 2). The effect of the incentives on HIV testing was more pronounced in the sensitivity analysis, where individual children in the $2 group and the lottery group were compared with children in the control group. The adjusted OR were 4·86 (3·84–6·17) in the $2 incentive group and 3·23 (2·53–4·13) in the lottery group (table 2; appendix). No adverse events were reported.Table 2 Effect of provision of and type of incentives on uptake of HIV testing at household level At least one child went to clinic Crude OR (95% CI) p value Adjusted OR (95% CI)* p value No incentive (N=472) 93 (20%) 1 .. 1 .. US$2 (N=654) 316 (48%) 3·81 (2·90–5·01) <0·0001 3·67 (2·77–4·85) <0·0001 Lottery (N=562) 223 (40%) 2·68 (2·02–3·56) <0·0001 2·66 (2·00–3·55) <0·0001 OR=odds ratio. * Adjusted for community and number of children in household as fixed effects and for research assistant as a random effect.

At least one child went to clinic Crude OR (95% CI) p value Adjusted OR (95% CI)* p value No incentive (N=472) 93 (20%) 1 .. 1 .. US$2 (N=654) 316 (48%) 3·81 (2·90–5·01) <0·0001 3·67 (2·77–4·85) <0·0001 Lottery (N=562) 223 (40%) 2·68 (2·02–3·56) <0·0001 2·66 (2·00–3·55) <0·0001 OR=odds ratio. * Adjusted for community and number of children in household as fixed effects and for research assistant as a random effect. Factors associated with increased uptake of HIV testing in the control group included lower household income, smaller household size, and older age of the participants (table 3).Table 3 Household and individual level factors associated with HIV testing in the control group

* Adjusted for community and number of children in household as fixed effects and for research assistant as a random effect. Factors associated with increased uptake of HIV testing in the control group included lower household income, smaller household size, and older age of the participants (table 3).Table 3 Household and individual level factors associated with HIV testing in the control group Crude OR (95% CI)* p value Adjusted OR (95% CI)* p value Household level Does household own dwelling No 1 .. .. .. Yes 0·85 (0·55–1·230) 0·44 .. .. Household income No regular salary or <US$200 1 .. 1 .. US$200–500 0·61 (0·35–1·06) 0·080 0·59 (0·34–1·05) 0·075 >US$500 0·43 (0·21–0·91) 0·028 0·51 (0·24–1·11) 0·089 Children aged 8–17 years (reference category =1) 0·62 (0·48–0·79) <0·0001 0·61 (0·47–0·79) <0·0001 Age of household head (years) <30 1 .. .. .. 30–60 0·46 (0·16–1·32) 0·15 .. .. >60 0·94 (0·44–2·01) 0·88 .. .. Individual level Sex Male 1 .. .. .. Female 0·79 (0·52–1·20) 0·26 .. .. Age (years) 8–12 1 .. 1 .. 13–17 1·38 (0·91–2·09) 0·13 1·46 (0·94–2·25) 0·090 Orphan No 1 .. .. .. Single or double orphan 1·46 (0·80–2·64) 0·21 .. .. General health Good 1 .. 1 .. Fair/poor 1·94 (0·75–5·05) 0·17 1·59 (0·54–4·63) 0·41 Ever admitted to hospital No 1 .. .. .. Yes 0·76 (0·22–2·65) 0·67 .. .. Chronic skin conditions No 1 .. 1 .. Yes 1·94 (0·68–5·50) 0·21 1·61 (0·51–5·13) 0·42 Schooling (for age) ≤one grade behind for age 1 .. 1 .. >one grade behind for age 1·31 (0·83–2·06) 0·24 1·21 (0·76–1·95) 0·42 Caregiver Biological parent 1 .. .. .. Not biological parent 0·80 (0·47–1·36) 0·40 .. .. * Adjusted for household as a fixed effect and research assistant as a random effect.

61 (0·51–5·13) 0·42 Schooling (for age) ≤one grade behind for age 1 .. 1 .. >one grade behind for age 1·31 (0·83–2·06) 0·24 1·21 (0·76–1·95) 0·42 Caregiver Biological parent 1 .. .. .. Not biological parent 0·80 (0·47–1·36) 0·40 .. .. * Adjusted for household as a fixed effect and research assistant as a random effect. Discussion Uptake of HIV testing by children and adolescents in households that received a financial incentive was higher than in households that did not receive an incentive. A lottery with a one in eight probability of receiving an incentive had a similar effect on HIV testing as a fixed incentive of $2. Uptake of HIV testing in households that received no incentive was low (20%) despite HIV testing being free of charge. This could be because diagnostic HIV testing at the clinic was available during working hours only, and bringing children to the clinic for HIV testing necessitated caregivers taking time off work or looking after other children and possibly children missing school.12, 23 Diagnostic HIV testing was not done during the household visit because it could have affected participation in the prevalence survey, but dedicated research staff were available at the primary health-care clinics so that those attending for HIV testing did not have to wait in the routine clinic queue.

mselves as trial participants, resulting in differential outcome misclassification and possibly overestimation of the effect. Fourth, as previously discussed, the effect of incentives is context-specific. Although the broad principle might be generalisable to other settings, the size of the effect is less likely to be. UNAIDS has set ambitious 90-90-90 targets, whereby 90% of people living with HIV infection should be diagnosed, 90% of HIV-infected individuals should be receiving ART, and 90% of those receiving ART should be virologically suppressed by 2020.32 If achieved, this would lead to a 90% reduction in AIDS-related mortality and HIV incidence by 2030 and eliminate HIV as a public health threat. Reducing the burden of undiagnosed HIV is the crucial first step to realising the UNAIDS targets. Existing strategies are clearly inadequate to address the substantial burden of undiagnosed HIV infection in adolescents, and novel approaches will be necessary if the targets are to be met in this age group. Our findings show that incentives targeted at caregivers substantially improve HIV testing rates in adolescents. Looking forward, the cost-effectiveness of this approach must be studied, and careful thought must be given to the social and cultural context if strategies such as this are to be brought to scale. Supplementary Material Supplementary appendix

UNAIDS has set ambitious 90-90-90 targets, whereby 90% of people living with HIV infection should be diagnosed, 90% of HIV-infected individuals should be receiving ART, and 90% of those receiving ART should be virologically suppressed by 2020.32 If achieved, this would lead to a 90% reduction in AIDS-related mortality and HIV incidence by 2030 and eliminate HIV as a public health threat. Reducing the burden of undiagnosed HIV is the crucial first step to realising the UNAIDS targets. Existing strategies are clearly inadequate to address the substantial burden of undiagnosed HIV infection in adolescents, and novel approaches will be necessary if the targets are to be met in this age group. Our findings show that incentives targeted at caregivers substantially improve HIV testing rates in adolescents. Looking forward, the cost-effectiveness of this approach must be studied, and careful thought must be given to the social and cultural context if strategies such as this are to be brought to scale. Supplementary Material Supplementary appendix Acknowledgments The study was funded by the Wellcome Trust (095878/Z/11/Z). RAF's institution received a grant from the Wellcome Trust (095878/Z/11/Z). Salary support for VS and HAW was in part from a grant from the Medical Research Council (MRC) and the Department for International Development (DFID UK) under the MRC/DFID Concordat (K012126/1). We thank all participants and their caregivers.

RAF's institution received a grant from the Wellcome Trust (095878/Z/11/Z). Salary support for VS and HAW was in part from a grant from the Medical Research Council (MRC) and the Department for International Development (DFID UK) under the MRC/DFID Concordat (K012126/1). We thank all participants and their caregivers. Contributors KK and RAF designed the study. VS and TB were responsible for data management. VS analysed the data with input from KK and HAW. ED, SD, and GM coordinated the trial. HM and GM provided clinical advice and contributed to patient management. SM, PC, and HM contributed to the study design and study logistics. All authors contributed to writing the report and have seen and approved the final draft. Declaration of interests We declare no competing interests.

Introduction Interest in HIV self-testing—an approach to increase access to HIV testing—has increased since 2014.1 As a discreet and convenient approach, HIV self-testing might be most useful in reaching people who are reluctant or unable to access existing HIV testing services because of concerns about privacy, stigma, discrimination, and, in some contexts, criminalisation. According to various studies,2, 3, 4, 5, 6, 7 HIV self-testing is highly acceptable among many different population groups, including those with low testing coverage and who report barriers to and low uptake of existing HIV testing services. Despite this, some policy makers and users have raised concerns that self-testers might not be able to do the test or interpret the test results correctly.1 We did a systematic review to assess the reliability and performance of HIV rapid diagnostic tests when used by self-testers, compared with health-care workers. Although previous reviews assessed the accuracy of rapid diagnostic tests for self-testing,8, 9 they primarily focused on sensitivity and specificity and did not consider the validity of the reference standard. Thus, we systematically measure and report test concordance between self-testers and health-care workers to account for imperfect reference standards to establish the reliability and performance of rapid diagnostic tests used for self-testing. Research in context Evidence before this study

We did a systematic review to assess the reliability and performance of HIV rapid diagnostic tests when used by self-testers, compared with health-care workers. Although previous reviews assessed the accuracy of rapid diagnostic tests for self-testing,8, 9 they primarily focused on sensitivity and specificity and did not consider the validity of the reference standard. Thus, we systematically measure and report test concordance between self-testers and health-care workers to account for imperfect reference standards to establish the reliability and performance of rapid diagnostic tests used for self-testing. Research in context Evidence before this study To diagnose HIV, at least two or three tests, depending on the HIV prevalence among the population being tested, are needed. The validity of using a single test as a reference standard is imperfect. Previous systematic reviews focused on sensitivity and specificity of HIV rapid diagnostic tests for self-testing. An initial search of PubMed, for studies published from Jan 1, 1995, to Jan 26, 2016, with the search terms “HIV self-testing” and “review”, indicated that the validity of the reference standard has not been considered previously. Added value of this study

To diagnose HIV, at least two or three tests, depending on the HIV prevalence among the population being tested, are needed. The validity of using a single test as a reference standard is imperfect. Previous systematic reviews focused on sensitivity and specificity of HIV rapid diagnostic tests for self-testing. An initial search of PubMed, for studies published from Jan 1, 1995, to Jan 26, 2016, with the search terms “HIV self-testing” and “review”, indicated that the validity of the reference standard has not been considered previously. Added value of this study To inform a WHO recommendation, we assessed the reliability and performance of HIV rapid diagnostic tests used by self-testers compared with health-care workers, by calculating statistics on test concordance to account for the imperfect reference standard. We included studies that used products designed for self-testing in diverse country settings. Previous reviews were done when HIV self-testing was emerging; these reviews primarily drew from US and European studies that used professional-use products or prototypes that have since been adapted for HIV self-testing. Implications of all the available evidence

To inform a WHO recommendation, we assessed the reliability and performance of HIV rapid diagnostic tests used by self-testers compared with health-care workers, by calculating statistics on test concordance to account for the imperfect reference standard. We included studies that used products designed for self-testing in diverse country settings. Previous reviews were done when HIV self-testing was emerging; these reviews primarily drew from US and European studies that used professional-use products or prototypes that have since been adapted for HIV self-testing. Implications of all the available evidence Self-testers could reliably and accurately do an HIV rapid diagnostic test, whether assistance was provided or not, when compared with a trained health-care worker. Errors in the test procedure might be reduced by refinement of the design of rapid diagnostic tests for self-testing, improvement of manufacturer labelling and instructions for use, and provision of additional support with instructional videos. Modifications should always be the responsibility of the manufacturer.

ker. Errors in the test procedure might be reduced by refinement of the design of rapid diagnostic tests for self-testing, improvement of manufacturer labelling and instructions for use, and provision of additional support with instructional videos. Modifications should always be the responsibility of the manufacturer. Methods Search strategy and selection criteria This systematic review and meta-analysis followed the PRISMA standards (appendix pp 1–2). We searched PubMed, PopLine, and Embase for studies published from Jan 1, 1995, to April 30, 2016. We also reviewed six electronic HIV/AIDS conference databases (ie, Conference on Retroviruses and Opportunistic Infections, International AIDS Conference, International AIDS Society, American Public Health Association, National HIV Prevention Conference, and the HIV Diagnostics Conference) for all available years (appendix p 3). We searched for grey literature through Google Scholar (first 100 titles of 201 results). We screened bibliographies of included articles and purposely selected and contacted experts (ie, academic researchers with ongoing studies on HIV self-testing) to identify additional sources. We contacted authors of relevant studies (up to two attempts) to retrieve relevant study information. We placed no language, age, study type, or geographical limitations on the search.

sely selected and contacted experts (ie, academic researchers with ongoing studies on HIV self-testing) to identify additional sources. We contacted authors of relevant studies (up to two attempts) to retrieve relevant study information. We placed no language, age, study type, or geographical limitations on the search. We included studies reporting the performance of rapid diagnostic tests by self-testers and those reporting the concordance or the sensitivity and specificity of rapid diagnostic tests compared with the results of testing done by a health-care worker. Two reviewers (CF, CJ) screened records independently and resolved disagreements through discussion and consensus. We defined HIV self-testing as a process where an individual collects his or her specimen, does a test, and interprets their own test result.1 In the directly assisted approach, self-testers received an in-person demonstration of how to do the test or to interpret the test result; in the unassisted approach, self-testers were provided only with manufacturers' instructions for use included in the kit. All self-testers, irrespective of type of approach used, could access or receive assistance over the phone, through the internet, or with additional instructions (eg, videos, animations, or diagrams).10 We did not consider HIV counselling, linkage to care, and referral information as part of HIV self-testing assistance.10 We considered observed studies when participants were directly observed or video recorded to evaluate their HIV self-testing performance.

itional instructions (eg, videos, animations, or diagrams).10 We did not consider HIV counselling, linkage to care, and referral information as part of HIV self-testing assistance.10 We considered observed studies when participants were directly observed or video recorded to evaluate their HIV self-testing performance. We excluded studies reporting on home specimen collection, concordance or sensitivity and specificity of self-testing, or self-monitoring devices for conditions other than HIV. We defined the testing strategy used to establish the reference result as any testing sequence used to identify HIV infection (appendix pp 4–7). We classified testing strategies as aligned or not aligned with WHO guidance on the basis of the 2015 Consolidated guidelines on HIV testing services.11 Data analysis We defined measures of concordance (inter-reader reliability) as the percentage agreement and Cohen's κ12 between the health-care worker and the self-tester. We defined measurements of accuracy as specificity and sensitivity. HIV positivity among participants was based on the number of HIV-positive participants with known status or who received an HIV-positive diagnosis during the study. HIV positivity was then categorised as high (≥5%) or low (<5%).11 Given the imperfect or absent reference standards among studies, to evaluate performance of HIV rapid diagnostic tests used by self-testers we first assessed whether the result of the index and the reference test agreed or disagreed.13 We then calculated the raw estimates of sensitivity and specificity.

We defined measurements of accuracy as specificity and sensitivity. HIV positivity among participants was based on the number of HIV-positive participants with known status or who received an HIV-positive diagnosis during the study. HIV positivity was then categorised as high (≥5%) or low (<5%).11 Given the imperfect or absent reference standards among studies, to evaluate performance of HIV rapid diagnostic tests used by self-testers we first assessed whether the result of the index and the reference test agreed or disagreed.13 We then calculated the raw estimates of sensitivity and specificity. We extracted data for true reactive, true non-reactive, false reactive, and false non-reactive results to calculate κ and raw estimates of sensitivity and specificity, and explored the effect for oral fluid and blood-based rapid diagnostic tests separately, by type of assistance (direct assistance or no assistance) and type of observation. We calculated raw estimates of sensitivity and specificity with Meta-DiSc software;14 we did not consider invalid or indeterminate values to avoid skewness of results.

ct for oral fluid and blood-based rapid diagnostic tests separately, by type of assistance (direct assistance or no assistance) and type of observation. We calculated raw estimates of sensitivity and specificity with Meta-DiSc software;14 we did not consider invalid or indeterminate values to avoid skewness of results. We determined quality of studies using the Standards for Reporting Studies of Diagnostic Accuracy (STARD) checklist,15 and the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2).16 We considered high risk of partial verification bias if more than 10% of study participants did not have their HIV test results and status confirmed, and if the selection of patients to receive the reference standard was not randomised. We considered a study to have a high risk of differential verification bias if more than 10% of patients received testing with a different reference standard. CF scored studies for quality in terms of risk of bias and concerns regarding applicability. Given the high study variability and the inclusion of multiple reference standards, we pooled κ estimates using a random-effects model with the R package metaphor, version 3.4.4.13 We assessed heterogeneity by visual inspection of forest plots and calculation of the I2 statistic (>25–50% moderate).17

We determined quality of studies using the Standards for Reporting Studies of Diagnostic Accuracy (STARD) checklist,15 and the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2).16 We considered high risk of partial verification bias if more than 10% of study participants did not have their HIV test results and status confirmed, and if the selection of patients to receive the reference standard was not randomised. We considered a study to have a high risk of differential verification bias if more than 10% of patients received testing with a different reference standard. CF scored studies for quality in terms of risk of bias and concerns regarding applicability. Given the high study variability and the inclusion of multiple reference standards, we pooled κ estimates using a random-effects model with the R package metaphor, version 3.4.4.13 We assessed heterogeneity by visual inspection of forest plots and calculation of the I2 statistic (>25–50% moderate).17 Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, writing of the report, or the decision to submit for publication. The corresponding author had full access to all study data and final responsibility for the decision to submit for publication.

f the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, writing of the report, or the decision to submit for publication. The corresponding author had full access to all study data and final responsibility for the decision to submit for publication. Results After screening and removing duplicates, we included 25 studies4, 6, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40 in the review (figure 1). All studies4, 6, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40 evaluated concordance, 15 studies4, 6, 18, 19, 20, 22, 26, 27, 28, 30, 31, 32, 34, 36, 38 evaluated sensitivity and specificity, and one25 only evaluated sensitivity. 15 studies4, 6, 18, 20, 24, 26, 28, 30, 31, 32, 33, 34, 35, 36, 38 used oral fluid-based rapid diagnostic tests, six21, 22, 25, 27, 37, 40 used blood-based rapid diagnostic tests, and four19, 23, 29, 39 used both.Figure 1 Study selection

6, 38 evaluated sensitivity and specificity, and one25 only evaluated sensitivity. 15 studies4, 6, 18, 20, 24, 26, 28, 30, 31, 32, 33, 34, 35, 36, 38 used oral fluid-based rapid diagnostic tests, six21, 22, 25, 27, 37, 40 used blood-based rapid diagnostic tests, and four19, 23, 29, 39 used both.Figure 1 Study selection 13 studies4, 19, 22, 23, 24, 25, 26, 27, 28, 31, 32, 33, 39 reported on unassisted HIV self-testing, 116, 18, 20, 29, 30, 34, 35, 36, 37, 38, 40 reported on directly assisted HIV self-testing, and one21 reported on both approaches (table 1). 23 of 25 studies were observational in design (three cohort,4, 6, 19 18 cross-sectional,21, 22, 23, 24, 25, 26, 27, 28, 29, 32, 33, 34, 35, 36, 37, 38, 39, 40 and two cross-sectional and qualitative20, 30), and two were randomised controlled trials.18, 31 Sample size varied from 22 to 5662 participants. HIV positivity among participants was available in 28 reports from 22 studies; 19 (68%) of 28 reports had a high HIV positivity,4, 6, 18, 20, 22, 25, 26, 27, 28, 29, 30, 31, 32, 36, 39 and eight (29%) had low HIV positivity.4, 19, 23, 24, 33, 34, 35, 38 Reference test strategy was not available in five of 25 studies, or was not aligned with WHO testing guidance in another five studies (appendix pp 4–7). 16 (64%) of 25 studies were considered to be at low risk of bias and applicability across all key domains for QUADAS-2 (appendix p 8). 17 (68%) of 25 studies also failed to fulfil at least 60% of the STARD criteria, with a mean of 16·2 available items out of 34 (appendix pp 9–34).Table 1 Characteristics of included studies

64%) of 25 studies were considered to be at low risk of bias and applicability across all key domains for QUADAS-2 (appendix p 8). 17 (68%) of 25 studies also failed to fulfil at least 60% of the STARD criteria, with a mean of 16·2 available items out of 34 (appendix pp 9–34).Table 1 Characteristics of included studies Setting Type of RDT specimen Sample size Male participants Type of population Age Participants ever tested for HIV Education Study design Directly assisted studies Prazuck et al (2016)37 France, urban Blood based 411* 54·5% GP (100%) .. 78·6% (367/411) .. Cross-sectional Majam et al (2016)40 South Africa, urban Blood based 60 46·7% GP (100%) .. .. 33% primary, 34% secondary, 33% tertiary Cross-sectional MacGowan et al (2014)29 USA, urban Both 22 100% KP (100%) .. .. 45% (10/22) college graduate or higher, 41% (9/22) some college, 14% (3/22) less than college Cross-sectional Choko et al (2015)6 Malawi, urban Oral fluid based 1649 .. GP (91·4%), PLHIV (8·5%) .. .. .. Cohort Choko et al (2011)20 Malawi, urban Oral fluid based 283 48·1% GP (92·7%), PLHIV (7·3%) 27 years (IQR 22–32) 62% (175/283) 40·3% (114/283) primary or less, 59·7% (169/283) higher than primary education Cross-sectional and qualitative Marley et al (2014)30 China, urban Oral fluid based 229 .. GP (100%), VCT clients .. .. .. Cross-sectional and qualitative Martínez Pérez et al (2016)36 South Africa, rural Oral fluid based 2198 33·7% GP (84·7%), PLHIV (15·3%) 27·5 years (IQR 22–36) 94·1% (2068/2198) .. Cross-sectional Sarkar et al (2016)38 India, rural Oral fluid based 202 0 Pregnant women (100%) .. .. .. Cross-sectional Pant Pai et al (2013)34 South Africa, urban Oral fluid based 251 21·1% HCW (100%) ..

th Africa, rural Oral fluid based 2198 33·7% GP (84·7%), PLHIV (15·3%) 27·5 years (IQR 22–36) 94·1% (2068/2198) .. Cross-sectional Sarkar et al (2016)38 India, rural Oral fluid based 202 0 Pregnant women (100%) .. .. .. Cross-sectional Pant Pai et al (2013)34 South Africa, urban Oral fluid based 251 21·1% HCW (100%) .. 86·8% (218/251) 59·8% (150/251) high school or less, 24·3% (61/251) college or technical school, 10·3% (26/251) university or higher, 4·8% (12/251) other Cross-sectional Pant Pai et al (2014)35 Canada, urban Oral fluid based 145 38·6% GP (100%) 22 years 49·4% (124/145) College 20·6% (30/145), vocational or trade school 13·1% (19/145), university or higher 66·2% (96/145) Cross-sectional Asiimwe et al (2014; observed arm)18 Uganda, rural Oral fluid based 123 62·6% GP (100%) 27 years (IQR 22–32) 78·1% (96/123) 70·7% (87/123) less than primary, 21·1% (26/123) primary complete, 8·1% (10/123) secondary or higher Randomised controlled trial Asiimwe et al (2014; unobserved arm)18 Uganda, rural Oral fluid based 123 52·1% GP (100%) 28 years (IQR 23–32) 78·9% (97/123) 75·6% (93/123) less than primary, 13·8% (17/123) primary complete, 10·6% (13/123) secondary or higher Randomised controlled trial Unassisted studies Lee et al (2007)27 Singapore, urban Blood based 350 89·4% GP (74·9%), PLHIV (25·1%) 33 years (IQR 27–41) 74·8% (262/350) 12% (40/350) primary, 28% (97/350) secondary, 60% (210/350) at least tertiary education Cross-sectional Gras et al (2014)25 France, urban Blood based 40 75·0% PLHIV (100%) .. .. 32·5% (13/40) primary, 35% (14/40) secondary, 32·5% (13/40) tertiary education Cross-sectional Dong et al (2014)22 South Africa, rural Blood based 233 28·8% GP (100%) .. 89·3% (208/233) Less than high school 63·5% (148/233), high school 29·2% (68/233), some tertiary education 7·3% (17/233) Cross-sectional Chavez et al (2016; oral fluid arm)19 USA, urban Both 818† 100% .. 27 years (range 18–54) 82% (671/818) 86% some college Cohort Gaydos et al (2011; oral fluid arm)23 USA, urban Oral fluid based 433 41·3% GP (100%) 38·5 years (12·7) .. .. Cross-sectional Gaydos et al (2011; blood-based arm)23 USA, urban Blood based 45 42·2% GP (100%) 37·2 years (13·0) .. .. Cross-sectional Spielberg et al (2003)39 USA, urban Both 340 .. PLHIV (100%) .. .. .. Cross-sectional Gaydos et al (2013)24 USA, urban Oral fluid based 467 40·4% GP (100%) 41 years .. ..

(12·7) .. .. Cross-sectional Gaydos et al (2011; blood-based arm)23 USA, urban Blood based 45 42·2% GP (100%) 37·2 years (13·0) .. .. Cross-sectional Spielberg et al (2003)39 USA, urban Both 340 .. PLHIV (100%) .. .. .. Cross-sectional Gaydos et al (2013)24 USA, urban Oral fluid based 467 40·4% GP (100%) 41 years .. .. Cross-sectional Kurth et al (2016)26 Kenya, urban Oral fluid based 239 67·4% GP (100%) 35·9 years (9·7) 90·7% (217/239) 12·04 years of education (3·13) Cross-sectional Li et al (2016)28 China, urban Oral fluid based 200 100% KP (100%) 29·6 years (8·66) 10% (10/200) primary or less; 44·5% (89/200) secondary, 45·5% (91/200) tertiary education Cross-sectional Nour et al (2012)33 USA, urban Oral fluid based 249 42·2% GP (100%) 41 years 0 (0/249)‡ .. Cross-sectional Mavedzenge et al (2015; urban arm)31 Zimbabwe, urban Oral fluid based 172 47·0% GP (91·1%), PLHIV (8·9%) 30 years (range 18–70) 80% (138/172) .. Randomised controlled trial Mavedzenge et al (2015; urban arm)31 Zimbabwe, rural Oral fluid based 62 47·0% GP (91·1%), PLHIV (8·9%) 29 years (range 18–70) 89% (55/62) .. Randomised controlled trial Ng et al (2012)32 Singapore, urban Oral fluid based 994 88·5% GP (63·7%), PLHIV (20%), KP (6·3%) 32·4 years (IQR 27·1–40·5) .. 32·8% (326/994) less than high school, 29·8% (296/994) high school, 37·4% (372/994) at least college Cross-sectional FDA phase 2b (2012; observed arm)4 USA, urban Oral fluid based 1031 66·9% GP (42·4%), PLHIV (51·3%) KP (6·3%) .. .. 19·1% (197/1031) low literate; 45·6% (470/1031) high school or less Cohort FDA phase 3 (2012; unobserved arm)4 USA, urban Oral fluid based 5662§ 51·3% GP (86·9%), KP (13·1%) .. .. Low literate 28·0% (1624/5662); high school or less 54·9% (3113/5662) Cohort Directly assisted and unassisted studies de la Fuente et al (2012; directly assisted arm)21 Spain, urban Blood based 208 58·2% GP (63·8%), KP (36·2%) .. 39·9% (83/208) 57·2% (119/208) at least university, 41·3% (86/208) less than university Cross-sectional de la Fuente et al (2012; unassisted arm)21 Spain, urban Blood based 313 70·5% GP (63·8%), KP (36·2%) .. 51·1% (160/313) 48·5% (150/313) at least university, 51·5% (159/313) less than university Cross-sectional Data are n, %, mean (SD), median (IQR), median, mean (range), or % (n/N). RDT=rapid diagnostic test. GP=general population. KP=key population (men who have sex with men, sex workers, people who inject drugs, transgender people, and people in prisons or closed settings). PLHIV=people living with HIV.

ity Cross-sectional Data are n, %, mean (SD), median (IQR), median, mean (range), or % (n/N). RDT=rapid diagnostic test. GP=general population. KP=key population (men who have sex with men, sex workers, people who inject drugs, transgender people, and people in prisons or closed settings). PLHIV=people living with HIV. VCT=voluntary counselling and testing. HCW=health-care worker. * Study was divided into two substudies: 264 participants performed the self-test, and 147 participants interpreted contrived pictures. † 515 participants had all three results (both self-tests and dried blood home collection), 622 reported the oral fluid-based result, 565 reported the blood-based result, and 548 had the dried blood spot cards processed. ‡ In the previous 6 months. § 163 participants had no self-test results. Of the 25 studies evaluating concordance between the result of an HIV rapid diagnostic test used by a self-tester compared with a result obtained by a health-care worker, 184, 6, 18, 19, 20, 21, 22, 23, 24, 28, 29, 31, 33, 34, 35, 37, 39, 40 reported raw percentage of agreement, three26, 27, 32 reported a κ statistic, and four23, 30, 36, 38 reported both (table 2).Table 2 HIVST concordance, reasons for disagreement, and errors in performance among studies (n=25)

by a health-care worker, 184, 6, 18, 19, 20, 21, 22, 23, 24, 28, 29, 31, 33, 34, 35, 37, 39, 40 reported raw percentage of agreement, three26, 27, 32 reported a κ statistic, and four23, 30, 36, 38 reported both (table 2).Table 2 HIVST concordance, reasons for disagreement, and errors in performance among studies (n=25) HIVST concordance* Reasons for disagreement† HIV positivity Type of observation Errors in performance Invalid results (invalid result/tests performed) Reasons for invalid result Directly assisted studies Prazuck et al (2016)‡37 97·1% (142/147) Non-reactive as reactive 2·7% (4/147), invalid as reactive 2·7% (4/147) or non-reactive 2·7% (4/147), non-reactive as invalid 1·4% (2/147) .. Observed .. 1% (2/264) .. Majam et al (2016)40 88% (53/60) Non-reactive as reactive 1·7% (1/60), non-reactive as invalid 1·7% (1/60), reactive as non-reactive 1·7% (1/60), invalid 6·7% (4/60) as reactive or non-reactive .. Observed 20 participants made mistakes; common errors were with blood collection and transferring and use of buffer .. .. MacGowan et al (2014; oral fluid arm)29 95% (21/22) Reactive as non-reactive 4·5% (1/22) 22·7% (5/22) Observed 13·6% (3/22) participants made mistakes, common errors were spilling buffer and incorrect time to read the results 4·5% (1/22) .. MacGowan et al (2014; blood-based arm)29 95% (20/21) One HIV-positive participant with an invalid result interpreted his result as reactive 4·8% (1/21) 19% (4/21) 23·8% (5/21) participants made mistakes; common errors were incorrectly pushing the device into test holder and incorrect timing to read the results; one participant broke the device 9·5% (2/21) Operational error Choko et al (2015)§6 99·4%, 98·9%–99·7% (1639/1649) Reactive as non-reactive 0·5% (9/1649), non-reactive as reactive 0·06% (1/1649) 8·6% (141/1649) Observed .. .. .. Choko et al (2011)¶20 99·2%, 97–100% (256/258)‖ One HIV-positive participant with a faint reactive result interpreted his result as uncertain, one HIV-positive participant had an invalid result 16·9% (48/283) Non-observed Common errors were touching collection pad, incorrect or incomplete swabbing, removing kit from developer too early, buffer spills, reading incorrectly, and fumbling vial or cap when opening developer fluid 0·4% (1/260) ..

sult as uncertain, one HIV-positive participant had an invalid result 16·9% (48/283) Non-observed Common errors were touching collection pad, incorrect or incomplete swabbing, removing kit from developer too early, buffer spills, reading incorrectly, and fumbling vial or cap when opening developer fluid 0·4% (1/260) .. Marley et al (2014)30 93·9% (215/229), κ 0·551, p=0·012 Reactive as invalid 3·1% (7/229), non-reactive as reactive 1·3% (3/229), invalid as non-reactive 1·3% (3/229), non-reactive as invalid 0·4% (1/229) 5·6% (13/229) Observed Common errors were unpreparedness before start 42% (94/229), inability to swab correctly 10% (23/229), buffer 15·9% (36/229), testing and reading test results 7·5% (17/229) 3·5% (8/229) Six participants used test paper incorrectly Martínez Pérez et al (2016)36 99·4% (2184/2198), κ 0·99** Reactive as non-reactive 0·2% (4/2181) 15·3% (337/2198) Observed Two participants had to repeat the self-test, they accidentally spilled buffer vial; excluding known people living with HIV, 0·18% (4/2181) interpreted their tests as negative whereas the HCW interpreted the result as positive 0·5% (11/2198) .. Sarkar et al (2016)38 98%, κ 0·566, p<0·001 Invalid as non-reactive 0·5% (1/202), non-reactive as invalid 0·9% (2/202) 0·9% (2/202) Observed .. 0·9% (2/202) .. Pant Pai (2013)34 98·8% (248/251) Reactive as non-reactive 1·2% (3/251), two of which had a faint reactive line 3·6% (9/251) Non-observed Errors were in conducting and interpreting results .. .. Pant Pai et al (2014)35 100% (145/145) No difference between self-tester and HCW interpretation 0 Non-observed .. .. .. Asiimwe et al (2014; observed arm)18 99·2% (122/123) Non-reactive as invalid 0·8% (1/123) 10·6% (13/123) Observed 19·5% (24/123) participants made mistakes; common errors were incorrect swabbing of gums, touching the collection pad and buffer spills 0·8% (1/123) .. Asiimwe et al (2014; unobserved arm)18 94·3% (116/123) .. 16·3% (20/123) Non-observed No errors were reported 0·8% (1/117)†† Unassisted studies Lee et al (2007)27 κ 0·277, p<0·001 Invalid as non-reactive 50·1% (176/350), invalid as reactive 4·6% (16/350) and reactive as non-reactive 0·3% (1/350) 25% (88/350) Observed .. 56·3% (197/350) 85% failed to perform all steps correctly Gras et al (2014)25 100% No difference between self-tester and HCW interpretation 100% (40/40) Observed Common errors were insufficient blood, wrong lancet utilisation, and mixing of samples 5·7% (2/35) ..

ve as non-reactive 0·3% (1/350) 25% (88/350) Observed .. 56·3% (197/350) 85% failed to perform all steps correctly Gras et al (2014)25 100% No difference between self-tester and HCW interpretation 100% (40/40) Observed Common errors were insufficient blood, wrong lancet utilisation, and mixing of samples 5·7% (2/35) .. Dong et al (2014)22 98·7% (230/233) Reactive as non-reactive 0·5% (1/195), invalid as non-reactive 0·5% (1/195), non-reactive as invalid 0·5% (1/195) 18·9% (44/233) Observed (video recorded) .. 0·4% (1/233) .. Chavez et al (2016; oral fluid arm)‡‡19 98% (500/511) Non-reactive as reactive 1·4%, non-reactive as invalid 0·8% 2% (11/622) Non-observed .. .. .. Chavez et al (2016; blood-based arm)‡‡19 99% (506/511) Non-reactive as reactive 0·6%, non-reactive as invalid 0·4% 1% (7/565) 4·6% (26/565) Operational error Gaydos et al (2011)§§23 99·6%, 0·41–1·00 (476/478) weighted κ 0·75 Reactive as non-reactive 0·2% (1/478) 0·8% (4/478) Observed Difficulties were interpreting results, reading result chart, reading or following instructions, swabbing or pricking properly, or both, and opening the kit 0·2% (1/478) Insufficient blood Spielberg et al (2003; oral fluid arm)39 95% .. 100% (340/340) Non-observed Difficulties performing test decreased through changes made to instructions and labelling from 4·3% to 4% 4·1% (14/340) Failure to put the test device in the vial with developer solution Spielberg et a; (2003; blood-based arm)39 97% 100% (340/340) Difficulties performing test decreased through changes made to instructions and labelling from 14% to 9% 7·9% (27/340) Gaydos et al (2013)24 100% No difference between self-tester and HCW interpretation 0·2% (1/467) Observed .. .. .. Kurth et al (2016)26 κ 0·92 (0·84–0·99) Non-reactive as invalid 12·5% (30/239), reactive as non-reactive 1·2% (3/239), non-reactive as reactive 0·4% (1/239) 14·6% (35/239) Observed (video recorded) Common errors were difficulty opening bottle, incorrect or incomplete swab of gums, and incorrect time to read the results; some individuals could have made multiple errors 15·1% (36/239) All individuals recognised something went wrong with their test Li et al (2016)28 95% (190/200) Non-reactive as invalid 2·5% (5/200), reactive as non-reactive 1·5% (3/200), non-reactive as reactive 0·5% (1/200) 27·5% (55/200) Observed Common errors were incorrect or incomplete swab of gums, incorrect time to read the results, touching the collection pad, and buffer spills 3% (6/200) ..

t al (2016)28 95% (190/200) Non-reactive as invalid 2·5% (5/200), reactive as non-reactive 1·5% (3/200), non-reactive as reactive 0·5% (1/200) 27·5% (55/200) Observed Common errors were incorrect or incomplete swab of gums, incorrect time to read the results, touching the collection pad, and buffer spills 3% (6/200) .. Nour et al (2012)33 100% No difference between self-tester and HCW interpretation 1·6% (4/249) Observed .. .. .. Mavedzenge et al (2015; urban arm)¶¶31 93% (160/172) Non-reactive as invalid 2% (3/172) 9% (16/172) Observed (video recorded) Common errors were confusion with desiccant, buffer spills, dipping test device in developer before collecting sample, incorrect sampling, and incorrect time to read the results. 2·9% (5/172) Participants with invalid results typically did not follow instructions Mavedzenge et al (2015; rural arm)31 90% (56/62) Non-reactive as reactive 4·8% (3/62) 8% (5/62) 3·2% (2/62) Ng et al (2012)32 κ 0·97, 0·95–0·99 Reactive as non-reactive 2·6% (5/983), reactive as invalid 0·5% (1/983), non-reactive as invalid 0·3% (2/983) and non-reactive as reactive 0·1% (1/983) 19·3% (192/994) Observed Common errors were incorrect or incomplete swab of gums, touching the collection pad during removal from packaging, or buffer spills 0·3% (3/983) .. FDA phase 2b (2012; observed arm)‖‖4 93% (942/1013) Reactive as non-reactive 0·9% (10/1013), non-reactive as reactive 0·1% (1/1013) 2·1% (120/5662) Observed Common errors were interpreting results (11/986), dipping device in developer prior to swabbing gums (11/986), buffer spills (4/986), incorrect swabbing (5/986), and could not find developer (2/986) 3·3% (33/986) Operational errors FDA phase 3 (2012; unobserved arm)***4 99·8% (5490/5499) Reactive as non-reactive 0·1% (8/5499), non-reactive as reactive 0·01% (1/5499) 51% (526/1031) Non-observed Not understanding where to place the test stick after sample collection (1/4999) 0·6% (31/4999) .. Directly assisted and unassisted studies de la Fuente et al (2012; directly assisted arm)21 85·4% (445/521) Invalid as reactive 2·8%, non-reactive as reactive 2·7%, non-reactive as invalid 2%, reactive as invalid 1·9%, invalid as non-reactive 1·5% and reactive as non-reactive 1·1% .. Observed ..

4999) 0·6% (31/4999) .. Directly assisted and unassisted studies de la Fuente et al (2012; directly assisted arm)21 85·4% (445/521) Invalid as reactive 2·8%, non-reactive as reactive 2·7%, non-reactive as invalid 2%, reactive as invalid 1·9%, invalid as non-reactive 1·5% and reactive as non-reactive 1·1% .. Observed .. 0·9% (2/208) Most difficult step was obtaining blood and depositing it in the correct place de la Fuente et al (2012; directly assisted arm)21 85·4% (445/521) Invalid as reactive 2·8%, non-reactive as reactive 2·7%, non-reactive as invalid 2·1%, reactive as invalid 1·9%, invalid as non-reactive 1·5% and reactive as non-reactive 1·1% .. Observed .. 0·9% (2/208) Most difficult step was obtaining blood and depositing it in the correct place Data are % (n/N); %, 95% CI, (n/N), κ, p value; or κ (95%CI). HIVST=HIV self-test. HCW=health-care worker. FDA=Food and Drug Administration. * Reported as percentage of agreement or κ. † Reason for disagreement assumes the self-tester perspective compared with the HCW. ‡ The study was divided into two substudies: 264 participants performed the self-test, and 147 participants interpreted contrived pictures. § Four participants were on antiretrovirals. ¶ 260 of 283 participants self-tested. ‖ Two participants had no confirmatory results. ** 17 known people living with HIV were not considered to calculate the κ. †† Six participants had no results.

‡ The study was divided into two substudies: 264 participants performed the self-test, and 147 participants interpreted contrived pictures. § Four participants were on antiretrovirals. ¶ 260 of 283 participants self-tested. ‖ Two participants had no confirmatory results. ** 17 known people living with HIV were not considered to calculate the κ. †† Six participants had no results. ‡‡ 515 participants had all three results (both self-tests and dried blood home collection [dried blood spot]), 622 reported the oral fluid-based result, 565 reported the blood-based result, and 548 had the dried blood spot cards processed. §§ Disaggregated results by type of specimen were not available. One participant was on antiretrovirals with undetectable viral load. ¶¶ One participant in the urban arm was on antiretrovirals. ‖‖ 1013 of 1031 participants completed the study. *** 18 positives and 482 negatives were excluded from the accuracy analysis.

‡‡ 515 participants had all three results (both self-tests and dried blood home collection [dried blood spot]), 622 reported the oral fluid-based result, 565 reported the blood-based result, and 548 had the dried blood spot cards processed. §§ Disaggregated results by type of specimen were not available. One participant was on antiretrovirals with undetectable viral load. ¶¶ One participant in the urban arm was on antiretrovirals. ‖‖ 1013 of 1031 participants completed the study. *** 18 positives and 482 negatives were excluded from the accuracy analysis. Reported κ ranged from fair (κ 0·277, p<0·001) to almost perfect (κ 0·99).23, 26, 27, 30, 32, 36, 38 The raw proportion of agreement was high, ranging from 85·4% to 100%. Overall, our estimates of pooled agreement across studies were almost perfect for both types of approaches (directly assisted κ 0·98, 95% CI 0·96 to 0·99; unassisted κ 0·97, 0·96 to 0·98; I2=34·5%, 0 to 97·8; figure 2). Pooled estimates according to whether HIV self-testing was observed or not also had almost perfect agreement (observed 0·98, 0·96 to 0·99; unobserved 0·96, 0·94 to 0·99; I2=43·0%, 38·8 to 98·4; figure 2). The lowest estimated agreement (κ 0·47, −0·04 to 0·97) was in rural Zimbabwe; the study investigators attributed poor performance to low literacy in the population tested, and verbose instructions that needed further optimisation.31Figure 2 Cohen's κ across studies by method of observation (A) and type of approach (B) TR=true reactive result. FR=false reactive result. FN=false non-reactive result. TN=true non-reactive result.

Reported κ ranged from fair (κ 0·277, p<0·001) to almost perfect (κ 0·99).23, 26, 27, 30, 32, 36, 38 The raw proportion of agreement was high, ranging from 85·4% to 100%. Overall, our estimates of pooled agreement across studies were almost perfect for both types of approaches (directly assisted κ 0·98, 95% CI 0·96 to 0·99; unassisted κ 0·97, 0·96 to 0·98; I2=34·5%, 0 to 97·8; figure 2). Pooled estimates according to whether HIV self-testing was observed or not also had almost perfect agreement (observed 0·98, 0·96 to 0·99; unobserved 0·96, 0·94 to 0·99; I2=43·0%, 38·8 to 98·4; figure 2). The lowest estimated agreement (κ 0·47, −0·04 to 0·97) was in rural Zimbabwe; the study investigators attributed poor performance to low literacy in the population tested, and verbose instructions that needed further optimisation.31Figure 2 Cohen's κ across studies by method of observation (A) and type of approach (B) TR=true reactive result. FR=false reactive result. FN=false non-reactive result. TN=true non-reactive result. The proportion of disagreements, assuming the self-tester perspective compared with health-care worker, ranged from 0% to 14·6%. Across 29 reports from 25 studies, four reports24, 25, 33, 35 found no difference in interpretation between self-testers and health-care workers. Most reported differences resulted from the interpretation of a reactive result as non-reactive (0·01–4·8%, 13 of 29 reports), a reactive result as invalid (2·7–6·7%, five of 29), a non-reactive result as reactive (0·1–4·5%, 14 of 29), an invalid result as reactive (0·5–3·1%, three of 29), an invalid result as non-reactive (0·3–12·5%, 13 of 29), or a non-reactive result as invalid (0·5–50%, seven of 29).

reactive (0·01–4·8%, 13 of 29 reports), a reactive result as invalid (2·7–6·7%, five of 29), a non-reactive result as reactive (0·1–4·5%, 14 of 29), an invalid result as reactive (0·5–3·1%, three of 29), an invalid result as non-reactive (0·3–12·5%, 13 of 29), or a non-reactive result as invalid (0·5–50%, seven of 29). Reasons for disagreements were higher in directly assisted studies (2·7–6·7%) than unassisted studies (4·6%) when interpreting a reactive result as invalid, and were higher in unassisted studies (0·01–4·8%) than directly assisted studies (0·06–2·7%) when interpreting a reactive result as non-reactive.

reactive (0·01–4·8%, 13 of 29 reports), a reactive result as invalid (2·7–6·7%, five of 29), a non-reactive result as reactive (0·1–4·5%, 14 of 29), an invalid result as reactive (0·5–3·1%, three of 29), an invalid result as non-reactive (0·3–12·5%, 13 of 29), or a non-reactive result as invalid (0·5–50%, seven of 29). Reasons for disagreements were higher in directly assisted studies (2·7–6·7%) than unassisted studies (4·6%) when interpreting a reactive result as invalid, and were higher in unassisted studies (0·01–4·8%) than directly assisted studies (0·06–2·7%) when interpreting a reactive result as non-reactive. Across 20 reports from 16 studies,4, 6, 18, 19, 20, 22, 25, 26, 27, 28, 30, 31, 32, 34, 36, 38 16 (80%) of 20 reports had specificity of more than 98%. Sensitivity varied substantially; 18 (90%) of 20 reports had sensitivity of at least 80%. Two studies reported sensitivity of less than 80%: one34 had insufficient information on how to interpret faint positive lines, and the other31 suggested lengthy instructions were a barrier to participants in the rural arm, in which literacy levels were lower than the urban arm. Excluding these two studies,31, 34 sensitivity estimates were higher for blood-based rapid diagnostic tests (96·2–100%)19, 22, 25, 27 than oral fluid-based rapid diagnostic tests (80–100%),4, 6, 18, 20, 26, 28, 30, 31, 32, 36, 38 as were specificity estimates (blood-based 99·5–100% vs oral fluid 95·1–100%). Studies4, 6, 18, 22, 25, 26, 27, 28, 30, 31, 32, 36, 38 in which testing was observed reported a modest difference in sensitivity (80–100%) compared with unobserved studies4, 18, 19, 20, 34 (88·9–100%; table 3).Table 3 Sensitivity and specificity of RDTs used for self-testing (n=16) by type of observation and approach

Studies4, 6, 18, 22, 25, 26, 27, 28, 30, 31, 32, 36, 38 in which testing was observed reported a modest difference in sensitivity (80–100%) compared with unobserved studies4, 18, 19, 20, 34 (88·9–100%; table 3).Table 3 Sensitivity and specificity of RDTs used for self-testing (n=16) by type of observation and approach Sensitivity TR/(TR+FN) Specificity TN/(TN+FR) HIV positivity Type of population Unobserved studies Pant Pai et al (2013)*34 66·7% (29·9–92·5) 6/(6+3) 100% (98·5–100) 242/(242+0) 3·6% (9/251) HCW (100%) Asiimwe et al (2014; unobserved arm)*18 90·0% (68·3–98·8) 18/(18+2) 95·1% (89·0–98·4) 98/(98+5) 17·2% (20/116) GP (100%) Chavez et al (2016; blood-based arm)†19 100% (54·1–100) 6/(6+0) 100% (99·2–100) 486/(486+0) 1·7% (9/515) KP (100%) Chavez et al (2016; oral fluid arm)*19 88·9% (51·8–99·7) 8/(8+1) 100% (99·3–100) 501/(501+0) 1·7% (9/515) KP (100%) FDA phase 3 (2012)*4 91·7% (84·2–96·3) 88/(88+8) 100% (99·9–100) 4902/(4902+1) 1·9% (96/4903) GP (86·9%), KP (13·1%) Choko et al (2011)*20 97·9% (88·9–99·9) 47/(47+1) 100% (98·3–100) 210/(210+0) 16·9% (48/283) GP (100%) Observed studies Gras et al (2014)†25 96·2% (80·4–99·9) 25/(25+1) .. ..

01+0) 1·7% (9/515) KP (100%) FDA phase 3 (2012)*4 91·7% (84·2–96·3) 88/(88+8) 100% (99·9–100) 4902/(4902+1) 1·9% (96/4903) GP (86·9%), KP (13·1%) Choko et al (2011)*20 97·9% (88·9–99·9) 47/(47+1) 100% (98·3–100) 210/(210+0) 16·9% (48/283) GP (100%) Observed studies Gras et al (2014)†25 96·2% (80·4–99·9) 25/(25+1) .. .. 100% (26/26) PLHIV (100%) Lee (2007)†27 98·8% (93·5–100) 83/(83+1) 99·6% (97·9–100) 260/(260+1) 24·3% (84/345) GP (90%), KP (10%) Dong et al (2014)†22 97·7% (88·0–99·9) 43/(43+1) 99·5% (97·1–100) 186/(186+1) 19·0% (44/231) GP (100%) Sarkar et al (2016)*38 100% (15·8–100) 2/(2+0) 100% (98·1–100) 197/(197+0) 0·9% (2/202) Pregnant women (100%) Marley et al (2014)‡30 100% (54·1–100) 6/(6+0) 98·6% (95·9–99·7) 209/(209+3) 5·8% (13/222) GP (100%), VCT clients Choko et al (2015)§6 93·6% (88·2–97·0) 132/(132+9) 99·9% (99·6–100) 1507/(1507+1) 8·6% (141/1649) GP (100%) Asiimwe et al (2014; observed arm)*18 100% (75·3–100) 13/(13+0) 99·1% (95·0–100) 109/(109+4) 10·6% (13/122) GP (100%) Martínez Pérez et al (2016)*36 98·8% (96·9–99·7) 323/(323+4) 100% (99·8–100) 1860/(1860+0) 14·9% (327/2187) GP (100%) Li et al (2016)*28 94·4% (84·6–98·8) 51/(51+3) 99·3% (96·1–100) 139/(139+1) 28·9% (55/190) KP (100%) Kurth et al (2016)*26 89·7% (72·6–97·8) 26/(26+3) 99·4% (96·8–100) 173/(173+1) 14·3% (29/203) GP (100%) Mavedzenge et al (2015; rural arm)*31 66·7% (9·4–99·2) 2/(2+1) 94·7% (85·4–98·9) 54/(54+3) 8% (5/62) GP (100%) Mavedzenge et al (2015; urban arm)*‡31 80·0% (28·4–99·5) 4/(4+1) 97·8% (88·5–99·9) 145/(145+1) 9% (16/172) GP (100%) Ng et al (2012)*32 97·4% (94·0–99·1) 186/(186+5) 99·9% (99·3–100) 791/(791+1) 19·3% (192/994) GP (63·7%), PLHIV (20%), KP (16·3%) FDA phase 2b (2012)*4 97· 9% (96·2–99·0) 470/(470+10) 99·8% (98·8–100) 472/(472+1) 51·9% (526/1013) GP (42·4%), PLHIV (513%), KP (6·3%) Directly assisted studies Pant Pai et al (2013)*34 66·7% (29·9–92·5) 6/(6+3) 100% (98·5–100) 242/(242+0) 3·6% (9/251) HCW (100%) Sarkar et al (2016)*38 100% (15·8–100) 2/(2+0) 100% (98·1–100) 197/(197+0) 0·9% (2/202) Pregnant women (100%) Choko et al (2011)*20 97·9% (88·9–99·9) 47/(47+1) 100% (98·3–100) 210/(210+0) 16·9% (48/283) GP (100%) Choko et al (2015)*§6 93·6% (88·2–97·0) 132/(132+9) 99·9% (99·6–100) 1507/(1507+1) 8·6% (141/1649) GP (100%) Marley et al (2014)*‡30 100% (54·1–100) 6/(6+0) 98·6% (95·9–99·7) 209/(209+3) 5·8% (13/222) GP (29%) Asiimwe et al (2014; observed arm) *18 100% (75·3–100) 13/(13+0) 99·1% (95·0–100) 109/(109+4) 10·6% (13/122) GP (100%) Asiimwe et al (2014; unobserv

) 132/(132+9) 99·9% (99·6–100) 1507/(1507+1) 8·6% (141/1649) GP (100%) Marley et al (2014)*‡30 100% (54·1–100) 6/(6+0) 98·6% (95·9–99·7) 209/(209+3) 5·8% (13/222) GP (29%) Asiimwe et al (2014; observed arm) *18 100% (75·3–100) 13/(13+0) 99·1% (95·0–100) 109/(109+4) 10·6% (13/122) GP (100%) Asiimwe et al (2014; unobserv ed arm)*18 90·0% (68·3–98·8) 18/(18+2) 95·1% (89·0–98·4) 98/(98+5) 17·2% (20/116) GP (100%) Martínez Pérez et al (2016)*36 98·8% (96·9–99·7) 323/(323+4) 100% (99·8–100) 1860/(1860+0) 14·9% (327/2187) GP (100%) Unassisted studies Gras et al (2014)†25 96·2% (80·4–99·9) 25/(25+1) ..

) 132/(132+9) 99·9% (99·6–100) 1507/(1507+1) 8·6% (141/1649) GP (100%) Marley et al (2014)*‡30 100% (54·1–100) 6/(6+0) 98·6% (95·9–99·7) 209/(209+3) 5·8% (13/222) GP (29%) Asiimwe et al (2014; observed arm) *18 100% (75·3–100) 13/(13+0) 99·1% (95·0–100) 109/(109+4) 10·6% (13/122) GP (100%) Asiimwe et al (2014; unobserv ed arm)*18 90·0% (68·3–98·8) 18/(18+2) 95·1% (89·0–98·4) 98/(98+5) 17·2% (20/116) GP (100%) Martínez Pérez et al (2016)*36 98·8% (96·9–99·7) 323/(323+4) 100% (99·8–100) 1860/(1860+0) 14·9% (327/2187) GP (100%) Unassisted studies Gras et al (2014)†25 96·2% (80·4–99·9) 25/(25+1) .. .. 100% (26/26) PLHIV (100%) Lee et al (2007)†27 98·8% (93·5–100) 83/(83+1) 99·6% (97·9–100) 260/(260+1) 24·3% (84/345) GP (90%), KP (10%) Dong et al (2014)†22 97·7% (88·0–99·9) 43/(43+1) 99·5% (97·1–100) 186/(186+1) 19·0% (44/231) GP (100%) Chavez et al (2016; blood-based arm)†19 100% (54·1–100) 6/(6+0) 100% (99·2–100) 486/(486+0) 1·7% (9/515) KP (100%) Chavez et al (2016; oral fluid arm) *19 88·9% (51·8–99·7) 8/(8+1) 100% (99·3–100) 501/(501+0) 1·7% (9/515) KP (100%) Li et al (2016)*28 94·4% (84·6–98·8) 51/(51+3) 99·3% (96·1–100) 139/(139+1) 28·9% (55/190) KP (100%) Kurth et al (2016)*26 89·7% (72·6–97·8) 26/(26+3) 99·4% (96·8–100) 173/(173+1) 14·3% (29/203) GP (100%) FDA phase 3 (2012)*4 91·7% (84·2–96·3) 88/(88+8) 100% (99·9–100) 4902/(4902+1) 1·9% (96/4903) GP (86·9%), KP (13·1%) Mavedzenge et al (2015; rural arm)31 66·7% (9·4–99·2) 2/(2+1) 94·7% (85·4–98·9) 54/(54+3) 8% (5/62) GP (100%) Mavedzenge et al (2015; urban arm)*¶31 80·0% (28·4–99·5) 4/(4+1) 97·8% (88·5–99·9) 45/(45+1) 9% (16/172) GP (100%) Ng et al (2012)*32 97·4% (94·0–99·1) 186/(186+5) 99·9% (99·3–100) 791/(791+1) 19·3% (192/994) GP (63·7%), PLHIV (20%), and KP (16·3%) FDA phase 2b (2012)*4 97·9% (96·2–99·0) 470/(470+10) 99·8% (98·8–100) 472/(472+1) 51·9% (526/1013) GP (42·4%), PLHIV (51·3%), and KP (6·3%) Data are % (95% CI) or n/(n+n). TR=true reactive result. FR=false reactive result. FN=false non-reactive result. TN=true non-reactive result. HCW=health-care worker. GP=general population. KP=key population. FDA=US Food and Drug Administration. PLHIV=people living with HIV. VCT=voluntary counselling and testing

(6·3%) Data are % (95% CI) or n/(n+n). TR=true reactive result. FR=false reactive result. FN=false non-reactive result. TN=true non-reactive result. HCW=health-care worker. GP=general population. KP=key population. FDA=US Food and Drug Administration. PLHIV=people living with HIV. VCT=voluntary counselling and testing * Oral fluid-based. † Finger stick-based or whole blood-based. ‡ This study assessed accuracy in a subsample of participants (229/800). § Four participants were on antiretrovirals; they tested negative via self-test and positive in confirmatory testing. ¶ One participant was on antiretrovirals; this person tested negative via self-test and positive in confirmatory testing. Heterogeneity: sensitivity I2 55·1%; specificity I2 78·7%. Spearman correlation coefficient −0·259, p 0·285. A study31 from Zimbabwe with oral fluid-based rapid diagnostic tests, with data disaggregated by setting, found that urban populations had higher sensitivity (80%, 95% CI 28·4–99·5) than rural populations with lower literacy (66·7%, 9·4–99·2), and that this was also the case for specificity (urban 97·8%, 88·5–99·9 vs rural 94·7%, 85·4–98·9). All studies6, 18, 20, 30, 34, 36, 38 addressing directly assisted HIV self-testing used oral fluid-based rapid diagnostic tests. The estimated sensitivity was similar to that in studies4, 19, 26, 28, 31, 32 with oral fluid rapid diagnostic tests within the unassisted approach (table 3).

A study31 from Zimbabwe with oral fluid-based rapid diagnostic tests, with data disaggregated by setting, found that urban populations had higher sensitivity (80%, 95% CI 28·4–99·5) than rural populations with lower literacy (66·7%, 9·4–99·2), and that this was also the case for specificity (urban 97·8%, 88·5–99·9 vs rural 94·7%, 85·4–98·9). All studies6, 18, 20, 30, 34, 36, 38 addressing directly assisted HIV self-testing used oral fluid-based rapid diagnostic tests. The estimated sensitivity was similar to that in studies4, 19, 26, 28, 31, 32 with oral fluid rapid diagnostic tests within the unassisted approach (table 3). Three studies included some participants taking antiretroviral drugs. In two studies,6, 31 participants had non-reactive test results, but later received confirmatory testing and were diagnosed or disclosed their HIV statuses afterward. In the third study,23 self-testers and health-care workers both obtained non-reactive results because they used the same oral test. We identified 25 reports from 20 studies with information on invalid results: seven reports20, 29, 30, 31, 32, 36, 40 used the directly assisted approach, six23, 25, 26, 28, 34, 39 used the unassisted approach and two4, 18 used both approaches. Invalid results ranged from one (0·2%) of 478 tests to 197 (56·3%) of 350 tests (table 2).4, 18, 19, 20, 21, 22, 23, 25, 26, 27, 28, 29, 30, 31, 32, 36, 37, 38, 39

rts20, 29, 30, 31, 32, 36, 40 used the directly assisted approach, six23, 25, 26, 28, 34, 39 used the unassisted approach and two4, 18 used both approaches. Invalid results ranged from one (0·2%) of 478 tests to 197 (56·3%) of 350 tests (table 2).4, 18, 19, 20, 21, 22, 23, 25, 26, 27, 28, 29, 30, 31, 32, 36, 37, 38, 39 Although most participants were able to obtain a correct result, user errors among self-testers were noted in 15 reports. Of these reports, two found a high proportion of user error: one27 reported most users were unable to take blood samples or transfer blood specimens correctly (197 invalid results from 350 tests; 56·3%), and the other26 reported users were aware of making mistakes (36 invalid results from 239 tests; 15·1%). Excluding these studies, the proportion of invalid results was similar in studies20, 29, 30, 31, 32, 36, 37, 38, 40 of the directly assisted approach (0·3–9·5%) and studies19, 22, 23, 25, 28, 39 of the unassisted approach (0·2–7·9%). The proportion of invalid results was higher in studies4, 18, 21, 22, 23, 25, 28, 29, 30, 31, 32, 36, 37, 38 in which testing was observed (0·2–9·5%) when compared with unobserved studies4, 18, 19, 20, 39 (0·4–7·9%).

y assisted approach (0·3–9·5%) and studies19, 22, 23, 25, 28, 39 of the unassisted approach (0·2–7·9%). The proportion of invalid results was higher in studies4, 18, 21, 22, 23, 25, 28, 29, 30, 31, 32, 36, 37, 38 in which testing was observed (0·2–9·5%) when compared with unobserved studies4, 18, 19, 20, 39 (0·4–7·9%). The proportion of studies reporting invalid results among self-testers was greater in studies19, 21, 22, 25, 27, 29, 37, 39, 40 using blood-based rapid diagnostic tests (0·4–9·5%) than studies4, 18, 20, 23, 26, 28, 30, 31, 32, 36, 38 using oral fluid-based rapid diagnostic tests (0·2–4·5%). Excluding studies21, 25, 26, 27, 29, 39 with feasibility of less than 60%, the proportion of invalid results was less than 5% (0·2–4·6%), regardless of the approach or specimen. User errors described in studies of the directly assisted approach were incorrect or incomplete specimen collection (finger prick or oral swab),20, 21, 30, 31, 40 incorrect use or spillage of buffer,20, 29, 30, 31, 36, 40 incorrect transfer of blood specimen, and problems with the interpretation of results.4, 20, 23, 30, 34, 39, 40 Reported errors in studies of the unassisted approach included specimen collection (finger prick or oral swab),23, 26, 28 misinterpretation of test results,23, 34 incorrect time to read the results,26, 28 test kit opened incorrectly,23, 26 incorrect use or spillage of buffer,28 instructions not followed or read,23 or incorrect transfer of the blood specimen.25

howed high uptake and acceptability of HIV self-tests delivered by volunteer counsellors. We found little evidence of rigorous, randomised, controlled trials of strategies to distribute HIV self-tests. Further, there was little available evidence of secondary distribution of HIV self-tests outside of facility settings. Added value of this study This cluster randomised trial provided rigorous evidence that a 3-month intervention of door-to-door distribution of HIV self-tests increased knowledge of HIV status among adults aged 16 years and older. This effect was driven by increased knowledge of HIV status among men and young adults aged 16–29 years, with no between-group differences noted in women. This was the first trial to show that community-based door-to-door secondary distribution of HIV tests can increase HIV testing among individuals, primarily men, who are not at home during household visits from lay counsellors. Among men aged 30 years or more who HIV self-tested, more than a third self-tested through unsupervised HIV self-testing or secondary distribution of HIV self-tests. The trial also showed that the door-to-door distribution of HIV self-tests increased knowledge of HIV status among community residents whose HIV status was not known to lay counsellors despite 2 years of household delivery of HIV-related services, including HIV testing. Implications of all the available evidence

ed approach included specimen collection (finger prick or oral swab),23, 26, 28 misinterpretation of test results,23, 34 incorrect time to read the results,26, 28 test kit opened incorrectly,23, 26 incorrect use or spillage of buffer,28 instructions not followed or read,23 or incorrect transfer of the blood specimen.25 In general, reported errors in performance were similar by type of specimen; however, studies using oral fluid rapid diagnostic tests reported errors in the interpretation of test results and studies using blood-based rapid diagnostic tests reported errors in transfer of the blood specimen. Two studies4, 25 found that people with known HIV status had a higher proportion of errors (ie, when collecting the specimen) when self-testing compared with people with unknown HIV status (0·8% vs 0·2%), whereas a third study27 found that known HIV-positive participants were more likely to do the test correctly. Discussion Self-testers can achieve the same results as health-care workers when using HIV rapid diagnostic tests and diagnostic accuracy of rapid diagnostic tests for self-testing is high. Application of the estimated ranges of sensitivity (80–100%) and specificity (95·1–100%) to a hypothetical group of 100 000 people with 1% of HIV prevalence would result in 0–200 HIV-positive cases being missed, and 0–4851 HIV-negative individuals being misidentified with a reactive result, excluding two outliers.31, 34

pplication of the estimated ranges of sensitivity (80–100%) and specificity (95·1–100%) to a hypothetical group of 100 000 people with 1% of HIV prevalence would result in 0–200 HIV-positive cases being missed, and 0–4851 HIV-negative individuals being misidentified with a reactive result, excluding two outliers.31, 34 This systematic review and meta-analysis suggests that in the hands of self-testers, the sensitivity and specificity of blood-based rapid diagnostic tests were higher than oral fluid rapid diagnostic tests, although fewer studies used blood-based rapid diagnostic tests. The reduced sensitivity is probably explained by the lower quantity of HIV antibodies in oral fluid compared with whole blood, as observed in professional-use assessments.41 Although blood-based rapid diagnostic tests might have the potential to deliver more accurate results, more invalid results might occur because the greatest number of user errors was related to standard procedures when capillary tubes and pipettes were used.

h whole blood, as observed in professional-use assessments.41 Although blood-based rapid diagnostic tests might have the potential to deliver more accurate results, more invalid results might occur because the greatest number of user errors was related to standard procedures when capillary tubes and pipettes were used. Most studies had a high HIV positivity among participants where tests are expected to have a higher positive-predictive value than lower prevalence populations. Furthermore, imperfect reference standards might also decrease the degree of accuracy. We found wide variability in sensitivity estimates, which could be explained by the use of adapted rapid diagnostic tests not specifically designed for self-test use,21, 22, 23, 25, 27, 37, 39, 40 or used test kits before approval by national regulatory authorities, in 23 of 25 included studies.4, 19, 20, 24, 26, 28, 29, 30, 31, 32, 33, 34, 35, 36, 38 When we excluded six studies21, 25, 26, 27, 29, 39 with low feasibility, the proportion of invalid results met the minimum acceptable criteria for rapid diagnostic test performance (<5%);42 however, we still found no significant differences in the proportion of invalid results by type of approach, suggesting that use of a rapid diagnostic tests for self-testing without assistance will not increase the possibility of obtaining an invalid result.

table criteria for rapid diagnostic test performance (<5%);42 however, we still found no significant differences in the proportion of invalid results by type of approach, suggesting that use of a rapid diagnostic tests for self-testing without assistance will not increase the possibility of obtaining an invalid result. Most invalid results and errors in performance reported by studies included in this review related to user errors and manufacturing defects. These invalid results and errors can be mitigated with instructions for use because the complexity of the test procedure or the complexity of the instructions can increase the possibility of failure in performance and incorrect interpretation of a result. Recommendations include use of simple and clear language and well designed pictorial instructions, especially for the steps related to specimen collection, buffer use, and interpretation of results;43 easily identifiable kit components; reduction of the volume of specimen needed to do the test; and intuitive single-step test kits with controlled and automatic specimen collection, transfer, and buffer use.

tructions, especially for the steps related to specimen collection, buffer use, and interpretation of results;43 easily identifiable kit components; reduction of the volume of specimen needed to do the test; and intuitive single-step test kits with controlled and automatic specimen collection, transfer, and buffer use. In some settings, instructions could be adapted and validated for the cultural context and for less-skilled users, including individuals with low literacy or visual impairments. This could include translation in local languages, clear and large print instructions for use, detailed images and descriptions, or electronic documents or audio instructions. To improve performance of less-skilled users, instructions for use could be coupled with in-person or video demonstrations on how to do the test and interpret the result. Product labelling should clearly state that people with reactive or invalid test results should seek further testing at a health facility. The labelling should also include information on test limitations in detecting HIV infection during the window period, for people taking pre-exposure prophylaxis, or in people with a suppressed immune response, such as people on antiretroviral drugs. This is a crucial issue because reports show that people with HIV on antiretrovial therapy might be using HIV self-testing kits to check and reconfirm their HIV status, and could obtain a false-negative result.44, 45

re prophylaxis, or in people with a suppressed immune response, such as people on antiretroviral drugs. This is a crucial issue because reports show that people with HIV on antiretrovial therapy might be using HIV self-testing kits to check and reconfirm their HIV status, and could obtain a false-negative result.44, 45 Strengths of this study include completeness of the search strategy, explicit inclusion criteria, a systematic approach to data collection, and independent assessment of each included study. Among the limitations were that most included studies used oral fluid-based rapid diagnostic tests, and studies used different and imperfect reference standard tests to identify HIV-positive individuals. Most studies did not compare approaches or specimens head-to-head. Results were considered biased in studies where the reference test strategy was not aligned with WHO testing guidance. Our last search was done on April 30, 2016, and since then 11 studies reporting on HIV self-testing performance have been published,46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56 of which eight are abstracts. These studies reached the same conclusion as we did, reporting that most participants were able to do the self-test correctly, and, where reported, the raw proportion of agreement was also high, ranging from 84% to 99%.45, 46, 48, 49, 55 Six of 11 studies would not have met our inclusion criteria, and the four that did meet the inclusion criteria might not have influenced our findings. Furthermore, most studies used adapted test kits that were not specifically designed or packaged for self-testing, and in some studies participants did not interpret their own results, but interpreted contrived devices or pictures or photographs.

t did meet the inclusion criteria might not have influenced our findings. Furthermore, most studies used adapted test kits that were not specifically designed or packaged for self-testing, and in some studies participants did not interpret their own results, but interpreted contrived devices or pictures or photographs. No study provided information on people recently or acutely infected with HIV, and no study disaggregated data by individuals taking antiretroviral drugs. Because little data were available, we could not explore the influence of HIV prevalence, type of reference used, or study design. Selection bias is likely because most studies carefully selected participants; some studies included only known HIV-positive individuals. We did not assess publication bias because analytical methods are not well suited for testing observational data.57 Finally, although most studies were judged to be at low risk of bias, concerns remained about studies with small samples and the extent to which the findings can be generalised. In summary, self-testers can achieve a high level of agreement with the results obtained by a health-care worker when using an HIV rapid diagnostic test for self-testing, whether or not assistance was provided. Errors in performance of the test procedure might be reduced through improvement of the design of rapid diagnostic tests for self-testing, clearer product labels, inclusion of simple instructions for use, and provision of additional support, such as instructional videos. Supplementary Material Supplementary appendix

In summary, self-testers can achieve a high level of agreement with the results obtained by a health-care worker when using an HIV rapid diagnostic test for self-testing, whether or not assistance was provided. Errors in performance of the test procedure might be reduced through improvement of the design of rapid diagnostic tests for self-testing, clearer product labels, inclusion of simple instructions for use, and provision of additional support, such as instructional videos. Supplementary Material Supplementary appendix Acknowledgments The Bill & Melinda Gates Foundation, in collaboration with the STAR Consortium, which is supported by Unitaid (grant number PO#10140-0-600), funded the study. We are grateful for the technical input from the WHO Technical Working Group on HIV self-testing. Special thanks to the investigators who provided additional data and shared their studies with us. Thanks also to Mary Henderson for language editing and proofreading. Contributors CF, CJ, and RB developed the initial study design. CF and CJ searched for and extracted data. CF analysed the findings and wrote the first draft of the manuscript. All authors collaboratively discussed key decisions throughout the course of the review, provided critical feedback on preliminary drafts and interpretation of results, and approved the final manuscript. Declaration of interests We declare no competing interests.

This cluster randomised trial provided rigorous evidence that a 3-month intervention of door-to-door distribution of HIV self-tests increased knowledge of HIV status among adults aged 16 years and older. This effect was driven by increased knowledge of HIV status among men and young adults aged 16–29 years, with no between-group differences noted in women. This was the first trial to show that community-based door-to-door secondary distribution of HIV tests can increase HIV testing among individuals, primarily men, who are not at home during household visits from lay counsellors. Among men aged 30 years or more who HIV self-tested, more than a third self-tested through unsupervised HIV self-testing or secondary distribution of HIV self-tests. The trial also showed that the door-to-door distribution of HIV self-tests increased knowledge of HIV status among community residents whose HIV status was not known to lay counsellors despite 2 years of household delivery of HIV-related services, including HIV testing. Implications of all the available evidence Our findings suggest that the door-to-door distribution of HIV self-tests increases knowledge of HIV status among individuals who are underserved by currently available HIV testing services. Secondary distribution of HIV self-tests through this strategy was acceptable and might dentify a higher proportion of individuals testing HIV positive. Future research should explore targeted secondary distribution of HIV self-tests to partners of HIV-positive individuals to support reaching individuals at highest risk of infection. The door-to-door distribution of HIV self-tests is a promising strategy that complements currently available HIV testing strategies by accessing so-called harder to reach individuals, including men. Door-to-door secondary distribution of HIV self-tests in high-prevalence settings could support reaching older men, who are more likely to be HIV positive than younger men, and linking them to prevention or treatment services, and reach HIV Prevention 2020 targets.

so-called harder to reach individuals, including men. Door-to-door secondary distribution of HIV self-tests in high-prevalence settings could support reaching older men, who are more likely to be HIV positive than younger men, and linking them to prevention or treatment services, and reach HIV Prevention 2020 targets. HPTN 071 (PopART) is a cluster-randomised trial ongoing in 21 communities in South Africa and Zambia to estimate the effect of universal HIV testing and immediate treatment on HIV incidence.8 After one round of door-to-door delivery of HIV testing services in the PopART intervention communities in Zambia, the first 90 of the UNAIDS 90-90-90 targets was nearly reached among women, but among men coverage was approximately 10–15% below target.9 PopART has shown that, even with intensive household delivery of HIV testing and related services, challenges remain in reaching the ambitious first and second 90s of the UNAIDS 90-90-90 targets, particularly among men.9

Introduction Despite widespread availability of facility-based and community-based HIV testing services, an estimated 30% of all people living with HIV are unaware of their HIV-positive status.1 Furthermore, since 2010, the number of global new adult HIV infections has remained stable at around 1·9 million per year.2 Increasing coverage of HIV prevention services requires novel strategies to deliver HIV testing services and reach individuals who remain unaware of their HIV status.3 In Zambia, coverage of HIV testing services has increased substantially since 2007,4, 5, 6 yet there remain gaps. In 2013–14, 46% of women were tested and received the test result within the previous 12 months compared with 37% of men.6 During 2015–16, 70% of HIV-positive women knew their HIV-positive status compared with 63% of men.5, 6 Furthermore, adults older than 24 years are more likely to have tested than adolescents and young people aged 15–24 years.6, 7 Research in context Evidence before this study

Introduction Despite widespread availability of facility-based and community-based HIV testing services, an estimated 30% of all people living with HIV are unaware of their HIV-positive status.1 Furthermore, since 2010, the number of global new adult HIV infections has remained stable at around 1·9 million per year.2 Increasing coverage of HIV prevention services requires novel strategies to deliver HIV testing services and reach individuals who remain unaware of their HIV status.3 In Zambia, coverage of HIV testing services has increased substantially since 2007,4, 5, 6 yet there remain gaps. In 2013–14, 46% of women were tested and received the test result within the previous 12 months compared with 37% of men.6 During 2015–16, 70% of HIV-positive women knew their HIV-positive status compared with 63% of men.5, 6 Furthermore, adults older than 24 years are more likely to have tested than adolescents and young people aged 15–24 years.6, 7 Research in context Evidence before this study We searched PubMed and Medline for English-language publications on studies of strategies to increase HIV testing uptake through distribution of HIV self-tests published through to Sept 14, 2017. We used the search terms self-test* AND HIV infections AND Africa. Of the studies identified, many explored the acceptability and accuracy of HIV self-testing in Kenya, Malawi, and South Africa. These studies consistently reported that HIV self-testing is acceptable, and in Malawi is the preferred option for future HIV testing. Studies exploring the distribution of HIV self-tests included two studies to promote male partner HIV testing: a cohort study of secondary distribution by HIV-negative female sex workers and women receiving antenatal care in Kenya, and a trial of secondary distribution by women receiving HIV self-tests through antenatal care or postpartum care in Kenya. In the cohort study, a large proportion of the women receiving an HIV self-test distributed these to their sexual partners. In the trial, partner HIV testing was higher in the HIV self-test group than in the group offered an invitation for male partners to attend clinic-based HIV testing. In Malawi, a community-based study of HIV self-test distribution showed high uptake and acceptability of HIV self-tests delivered by volunteer counsellors. We found little evidence of rigorous, randomised, controlled trials of strategies to distribute HIV self-tests. Further, there was little available evidence of secondary distribution of HIV self-tests outside of facility settings.

gets was nearly reached among women, but among men coverage was approximately 10–15% below target.9 PopART has shown that, even with intensive household delivery of HIV testing and related services, challenges remain in reaching the ambitious first and second 90s of the UNAIDS 90-90-90 targets, particularly among men.9 HIV self-testing is a novel strategy that has the potential to reach individuals underserved by currently available HIV testing strategies.10 A systematic review11 showed that HIV self-testing is feasible for populations in high HIV-prevalence settings. Since 2016, WHO has recommended evidence-based approaches to delivering HIV self-testing services to reach men and other key populations.12 Evidence from Malawi and Kenya suggests that oral HIV self-testing is acceptable and accurate, and can potentially increase community levels of HIV testing and promote male partner testing.11, 13 Although feasible and acceptable, there remains a need for evidence of how to deliver HIV self-testing services to increase knowledge of HIV status, particularly among men and younger adults. Here, we report results from a cluster-randomised trial of HIV self-testing services nested within the HPTN 071 (PopART) trial. The nested trial offered an opportunity to evaluate whether the door-to-door offer of the option of oral HIV self-testing alongside the offer of home-based finger-prick rapid diagnostic testing (finger-prick RDT) by lay counsellors increased current knowledge of HIV status among the general adult and adolescent population in four urban communities in Zambia.

to evaluate whether the door-to-door offer of the option of oral HIV self-testing alongside the offer of home-based finger-prick rapid diagnostic testing (finger-prick RDT) by lay counsellors increased current knowledge of HIV status among the general adult and adolescent population in four urban communities in Zambia. Methods Study design and participants We nested this HIV self-testing cluster-randomised trial in four urban communities in two northern provinces of Zambia (figure 1).8 Details of the HPTN 071 (PopART) trial are reported elsewhere.8 Briefly, this is a cluster-randomised trial done in 21 communities in Zambia and South Africa to estimate the effect of a household combination HIV prevention package (PopART intervention), which includes the door-to-door offer of HIV testing services (finger-prick RDT), immediate treatment for HIV-positive individuals regardless of CD4 cell count, and promotion of male circumcision for HIV-negative men, on HIV incidence.8 Community HIV care providers did annual rounds within a defined geographical area (called a zone), during which they attempt to visit all households, enumerate all household members, and offer the PopART intervention services to all individuals, regardless of previous participation in previous rounds. Throughout an annual round, community HIV care providers return to households to offer HIV testing services to individuals absent at the first household visit, and support linkage to and retention in care for individuals testing HIV-positive.8 At the time of the HIV self-testing study, the community HIV care providers were doing their third PopART annual round.8Figure 1 Map showing the randomised zones and location of 12 PopART clusters in Zambia, 2013

first household visit, and support linkage to and retention in care for individuals testing HIV-positive.8 At the time of the HIV self-testing study, the community HIV care providers were doing their third PopART annual round.8Figure 1 Map showing the randomised zones and location of 12 PopART clusters in Zambia, 2013 Blue areas are zones randomly allocated to the HIV self-testing intervention. Red cross indicates location of the health facility within the community. Black closed circles are towns. Red triangles are HPTN 071 (PopART) communities. Grey areas are PopART districts. Green areas are provinces. During household visits, data on household enumeration, consent to participate in PopART, and uptake of HIV testing services and other PopART services are collected by community HIV care providers using electronic data capture devices. The four HIV self-testing trial communities were PopART intervention communities selected by the study team to reflect the range and pattern of uptake of HIV testing services in the other PopART intervention communities. Each community had one public health facility, situated a maximum of 1·93–3·67 km from households. Each community had been divided into zones (clusters) that comprised approximately 450–500 households, with an estimated average population of approximately 1400 individuals aged at least 16 years per zone. There were 66 zones across the communities, with an estimated total population of around 90 000 individuals aged at least 16 years. A pair of community HIV care providers managed each zone.

mately 450–500 households, with an estimated average population of approximately 1400 individuals aged at least 16 years per zone. There were 66 zones across the communities, with an estimated total population of around 90 000 individuals aged at least 16 years. A pair of community HIV care providers managed each zone. All adolescents and adults aged 16 years and older and resident in the 66 zones were eligible to participate in the HIV self-testing study. During the HIV self-testing study implementation period of approximately 3 months, individuals at home during community HIV care providers' household visits were asked for verbal consent to participate in PopART. If the individual consented and did not report an HIV-positive status, they were eligible for an offer of HIV testing services. The study was approved by the University of Zambia Biomedical Research Ethics Committee and London School of Hygiene and Tropical Medicine ethics committee. Permission to do the study was also granted by the Division of AIDS at the National Institutes of Health (MD, USA), the Zambia National Health Research Authority, and the Zambia Medicines Regulatory Authority.

mbia Biomedical Research Ethics Committee and London School of Hygiene and Tropical Medicine ethics committee. Permission to do the study was also granted by the Division of AIDS at the National Institutes of Health (MD, USA), the Zambia National Health Research Authority, and the Zambia Medicines Regulatory Authority. Randomisation and masking The 66 zones were randomly assigned (1:1) to either oral HIV self-testing plus the PopART standard of care of routine door-to-door HIV testing services (HIV self-testing group) or door-to-door HIV testing services alone (non-HIV self-testing group) over a 3-month period. To achieve balance across clusters in factors that were likely to affect the primary outcome, we stratified randomisation by community.14 We also restricted the randomisation, first within each community and second across all four communities, to achieve balance by trial group on average values of key outcomes measured during round 2, including: the percentage of adults whose HIV status was known to the community HIV care providers by the end of round 2; the percentage accepting the offer of HIV-testing among those eligible to test in round 2, overall and separately for men, women, adults aged 18–29 years, and those resident in rounds 1 and 2; the percentage of men not contacted during round 2, and the average number of adults per zone. SF generated the randomisation protocol and, from several billion possible allocations that met restriction criteria, drew a computer-generated random sample from combinations of 10 000 possible allocations.

n rounds 1 and 2; the percentage of men not contacted during round 2, and the average number of adults per zone. SF generated the randomisation protocol and, from several billion possible allocations that met restriction criteria, drew a computer-generated random sample from combinations of 10 000 possible allocations. In December, 2016, we held a randomisation ceremony with community HIV care providers, their supervisors, and members of the PopART community advisory boards to allocate the zones to the HIV self-testing or non-HIV self-testing groups. Using the randomly selected list of allocations numbered from 0000 to 9999, four individuals selected four numbered balls from a bag. This four-digit number corresponded to an allocation number in the list of 10 000, and determined, for each zone, whether it was to be allocated to group 0 or 1. In a second stage, a similar process was used to randomly assign the zones numbered 0 and 1 to either the HIV self-testing or non-HIV self-testing groups (figure 1). As a cluster-randomised trial of a strategy to deliver HIV self-testing services to all households within a cluster, blinding of participants and community HIV care providers was not feasible.

as used to randomly assign the zones numbered 0 and 1 to either the HIV self-testing or non-HIV self-testing groups (figure 1). As a cluster-randomised trial of a strategy to deliver HIV self-testing services to all households within a cluster, blinding of participants and community HIV care providers was not feasible. Procedures The intervention was implemented between 18 Jan, and 30 April, 2017. The first 2 weeks of implementation were considered a priori pre-full-implementation weeks. As such, the evaluation period was from Feb 1, to April 30, 2017. In zones randomised to the HIV self-testing intervention, individuals contacted during community HIV care providers household visits who consented to participate in PopART and were eligible for HIV testing were offered a choice of using oral HIV self-testing or a finger-prick sampling of whole blood and rapid HIV testing (finger-prick RDT). This was done according to PopART procedures and the Zambian national HIV testing algorithm, using Alere Determine HIV-1/2 (Chiba, Japan) as the screening test and Uni-Gold HIV test (ref 1206502, Trinity Biotech, Ireland) as the confirmatory test. For individuals choosing HIV self-testing, community HIV care providers demonstrated how to do the self-test and read the result. Because HIV self-testing was a novel procedure in the four communities, individuals were offered either supervised (in the presence of a community HIV care provider) or unsupervised (in the absence of a community HIV care provider) HIV self-testing. The level of supervision and support provided was dependent on individual preference. Individuals opting for HIV self-testing were given a package consisting of the OraQuick self-test kit (Bangkok, Thailand),15 user-friendly pictorial instructions for use provided by the manufacturer (translated into local languages), a card with the community HIV care provider's phone number, a self-complete results form, and an envelope for returning the used self-test.

package consisting of the OraQuick self-test kit (Bangkok, Thailand),15 user-friendly pictorial instructions for use provided by the manufacturer (translated into local languages), a card with the community HIV care provider's phone number, a self-complete results form, and an envelope for returning the used self-test. Individuals aged 18 years or older with a partner living in the same household but absent at the time of the community HIV care provider's visit were offered an HIV self-test kit for their absent partner. Individuals accepting an HIV self-test kit for secondary distribution were asked to sign an agreement stating that the kit would only be given to the intended individual, that the individual would not be coerced into using the HIV self-test kit, and that the information required for the individual to use the HIV self-test kit would be communicated to them. The community HIV care provider left a card with their phone number to allow the absent individual to contact them should additional support be required and for linkage to services, including confirmatory testing.

at the information required for the individual to use the HIV self-test kit would be communicated to them. The community HIV care provider left a card with their phone number to allow the absent individual to contact them should additional support be required and for linkage to services, including confirmatory testing. For individuals choosing unsupervised HIV self-testing or where an HIV self-test kit was left for an absent partner, community HIV care providers did a follow-up visit within 7 days to verify use and offer follow-up services. Individuals also had the option of returning the HIV self-test package to the clinic. During follow-up visits, community HIV care providers collected HIV self-test kits and self-completed results forms where available. The community HIV care providers also attempted to meet individuals who received an HIV self-test kit via their partner. Individuals whose HIV self-test results were recorded in their absence were considered to have participated in the PopART intervention in round 3. Where the HIV self-test kit was available, the community HIV care providers checked the results of the test, and for individuals who they met in person and whose HIV self-test result was reactive, they offered confirmatory HIV testing using parallel testing with the same rapid diagnostic tests as HIV finger-prick RDT. Post-test counselling was provided, and HIV-positive individuals were referred to treatment and care irrespective of whether they accepted the offer of confirmatory testing. HIV-negative individuals were counselled and referred to HIV prevention services. Individuals who were left an HIV self-test kit during the intervention period were followed up until Sept 30, 2017 to provide support on doing the HIV self-testing, confirmatory testing, and linking to services as required.

atory testing. HIV-negative individuals were counselled and referred to HIV prevention services. Individuals who were left an HIV self-test kit during the intervention period were followed up until Sept 30, 2017 to provide support on doing the HIV self-testing, confirmatory testing, and linking to services as required. Zones randomised to the non-HIV self-testing group continued to receive the standard PopART combination package of interventions, including the offer of home-based HIV finger-prick RDT. For individuals testing HIV positive, community HIV care providers provided post-test counselling, and referral to HIV treatment and care services. HIV-negative individuals were counselled and referred to HIV prevention services.

bination package of interventions, including the offer of home-based HIV finger-prick RDT. For individuals testing HIV positive, community HIV care providers provided post-test counselling, and referral to HIV treatment and care services. HIV-negative individuals were counselled and referred to HIV prevention services. Outcomes The primary outcome was current knowledge of HIV status, defined as an individual self-reporting knowing their HIV-positive status to a community HIV care provider, or accepting an offer of either HIV self-testing or finger-prick RDT and the HIV test result was recorded by community HIV care providers. Where an HIV self-test kit was distributed for secondary use, the test was recorded as used if the index individual reported the result to the community HIV care provider or the intended user of the HIV self-test kit later met the community HIV care provider and reported that they had used the test. We measured the outcome among households that were first visited in round 3 between Feb 1, and April 30, 2017, restricted to individual household members aged 16 years or older who were first enumerated or (re-)enumerated as a household member during this period. Information on outcomes used data collected from household visits, and re-visits (where an HIV self-test kit was left for unsupervised or a secondary distribution), done between Feb 1, and June 30, 2017. Individuals aged 16 years or older enumerated as a household member but not seen by the community HIV care provider during the study period were assumed not to know their HIV status in round 3.

its (where an HIV self-test kit was left for unsupervised or a secondary distribution), done between Feb 1, and June 30, 2017. Individuals aged 16 years or older enumerated as a household member but not seen by the community HIV care provider during the study period were assumed not to know their HIV status in round 3. Secondary outcomes were consent to participate in PopART; uptake of HIV testing services among individuals who consented to PopART and were eligible for an offer of HIV-testing services; linkage to confirmatory testing in the HIV self-testing group; programmatic costs and incremental cost-effectiveness of adding HIV self-testing to the PopART intervention; and qualitatively describing HIV self-test kit distribution and social harms. Social harms were identified during observations of HIV self-test kit distribution and through community engagement mechanisms, such as stakeholder and community advisory board meetings, and were categorised between incidents related or unrelated to the study and between serious and non-serious incidents. In this Article, we describe only social harms reported during the study period. Further analysis on secondary distribution and linkage to confirmatory testing is ongoing and will be reported elsewhere.

d were categorised between incidents related or unrelated to the study and between serious and non-serious incidents. In this Article, we describe only social harms reported during the study period. Further analysis on secondary distribution and linkage to confirmatory testing is ongoing and will be reported elsewhere. We did sensitivity analyses of the primary outcome to include individuals whose HIV-positive status was known to the community HIV care provider in round 1 or round 2 (or both) of PopART, who were enumerated during the 3-month implementation phase of this trial, but who did not participate, either because they were absent or they did not consent to PopART. Statistical analysis The study was powered to show an overall reduction of 5% in the percentage of adults who did not know their HIV status in the HIV self-testing group compared with the non-HIV self-testing group, assuming that the percentage who did not know their HIV status in the non-HIV self-testing group was in the range 35–40%. Study power was greater than 90% if the between-zone coefficient of variation k was 0·15, and around 70–80% if k was 0·20, assuming an average of approximately 400 adults enumerated per zone. For subgroup analyses by sex, study power was in the range of around 60–90% to show a 5% reduction in the percentage of individuals who did not know their HIV status in the HIV self-testing group, assuming that the percentage who did not know their HIV status in the non-HIV self-testing group was in the range 30–45% and that k was 0·2.

yses by sex, study power was in the range of around 60–90% to show a 5% reduction in the percentage of individuals who did not know their HIV status in the HIV self-testing group, assuming that the percentage who did not know their HIV status in the non-HIV self-testing group was in the range 30–45% and that k was 0·2. To estimate the effect of the HIV self-testing intervention, we analysed data at the individual level, using population-average logistic regression models to account for clustering by zone, to adjust for community to explain some of the between-zone variation, and a priori for age group and sex. Prespecified sub-group analyses for the primary and secondary outcomes included analyses by sex, age (individuals aged 16–29 years, adults aged ≥30 years), and individuals who were resident during rounds 1 and 2 of PopART but who did not participate in either round. We analysed the data with Stata version 15.0.

nd sex. Prespecified sub-group analyses for the primary and secondary outcomes included analyses by sex, age (individuals aged 16–29 years, adults aged ≥30 years), and individuals who were resident during rounds 1 and 2 of PopART but who did not participate in either round. We analysed the data with Stata version 15.0. We did a prospective economic evaluation, from the provider's perspective, to comparatively calculate unit costs of HIV testing services in both groups and calculated the incremental cost of delivering HIV testing services in the HIV self-testing group. Full annual financial and economic costs were calculated. Financial costs included all expenditures for resources in both groups, whereas economic costs captured the full value of all resources used to deliver HIV testing services in both groups, including the valuation of donated goods or services and individual time to deliver services.16 Cost inputs included equipment, HIV testing supplies, general supplies, transportation and travel, administration, and personnel resources (appendix p 3). In this analysis, we used landed costs of US$3·00 per HIV self-test kit, which accounted for purchase, shipment, and landing taxes. Resource use data were collected between Dec 1, 2016, and June 30, 2017. Costs were adjusted to 2017 US$ using an assumed exchange rate of ZMW 9·50.

and personnel resources (appendix p 3). In this analysis, we used landed costs of US$3·00 per HIV self-test kit, which accounted for purchase, shipment, and landing taxes. Resource use data were collected between Dec 1, 2016, and June 30, 2017. Costs were adjusted to 2017 US$ using an assumed exchange rate of ZMW 9·50. Data sources included financial records, the community HIV care provider's electronic data capture device, and interviews with the HIV self-testing intervention team. In the cost analysis, we calculated the total cost of implementing HIV testing service activities, and cost per person who was: enumerated, tested, and newly diagnosed HIV positive for both groups. Role of the funding source The International Initiative for Impact Evaluation (3ie) reviewed and provided non-binding comments on the draft manuscript before submission. The other funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. All authors had final responsibility for the decision to submit for publication. Results Between Feb 1, and April 30, 2017, the community HIV care providers enumerated 13 267 eligible individuals in the HIV self-testing group and 13 706 in the non-HIV self-testing group (table 1, figure 2). In both groups, half the individuals were aged 16–29 years, and a similar proportion were absent during the community HIV care provider's household visit (table 1).Table 1 Baseline characteristics of the study population

als in the HIV self-testing group and 13 706 in the non-HIV self-testing group (table 1, figure 2). In both groups, half the individuals were aged 16–29 years, and a similar proportion were absent during the community HIV care provider's household visit (table 1).Table 1 Baseline characteristics of the study population HIV self-testing (n=13 267) Non-HIV self-testing (n=13 706) Male sex 6368 (48%) 6486 (47%) Age group (years) 16–19 2176 (16%) 2190 (16%) 20–24 2653 (20%) 2804 (21%) 25–29 1940 (15%) 2008 (15%) 30–34 1651 (12%) 1641 (12%) 35–44 2355 (18%) 2345 (17%) ≥45 2492 (19%) 2718 (20%) Absent during community HIV care-provider's visit Total 2782 (21%) 3018 (22%) Men 1942 (70%) 2140 (71%) Women 840 (30%) 878 (29%) Self-reported HIV positive (percentage of those present) 950 (9%) 1152 (11%) Eligible for HIV testing 9340 (91%) 9304 (89%) Previously participated in PopART (in same community HIV care-provider zone) 8093 (61%) 8745 (64%) Previously resident in PopART annual rounds 1 or 2 (in same community HIV care-provider zone) 9376 (71%) 9946 (73%) Data are n (%). The HIV self-testing group received oral HIV self-testing plus routine door-to-door HIV testing. The non-HIV self-testing group received only routine door-to-door HIV testing. Figure 2 Enumeration and uptake of HIV testing in the HIV self-test and non-HIV self-test groups

HIV self-testing (n=13 267) Non-HIV self-testing (n=13 706) Male sex 6368 (48%) 6486 (47%) Age group (years) 16–19 2176 (16%) 2190 (16%) 20–24 2653 (20%) 2804 (21%) 25–29 1940 (15%) 2008 (15%) 30–34 1651 (12%) 1641 (12%) 35–44 2355 (18%) 2345 (17%) ≥45 2492 (19%) 2718 (20%) Absent during community HIV care-provider's visit Total 2782 (21%) 3018 (22%) Men 1942 (70%) 2140 (71%) Women 840 (30%) 878 (29%) Self-reported HIV positive (percentage of those present) 950 (9%) 1152 (11%) Eligible for HIV testing 9340 (91%) 9304 (89%) Previously participated in PopART (in same community HIV care-provider zone) 8093 (61%) 8745 (64%) Previously resident in PopART annual rounds 1 or 2 (in same community HIV care-provider zone) 9376 (71%) 9946 (73%) Data are n (%). The HIV self-testing group received oral HIV self-testing plus routine door-to-door HIV testing. The non-HIV self-testing group received only routine door-to-door HIV testing. Figure 2 Enumeration and uptake of HIV testing in the HIV self-test and non-HIV self-test groups Pending refers to individuals who made appointments but were not yet seen by a community HIV care provider as of June 30, 2017. *3/323 individuals were identified as knowing their HIV positive status before self-testing in the absence of the community HIV care provider after they were subsequently contacted in PopART annual round 3.

to individuals who made appointments but were not yet seen by a community HIV care provider as of June 30, 2017. *3/323 individuals were identified as knowing their HIV positive status before self-testing in the absence of the community HIV care provider after they were subsequently contacted in PopART annual round 3. After the intervention period, 9027 (68%) of 13 267 in the HIV self-testing group knew their HIV status compared with 8952 (65%) of 13 706 individuals in the non-HIV self-testing group (adjusted odds ratio [OR] 1·30, 95% CI 1·03–1·65; p=0·03; table 2).Table 2 Knowledge of HIV status HIV self-testing Non-HIV self-testing Adjusted odds ratio*(95% CI) p value Overall 9027/13267 (68%) 8952/13706 (65%) 1·30 (1·03–1·65) 0·03 Men 3843/6368 (60%) 3571/6486 (55%) 1·31 (1·07–1·60) 0·01 Women 5184/6899 (75%) 5381/7220 (75%) 1·05 (0·86–1·30) 0·62 Young adults (age 16–29 years) 4972/6769 (74%) 4917/7002 (70%) 1·31 (1·05–1·63) 0·02 Older adults (≥30 years) 4055/6498 (62%) 4035/6704 (60%) 1·22 (0·98–1·52) 0·07 Resident in PopART annual rounds 1 and 2, but did not participate in either round 173/583 (30%) 117/567 (21%) 1·63 (1·15–2·31) 0·01 Participated in PopART annual rounds 1 or 2 or both, but declined door-to-door HIV testing 600/1344 (45%) 587/1425 (41%) 1·29 (0·95–1·76) 0·11 Data are n/N (%). The HIV self-testing group received oral HIV self-testing plus routine door-to-door HIV testing. The non-HIV self-testing group received only routine door-to-door HIV testing. * Adjusted for sex, age, community, and clustering by zones.

HIV self-testing Non-HIV self-testing Adjusted odds ratio*(95% CI) p value Overall 9027/13267 (68%) 8952/13706 (65%) 1·30 (1·03–1·65) 0·03 Men 3843/6368 (60%) 3571/6486 (55%) 1·31 (1·07–1·60) 0·01 Women 5184/6899 (75%) 5381/7220 (75%) 1·05 (0·86–1·30) 0·62 Young adults (age 16–29 years) 4972/6769 (74%) 4917/7002 (70%) 1·31 (1·05–1·63) 0·02 Older adults (≥30 years) 4055/6498 (62%) 4035/6704 (60%) 1·22 (0·98–1·52) 0·07 Resident in PopART annual rounds 1 and 2, but did not participate in either round 173/583 (30%) 117/567 (21%) 1·63 (1·15–2·31) 0·01 Participated in PopART annual rounds 1 or 2 or both, but declined door-to-door HIV testing 600/1344 (45%) 587/1425 (41%) 1·29 (0·95–1·76) 0·11 Data are n/N (%). The HIV self-testing group received oral HIV self-testing plus routine door-to-door HIV testing. The non-HIV self-testing group received only routine door-to-door HIV testing. * Adjusted for sex, age, community, and clustering by zones. There was evidence that the effect of the intervention differed by sex (pinteraction=0·01). Among men, knowledge of HIV status was higher in the HIV self-testing group than in the non-HIV self-testing group whereas among women, knowledge of HIV status did not differ between groups (table 2). The effect did not seem to differ by age group (pinteraction=0·44), but more adults aged 16–29 years and 30 years or older in the HIV self-testing group knew their HIV status compared with the non-HIV self-testing group (table 2).

oup whereas among women, knowledge of HIV status did not differ between groups (table 2). The effect did not seem to differ by age group (pinteraction=0·44), but more adults aged 16–29 years and 30 years or older in the HIV self-testing group knew their HIV status compared with the non-HIV self-testing group (table 2). Similarly, among individuals who were resident during PopART rounds 1 and 2, but did not participate in either round, a greater number in the HIV self-testing group had knowledge of their HIV status compared with the non-HIV self-testing group (table 2). Stratified by age and sex, knowledge of HIV status was higher in the HIV self-testing group than in the non-HIV self-testing group among younger and older men, men who were resident during PopART rounds 1 and 2 but who did not participate in either round, and men who participated in previous rounds of PopART but declined HIV testing (table 3). There was weak evidence that the intervention had an effect among the small number of women resident in PopART rounds 1 and 2 but who did not participate in either round, with little evidence of an effect among women who participated in previous rounds of PopART but declined HIV testing (table 3).Table 3 Knowledge of HIV status stratified by sex

t the intervention had an effect among the small number of women resident in PopART rounds 1 and 2 but who did not participate in either round, with little evidence of an effect among women who participated in previous rounds of PopART but declined HIV testing (table 3).Table 3 Knowledge of HIV status stratified by sex Men Women HIV self-testing Non-HIV self-testing Adjusted odds ratio* (95% CI) p value HIV self-testing Non-HIV self-testing Adjusted odds ratio* (95% CI) p value Overall 3843/6368 (60%) 3571/6486 (55%) 1·31 (1·07–1·60) 0·01 5184/6899 (75%) 5381/7220 (75%) 1·05 (0·86–1·30) 0·62 Young adults (age 16–29 years) 2091/3129 (67%) 1979/3233 (61%) 1·31 (1·04–1·65) 0·02 2881/3640 (79%) 2938/3769 (78%) 1·15 (0·93–1·44) 0·21 Older adults (age ≥30 years) 1752/3239 (54%) 1592/3253 (49%) 1·37 (1·10–1·72) 0·01 2303/3259 (71%) 2443/3451 (71%) 1·01 (0·80–1·27) 0·96 Resident in PopART annual rounds 1 and 2, but did not participate in either round 128/441 (29%) 86/427 (20%) 1·64 (1·15–2·35) 0·01 45/142 (32%) 31/140 (22%) 1·68 (1·02–2·77) 0·04 Participated in PopART annual rounds 1 or 2 or both, but declined door-to-door HIV testing 282/656 (43%) 236/665 (36%) 1·47 (1·03–2·09) 0·03 318/688 (46%) 351/760 (46%) 1·05 (0·78–1·41) 0·73 Data are n/N (%). The HIV self-testing group received oral HIV self-testing plus routine door-to-door HIV testing. The non-HIV self-testing group received only routine door-to-door HIV testing. * Adjusted for sex, age, community, and clustering by zones.

Men Women HIV self-testing Non-HIV self-testing Adjusted odds ratio* (95% CI) p value HIV self-testing Non-HIV self-testing Adjusted odds ratio* (95% CI) p value Overall 3843/6368 (60%) 3571/6486 (55%) 1·31 (1·07–1·60) 0·01 5184/6899 (75%) 5381/7220 (75%) 1·05 (0·86–1·30) 0·62 Young adults (age 16–29 years) 2091/3129 (67%) 1979/3233 (61%) 1·31 (1·04–1·65) 0·02 2881/3640 (79%) 2938/3769 (78%) 1·15 (0·93–1·44) 0·21 Older adults (age ≥30 years) 1752/3239 (54%) 1592/3253 (49%) 1·37 (1·10–1·72) 0·01 2303/3259 (71%) 2443/3451 (71%) 1·01 (0·80–1·27) 0·96 Resident in PopART annual rounds 1 and 2, but did not participate in either round 128/441 (29%) 86/427 (20%) 1·64 (1·15–2·35) 0·01 45/142 (32%) 31/140 (22%) 1·68 (1·02–2·77) 0·04 Participated in PopART annual rounds 1 or 2 or both, but declined door-to-door HIV testing 282/656 (43%) 236/665 (36%) 1·47 (1·03–2·09) 0·03 318/688 (46%) 351/760 (46%) 1·05 (0·78–1·41) 0·73 Data are n/N (%). The HIV self-testing group received oral HIV self-testing plus routine door-to-door HIV testing. The non-HIV self-testing group received only routine door-to-door HIV testing. * Adjusted for sex, age, community, and clustering by zones. In a sensitivity analysis, including individuals whose HIV-positive status was known to community HIV care providers in PopART rounds 1 or 2, or both, but who had not participated in this HIV self-testing trial, 9179 (69%) of 13 267 knew their HIV status in the HIV self-testing group compared with 9079 (66%) of 13 706 in the non-HIV self-testing group (adjusted OR 1·34, 95% CI 1·07–1·69; p=0·01). Among men, 3910 (61%) of 6368 knew their HIV status in the HIV self-testing group compared with 3622 (56%) of 6486 in the non-HIV self-testing group (adjusted OR 1·34, 95% CI 1·10–1.64, p=0·004). After accounting for variation explained by differences among the four communities and the effect of the HIV self-testing intervention, the between-zone coefficient of variation (k) for the percentage of individuals who did not know their HIV status by the end of this trial was 0·23.

R 1·34, 95% CI 1·10–1.64, p=0·004). After accounting for variation explained by differences among the four communities and the effect of the HIV self-testing intervention, the between-zone coefficient of variation (k) for the percentage of individuals who did not know their HIV status by the end of this trial was 0·23. Participation in PopART was slightly higher in the HIV self-testing group, with 10 290 (78%) of 13 267 enumerated, either because they were contacted and consented to participate, or because they self-tested through secondary distribution and their results were reported to and recorded by the community HIV care providers. In the non-HIV self-testing group, 10 456 (76%) of 13 706 enumerated individuals participated in PopART (adjusted OR 1·40, 95% CI 0·98–1·99; p=0·06). Stratified by sex, more men in the HIV self-testing group than in the non-HIV self-testing group participated in PopART, but there were no between-group differences in women (table 4). By sex and age, there was little evidence of an effect among women of either age group, or among men aged 16–29 years.Table 4 Participation in the PopART intervention and accepting an offer of HIV testing

non-HIV self-testing group participated in PopART, but there were no between-group differences in women (table 4). By sex and age, there was little evidence of an effect among women of either age group, or among men aged 16–29 years.Table 4 Participation in the PopART intervention and accepting an offer of HIV testing Men Women HIV self-testing Non-HIV self-testing Adjusted odds ratio (95% CI)* p value HIV self-testing Non-HIV self-testing Adjusted odds ratio (95% CI)* p value Participation in the PopART intervention Overall 4331/6368 (68%) 4219/6486 (65%) 1·27 (0·99–1·63) 0·06 5959/6899 (86%) 6237/7220 (86%) 1·00 (0·77–1·30) 0·99 Young adults (age 16–29 years) 2281/3129 (73%) 2273/3233 (70%) 1·18 (0·91–1·52) 0·21 3176/3640 (87%) 3297/3769 (88%) 1·04 (0·82–1·33) 0·74 Older adults (age ≥30 years) 2050/3239 (63%) 1946/3253 (60%) 1·29 (0·98–1·69) 0·07 2783/3259 (85%) 2940/3451 (85%) 0·97 (0·71–1·33) 0·86 Resident in PopART annual rounds 1 and 2, but did not participate in either round 148/441 (34%) 105/427 (25%) 1·59 (1·08–2·33) 0·02 64/142 (45%) 54/140 (39%) 1·20 (0·71–2·03) 0·50 Participated in PopART annual rounds 1 or 2 or both, but declined door-to-door HIV testing 402/656 (61%) 390/665 (59%) 1·26 (0·85–1·87) 0·24 547/688 (80%) 604/760 (80%) 0·99 (0·66–1·48) 0·95 Accepting an offer of HIV testing services among individuals eligible for HIV testing Overall† 3581/4069 (88%) 3242/3890 (83%) 1·42 (1·10–1·85) 0·01 4496/5271 (85%) 4558/5414 (84%) 1·05 (0·82–1·35) 0·68 Young adults (age 16–29 years) 2063/2253 (92%) 1945/2239 (87%) 1·56 (1·15–2·12) 0·01 2714/3009 (90%) 2731/3090 (88%) 1·29 (0·95–1·74) 0·10 Older adults (age ≥30 years) 1518/1816 (84%) 1297/1651 (79%) 1·44 (1·07–1·94) 0·02 1782/2262 (79%) 1827/2324 (79%) 1·02 (0·79–1·33) 0·85 Resident in PopART annual rounds 1 and 2, but did not participate in either round 122/142 (86%) 83/102 (81%) 1·33 (0·66–2·66) 0·42 43/62 (69%) 26/49 (53%) 1·74 (0·76–3·98) 0·19 Participated in PopART annual rounds 1 or 2 or both, but declined door-to-door HIV testing 276/396 (70%) 223/377 (59%) 1·64 (1·05–2·54) 0·03 300/529 (57%) 327/580 (56%) 1·02 (0·71–1·45) 0·93 Data are n/N (%).

in either round 122/142 (86%) 83/102 (81%) 1·33 (0·66–2·66) 0·42 43/62 (69%) 26/49 (53%) 1·74 (0·76–3·98) 0·19 Participated in PopART annual rounds 1 or 2 or both, but declined door-to-door HIV testing 276/396 (70%) 223/377 (59%) 1·64 (1·05–2·54) 0·03 300/529 (57%) 327/580 (56%) 1·02 (0·71–1·45) 0·93 Data are n/N (%). * Adjusted for sex, age, community, and clustering by zones. † Three individuals who first used an HIV self-test kit in the absence of a community HIV care provider but later self-reported being HIV positive were not included. The intervention increased acceptance of an offer of HIV testing services among men but not women (table 4, appendix pp 5, 6). The effect was similar among young men aged 16–29 years and men aged 30 years and older. There were no between-group differences in accepting HIV testing services in either men or women who were previously resident in PopART rounds 1 and 2 but did not participate in either round. Among men who participated in previous rounds of PopART but declined HIV testing, HIV testing uptake was greater in the HIV self-testing group (table 4). There was no evidence of an effect among women (table 4).

men or women who were previously resident in PopART rounds 1 and 2 but did not participate in either round. Among men who participated in previous rounds of PopART but declined HIV testing, HIV testing uptake was greater in the HIV self-testing group (table 4). There was no evidence of an effect among women (table 4). Among individuals who opted for HIV self-testing in the HIV self-testing group, most chose supervised HIV self-testing, with the number of supervised HIV self-tests higher in women than men (appendix p 2). Among women, the method of HIV self-testing differed little by age group (appendix p 2). Among men, the type of HIV self-testing differed by age: 325 (38%) of 847 men aged 30 years and older who self-tested used unsupervised or secondary distribution HIV self-test kits, compared with 199 (17%) of 1161 men aged 16–29 years (p=0·004). Among the 148 individuals whose first HIV self-testing result was reactive (figure 2), seven (5%) subsequently reported they had known they were HIV positive before self-testing, and three (2%) who had tested via secondary distribution subsequently did a repeat HIV self-testing after meeting the community HIV care providers and the test result was negative. Of the remaining 138 who were eligible for confirmatory HIV testing, 105 (76%) linked to confirmatory testing, of whom 102 (97%) were confirmed HIV positive.

had tested via secondary distribution subsequently did a repeat HIV self-testing after meeting the community HIV care providers and the test result was negative. Of the remaining 138 who were eligible for confirmatory HIV testing, 105 (76%) linked to confirmatory testing, of whom 102 (97%) were confirmed HIV positive. The total cost of delivering HIV testing services was US$243 745 in the HIV self-testing group and US$172 069 in the non-HIV self-testing group. Personnel costs formed the largest proportion of the total costs in both groups followed by testing supplies, with the remaining costs being resource inputs at less than 10% in each group (appendix p 4). HIV self-testing activities accounted for $84 135 (35%) of the $243 745 cost of implementing HIV testing services in the HIV self-testing group. The cost per person tested in the HIV self-testing group was 1·37 times higher than in the non-HIV self-testing group. The incremental costs of distributing HIV self-test kits door-to-door alongside routine door-to-door HIV testing services was estimated at $71 675·78, which resulted in an incremental cost per additional person tested of $255·98 ($71 675·78 total costs for 280 individuals). Incremental costs per individual confirmed HIV self-test-positive was calculated at $771·88 ($84 135 total costs for 109 individuals; appendix p 4).

V testing services was estimated at $71 675·78, which resulted in an incremental cost per additional person tested of $255·98 ($71 675·78 total costs for 280 individuals). Incremental costs per individual confirmed HIV self-test-positive was calculated at $771·88 ($84 135 total costs for 109 individuals; appendix p 4). 13 social harms occurred in the HIV self-testing group. These ranged from invasion of privacy, emotional distress, being deceived or forced into doing HIV testing, threatening violence, to actual violence, and separation of couples (appendix p 7). Some social harms were exacerbated by pre-existing conditions within a couple, such as alcohol abuse and a history of gender-based violence. Discussion Our findings showed that a 3-month intervention of a door-to-door offer of HIV self-testing as an option for HIV testing had small but significant effects on current knowledge of HIV status among the general population aged 16 years or older in four communities in Zambia. The intervention had an overall effect among men, but not women, and an effect among men and women resident in the communities during rounds 1 and 2 of PopART but who did not participate in either round. Participation in PopART increased among men but not women, with an increase in HIV testing among men who previously declined HIV testing services and a small increase among women who previously declined participation in PopART.

the communities during rounds 1 and 2 of PopART but who did not participate in either round. Participation in PopART increased among men but not women, with an increase in HIV testing among men who previously declined HIV testing services and a small increase among women who previously declined participation in PopART. In Zambia, like other southern African countries, men are harder to reach with HIV testing services.17 HIV self-testing has been proposed as a strategy to reach men.18, 19 We found that the door-to-door offer of the HIV self-test option increased men's knowledge of HIV status and uptake of HIV testing services. This effect was driven by increased acceptability of HIV self-testing compared with standard finger-prick RDT, and by secondary distribution of HIV self-testing kits. Almost all individuals reached through secondary distribution were male, and almost half the men aged 30 years and older who self-tested did so through unsupervised or secondary distribution of HIV self-testing kits. This study is among the first to evaluate community-based secondary distribution of HIV self-test kits. The available evidence shows that secondary distribution is feasible and effective when offered to antenatal care attendees, postpartum women, or female sex workers.11, 20, 21 A Kenyan trial22 comparing secondary distribution of HIV self-test kits with an invitation for partners of antenatal care attendees and postpartum women to attend facility-based HIV-testing services showed that secondary distribution increased partner testing by 39%. Our study adds to the available evidence suggesting that community-based secondary distribution is effective at reaching men in communities exposed to 3 years of door-to-door delivery of HIV testing services.

attend facility-based HIV-testing services showed that secondary distribution increased partner testing by 39%. Our study adds to the available evidence suggesting that community-based secondary distribution is effective at reaching men in communities exposed to 3 years of door-to-door delivery of HIV testing services. Similar to findings from a Malawian HIV study,13 in which 25% of individuals self-testing after 2 years of service promotion were aged younger than 20 years, we showed that the HIV self-testing intervention had small but significant effects on knowledge of HIV status and uptake of HIV testing services among younger individuals aged 16–29 years.23, 24 HIV incidence is high among adolescents and young people, particularly among females aged 15–24 years who are at highest risk of HIV in sub-Saharan Africa.2 Adolescents and young people, particularly younger men, are harder to reach with HIV services, and less likely to be engaged in all steps of the HIV care cascade.24 In PopART, door-to-door HIV testing services increased HIV testing uptake and knowledge of HIV status among adolescents aged 15–19 years; however, 28% of adolescents were not reached, mainly because men and younger age groups were absent during household visits.25 HIV self-testing provides a crucial opportunity to reach this underserved population.

sting services increased HIV testing uptake and knowledge of HIV status among adolescents aged 15–19 years; however, 28% of adolescents were not reached, mainly because men and younger age groups were absent during household visits.25 HIV self-testing provides a crucial opportunity to reach this underserved population. Our findings suggest that HIV self-testing reached previously unreached individuals, including individuals previously resident in the community but who did not participate in two rounds of PopART.9, 26 Most individuals who self-tested opted for supervised HIV self-testing. HIV self-testing probably displaced use of finger-prick RDT by some individuals who would have otherwise accepted finger-prick HIV-testing services, but the use of HIV self-testing was new in these communities and testing preferences would probably change over time with increased familiarity. The OraQuick HIV self-test used in this study has a sensitivity of 95·5% (95% CI 89·7–98·5) when compared with the Zambian national rapid diagnostic test algorithm and is more expensive than standard blood-based rapid diagnostic tests.15 Our cost analysis showed that the economic cost per HIV tester was higher in the HIV self-testing group than in the non-HIV self-testing group. As lay counsellors become more familiar with offering HIV self-testing and communities more aware of HIV self-testing, unit costs are likely to decrease with time. Costs of accessing harder-to-reach individuals might, however, be higher still in settings where there has been little access to HIV testing services. In this study setting, many HIV-positive individuals were reached after 3 years of PopART service delivery, and therefore HIV positivity in testers was lower than in a population naive to HIV testing.

ach individuals might, however, be higher still in settings where there has been little access to HIV testing services. In this study setting, many HIV-positive individuals were reached after 3 years of PopART service delivery, and therefore HIV positivity in testers was lower than in a population naive to HIV testing. The HIV positivity of supervised HIV self-testing, unsupervised HIV self-testing, and finger-prick RDT were similar. Through secondary distribution, HIV positivity was slightly higher than primary HIV self-test kit distribution and finger-prick RDT. Although the numbers were small, these findings suggest that offering HIV self-testing to individuals who are home during household visits reaches a similar population as finger-prick RDT. Secondary distribution and unsupervised HIV self-testing among men, however, seems to reach individuals that would have been missed had only finger-prick RDT been available. However, confirming use of an HIV self-test kit, and measuring linkage to care is challenging with unsupervised HIV self-testing and secondary distribution. In a Kenyan trial of truck drivers, there was no difference in HIV testing uptake comparing the choice of finger-prick RDT or supervised HIV self-testing with an offer of finger-prick RDT.27 However, when including men who took an HIV self-testing kit home and self-reported use via the telephone, there was evidence that the choice group had higher testing uptake.27 In our study, among individuals who used a secondary distribution HIV self-test kit and for whom community HIV care providers recorded a reactive result, more than a third were later seen by the community HIV care provider and confirmed HIV positive. For many, HIV self-testing is probably appealing as it is private and confidential. This benefit, however, poses challenges for public health research to measure the effect of HIV self-testing on the uptake of HIV testing services, linkage to care, or prevention services. Novel strategies to measure uptake and linkage, or non-financial incentives to encourage return of used HIV self-test kits, could be explored in future studies.

challenges for public health research to measure the effect of HIV self-testing on the uptake of HIV testing services, linkage to care, or prevention services. Novel strategies to measure uptake and linkage, or non-financial incentives to encourage return of used HIV self-test kits, could be explored in future studies. The expansion of HIV self-testing has been met with concerns regarding potential for social harm.11, 28 A review of studies of HIV self-testing found little published evidence of social harms associated with HIV self-testing.28 We noted little evidence of serious adverse events attributable to HIV self-testing in our study. We consider this to be because community HIV care providers were careful about how they introduced HIV self-testing and, in the case of secondary distribution, informed individuals to be cautious when introducing HIV self-test kits to partners. Further, HIV self-test kits were only left for absent partners of individuals aged 18 years and older. This could, however, also be partly due to a reluctance of community members to discuss negative social experiences with researchers. The occurrence and reporting of less severe social harms, such as coerced HIV testing, suggests that developing mechanisms for detecting and reporting social harms earlier in a study might be beneficial.

er, also be partly due to a reluctance of community members to discuss negative social experiences with researchers. The occurrence and reporting of less severe social harms, such as coerced HIV testing, suggests that developing mechanisms for detecting and reporting social harms earlier in a study might be beneficial. Our study had limitations. The HIV self-testing intervention was done in established PopART communities where trained community HIV care providers have built good rapport with the community and have been providing HIV testing services since 2013. This exposure to door-to-door testing services might have affected uptake of HIV self-testing and might limit the generalisability of the findings. It is also likely that providing HIV self-test kits in this setting increased the costs because a more pragmatic approach would be taken in a real-life setting. By its nature, an HIV self-test kit is meant to be used in private. As such, the results of use of secondary distribution HIV self-test kits might have inherent biases. However, we believe reporting bias was minimised given the established relationships between the community HIV care providers and the community. Although blinding of the participants was not feasible, performance bias was minimised by allocating the community HIV care providers permanently to either the HIV self-testing group or the non-HIV self-testing group throughout the study.

blished relationships between the community HIV care providers and the community. Although blinding of the participants was not feasible, performance bias was minimised by allocating the community HIV care providers permanently to either the HIV self-testing group or the non-HIV self-testing group throughout the study. In conclusion, a 3-month intervention of the addition of HIV self-testing to door-to-door offer of finger-prick RDT had small but significant effects on knowledge of HIV status and uptake of HIV testing services in four communities in Zambia. This effect was seen among men, but also among community residents who previously declined participation in PopART. Community-based secondary distribution of HIV self-test kits might be an effective strategy to provide HIV testing to reach individuals underserved by HIV-testing services in settings exposed to door-to-door delivery of HIV-testing services. To maximise the effect and reduce costs, any future rollout plan should target services more efficiently to reach men and other populations who are not currently accessing available HIV-testing services. Supplementary Material Supplementary appendix

In conclusion, a 3-month intervention of the addition of HIV self-testing to door-to-door offer of finger-prick RDT had small but significant effects on knowledge of HIV status and uptake of HIV testing services in four communities in Zambia. This effect was seen among men, but also among community residents who previously declined participation in PopART. Community-based secondary distribution of HIV self-test kits might be an effective strategy to provide HIV testing to reach individuals underserved by HIV-testing services in settings exposed to door-to-door delivery of HIV-testing services. To maximise the effect and reduce costs, any future rollout plan should target services more efficiently to reach men and other populations who are not currently accessing available HIV-testing services. Supplementary Material Supplementary appendix Acknowledgments The HIVST study was funded by the international Initiative for Impact Evaluation, (3ie, grant no. TW2.2.18), with support from the Bill & Melinda Gates Foundation, the National Institute of Allergy and Infectious Diseases (NIAID), the National Institute on Drug Abuse (NIDA), and the National Institute of Mental Health (NIMH), all part of the National Institutes of Health (NIH). Zambart was the sponsor of the HIVST sub-study. HPTN 071 is sponsored by NIAID under Cooperative Agreements UM1-AI068619, UM1-AI068617, and UM1-AI068613, with funding from the US President's Emergency Plan for AIDS Relief (PEPFAR). Additional funding is provided by the International Initiative for Impact Evaluation (3ie) with support from the Bill & Melinda Gates Foundation, NIAID, NIDA, and NIMH. We thank the many people who have helped with this research. This includes the community HIV care providers, their supervisors, and the District Intervention Coordinators for their work in delivering the intervention; staff at the health facility who participated in the implementation of this study and the study participants; the social science research assistants Lwindi Gwaba and Able Han'gandu for the collection of the qualitative data; Brighton Phiri and the data team for their support modifying the electronic data capture tools, and Sarah Kanema for contributions to the economics data collection and analysis. We also thank the community engagement team and mobilisers for their delivery of the engagement activities, and the residents and the community advisory boards of the four communities included in this study. The views expressed in this article do not necessarily represent the official views of 3ie, the Bill & Melinda Gates Foundation, NIAID, NIDA, NIMH, PEPFAR, or the HPTN 071.

or their delivery of the engagement activities, and the residents and the community advisory boards of the four communities included in this study. The views expressed in this article do not necessarily represent the official views of 3ie, the Bill & Melinda Gates Foundation, NIAID, NIDA, NIMH, PEPFAR, or the HPTN 071. Contributors SFl, AJS, RH, and BH were involved in the randomisation, programming of the study tools, and statistical analysis. LM collected cost data and did the cost analysis. CM, CB, MMP, and CRP were involved with leading field data collection. RH, SFi, AM, and HA provided expert knowledge and oversaw the study. All authors were involved in the design of the study, contributed to the writing of the paper, and read and approved the final version.

llected cost data and did the cost analysis. CM, CB, MMP, and CRP were involved with leading field data collection. RH, SFi, AM, and HA provided expert knowledge and oversaw the study. All authors were involved in the design of the study, contributed to the writing of the paper, and read and approved the final version. HPTN 071 (PopART) Study Team Richard Hayes (London School of Hygiene & Tropical Medicine [LSHTM], London, UK), Sarah Fidler (Imperial College, London, UK), Nulda Beyer (Desmond Tutu Tuberculosis Centre, Stellenbosch University, Stellenbosch, South Africa), Helen Ayles (LSHTM and Zambart, University of Zambia School of Medicine, Lusaka, Zambia), Peter Bock (Desmond Tutu Tuberculosis Centre), Wafaa El-Sadr [HIV Prevention Trials Network Leadership and Operations Center (HPTN LOC), London, UK], Myron Cohen (HPTN LOC), Susan Eshleman (HPTN Laboratory Centre [LC], London, UK), Yaw Agyei (HPTN LC), Estelle Piwowar-Manning (HPTN LC), Virginia Bond (LSHTM and Zambart), Graeme Hoddinott (Desmond Tutu Tuberculosis Centre), Deborah Donnell (HPTN Statistical and Data Management Center [SDMC], London, UK), Sian Floyd (LSHTM), Ethan Wilson (HPTN SDMC), Lynda Emel (HPTN SDMC), Heather Noble (HPTN SDMC), Dave Macleod (LSHTM), David Burns (National Institute of Allergy and Infectious Diseases, Bethesda, MD, USA), Christophe Fraser (Oxford University, Oxford, UK), Anne Cori (Imperial College), Niru Sista (HPTN LOC), Sam Griffith (HPTN LOC), Ayana Moore (HPTN LOC), Tanette Headen (HPTN LOC), Rhonda White (HPTN LOC), Eric Miller (HPTN LOC), James Hargreaves (LSHTM), Katharina Hauck (Imperial College), Ranjeeta Thomas (Imperial College), Mohammed Limbada (Zambart), Justin Bwalya (Zambart), Alwyn Mwinga (Zambart), Michael Pickles (University of Manitoba, Winnipeg, MB, Canada), Kalpana Sabapathy (LSHTM), Albertus J Schaap (Zambart), Rory Dunbar (Desmond Tutu Tuberculosis Centre), Kwame Shanaube (Zambart), Blia Yang (Desmond Tutu Tuberculosis Centre), Musonda Simwinga (Zambart), Peter C Smith (Imperial College Business School, London, UK), Sten Vermund (HPTN), Nomtha Mandla (Desmond Tutu Tuberculosis Centre), Nozizwe Makola (Desmond Tutu Tuberculosis Centre), Anneen van Deventer (Desmond Tutu Tuberculosis Centre), Anelet James (Desmond Tutu Tuberculosis Centre), Karen Jennings (City Health Department, City of Cape Town, Cape Town, South Africa), James Kruger (Department of Health, Western Cape, South Africa), Mwelwa Phiri (Zambart), Barry Kosloff (Zambart), Lawrence Mwenge (Zambart), S

eventer (Desmond Tutu Tuberculosis Centre), Anelet James (Desmond Tutu Tuberculosis Centre), Karen Jennings (City Health Department, City of Cape Town, Cape Town, South Africa), James Kruger (Department of Health, Western Cape, South Africa), Mwelwa Phiri (Zambart), Barry Kosloff (Zambart), Lawrence Mwenge (Zambart), S arah Kanema (Zambart), Rafael Sauter (Oxford University), Will Probert (Oxford University), Ramya Kumar (Zambart), Ephraim Sakala (Zambart), Andrew Silumesi (Ministry of Health, Zambia), Tim Skalland (HPTN SDMC), Krista Yuhas (HPTN SDMC). Declaration of interests SFi reports grants from Merck Sharp & Dohme and GlaxoSmithKline. All other authors declare no competing interests.

Introduction Between 2000 and 2015, excitement around the Millennium Development Goals (MDGs) galvanised more than US$500 billion in spending on prevention, care, and treatment for HIV/AIDS globally.1 Despite the subsequent decrease in overall HIV-related mortality, more than 36 million people still live with HIV/AIDS, which continues to be the underlying cause of death for almost 1 million people every year, concentrated in sub-Saharan Africa.2, 3 Recognising the sustained threat, UNAIDS set targets for the years 2020 and 2030 with the aim of ending the epidemic by 2030.4, 5 In this study, we estimate the current and future burden of HIV/AIDS and track progress towards meeting these targets. Research in context Evidence before this study The levels and trends of the global HIV/AIDS epidemic have been estimated by two groups: the Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) and UNAIDS. We searched PubMed with the search terms hiv[MeSH] AND (“mortality” OR “incidence” OR “prevalence”) AND “global” AND (trend*), with no language restrictions, for articles published since database inception until Nov 7, 2018. We did not identify any additional studies that provided comparable evaluations of the global trends in the HIV/AIDS epidemic across countries. The last GBD on HIV was in 2015; however, it did not include assessment of achieving UNAIDS targets using forecasts of past trends and associations in the data. Added value of this study

The levels and trends of the global HIV/AIDS epidemic have been estimated by two groups: the Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) and UNAIDS. We searched PubMed with the search terms hiv[MeSH] AND (“mortality” OR “incidence” OR “prevalence”) AND “global” AND (trend*), with no language restrictions, for articles published since database inception until Nov 7, 2018. We did not identify any additional studies that provided comparable evaluations of the global trends in the HIV/AIDS epidemic across countries. The last GBD on HIV was in 2015; however, it did not include assessment of achieving UNAIDS targets using forecasts of past trends and associations in the data. Added value of this study For GBD 2017, the main inputs for our estimation of global HIV trends were systematically updated. These updates include a comprehensive update of population estimates that are internally consistent with fertility and mortality estimates for GBD 2017, and incorporate new prevalence data from national surveys and antenatal care clinics. Additionally, we made improvements in our estimation of paediatric HIV via modelling of natural disease progression and incorporating cohort data on child antiretroviral therapy (ART) initiation and mortality. We also better reflected geographical differences in the sex-specific distribution of HIV burden on the basis of a model fit to the sex ratio of prevalence observed in countries with representative surveys. Finally, we used forecasting methods to generate country-level estimates towards achieving global targets related to ART coverage, HIV incidence, and HIV-related mortality by 2020 and 2030.

bution of HIV burden on the basis of a model fit to the sex ratio of prevalence observed in countries with representative surveys. Finally, we used forecasting methods to generate country-level estimates towards achieving global targets related to ART coverage, HIV incidence, and HIV-related mortality by 2020 and 2030. Implications of all the available evidence By improving and extending existing HIV/AIDS burden estimates, this study provides valuable insight into progress towards Sustainable Development Goal 3's target to end the AIDS epidemic by 2030 and the fast-track strategy to do so. Relative to incidence and mortality, more countries are on track to meet ART coverage targets of 81% (90% started, 90% retained) by 2020 and 90% (95% started, 95% retained) by 2030. The relative progress necessary to achieve the 2020 and 2030 targets for reduction in incidence and mortality is not on pace in most countries. Renewed attention and investment in HIV prevention initiatives could help to restore global propensity to meet these targets. This study's assessment of current trends and progress towards ambitious global targets provides evidence for decision makers to respond to current needs and plan for a future free of HIV/AIDS.

Renewed attention and investment in HIV prevention initiatives could help to restore global propensity to meet these targets. This study's assessment of current trends and progress towards ambitious global targets provides evidence for decision makers to respond to current needs and plan for a future free of HIV/AIDS. Complementing the ambitious Sustainable Development Goal (SDG) to end the HIV/AIDS epidemic by 2030, UNAIDS' 90-90-90 targets (90% of people living with HIV diagnosed, of whom 90% are on treatment, of whom 90% are virally suppressed) have been set for 2020, and 95-95-95 targets (95% of people living with HIV diagnosed, of whom 95% are on treatment, of whom 95% are virally suppressed) for 2030.5 In accordance with this fast-track initiative to achieve the SDG goal, UNAIDS has since set targets for reducing the number of HIV incident cases and deaths between 2010 and 2020 by 75% and between 2010 and 2030 by 90% for each country.4 Although these latest targets have helped to renew focus on the epidemic, measuring patterns in HIV/AIDS incidence, prevalence, and mortality is challenging, in part because of poor vital registration data and incomplete disease notification systems in high-burden areas, and complex disease modelling strategies and methodological limitations.6 Still, comprehensive global estimates are needed to track progress and understand future burden.

ence, and mortality is challenging, in part because of poor vital registration data and incomplete disease notification systems in high-burden areas, and complex disease modelling strategies and methodological limitations.6 Still, comprehensive global estimates are needed to track progress and understand future burden. In this Article, we present results from the Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) 2017. We address several methodological and data-related challenges associated with estimating HIV burden to provide a comprehensive and robust assessment of trends in HIV incidence, prevalence, and mortality across 195 countries and territories from 1980 to 2017. Building on previous iterations, we extensively updated population estimates and incorporated new prevalence data from national surveys and antenatal care clinics. Additionally, we generated country-level forecasts towards achieving targets associated with antiretroviral therapy (ART) coverage, HIV incidence, and HIV-related mortality. These forecasts enable us to report country-specific progress towards achieving the following targets: a reduction in the number of HIV incident cases of 75% between 2010 and 2020 and 90% between 2010 and 2030; a reduction in the number of HIV deaths of 75% between 2010 and 2020 and 90% between 2010 and 2030; 81% (90% started, 90% retained) ART coverage by 2020 and 90% (95% started, 95% retained) coverage by 2030.4, 5

: a reduction in the number of HIV incident cases of 75% between 2010 and 2020 and 90% between 2010 and 2030; a reduction in the number of HIV deaths of 75% between 2010 and 2020 and 90% between 2010 and 2030; 81% (90% started, 90% retained) ART coverage by 2020 and 90% (95% started, 95% retained) coverage by 2030.4, 5 Methods Study design and modelling strategy GBD is a systematic, scientific effort to quantify the comparative magnitude of health loss due to diseases and injuries by age, sex, and geography over time. GBD 2017 includes 195 countries and territories, 16 of which (Brazil, China, Ethiopia, India, Indonesia, Iran, Japan, Kenya, Mexico, New Zealand, Norway, Russia, South Africa, Sweden, the UK, and the USA) were analysed at the subnational level. The conceptual and analytical framework for GBD, hierarchy of causes, and detailed methods have been published elsewhere.2, 3, 7 The GBD protocol is also available online. Herein we describe the specific methods used for analysing the burden of HIV for GBD 2017. Input data for modelling HIV morbidity and mortality include vital registration data, household seroprevalence surveys, data from antenatal care clinics, demographic estimates (population, fertility, migration, and HIV-free survival rates from GBD 2017), intervention coverage data reported to UNAIDS including ART, prevention of mother-to-child transmission, HIV mortality on and off ART, and rates of disease progression from a systematic review (appendix 1 pp 2–4).

nics, demographic estimates (population, fertility, migration, and HIV-free survival rates from GBD 2017), intervention coverage data reported to UNAIDS including ART, prevention of mother-to-child transmission, HIV mortality on and off ART, and rates of disease progression from a systematic review (appendix 1 pp 2–4). The GBD framework for HIV/AIDS aims to produce internally consistent estimates for HIV incidence, prevalence, and mortality and relies on two established estimation models. We used the Estimation and Projection Package (EPP), an HIV epidemic model originally developed by the UNAIDS Reference Group on Estimates, Modelling, and Projections.8 EPP uses Bayesian methods to infer force of infection from trends in HIV prevalence data. EPP generates incidence and prevalence estimates for individuals aged 15–49 years for both sexes combined. We also used a modified version of Spectrum, a compartmental model used by UNAIDS that ages a population over time while applying HIV incidence, progression, and mortality to produce age-sex-specific HIV incidence, prevalence, and mortality.8 Multiple methodological improvements to both EPP and Spectrum were made for GBD estimation, including developing a model of ART coverage distribution as a function of income, age, sex, and disease progression that we used in Spectrum. Full details of modifications to EPP and Spectrum are in appendix 1 (pp 7–13).

The GBD framework for HIV/AIDS aims to produce internally consistent estimates for HIV incidence, prevalence, and mortality and relies on two established estimation models. We used the Estimation and Projection Package (EPP), an HIV epidemic model originally developed by the UNAIDS Reference Group on Estimates, Modelling, and Projections.8 EPP uses Bayesian methods to infer force of infection from trends in HIV prevalence data. EPP generates incidence and prevalence estimates for individuals aged 15–49 years for both sexes combined. We also used a modified version of Spectrum, a compartmental model used by UNAIDS that ages a population over time while applying HIV incidence, progression, and mortality to produce age-sex-specific HIV incidence, prevalence, and mortality.8 Multiple methodological improvements to both EPP and Spectrum were made for GBD estimation, including developing a model of ART coverage distribution as a function of income, age, sex, and disease progression that we used in Spectrum. Full details of modifications to EPP and Spectrum are in appendix 1 (pp 7–13). To ensure appropriate modelling techniques, we grouped countries on the basis of availability and quality of data. Group 1 includes countries with HIV prevalence data from antenatal care clinics or representative population-based seroprevalence surveys. Group 1A includes countries with a peak of at least 0·5% prevalence, and group 1B includes countries with a peak prevalence of at least 0·25% plus vital registration completeness less than 65%. Group 2 includes all other countries, which are further classified as groups 2A, 2B, and 2C on the basis of availability of vital registration data. Group 2A locations have high-quality data, group 2B locations have at least some data, and group 2C locations have no data on HIV-specific mortality.2 The modelling framework by country grouping is shown in appendix 1 (pp 5, 6).

lassified as groups 2A, 2B, and 2C on the basis of availability of vital registration data. Group 2A locations have high-quality data, group 2B locations have at least some data, and group 2C locations have no data on HIV-specific mortality.2 The modelling framework by country grouping is shown in appendix 1 (pp 5, 6). This study was approved by the University of Washington Institutional Review Board (application 46665).

lassified as groups 2A, 2B, and 2C on the basis of availability of vital registration data. Group 2A locations have high-quality data, group 2B locations have at least some data, and group 2C locations have no data on HIV-specific mortality.2 The modelling framework by country grouping is shown in appendix 1 (pp 5, 6). This study was approved by the University of Washington Institutional Review Board (application 46665). Incidence and prevalence estimation For group 1 countries, we used EPP to estimate incidence and prevalence for individuals aged 15–49 years, for both sexes combined, using population-based surveys and antenatal care clinic data. To account for bias created by the differences in HIV prevalence between pregnant women who attended an antenatal care clinic and the general population, we extracted data from available Demographic and Health Surveys on HIV prevalence among pregnant women who gave birth in the past year and who attended an antenatal care clinic. For antenatal care bias adjustment, we input this data into a regression model with regional random effects to generate country-specific prior distributions where surveys were available and regional prior distributions for locations without a survey. We then used the incidence and prevalence results from EPP as inputs in Spectrum to further disaggregate to age-sex-specific HIV incidence and prevalence. We used the sex ratio of prevalence from population-based surveys to inform the sex-splitting assumptions for adults in Spectrum, and applied default age-splitting assumptions from Spectrum.8 We calculated vertical transmission as a function of prevention of mother-to-child transmission inputs and age-specific fertility rates adjusted to account for differential fertility among women who were HIV positive.

ting assumptions for adults in Spectrum, and applied default age-splitting assumptions from Spectrum.8 We calculated vertical transmission as a function of prevention of mother-to-child transmission inputs and age-specific fertility rates adjusted to account for differential fertility among women who were HIV positive. For group 2 countries, we developed a process called cohort incidence bias adjustment to estimate incidence and prevalence using mortality data. We ran a first stage of Spectrum to generate initial incidence, prevalence, and mortality curves, along with incidence cohort survival. We then calculated the bias between Spectrum mortality estimates and smoothed vital registration data for each year, which we used along with Spectrum cohort survival estimates to adjust incidence (appendix 1 pp 11–13). To account for sensitivity in our estimates to input incidence, we ran the first stage of Spectrum using various input incidence curves and selected the option with the smallest resulting bias in mortality estimates. We ran a second stage of Spectrum using adjusted incidence to produce age-sex-specific incidence and prevalence estimates. In countries with high-quality case notification data, we scaled incidence results to align with case reports after accounting for an assumed average of 5 years' lag between infection and diagnosis.9

ran a second stage of Spectrum using adjusted incidence to produce age-sex-specific incidence and prevalence estimates. In countries with high-quality case notification data, we scaled incidence results to align with case reports after accounting for an assumed average of 5 years' lag between infection and diagnosis.9 Mortality estimation We undertook a meta-analysis of cohort studies to derive on-ART and off-ART mortality as inputs into Spectrum and EPP. We estimated age-sex-specific, CD4-specific, region-specific, and duration-specific on-ART mortality using cohort data after correcting for loss to follow-up (appendix 1 pp 4–7). We jointly estimated off-ART mortality and CD4 progression via an optimisation process that found a best fit to survival curves from cohort studies.

mated age-sex-specific, CD4-specific, region-specific, and duration-specific on-ART mortality using cohort data after correcting for loss to follow-up (appendix 1 pp 4–7). We jointly estimated off-ART mortality and CD4 progression via an optimisation process that found a best fit to survival curves from cohort studies. For group 1 countries, we generated age-sex-specific HIV mortality estimates in Spectrum using the incidence and prevalence estimated in EPP. For group 2 countries, we adjusted vital registration data for incompleteness and garbage coding (ie, deaths coded to an intermediate, immediate, or poorly defined cause).2 We further corrected the data for HIV misclassification by identifying causes of death that deviated from expected age patterns of mortality in years with known HIV epidemics, and excess deaths were attributed to HIV. We used spatiotemporal Gaussian process regression to smooth and complete the time series of adjusted vital registration data (appendix 1 p 11). For groups 2A and 2B, we used the smoothed vital registration data to inform Spectrum-estimated mortality through the cohort incidence bias adjustment process. In group 2C countries, we leveraged spatial information by sampling cohort incidence bias adjustment-generated incidence-adjustment scalars in the region, which were then input into Spectrum to create mortality estimates.

data to inform Spectrum-estimated mortality through the cohort incidence bias adjustment process. In group 2C countries, we leveraged spatial information by sampling cohort incidence bias adjustment-generated incidence-adjustment scalars in the region, which were then input into Spectrum to create mortality estimates. The GBD framework produced three distinct sources of HIV mortality estimates: HIV mortality results from Spectrum; estimated excess HIV mortality from the all-cause mortality process; and smoothed HIV-specific mortality from vital registration data.10 For group 1 countries, we used an ensemble approach to reconcile the differences between HIV mortality estimated by Spectrum and by the all-cause mortality process and generate final HIV mortality. In group 1 countries, EPP and Spectrum estimates were largely driven by HIV prevalence data and mortality estimates generated from cohort data, whereas the all-cause mortality process was primarily based on sibling survival data. For individuals aged 15 years and older, the ensemble model averaged HIV mortality estimates from the two processes with equal weights. For individuals younger than 15 years, we applied the fraction of deaths due to HIV in Spectrum to estimated all-cause mortality to generate HIV-specific mortality and mortality from all other causes (HIV-free mortality). In group 2A countries, we estimated mortality only from vital registration data, and for group 2B and 2C countries we only used Spectrum results.

ed the fraction of deaths due to HIV in Spectrum to estimated all-cause mortality to generate HIV-specific mortality and mortality from all other causes (HIV-free mortality). In group 2A countries, we estimated mortality only from vital registration data, and for group 2B and 2C countries we only used Spectrum results. Forecasting to 2030 We forecasted HIV incidence, prevalence, mortality, and treatment coverage through to 2030 in Spectrum using input parameters extended to 2030. We forecasted the adult ART coverage input on the basis of forecasted ART price, HIV spending on care and treatment, and lag-distributed income (ie, gross domestic product per capita that has been smoothed over the preceding 10 years). We modelled country-year-specific annual ART price per patient using Gaussian process regression with data from the Global Price Reporting Mechanism.11 We calculated the annualised rate of change of per-capita expenditure on HIV care and treatment in each country since 2010. We then forecast expenditure on HIV care and treatment for each country using the 50th percentile annualised rate of change across countries.1 We calculated annual dose-equivalents of ART by dividing spending by ART price, and we used logistic regression to model the association between annual dose-equivalents and ART coverage. We forecasted other treatment coverage inputs to Spectrum, such as child ART coverage and prevention of mother-to-child transmission, using the same approach based on the 50th percentile annualised rate of change observed across countries. Forecasting the incidence input had two steps. First, we calculated counterfactual incidence (ie, expected incidence in the absence of ART) using an assumption of 70% viral suppression among those on treatment,12 then we forecast counterfactual incidence using the 50th percentile annualised rate of change observed across countries in the previous 5 years. Because the forecasted incidence was derived from the counterfactual incidence using forecasted ART coverage, the final forecasted incidence changed in response to both the underlying secular trend and improvements in ART coverage. We used forecasted demographic inputs that were estimated for each location,13 then we ran Spectrum for all locations. Full details on the methods for forecasting are in appendix 1 (pp 17–22).

rage, the final forecasted incidence changed in response to both the underlying secular trend and improvements in ART coverage. We used forecasted demographic inputs that were estimated for each location,13 then we ran Spectrum for all locations. Full details on the methods for forecasting are in appendix 1 (pp 17–22). We used the mean values (rounded to the nearest integer) of the resultant HIV forecasts to determine whether countries were on track to meet the 2020 and 2030 UNAIDS targets. Uncertainty analysis We propagated uncertainty by generating 1000 draws of key inputs, including draw-level linkage of HIV-free mortality with the GBD all-cause mortality process. We ran EPP and Spectrum 1000 times per location to generate results for each draw. We present results with 95% uncertainty intervals (UIs). Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or the writing of the report. All authors had full access to the data in the study and had final responsibility for the decision to submit for publication

Uncertainty analysis We propagated uncertainty by generating 1000 draws of key inputs, including draw-level linkage of HIV-free mortality with the GBD all-cause mortality process. We ran EPP and Spectrum 1000 times per location to generate results for each draw. We present results with 95% uncertainty intervals (UIs). Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or the writing of the report. All authors had full access to the data in the study and had final responsibility for the decision to submit for publication Results Global deaths from HIV peaked in 2006 and have since decreased from 1·95 million (95% UI 1·87–2·04) deaths in 2006 to 0·95 million (0·91–1·01) in 2017 (figure 1). Global ART coverage increased from 2·98 million (2·44–3·58) in 2006 to 21·8 million (20·7–22·9) in 2017. The number of new HIV infection cases peaked in 1999 (3·16 million [2·79–3·67]) and has gradually decreased thereafter. Between 2007 and 2017, the global age-standardised annualised rate of change in HIV incidence decreased by 3·0% (1·5–4·5), with the number of new cases decreasing from 2·35 million (2·02–2·76) in 2007 to 1·94 million (1·63–2·29) in 2017 (table, figure 1). The confluence of these trends produces a steady increase in the total number of people living with HIV. Prevalence has increased from 8·74 million (7·90–9·68) people living with HIV in 1990 to 36·8 million (34·8–39·2) in 2017, of whom 40·5% (37·8–43·7) were not on ART.Figure 1 Global HIV incidence, prevalence, mortality, and people on ART, by sex, for all ages, 1980–2017

ase in the total number of people living with HIV. Prevalence has increased from 8·74 million (7·90–9·68) people living with HIV in 1990 to 36·8 million (34·8–39·2) in 2017, of whom 40·5% (37·8–43·7) were not on ART.Figure 1 Global HIV incidence, prevalence, mortality, and people on ART, by sex, for all ages, 1980–2017 Shaded areas are 95% uncertainty intervals. ART=antiretroviral therapy. Table Number of HIV incident cases and deaths in 2017 by sex and annualised rate of change of HIV incident cases and deaths for 1990–2007 and 2007–17, for 21 GBD regions and 195 countries and territories

ase in the total number of people living with HIV. Prevalence has increased from 8·74 million (7·90–9·68) people living with HIV in 1990 to 36·8 million (34·8–39·2) in 2017, of whom 40·5% (37·8–43·7) were not on ART.Figure 1 Global HIV incidence, prevalence, mortality, and people on ART, by sex, for all ages, 1980–2017 Shaded areas are 95% uncertainty intervals. ART=antiretroviral therapy. Table Number of HIV incident cases and deaths in 2017 by sex and annualised rate of change of HIV incident cases and deaths for 1990–2007 and 2007–17, for 21 GBD regions and 195 countries and territories New HIV infections, 2017 HIV deaths, 2017 Annualised rate of change in new infections Annualised rate of change in HIV deaths Females Males Total Females Males Total 1990–2007 2007–17 1990–2007 2007–17 Global 966 000 (786 000 to 1 180 000) 976 000 (835 000 to 1 120 000) 1 940 000 (1 630 000 to 2 290 000) 446 000 (417 000 to 479 000) 508 000 (483 000 to 540 000) 954 000 (907 000 to 1 010 000) −0·4% (−1·2 to 0·3) −3·0% (−4·5 to −1·5) 8·5% (7·8 to 9·1) −8·3% (−8·7 to −7·9) Low SDI 259 000 (176 000 to 378 000) 177 000 (124 000 to 257 000) 436 000 (303 000 to 627 000) 132 000 (120 000 to 147 000) 131 000 (121 000 to 143 000) 262 000 (244 000 to 286 000) −5·2% (−6·2 to −4·2) −5·9% (−9·4 to −2·0) 4·6% (3·6 to 5·7) −12·4% (−13·0 to −11·8) Low-middle SDI 359 000 (271 000 to 459 000) 278 000 (212 000 to 356 000) 636 000 (487 000 to 808 000) 192 000 (170 000 to 217 000) 184 000 (163 000 to 207 000) 375 000 (338 000 to 416 000) −1·6% (−2·8 to −0·5) −4·1% (−6·1 to −2·0) 11·3% (10·0 to 12·4) −8·5% (−9·1 to −7·8) Middle SDI 240 000 (196 000 to 287 000) 280 000 (245 000 to 317 000) 521 000 (450 000 to 591 000) 105 000 (94 500 to 119 000) 153 000 (143 000 to 165 000) 258 000 (241 000 to 278 000) 6·0% (4·0 to 8·2) −3·7% (−5·3 to −2·1) 18·6% (17·9 to 19·3) −7·8% (−8·5 to −7·2) High-middle SDI 80 800 (69 100 to 99 400) 167 000 (139 000 to 198 000) 247 000 (210 000 to 297 000) 14 200 (13 700 to 14 800) 31 500 (30 800 to 32 700) 45 700 (44 600 to 47 500) 3·7% (3·2 to 4·3) 7·5% (5·6 to 9·2) 4·6% (4·5 to 4·8) 0·7% (0·5 to 1·1) High SDI 26 700 (15 300 to 39 200) 73 100 (42 900 to 107 000) 99 800 (58 300 to 146 000) 3170 (3140 to 3200) 8700 (8610 to 8800) 11 900 (11 800 to 12 000) −2·5% (−3·9 to −1·0) 1·9% (−0·8 to 3·2) −5·7% (−5·8 to −5·7) −5·3% (−5·4 to −5·2) Central Europe, eastern Europe, and central Asia 59 300 (49 400 to 74 000) 121 000 (94 800 to 149 000) 180 000 (146 000 to 222 000) 8260 (8170 to 8350) 18 700 (18 400 to 18 900) 26 900 (26 700 to 27 100) 8·1% (7·0 to 9·2) 11·7% (9·1 to 13·5) 7·7% (7·6 to 7·8) 1·8% (1·7 to 2·0) Central Asia 3280 (2200 to 4480) 4030 (2950 to 5110) 7300 (5200 to 9550) 398 (377 to 419) 902 (869 to 935) 1300 (1260 to 1340) 4·2% (2·2 to 6·3) 7·2% (3·7 t

6 000 to 222 000) 8260 (8170 to 8350) 18 700 (18 400 to 18 900) 26 900 (26 700 to 27 100) 8·1% (7·0 to 9·2) 11·7% (9·1 to 13·5) 7·7% (7·6 to 7·8) 1·8% (1·7 to 2·0) Central Asia 3280 (2200 to 4480) 4030 (2950 to 5110) 7300 (5200 to 9550) 398 (377 to 419) 902 (869 to 935) 1300 (1260 to 1340) 4·2% (2·2 to 6·3) 7·2% (3·7 t o 9·8) 6·8% (6·6 to 7·2) −0·6% (−1·0 to −0·2) Armenia 65·3 (46·1 to 91·6) 99·6 (67·6 to 154) 165 (120 to 241) 5·7 (5·2 to 6·2) 18·0 (17·1 to 19·0) 23·7 (22·7 to 24·8) 21·1% (11·6 to 59·7) 7·5% (4·1 to 12·7) 11·7% (11·3 to 12·0) 10·7% (10·0 to 11·3) Azerbaijan 91·2 (60·5 to 127) 260 (157 to 416) 351 (222 to 535) 7·8 (5·5 to 9·8) 23·7 (17·9 to 28·3) 31·5 (23·8 to 37·8) 2·4% (−5·8 to 5·8) 9·5% (5·2 to 12·9) 6·4% (3·5 to 14·8) −8·0% (−11·2 to −5·0) Georgia 108 (74·0 to 160) 283 (181 to 491) 391 (256 to 641) 8·5 (7·9 to 9·1) 25·9 (24·5 to 27·4) 34·4 (32·9 to 35·9) 18·4% (12·8 to 25·9) 2·2% (−0·0 to 4·4) 8·7% (8·3 to 9·1) 17·6% (17·0 to 18·3) Kazakhstan 988 (640 to 1440) 1510 (1000 to 1880) 2500 (1670 to 3220) 72·6 (67·2 to 78·2) 186 (175 to 198) 258 (246 to 271) 2·5% (0·9 to 4·2) 12·1% (8·0 to 14·2) 7·8% (7·5 to 8·2) −2·1% (−2·7 to −1·6) Kyrgyzstan 457 (291 to 719) 447 (275 to 707) 904 (575 to 1390) 53·9 (50·2 to 57·6) 133 (126 to 140) 187 (179 to 194) 8·9% (7·3 to 11·1) 6·2% (1·1 to 10·7) 9·2% (8·8 to 9·5) 1·6% (1·0 to 2·2) Mongolia 11·5 (2·3 to 29·5) 48·7 (11·3 to 105) 60·2 (14·9 to 132) 3·8 (0·3 to 11·1) 16·1 (1·6 to 32·9) 19·9 (2·1 to 44·0) 25·1% (47·5 to 58·1) 6·7% (0·2 to 11·6) 35·4% (43·1 to 49·1) 8·2% (−11·0 to 13·4) Tajikistan 419 (184 to 633) 156 (95·4 to 224) 575 (280 to 819) 29·8 (21·3 to 44·2) 14·7 (12·2 to 18·1) 44·5 (34·3 to 60·5) 2·4% (−0·6 to 4·5) 9·1% (2·4 to 13·3) 14·1% (11·0 to 18·8) −13·5% (−16·5 to −9·5) Turkmenistan 44·6 (32·4 to 69·4) 116 (88·5 to 169) 160 (122 to 236) 26·4 (24·4 to 28·8) 73·0 (68·7 to 77·6) 99·5 (94·6 to 105) −6·4% (−10·1 to −1·7) 2·2% (0·3 to 4·4) 2·5% (2·2 to 2·9) −2·5% (−3·1 to −1·8) Uzbekistan 1090 (330 to 1910) 1100 (305 to 2000) 2200 (650 to 3780) 189 (174 to 204) 413 (390 to 436) 601 (575 to 628) 5·0% (0·8 to 11·4) 5·1% (−7·3 to 11·6) 5·1% (4·8 to 5·4) 1·0% (0·3 to 1·5) Central Europe 529 (401 to 669) 1600 (1260 to 2120) 2130 (1670 to 2770) 126 (118 to 144) 395 (356 to 462) 521 (476 to 604) 6·3% (4·4 to 8·0) −0·1% (−2·1 to 1·4) −0·0% (−0·4 to 0·8) −0·9% (−1·7 to −0·1) Albania 0·9 (0·7 to 1·4) 2·1 (1·5 to 3·0) 3·0 (2·2 to 4·3) 0·4 (0·4 to 0·5) 0·9 (0·8 to 1·0) 1·3 (1·1 to 1·5) 0·4% (−0·4 to 1·3) 2·0% (0·2 to 3·7) 2·9% (

0 (1260 to 2120) 2130 (1670 to 2770) 126 (118 to 144) 395 (356 to 462) 521 (476 to 604) 6·3% (4·4 to 8·0) −0·1% (−2·1 to 1·4) −0·0% (−0·4 to 0·8) −0·9% (−1·7 to −0·1) Albania 0·9 (0·7 to 1·4) 2·1 (1·5 to 3·0) 3·0 (2·2 to 4·3) 0·4 (0·4 to 0·5) 0·9 (0·8 to 1·0) 1·3 (1·1 to 1·5) 0·4% (−0·4 to 1·3) 2·0% (0·2 to 3·7) 2·9% ( 2·0 to 4·2) −3·6% (−4·8 to −2·4) Bosnia and Herzegovina 1·2 (0·8 to 1·6) 2·2 (1·6 to 3·2) 3·4 (2·5 to 4·4) 0·5 (0·5 to 0·6) 1·0 (0·9 to 1·2) 1·5 (1·4 to 1·7) −3·9% (−5·1 to −2·5) 4·5% (2·6 to 6·1) −2·0% (−3·4 to 0·0) −3·9% (−4·7 to −3·1) Bulgaria 28·1 (22·9 to 33·9) 100 (66·9 to 126) 128 (93·7 to 156) 12·5 (11·8 to 13·3) 42·1 (39·9 to 44·3) 54·6 (52·3 to 57·0) 1·5% (−2·5 to 56·2) 3·3% (1·2 to 4·8) 6·7% (6·4 to 7·0) −4·7% (−5·2 to −4·2) Croatia 10·9 (5·5 to 15·1) 34·1 (17·9 to 48·0) 44·9 (23·6 to 61·9) 2·1 (2·0 to 2·3) 6·8 (6·5 to 7·2) 9·0 (8·7 to 9·3) 1·7% (0·0 to 3·2) 1·9% (−3·9 to 6·0) −2·6% (−2·9 to −2·2) 6·7% (6·2 to 7·3) Czech Republic 17·6 (9·3 to 28·4) 87·8 (38·0 to 148) 105 (48·0 to 165) 7·4 (6·9 to 7·9) 16·4 (15·6 to 17·2) 23·7 (22·9 to 24·6) 9·2% (8·0 to 10·3) 2·8% (−3·7 to 5·6) 3·9% (3·6 to 4·2) 5·4% (4·9 to 5·9) Hungary 28·8 (17·5 to 50·9) 99·4 (63·4 to 158) 128 (86·0 to 193) 8·1 (7·5 to 8·7) 31·8 (30·2 to 33·7) 39·8 (38·1 to 41·7) −1·5% (−3·0 to 0·2) 3·1% (1·2 to 5·6) −4·2% (−4·5 to −3·8) −7·6% (−8·2 to −7·0) Macedonia 4·0 (3·0 to 5·2) 5·4 (4·0 to 7·4) 9·3 (7·1 to 12·2) 0·7 (0·5 to 0·9) 1·0 (0·8 to 1·1) 1·7 (1·4 to 2·0) 9·9% (7·6 to 11·8) 1·8% (−0·7 to 4·0) 8·6% (7·3 to 10·1) −2·2% (−4·2 to −0·9) Montenegro 1·9 (1·4 to 2·4) 4·7 (3·6 to 6·2) 6·6 (5·1 to 8·4) 0·6 (0·4 to 0·9) 0·7 (0·6 to 0·8) 1·4 (1·1 to 1·7) 3·9% (1·8 to 5·7) 3·1% (1·0 to 4·8) 4·9% (3·2 to 6·3) −2·8% (−6·1 to −0·0) Poland 151 (78·7 to 246) 532 (239 to 981) 683 (319 to 1220) 32·0 (30·2 to 33·9) 112 (106 to 118) 144 (138 to 150) 5·2% (2·8 to 7·6) 0·3% (−5·6 to 2·7) 5·3% (5·0 to 5·6) −2·2% (−2·7 to −1·6) Romania 258 (185 to 323) 591 (417 to 778) 849 (634 to 1050) 47·9 (45·1 to 51·2) 118 (112 to 123) 165 (160 to 171) 10·0% (7·6 to 12·6) −0·1% (−2·4 to 1·7) −2·8% (−3·2 to −2·4) 2·7% (2·1 to 3·2) Serbia 18·5 (11·6 to 30·9) 112 (73·2 to 181) 130 (85·9 to 209) 12·3 (6·2 to 28·9) 60·2 (23·9 to 126) 72·5 (30·3 to 156) 8·3% (4·6 to 60·9) −4·5% (−7·3 to −1·6) 3·7% (−0·3 to 54·4) 1·9% (−4·5 to 6·9) Slovakia 6·5 (4·7 to 8·3) 16·5 (11·2 to 25·2) 23·0 (16·2 to 33·5) 1·2 (0·8 to 1·7) 2·7 (2·1 to 3·3) 3·9 (3·0 to 4·9) 4·2% (−1·5 to 48·2) 2·0% (−0·1 to 4·3) 7·7% (4·1 to 38·6) −0·1% (−3·7 to 2·0) Sloven

) 60·2 (23·9 to 126) 72·5 (30·3 to 156) 8·3% (4·6 to 60·9) −4·5% (−7·3 to −1·6) 3·7% (−0·3 to 54·4) 1·9% (−4·5 to 6·9) Slovakia 6·5 (4·7 to 8·3) 16·5 (11·2 to 25·2) 23·0 (16·2 to 33·5) 1·2 (0·8 to 1·7) 2·7 (2·1 to 3·3) 3·9 (3·0 to 4·9) 4·2% (−1·5 to 48·2) 2·0% (−0·1 to 4·3) 7·7% (4·1 to 38·6) −0·1% (−3·7 to 2·0) Sloven ia 1·6 (1·1 to 2·5) 9·9 (6·7 to 15·2) 11·5 (7·9 to 17·8) 0·4 (0·3 to 0·4) 2·1 (2·0 to 2·2) 2·5 (2·4 to 2·6) −1·2% (−4·7 to 48·3) 4·1% (2·3 to 5·8) −1·8% (−2·1 to −1·4) −4·8% (−5·3 to −4·2) Eastern Europe 55 500 (46 300 to 68 700) 115 000 (90 500 to 142 000) 171 000 (139 000 to 209 000) 7730 (7650 to 7820) 17 400 (17 100 to 17 600) 25 100 (24 800 to 25 300) 8·8% (7·6 to 10·1) 12·4% (9·8 to 14·3) 8·4% (8·3 to 8·5) 2·1% (2·0 to 2·3) Belarus 1280 (781 to 1930) 1260 (935 to 1680) 2540 (1730 to 3610) 94·1 (89·0 to 99·4) 210 (194 to 226) 304 (288 to 320) 7·8% (0·9 to 66·8) 9·8% (6·8 to 12·6) 10·6% (10·1 to 11·0) 0·3% (−0·4 to 1·0) Estonia 41·6 (30·5 to 61·3) 165 (119 to 238) 207 (152 to 298) 7·8 (7·4 to 8·2) 29·4 (27·3 to 31·7) 37·2 (35·0 to 39·5) 27·4% (24·5 to 30·8) −1·8% (−5·3 to 2·9) 25·5% (25·1 to 25·9) 0·2% (−0·6 to 1·1) Latvia 203 (134 to 238) 255 (204 to 313) 458 (345 to 528) 57·1 (54·4 to 59·7) 69·4 (64·8 to 74·0) 126 (121 to 132) 4·0% (2·9 to 4·9) 7·1% (3·8 to 8·5) 5·5% (5·1 to 5·9) 5·1% (4·5 to 5·8) Lithuania 41·1 (10·0 to 61·9) 94·0 (26·1 to 146) 135 (36·0 to 196) 42·3 (40·3 to 44·6) 24·7 (23·1 to 26·4) 67·0 (64·5 to 69·7) 6·4% (4·3 to 62·5) −0·1% (−12·3 to 2·7) 9·2% (8·9 to 9·6) 0·6% (0·0 to 1·2) Moldova 377 (267 to 548) 398 (287 to 560) 775 (569 to 1080) 49·5 (46·8 to 52·1) 124 (116 to 132) 173 (164 to 182) 8·5% (0·9 to 14·6) 1·8% (−0·5 to 3·6) 10·7% (10·2 to 11·1) −3·1% (−3·9 to −2·4) Russia 41 700 (34 600 to 51 600) 97 100 (78 500 to 122 000) 139 000 (115 000 to 171 000) 5480 (5430 to 5520) 13 800 (13 700 to 14 000) 19 300 (19 100 to 19 500) 10·0% (8·8 to 12·3) 13·2% (10·3 to 15·5) 7·0% (6·9 to 7·0) 5·9% (5·8 to 6·0) Ukraine 11 800 (7920 to 16 500) 16 100 (9210 to 20 500) 27 900 (19 000 to 35 300) 2000 (1930 to 2080) 3070 (2930 to 3220) 5080 (4920 to 5240) 5·9% (2·8 to 8·1) 9·7% (6·9 to 12·6) 10·7% (10·5 to 11·0) −5·9% (−6·3 to −5·5) High-income 34 600 (19 800 to 51 000) 88 600 (52 700 to 130 000) 123 000 (72 600 to 179 000) 3750 (3720 to 3790) 10 200 (10 100 to 10 300) 13 900 (13 800 to 14 000) −1·8% (−3·1 to −0·4) 2·0% (−0·2 to 3·2) −5·0% (−5·1 to −5·0) −4·9% (−4·9 to −4·8) Australasia 417 (228 to 745) 1160 (691 to 1760) 1580 (950 to 2270) 20·6 (19·8

income 34 600 (19 800 to 51 000) 88 600 (52 700 to 130 000) 123 000 (72 600 to 179 000) 3750 (3720 to 3790) 10 200 (10 100 to 10 300) 13 900 (13 800 to 14 000) −1·8% (−3·1 to −0·4) 2·0% (−0·2 to 3·2) −5·0% (−5·1 to −5·0) −4·9% (−4·9 to −4·8) Australasia 417 (228 to 745) 1160 (691 to 1760) 1580 (950 to 2270) 20·6 (19·8 to 21·4) 65·9 (63·8 to 68·1) 86·6 (84·4 to 88·9) −0·3% (−2·8 to 1·9) 2·5% (1·2 to 4·0) −8·4% (−8·7 to −8·2) −5·9% (−6·2 to −5·6) Australia 378 (202 to 698) 1020 (591 to 1560) 1390 (834 to 2030) 18·7 (17·9 to 19·5) 57·9 (55·8 to 60·0) 76·6 (74·4 to 78·9) −1·0% (−3·7 to 1·4) 2·8% (1·5 to 4·3) −8·2% (−8·4 to −7·9) −6·3% (−6·7 to −6·0) New Zealand 38·9 (20·7 to 59·8) 147 (68·4 to 233) 186 (89·6 to 293) 1·9 (1·9 to 2·0) 8·1 (7·8 to 8·3) 10·0 (9·7 to 10·3) 5·7% (4·1 to 8·9) 0·6% (−4·2 to 3·7) −10·4% (−10·7 to −10·2) −2·6% (−3·0 to −2·2) High-income Asia Pacific 669 (396 to 955) 2310 (1240 to 3660) 2970 (1680 to 4530) 79·9 (78·1 to 81·9) 312 (302 to 322) 392 (382 to 402) 5·0% (2·4 to 8·5) 0·2% (−4·5 to 2·7) 4·5% (4·3 to 4·7) −2·2% (−2·5 to −1·9) Brunei 7·9 (4·3 to 12·8) 27·8 (15·1 to 45·9) 35·7 (19·8 to 57·7) 1·4 (1·1 to 1·6) 3·3 (2·8 to 3·9) 4·7 (4·0 to 5·4) 5·3% (3·9 to 6·6) 3·7% (0·4 to 5·8) 6·4% (5·0 to 8·0) −1·6% (−3·9 to 0·3) Japan 426 (212 to 624) 1040 (501 to 1510) 1460 (724 to 2130) 59·4 (58·1 to 60·6) 147 (144 to 151) 207 (203 to 210) 4·4% (2·8 to 6·1) 1·9% (−2·1 to 3·5) 5·8% (5·8 to 6·0) −2·9% (−3·0 to −2·7) Singapore 56·0 (22·3 to 90·6) 113 (62·9 to 196) 169 (90·1 to 268) 3·1 (2·8 to 3·5) 29·6 (27·5 to 31·7) 32·7 (30·6 to 34·9) −5·6% (−8·1 to −4·0) −1·1% (−3·8 to 1·6) 7·5% (6·9 to 8·1) −6·7% (−7·7 to −5·8) South Korea 180 (50·8 to 296) 1130 (273 to 2280) 1310 (327 to 2530) 16·1 (14·9 to 17·4) 131 (122 to 142) 148 (138 to 157) 8·3% (1·7 to 62·6) −2·6% (−13·6 to 2·2) 1·0% (0·5 to 1·4) −0·7% (−1·5 to 0·1) High-income North America 15 600 (6270 to 24 700) 39 200 (16 100 to 62 300) 54 800 (22 600 to 86 200) 2140 (2110 to 2170) 5480 (5400 to 5550) 7620 (7540 to 7690) −2·3% (−3·8 to 0·9) 1·6% (−4·7 to 3·6) −6·2% (−6·2 to −6·1) −6·0% (−6·2 to −5·9) Canada 942 (434 to 1490) 2590 (1050 to 4310) 3530 (1520 to 5660) 72·2 (68·5 to 76·3) 202 (189 to 215) 274 (261 to 287) −1·8% (−3·5 to 1·9) 3·9% (−1·7 to 6·8) −4·8% (−5·1 to −4·6) −5·6% (−6·2 to −5·0) Greenland 6·2 (2·2 to 11·1) 4·7 (2·1 to 7·7) 10·9 (4·6 to 18·1) 0·8 (0·6 to 1·0) 0·9 (0·6 to 1·1) 1·6 (1·3 to 2·0) −0·8% (−2·5 to 2·2) 4·2% (−2·2 to 6·6) 0·1% (−1·3 to 1·7) −5·0% (−7·6 to −3·4) USA 14 700 (5440 to 23 600) 36

4 (261 to 287) −1·8% (−3·5 to 1·9) 3·9% (−1·7 to 6·8) −4·8% (−5·1 to −4·6) −5·6% (−6·2 to −5·0) Greenland 6·2 (2·2 to 11·1) 4·7 (2·1 to 7·7) 10·9 (4·6 to 18·1) 0·8 (0·6 to 1·0) 0·9 (0·6 to 1·1) 1·6 (1·3 to 2·0) −0·8% (−2·5 to 2·2) 4·2% (−2·2 to 6·6) 0·1% (−1·3 to 1·7) −5·0% (−7·6 to −3·4) USA 14 700 (5440 to 23 600) 36  600 (14 300 to 58 900) 51 300 (19 700 to 81 300) 2070 (2040 to 2100) 5270 (5200 to 5340) 7340 (7260 to 7420) −2·3% (−3·9 to 1·0) 1·5% (−5·6 to 3·6) −6·2% (−6·3 to −6·2) −6·0% (−6·2 to −5·9) Southern Latin America 6630 (3020 to 12 300) 10 500 (5830 to 17 100) 17 100 (9040 to 28 300) 673 (657 to 690) 1730 (1700 to 1770) 2400 (2370 to 2440) 1·7% (0·3 to 3·2) 1·8% (0·8 to 3·0) 5·0% (4·9 to 5·1) −0·9% (−1·1 to −0·7) Argentina 5750 (2490 to 11 300) 6430 (3330 to 10 900) 12 200 (6000 to 21 000) 548 (533 to 565) 1170 (1140 to 1200) 1720 (1680 to 1750) 1·0% (−0·6 to 2·6) 1·0% (−0·3 to 2·7) 4·7% (4·6 to 4·8) −1·4% (−1·7 to −1·2) Chile 682 (339 to 1180) 3440 (1850 to 5490) 4120 (2220 to 6690) 85·3 (82·9 to 87·7) 419 (407 to 430) 505 (493 to 516) 5·0% (3·8 to 6·2) 3·7% (2·8 to 4·7) 6·3% (6·1 to 6·5) 0·6% (0·3 to 0·9) Uruguay 193 (131 to 279) 635 (527 to 781) 828 (671 to 1040) 39·8 (38·8 to 40·8) 141 (137 to 144) 181 (177 to 184) 0·4% (−6·0 to 4·8) 5·9% (3·5 to 8·5) 7·1% (7·0 to 7·3) −0·5% (−0·8 to −0·2) Western Europe 11 200 (6590 to 16 200) 35 500 (21 200 to 51 600) 46 700 (27 800 to 67 200) 836 (818 to 856) 2610 (2550 to 2670) 3450 (3390 to 3510) −2·7% (−4·8 to −1·1) 2·0% (0·2 to 3·4) −4·9% (−5·0 to −4·8) −5·4% (−5·6 to −5·2) Andorra 0·6 (0·0 to 2·8) 4·0 (0·2 to 19·1) 4·6 (0·3 to 21·4) 0·1 (0·0 to 0·4) 0·3 (0·0 to 1·5) 0·4 (0·0 to 1·9) −0·6% (−11·1 to 7·0) 0·4% (−6·3 to 7·3) −3·6% (−10·2 to 1·4) −8·5% (−12·8 to −3·2) Austria 169 (93·4 to 248) 518 (273 to 798) 688 (369 to 1030) 8·5 (8·0 to 9·0) 29·5 (27·9 to 31·2) 38·0 (36·3 to 39·8) −0·0% (−6·1 to 65·4) 0·8% (−2·0 to 4·5) −2·0% (−2·4 to −1·6) −4·7% (−5·4 to −4·1) Belgium 350 (153 to 541) 787 (298 to 1310) 1140 (444 to 1850) 20·0 (19·0 to 21·2) 41·8 (39·7 to 44·3) 61·9 (59·4 to 64·7) 1·6% (−0·9 to 3·6) −0·2% (−6·0 to 3·4) −4·3% (−4·6 to −4·0) −2·5% (−3·0 to −1·9) Cyprus 15·2 (10·4 to 21·0) 96·0 (68·3 to 138) 111 (79·5 to 158) 0·4 (0·4 to 0·5) 1·4 (1·2 to 1·7) 1·9 (1·6 to 2·1) 6·1% (3·6 to 9·2) 5·0% (2·7 to 7·6) 4·4% (2·5 to 6·0) −3·7% (−5·2 to −2·1) Denmark 155 (53·8 to 264) 253 (125 to 457) 408 (206 to 643) 7·7 (7·2 to 8·2) 19·2 (18·1 to 20·2) 26·9 (25·8 to 28·1) −2·3% (−4·0 to −0·8) 1·1% (−5·0 to 3·8) −5·9% (−6·3 to −5·

111 (79·5 to 158) 0·4 (0·4 to 0·5) 1·4 (1·2 to 1·7) 1·9 (1·6 to 2·1) 6·1% (3·6 to 9·2) 5·0% (2·7 to 7·6) 4·4% (2·5 to 6·0) −3·7% (−5·2 to −2·1) Denmark 155 (53·8 to 264) 253 (125 to 457) 408 (206 to 643) 7·7 (7·2 to 8·2) 19·2 (18·1 to 20·2) 26·9 (25·8 to 28·1) −2·3% (−4·0 to −0·8) 1·1% (−5·0 to 3·8) −5·9% (−6·3 to −5· 5) −3·4% (−4·1 to −2·9) Finland 34·5 (9·7 to 59·1) 127 (34·0 to 223) 161 (43·7 to 280) 1·7 (1·6 to 1·8) 4·8 (4·5 to 5·1) 6·5 (6·2 to 6·8) 2·6% (0·5 to 4·7) −1·2% (−10·7 to 4·0) −0·2% (−0·6 to 0·1) −10·2% (−10·8 to −9·6) France 1620 (734 to 2880) 3570 (1660 to 6480) 5190 (2670 to 8610) 134 (126 to 142) 358 (337 to 380) 491 (469 to 514) −3·1% (−6·1 to −1·1) 0·3% (−1·6 to 2·0) −8·7% (−8·9 to −8·5) −6·1% (−6·7 to −5·6) Germany 1460 (884 to 2050) 4980 (2920 to 7180) 6430 (3810 to 9130) 102 (96·3 to 107) 363 (345 to 381) 465 (446 to 483) 1·3% (−0·2 to 2·2) 2·7% (−0·2 to 6·7) −7·0% (−7·2 to −6·7) −2·5% (−3·0 to −2·1) Greece 54·8 (35·3 to 75·5) 459 (280 to 805) 514 (326 to 854) 6·0 (5·6 to 6·4) 23·6 (22·3 to 24·8) 29·6 (28·2 to 30·9) 1·9% (−0·1 to 4·0) 0·1% (−2·1 to 2·6) −4·0% (−4·4 to −3·6) 3·3% (2·7 to 3·9) Iceland 0·0 (0·0 to 0·1) 20·1 (8·9 to 30·3) 20·1 (8·9 to 30·3) 0·0 (0·0 to 0·0) 0·8 (0·8 to 0·9) 0·9 (0·8 to 0·9) −0·1% (−2·5 to 1·9) −1·2% (−8·5 to 3·0) −4·3% (−4·8 to −4·0) −4·4% (−5·2 to −3·7) Ireland 148 (56·6 to 225) 414 (153 to 665) 562 (206 to 884) 2·7 (2·5 to 2·9) 7·1 (6·7 to 7·6) 9·8 (9·3 to 10·3) −0·9% (−2·8 to 0·9) 4·1% (−4·3 to 6·8) −0·5% (−0·9 to −0·1) −5·4% (−6·0 to −4·7) Israel 202 (106 to 288) 485 (253 to 695) 688 (377 to 970) 12·8 (12·2 to 13·5) 28·6 (27·1 to 30·1) 41·4 (39·8 to 43·0) 3·1% (0·6 to 5·8) 0·0% (−4·8 to 4·8) −1·2% (−1·5 to −0·9) −4·7% (−5·3 to −4·2) Italy 1580 (962 to 2300) 5560 (3360 to 8330) 7140 (4260 to 10 600) 156 (146 to 166) 505 (479 to 533) 661 (634 to 692) −5·5% (−11·6 to 2·7) 3·7% (0·8 to 7·0) −5·3% (−5·6 to −5·1) −5·1% (−5·6 to −4·6) Luxembourg 19·7 (9·6 to 28·9) 75·9 (31·9 to 115) 95·6 (41·5 to 141) 0·6 (0·6 to 0·7) 1·8 (1·7 to 1·9) 2·4 (2·3 to 2·5) −3·7% (−5·6 to −2·3) 2·5% (−4·8 to 5·9) −4·8% (−5·2 to −4·3) −4·8% (−5·5 to −4·2) Malta 16·7 (9·1 to 23·9) 81·7 (43·7 to 133) 98·4 (53·3 to 155) 0·4 (0·4 to 0·5) 1·2 (1·1 to 1·3) 1·6 (1·6 to 1·7) 3·7% (1·6 to 5·1) 9·1% (7·7 to 10·2) −2·1% (−2·5 to −1·8) −2·7% (−3·4 to −2·0) Netherlands 187 (95·1 to 283) 344 (166 to 526) 531 (260 to 804) 17·1 (16·1 to 18·1) 38·5 (36·5 to 40·7) 55·5 (53·3 to 58·0) 3·9% (2·5 to 5·2) −7·3% (−13·2 to −3·3) −8·5% (−8·8 to −8·2) −4·7% (−5·3 to −4·2) No

·2 (1·1 to 1·3) 1·6 (1·6 to 1·7) 3·7% (1·6 to 5·1) 9·1% (7·7 to 10·2) −2·1% (−2·5 to −1·8) −2·7% (−3·4 to −2·0) Netherlands 187 (95·1 to 283) 344 (166 to 526) 531 (260 to 804) 17·1 (16·1 to 18·1) 38·5 (36·5 to 40·7) 55·5 (53·3 to 58·0) 3·9% (2·5 to 5·2) −7·3% (−13·2 to −3·3) −8·5% (−8·8 to −8·2) −4·7% (−5·3 to −4·2) No rway 95·3 (50·8 to 141) 197 (99·0 to 300) 292 (150 to 441) 4·3 (4·2 to 4·4) 10·5 (10·2 to 10·7) 14·8 (14·5 to 15·1) −1·0% (−3·4 to 0·9) 0·7% (−2·8 to 3·6) −5·2% (−5·4 to −5·0) −1·0% (−1·3 to −0·8) Portugal 1550 (776 to 2320) 6080 (2930 to 9150) 7630 (3730 to 11 200) 108 (102 to 115) 394 (372 to 418) 502 (479 to 528) 2·5% (−0·5 to 5·3) 2·5% (−2·1 to 6·1) 7·9% (7·6 to 8·2) −6·2% (−6·8 to −5·7) Spain 1150 (464 to 2120) 5380 (2160 to 9960) 6530 (2820 to 11 800) 146 (136 to 157) 563 (531 to 598) 708 (675 to 743) −10·3% (−12·1 to −8·6) 5·8% (1·8 to 8·4) −4·7% (−5·0 to −4·5) −7·3% (−7·9 to −6·8) Sweden 243 (109 to 379) 383 (185 to 593) 626 (305 to 962) 8·2 (7·8 to 8·6) 14·6 (14·0 to 15·1) 22·8 (22·1 to 23·4) −5·1% (−7·3 to −3·5) 1·5% (−3·9 to 3·6) −4·6% (−4·8 to −4·3) −5·3% (−5·7 to −5·0) Switzerland 194 (74·5 to 288) 460 (152 to 704) 654 (226 to 987) 10·5 (9·9 to 11·1) 27·9 (26·4 to 29·6) 38·3 (36·7 to 40·2) −7·0% (−9·3 to −4·0) −0·1% (−6·7 to 2·1) −7·0% (−7·4 to −6·6) −6·4% (−7·0 to −5·8) UK 1950 (1120 to 2800) 5200 (3010 to 7410) 7150 (4110 to 10 100) 89·2 (88·5 to 89·9) 175 (174 to 176) 264 (263 to 266) 3·7% (1·7 to 5·1) 1·1% (−0·8 to 2·5) −2·4% (−2·5 to −2·4) −3·9% (−3·9 to −3·8) Latin America and Caribbean 52 200 (44 900 to 60 600) 110 000 (94 300 to 129 000) 162 000 (140 000 to 188 000) 14 100 (12 800 to 16 400) 28 400 (26 700 to 31 100) 42 500 (39 800 to 47 300) −0·5% (−1·8 to 0·6) 2·1% (0·6 to 3·2) 4·1% (3·3 to 4·5) −2·9% (−3·4 to −2·1) Andean Latin America 3950 (2520 to 6980) 8440 (5850 to 13 200) 12 400 (8420 to 19 800) 1600 (803 to 3610) 3060 (1830 to 5640) 4660 (2630 to 9190) 5·6% (1·5 to 8·4) 2·0% (−0·2 to 4·5) 6·6% (3·1 to 9·6) 0·9% (−2·8 to 4·6) Bolivia 905 (77·5 to 3880) 1590 (168 to 5750) 2490 (260 to 9880) 458 (11·3 to 2490) 709 (31·3 to 3220) 1170 (46·3 to 5690) 0·2% (−10·4 to 8·7) 2·3% (−4·8 to 7·0) 4·0% (−4·3 to 15·5) 0·5% (−20·2 to 7·9) Ecuador 1050 (708 to 1410) 2060 (1740 to 2580) 3100 (2500 to 3980) 271 (252 to 293) 822 (775 to 874) 1090 (1040 to 1150) 9·2% (4·5 to 12·0) 1·4% (0·2 to 2·6) 11·9% (11·6 to 12·2) 0·4% (−0·2 to 1·0) Peru 1990 (1390 to 3030) 4790 (3320 to 7700) 6780 (4730 to 10 600) 868 (430 to 1320) 1530 (830 to 2370) 2400 (1270 to

7·9) Ecuador 1050 (708 to 1410) 2060 (1740 to 2580) 3100 (2500 to 3980) 271 (252 to 293) 822 (775 to 874) 1090 (1040 to 1150) 9·2% (4·5 to 12·0) 1·4% (0·2 to 2·6) 11·9% (11·6 to 12·2) 0·4% (−0·2 to 1·0) Peru 1990 (1390 to 3030) 4790 (3320 to 7700) 6780 (4730 to 10 600) 868 (430 to 1320) 1530 (830 to 2370) 2400 (1270 to 3710) 7·1% (5·1 to 8·4) 2·0% (−0·9 to 5·7) 6·4% (5·0 to 7·9) 1·1% (−3·7 to 4·0) Caribbean 8830 (6730 to 11 000) 9990 (8020 to 12 200) 18 800 (15 600 to 22 300) 3460 (2670 to 4410) 5100 (4300 to 6200) 8560 (7470 to 9950) −4·4% (−5·8 to −2·9) −2·5% (−4·1 to −0·9) 6·5% (4·1 to 8·1) −7·5% (−8·4 to −6·6) Antigua and Barbuda 2·6 (2·1 to 3·4) 6·3 (5·3 to 8·2) 8·9 (7·5 to 11·6) 2·3 (2·2 to 2·3) 5·2 (5·1 to 5·4) 7·5 (7·4 to 7·6) −6·8% (−8·4 to −5·6) 2·7% (0·4 to 5·8) −2·6% (−2·8 to −2·5) −0·8% (−1·1 to −0·6) The Bahamas 137 (115 to 186) 227 (186 to 286) 365 (321 to 463) 47·2 (46·1 to 48·5) 67·0 (65·3 to 68·8) 114 (112 to 116) −3·2% (−4·2 to −1·9) 3·4% (2·4 to 5·0) 1·8% (1·6 to 2·0) −3·0% (−3·3 to −2·8) Barbados 26·1 (18·7 to 37·2) 43·2 (30·3 to 61·5) 69·3 (49·6 to 98·4) 8·8 (8·6 to 8·9) 16·2 (15·9 to 16·6) 25·0 (24·6 to 25·4) −1·5% (−2·6 to 1·0) −0·6% (−4·0 to 3·2) −1·9% (−2·1 to −1·8) −3·6% (−3·9 to −3·4) Belize 113 (86·6 to 148) 138 (122 to 167) 251 (210 to 313) 28·6 (27·9 to 29·3) 52·6 (51·2 to 53·9) 81·2 (79·5 to 82·6) 0·3% (−1·0 to 1·8) 3·7% (2·5 to 5·2) 0·3% (0·1 to 0·5) −1·1% (−1·4 to −0·9) Bermuda 1·4 (1·2 to 1·6) 4·4 (3·9 to 5·0) 5·8 (5·1 to 6·6) 1·6 (1·5 to 1·6) 5·3 (5·2 to 5·4) 6·8 (6·7 to 7·0) −8·3% (−9·5 to −7·5) 1·0% (−0·8 to 3·4) −4·6% (−4·7 to −4·4) −2·0% (−2·3 to −1·7) Cuba 684 (295 to 1180) 2130 (1140 to 3350) 2810 (1540 to 4330) 58·6 (57·3 to 60·0) 303 (295 to 310) 361 (354 to 369) 11·5% (9·8 to 13·1) 5·8% (4·1 to 7·3) 4·5% (4·4 to 4·7) 7·7% (7·5 to 8·0) Dominica 2·0 (1·4 to 3·2) 4·7 (3·8 to 6·2) 6·6 (5·3 to 9·4) 1·2 (1·2 to 1·2) 3·9 (3·8 to 3·9) 5·0 (4·9 to 5·1) −5·3% (−6·4 to −4·4) 1·8% (−0·1 to 4·4) −1·4% (−1·6 to −1·2) −1·0% (−1·3 to −0·7) Dominican Republic 1740 (787 to 2930) 2220 (1210 to 3600) 3960 (2410 to 6000) 853 (332 to 1510) 1710 (1140 to 2520) 2560 (1800 to 3630) 0·2% (−2·4 to 3·0) −3·1% (−5·7 to −0·5) 21·5% (17·9 to 24·9) −6·2% (−8·1 to −4·5) Grenada 1·9 (1·4 to 2·6) 4·6 (4·0 to 5·7) 6·5 (5·5 to 8·2) 1·7 (1·6 to 1·7) 4·5 (4·4 to 4·6) 6·2 (6·0 to 6·3) −6·2% (−7·6 to −5·0) −0·1% (−1·8 to 2·1) −2·8% (−3·0 to −2·7) −0·8% (−1·1 to −0·5) Guyana 335 (236 to 426) 265 (190 to 354) 600 (431 to 756) 79·4 (77·6 to 81·2) 116 (113 to 120) 196 (192 to

−6·2% (−8·1 to −4·5) Grenada 1·9 (1·4 to 2·6) 4·6 (4·0 to 5·7) 6·5 (5·5 to 8·2) 1·7 (1·6 to 1·7) 4·5 (4·4 to 4·6) 6·2 (6·0 to 6·3) −6·2% (−7·6 to −5·0) −0·1% (−1·8 to 2·1) −2·8% (−3·0 to −2·7) −0·8% (−1·1 to −0·5) Guyana 335 (236 to 426) 265 (190 to 354) 600 (431 to 756) 79·4 (77·6 to 81·2) 116 (113 to 120) 196 (192 to 199) 5·0% (1·2 to 8·1) −4·2% (−5·7 to −2·3) 5·9% (5·7 to 6·0) −3·4% (−3·7 to −3·2) Haiti 4590 (2990 to 6280) 3320 (2170 to 4910) 7910 (5770 to 10 500) 1860 (1290 to 2550) 1990 (1570 to 2720) 3850 (3150 to 4870) −8·5% (−10·1 to −6·5) −6·1% (−8·3 to −4·2) 6·4% (2·5 to 9·7) −12·6% (−14·0 to −11·2) Jamaica 529 (369 to 759) 751 (590 to 968) 1280 (965 to 1700) 187 (182 to 191) 253 (247 to 259) 440 (432 to 447) −0·5% (−5·6 to 2·9) 1·2% (−1·6 to 3·7) 3·3% (3·2 to 3·5) −2·4% (−2·6 to −2·1) Puerto Rico 53·7 (45·2 to 62·8) 144 (118 to 171) 198 (165 to 231) 66·8 (65·3 to 68·4) 171 (167 to 176) 238 (234 to 243) −14·6% (−16·0 to −13·7) 2·8% (1·1 to 5·0) −5·7% (−5·8 to −5·5) −5·9% (−6·2 to −5·6) Saint Lucia 3·4 (2·7 to 4·8) 5·4 (4·5 to 7·1) 8·8 (7·3 to 11·7) 2·9 (2·8 to 3·0) 4·9 (4·8 to 5·1) 7·8 (7·7 to 8·0) −6·8% (−8·1 to −5·7) 1·6% (−0·5 to 4·3) −2·2% (−2·3 to −2·0) −2·5% (−2·7 to −2·2) Saint Vincent and the Grenadines 7·2 (5·9 to 9·7) 13·7 (11·6 to 17·5) 20·9 (17·6 to 27·0) 6·8 (6·6 to 7·0) 13·3 (13·0 to 13·7) 20·1 (19·8 to 20·5) −6·2% (−8·1 to −4·4) 1·2% (−0·7 to 3·8) 0·8% (0·7 to 1·0) −2·3% (−2·5 to −2·0) Suriname 93·8 (74·7 to 127) 114 (91·2 to 148) 207 (169 to 271) 40·2 (39·3 to 41·2) 61·6 (60·0 to 63·1) 102 (99·9 to 104) −0·7% (−4·6 to 3·7) 0·2% (−1·9 to 3·1) 1·9% (1·8 to 2·1) −3·9% (−4·1 to −3·6) Trinidad and Tobago 187 (151 to 250) 227 (192 to 280) 414 (354 to 512) 83·8 (81·6 to 85·9) 136 (132 to 139) 219 (216 to 224) −1·0% (−2·9 to 1·5) −1·1% (−2·7 to 0·9) 4·3% (4·2 to 4·5) −2·3% (−2·6 to −2·1) Virgin Islands 3·9 (2·8 to 5·9) 10·7 (7·2 to 17·0) 14·7 (10·1 to 22·8) 3·0 (2·9 to 3·0) 6·0 (5·8 to 6·2) 9·0 (8·8 to 9·1) −4·1% (−8·2 to −1·5) 4·4% (1·4 to 7·4) 0·3% (0·1 to 0·4) −0·6% (−0·9 to −0·4) Central Latin America 10 500 (7800 to 14 100) 35 300 (27 800 to 45 900) 45 800 (35 700 to 58 900) 3330 (3140 to 3480) 9770 (9420 to 10 100) 13 100 (12 600 to 13 500) −0·9% (−2·4 to 0·6) 2·7% (1·0 to 4·2) 4·6% (4·5 to 4·8) −1·8% (−2·2 to −1·6) Colombia 2430 (1640 to 3480) 7090 (5400 to 9580) 9520 (7040 to 12 900) 661 (627 to 697) 1980 (1890 to 2070) 2640 (2530 to 2740) 2·9% (0·4 to 5·4) 5·0% (2·4 to 7·6) 8·4% (8·1 to 8·7) −2·3% (−2·8 to −1·9) Costa Rica 91·4 (62·4 to 121) 295 (21

0) −0·9% (−2·4 to 0·6) 2·7% (1·0 to 4·2) 4·6% (4·5 to 4·8) −1·8% (−2·2 to −1·6) Colombia 2430 (1640 to 3480) 7090 (5400 to 9580) 9520 (7040 to 12 900) 661 (627 to 697) 1980 (1890 to 2070) 2640 (2530 to 2740) 2·9% (0·4 to 5·4) 5·0% (2·4 to 7·6) 8·4% (8·1 to 8·7) −2·3% (−2·8 to −1·9) Costa Rica 91·4 (62·4 to 121) 295 (21 7 to 406) 386 (283 to 526) 38·5 (36·5 to 40·6) 115 (108 to 122) 153 (146 to 160) 0·9% (0·1 to 1·9) −1·5% (−3·9 to 0·3) 2·2% (1·9 to 2·6) −0·4% (−0·9 to 0·2) El Salvador 458 (276 to 660) 789 (487 to 1080) 1250 (770 to 1690) 247 (86·2 to 378) 481 (235 to 685) 728 (320 to 1050) 7·1% (3·0 to 10·5) −4·9% (−7·7 to −2·4) 11·5% (9·9 to 12·9) −1·0% (−6·9 to 1·3) Guatemala 830 (375 to 1910) 1480 (709 to 3230) 2310 (1100 to 5120) 228 (217 to 240) 472 (449 to 496) 700 (676 to 727) 2·7% (−2·8 to 6·8) −1·2% (−6·9 to 3·7) 3·2% (2·9 to 3·5) −5·4% (−5·9 to −5·0) Honduras 211 (127 to 334) 268 (179 to 393) 479 (320 to 683) 25·4 (17·4 to 39·1) 37·5 (29·1 to 47·8) 62·9 (47·0 to 83·2) 0·5% (−1·9 to 2·7) −3·0% (−7·1 to 0·3) −0·9% (−3·2 to 3·2) −8·2% (−11·8 to −4·8) Mexico 3930 (2930 to 5560) 15 900 (11 700 to 21 400) 19 800 (14 600 to 26 000) 1250 (1230 to 1260) 4330 (4270 to 4380) 5580 (5520 to 5630) −3·8% (−5·2 to −2·4) 2·8% (0·7 to 4·8) 3·7% (3·6 to 3·8) −2·4% (−2·5 to −2·2) Nicaragua 718 (490 to 1030) 1240 (794 to 1750) 1960 (1320 to 2680) 195 (103 to 275) 395 (229 to 582) 590 (333 to 835) 14·3% (10·7 to 16·5) 6·1% (3·1 to 9·1) 11·3% (9·8 to 12·8) 9·5% (4·6 to 13·4) Panama 426 (338 to 538) 1440 (1120 to 2000) 1870 (1490 to 2510) 128 (121 to 135) 398 (377 to 420) 526 (505 to 548) −1·0% (−2·2 to 0·8) 4·8% (3·3 to 6·5) 3·4% (3·0 to 3·8) −1·0% (−1·6 to −0·5) Venezuela 1370 (979 to 1930) 6830 (5480 to 8270) 8200 (6490 to 10 100) 557 (522 to 593) 1570 (1480 to 1670) 2130 (2030 to 2230) 3·9% (3·0 to 5·2) 2·8% (0·9 to 4·5) 3·1% (2·9 to 3·3) −0·1% (−0·7 to 0·4) Tropical Latin America 29 000 (20 800 to 35 000) 56 000 (41 500 to 66 700) 85 000 (62 600 to 101 000) 5720 (5530 to 5880) 10 500 (10 200 to 10 700) 16 200 (15 700 to 16 500) 2·7% (1·2 to 4·0) 3·4% (1·8 to 4·2) 1·7% (1·6 to 1·9) −0·9% (−1·2 to −0·7) Brazil 28 400 (20 400 to 34 300) 55 000 (40 700 to 65 400) 83 300 (61 500 to 99 100) 5430 (5330 to 5540) 9970 (9800 to 10 100) 15 400 (15 200 to 15 600) 2·6% (1·1 to 3·9) 3·5% (1·9 to 4·3) 1·6% (1·5 to 1·7) −1·2% (−1·4 to −1·0) Paraguay 638 (412 to 896) 1000 (648 to 1350) 1640 (1070 to 2220) 283 (140 to 397) 478 (261 to 666) 762 (412 to 1060) 11·8% (4·3 to 16·4) −0·1% (−2·7 to 2·4) 11·9%

00 to 99 100) 5430 (5330 to 5540) 9970 (9800 to 10 100) 15 400 (15 200 to 15 600) 2·6% (1·1 to 3·9) 3·5% (1·9 to 4·3) 1·6% (1·5 to 1·7) −1·2% (−1·4 to −1·0) Paraguay 638 (412 to 896) 1000 (648 to 1350) 1640 (1070 to 2220) 283 (140 to 397) 478 (261 to 666) 762 (412 to 1060) 11·8% (4·3 to 16·4) −0·1% (−2·7 to 2·4) 11·9% (9·5 to 16·0) 5·4% (1·2 to 8·7) North Africa and Middle East 8300 (4650 to 15 300) 9240 (5490 to 18 900) 17 500 (10 700 to 32 500) 4690 (3410 to 6700) 4750 (3380 to 7020) 9440 (7180 to 13 100) 4·2% (1·4 to 7·7) −0·4% (−4·0 to 3·0) 10·5% (8·1 to 12·7) 0·5% (−1·6 to 2·4) Afghanistan 353 (12·2 to 1650) 569 (26·6 to 2360) 922 (41·7 to 3660) 108 (1·3 to 592) 194 (3·4 to 961) 302 (5·7 to 1450) 2·0% (−8·6 to 12·2) 8·4% (−2·2 to 16·0) 4·3% (−4·5 to 14·9) 2·6% (−21·8 to 11·3) Algeria 396 (4·8 to 1800) 505 (6·2 to 1970) 901 (11·1 to 3490) 130 (8·6 to 988) 197 (7·2 to 939) 327 (14·9 to 1860) 8·7% (2·3 to 14·1) −1·0% (−42·9 to 7·8) 8·9% (1·7 to 13·1) −2·2% (−17·4 to 4·7) Bahrain 4·6 (3·8 to 5·9) 15·5 (13·2 to 18·4) 20·2 (17·3 to 23·9) 3·0 (2·4 to 3·7) 8·5 (7·4 to 10·0) 11·5 (9·8 to 13·6) −4·3% (−6·1 to −2·7) −0·6% (−1·9 to 1·0) 1·8% (0·9 to 3·1) −6·4% (−7·5 to −5·4) Egypt 154 (111 to 223) 403 (261 to 552) 557 (394 to 723) 22·5 (14·3 to 39·9) 42·3 (28·7 to 62·7) 64·8 (45·6 to 92·3) 0·6% (−1·7 to 2·7) 6·8% (3·5 to 9·3) 1·3% (−0·6 to 4·6) −7·2% (−10·2 to −3·6) Iran 1540 (937 to 2200) 1610 (1110 to 2270) 3150 (2130 to 4190) 322 (273 to 371) 470 (428 to 521) 792 (707 to 889) 8·7% (1·1 to 14·5) 10·3% (6·7 to 12·5) 10·2% (9·3 to 11·4) 5·3% (4·2 to 6·4) Iraq 124 (61·7 to 241) 104 (52·3 to 201) 229 (115 to 443) 60·6 (41·0 to 94·3) 55·1 (36·1 to 86·5) 116 (77·9 to 181) 4·9% (0·4 to 8·1) 5·0% (0·6 to 9·1) 9·3% (6·6 to 11·6) 2·7% (−1·0 to 5·9) Jordan 18·5 (12·3 to 26·7) 35·2 (23·9 to 47·1) 53·7 (37·1 to 69·7) 8·5 (7·1 to 9·9) 14·5 (9·1 to 20·7) 22·9 (16·9 to 29·7) 5·8% (2·8 to 8·2) 1·1% (−2·1 to 3·7) 6·9% (6·0 to 8·1) 2·6% (−1·5 to 5·4) Kuwait 4·8 (4·0 to 6·4) 10·6 (7·9 to 13·5) 15·4 (12·3 to 19·2) 2·9 (2·4 to 3·4) 2·5 (2·3 to 2·8) 5·4 (4·9 to 6·0) −3·1% (−5·0 to −1·7) 1·4% (−1·3 to 4·0) −2·2% (−3·0 to −1·3) −5·1% (−6·6 to −3·4) Lebanon 92·2 (4·0 to 403) 162 (8·5 to 802) 254 (13·1 to 1090) 48·6 (0·8 to 210) 63·9 (2·1 to 368) 113 (3·3 to 578) −2·4% (−11·3 to 5·4) 3·7% (−6·8 to 11·7) −0·6% (−10·4 to 5·5) −1·2% (−13·0 to 8·4) Libya 95·0 (4·8 to 421) 93·9 (5·1 to 479) 189 (10·9 to 845) 47·3 (3·5 to 201) 56·0 (4·2 to 261) 103 (8·8 to 436) 5·4% (−4·9 to 14·2) 5·2% (−5·5 to 13·1) 9·1% (0·9 to

254 (13·1 to 1090) 48·6 (0·8 to 210) 63·9 (2·1 to 368) 113 (3·3 to 578) −2·4% (−11·3 to 5·4) 3·7% (−6·8 to 11·7) −0·6% (−10·4 to 5·5) −1·2% (−13·0 to 8·4) Libya 95·0 (4·8 to 421) 93·9 (5·1 to 479) 189 (10·9 to 845) 47·3 (3·5 to 201) 56·0 (4·2 to 261) 103 (8·8 to 436) 5·4% (−4·9 to 14·2) 5·2% (−5·5 to 13·1) 9·1% (0·9 to 16·7) 3·2% (−5·7 to 10·4) Morocco 557 (28·7 to 2580) 551 (28·5 to 2790) 1110 (59·0 to 4860) 246 (6·2 to 1740) 356 (8·4 to 2260) 602 (16·0 to 3660) 5·8% (−4·3 to 13·6) −4·0% (−15·1 to 3·9) 10·7% (−0·8 to 17·5) −4·6% (−19·4 to 2·6) Oman 54·4 (30·7 to 79·9) 541 (241 to 1040) 595 (273 to 1120) 11·4 (5·8 to 19·2) 123 (62·3 to 181) 135 (70·0 to 199) 9·8% (7·1 to 11·5) 4·3% (−2·1 to 8·4) 9·9% (8·6 to 11·1) 7·3% (0·1 to 10·1) Palestine 9·2 (5·7 to 14·5) 10·8 (7·0 to 16·3) 20·0 (12·7 to 30·8) 4·6 (3·8 to 5·5) 5·7 (4·8 to 6·7) 10·3 (8·6 to 12·1) 4·2% (2·5 to 5·5) 2·3% (−0·3 to 4·5) 9·9% (8·8 to 11·4) 0·7% (−0·5 to 1·9) Qatar 2·4 (1·8 to 3·1) 8·1 (6·4 to 10·5) 10·5 (8·5 to 13·3) 2·2 (1·5 to 2·9) 4·7 (4·1 to 5·7) 7·0 (5·8 to 8·1) −8·0% (−9·6 to −6·5) −2·4% (−3·9 to −0·7) −2·4% (−3·8 to −1·0) −5·0% (−6·3 to −2·9) Saudi Arabia 331 (187 to 706) 342 (198 to 652) 673 (397 to 1270) 245 (191 to 324) 281 (209 to 353) 526 (402 to 676) 1·0% (−3·6 to 4·2) −0·2% (−4·0 to 5·2) 6·4% (4·0 to 8·5) −1·4% (−4·5 to 0·8) Sudan 3760 (1610 to 7140) 2810 (1410 to 5420) 6570 (3440 to 12 200) 3110 (2130 to 4230) 2190 (1650 to 3010) 5300 (4150 to 6740) 3·7% (0·7 to 8·7) −5·7% (−11·3 to −0·3) 15·6% (12·3 to 18·0) 0·3% (−1·4 to 2·0) Syria 11·0 (7·2 to 16·8) 30·8 (16·5 to 52·2) 41·8 (28·7 to 62·9) 6·6 (4·8 to 9·7) 4·7 (3·2 to 8·1) 11·3 (8·2 to 17·7) 2·6% (−0·1 to 4·2) 1·6% (−1·1 to 5·7) 3·7% (2·2 to 4·7) −4·1% (−7·0 to −0·1) Tunisia 180 (19·3 to 750) 251 (26·5 to 905) 431 (50·0 to 1630) 80·4 (2·4 to 392) 113 (2·8 to 455) 194 (5·4 to 790) 9·8% (2·6 to 17·1) 4·3% (−4·4 to 11·1) 11·0% (2·9 to 16·5) 5·9% (−8·0 to 11·4) Turkey 171 (114 to 255) 303 (203 to 438) 474 (325 to 689) 70·6 (55·8 to 85·2) 130 (102 to 156) 200 (159 to 236) 12·3% (6·9 to 50·5) 1·5% (−1·1 to 3·5) 13·5% (9·0 to 43·3) 6·5% (3·3 to 9·1) United Arab Emirates 42·0 (2·1 to 191) 195 (8·7 to 1210) 237 (11·2 to 1290) 23·5 (1·9 to 98·8) 181 (11·0 to 967) 205 (12·5 to 1020) 11·2% (−3·9 to 22·2) 1·8% (−9·9 to 12·3) 10·2% (1·3 to 18·2) 20·3% (−5·0 to 34·3) Yemen 393 (13·5 to 1820) 689 (25·9 to 3110) 1080 (43·2 to 4940) 135 (1·9 to 699) 252 (4·7 to 1320) 387 (7·3 to 1940) 1·6% (−9·9 to 13·9) 5·5% (−7·6 to 13·7) 3·9% (−5·5 to 16·3) −0·0% (−22·1 to 9

98·8) 181 (11·0 to 967) 205 (12·5 to 1020) 11·2% (−3·9 to 22·2) 1·8% (−9·9 to 12·3) 10·2% (1·3 to 18·2) 20·3% (−5·0 to 34·3) Yemen 393 (13·5 to 1820) 689 (25·9 to 3110) 1080 (43·2 to 4940) 135 (1·9 to 699) 252 (4·7 to 1320) 387 (7·3 to 1940) 1·6% (−9·9 to 13·9) 5·5% (−7·6 to 13·7) 3·9% (−5·5 to 16·3) −0·0% (−22·1 to 9 ·6) South Asia 49 900 (29 500 to 82 500) 67 700 (39 100 to 116 000) 118 000 (69 100 to 195 000) 26 200 (22 500 to 35 100) 34 500 (28 300 to 51 000) 60 700 (51 400 to 84 900) 6·1% (1·1 to 11·6) −1·8% (−5·4 to 2·1) 30·2% (23·9 to 35·9) −12·4% (−14·0 to −9·4) Bangladesh 414 (6·6 to 1980) 542 (15·9 to 2330) 956 (21·7 to 4120) 266 (0·8 to 1520) 361 (1·8 to 1850) 627 (2·4 to 3170) 52·2% (31·4 to 60·7) −1·8% (−11·7 to 8·9) 43·8% (19·6 to 53·6) 8·8% (−11·1 to 18·6) Bhutan 41·0 (0·7 to 186) 94·2 (2·6 to 420) 135 (3·3 to 597) 24·1 (0·1 to 122) 57·2 (1·2 to 254) 81·3 (1·4 to 363) 3·9% (−10·9 to 19·1) 1·6% (−8·1 to 11·1) 8·9% (−2·9 to 22·1) 4·9% (−7·7 to 16·5) India 43 600 (27 600 to 65 800) 55 800 (34 600 to 82 900) 99 400 (62 900 to 148 000) 23 400 (21 800 to 25 100) 28 700 (27 100 to 30 500) 52 100 (49 200 to 55 400) 5·9% (1·1 to 11·4) −2·3% (−5·7 to 1·8) 34·0% (30·3 to 37·4) −13·6% (−14·3 to −12·9) Nepal 824 (12·9 to 3920) 1420 (37·3 to 6520) 2240 (52·3 to 9930) 762 (3·4 to 4880) 1860 (12·8 to 9710) 2620 (17·3 to 13 500) 30·1% (16·4 to 44·9) −11·1% (−21·4 to −1·3) 65·3% (52·9 to 81·6) −2·2% (−19·4 to 6·0) Pakistan 4990 (81·4 to 21 300) 9930 (306 to 42 500) 14 900 (445 to 62 900) 1780 (7·5 to 9830) 3530 (23·1 to 16 500) 5310 (31·3 to 25 900) 6·3% (−7·3 to 21·3) 7·8% (−1·7 to 18·8) 4·4% (−7·9 to 18·3) 12·9% (−11·2 to 25·5) Southeast Asia, east Asia, and Oceania 39 200 (30 600 to 50 500) 92 200 (69 600 to 127 000) 131 000 (103 000 to 175 000) 26 600 (22 900 to 32 800) 62 400 (58 300 to 68 100) 89 000 (81 900 to 99 600) 1·8% (0·3 to 3·8) −3·2% (−4·9 to −1·5) 13·5% (12·6 to 14·8) −0·6% (−1·8 to 0·5) East Asia 7320 (3270 to 12 800) 28 000 (11 300 to 48 800) 35 300 (14 500 to 61 500) 9070 (8330 to 10 400) 27 400 (25 400 to 30 500) 36 500 (33 800 to 40 900) 9·3% (5·8 to 11·9) −7·0% (−13·3 to −4·6) 8·6% (7·3 to 10·3) 4·7% (3·9 to 5·3) China 6860 (3020 to 11 400) 26 500 (10 900 to 45 900) 33 300 (13 800 to 57 300) 8640 (8080 to 9140) 26 200 (24 500 to 27 600) 34 800 (32 600 to 36 600) 9·2% (5·5 to 11·9) −7·2% (−13·3 to −4·9) 8·5% (7·3 to 10·2) 4·7% (3·9 to 5·2) North Korea 321 (2·4 to 1970) 869 (6·6 to 5270) 1190 (9·4 to 7690) 263 (3·1 to 1460) 634 (9·2 to 3590) 897 (13·6 to 5260

500 (10 900 to 45 900) 33 300 (13 800 to 57 300) 8640 (8080 to 9140) 26 200 (24 500 to 27 600) 34 800 (32 600 to 36 600) 9·2% (5·5 to 11·9) −7·2% (−13·3 to −4·9) 8·5% (7·3 to 10·2) 4·7% (3·9 to 5·2) North Korea 321 (2·4 to 1970) 869 (6·6 to 5270) 1190 (9·4 to 7690) 263 (3·1 to 1460) 634 (9·2 to 3590) 897 (13·6 to 5260 ) 11·0% (−0·2 to 27·3) −1·6% (−16·8 to 8·3) 14·5% (1·9 to 40·9) 5·7% (−2·6 to 15·8) Taiwan (province of China) 21·7 (10·1 to 35·8) 189 (94·4 to 340) 210 (106 to 378) 21·8 (19·8 to 23·9) 176 (161 to 193) 198 (182 to 215) 8·8% (5·3 to 11·8) −7·8% (−13·5 to −5·2) 7·0% (6·4 to 7·5) 4·0% (3·0 to 5·1) Oceania 1980 (247 to 7710) 1550 (214 to 6120) 3530 (461 to 13 200) 778 (391 to 2400) 729 (342 to 2050) 1510 (747 to 4310) 18·8% (9·9 to 26·3) −4·7% (−16·2 to 2·1) 27·4% (21·1 to 33·9) −7·9% (−11·8 to −3·6) American Samoa 0·2 (0·1 to 0·4) 0·5 (0·3 to 0·8) 0·7 (0·4 to 1·1) 0·1 (0·1 to 0·1) 0·2 (0·2 to 0·3) 0·3 (0·3 to 0·4) −1·8% (−8·6 to 4·5) 3·2% (−0·4 to 7·2) 7·0% (2·1 to 15·3) 0·6% (−1·8 to 2·8) Federated States of Micronesia 34·5 (0·2 to 235) 39·6 (0·4 to 246) 74·1 (0·6 to 434) 12·2 (0·1 to 51·6) 18·3 (0·3 to 83·6) 30·4 (0·5 to 131) 10·2% (−3·0 to 26·1) 7·3% (−7·3 to 18·0) 9·5% (−1·5 to 27·8) 12·7% (2·5 to 25·8) Fiji 24·8 (18·1 to 34·7) 22·1 (16·2 to 30·7) 46·9 (35·9 to 64·6) 5·5 (3·5 to 9·6) 4·9 (3·0 to 8·3) 10·4 (7·0 to 18·1) 2·5% (−3·9 to 7·4) 4·3% (2·8 to 6·0) 8·9% (5·8 to 14·1) −3·2% (−6·1 to 1·4) Guam 0·8 (0·4 to 1·4) 6·7 (3·8 to 11·3) 7·5 (4·2 to 12·6) 0·6 (0·4 to 0·8) 3·7 (2·9 to 4·7) 4·3 (3·3 to 5·5) −3·6% (−10·8 to 2·9) 4·0% (0·4 to 7·8) 5·0% (−0·0 to 13·9) 0·6% (−1·9 to 2·5) Kiribati 0·3 (0·2 to 0·7) 0·4 (0·2 to 0·6) 0·7 (0·4 to 1·3) 0·2 (0·2 to 0·3) 0·3 (0·2 to 0·3) 0·5 (0·4 to 0·6) −6·6% (−8·0 to −5·0) 0·9% (−3·5 to 6·5) 1·3% (−1·3 to 4·1) −3·5% (−4·3 to −2·6) Marshall Islands 3·9 (0·0 to 24·5) 3·8 (0·0 to 26·8) 7·7 (0·1 to 50·4) 1·4 (0·0 to 7·9) 1·3 (0·0 to 8·2) 2·7 (0·0 to 15·8) 0·1% (−11·7 to 15·0) 7·9% (−5·4 to 16·2) 5·8% (−4·2 to 20·9) 4·2% (−7·5 to 16·3) Northern Mariana Islands 0·2 (0·1 to 0·3) 0·6 (0·3 to 0·9) 0·7 (0·4 to 1·2) 0·1 (0·1 to 0·2) 0·3 (0·2 to 0·4) 0·4 (0·3 to 0·5) −0·0% (−6·6 to 5·8) 3·9% (0·5 to 7·5) 8·5% (4·0 to 16·3) 1·0% (−1·2 to 3·3) Papua New Guinea 1730 (155 to 7240) 1320 (108 to 5660) 3050 (268 to 12 000) 688 (339 to 2200) 636 (285 to 1880) 1320 (627 to 4030) 25·2% (19·0 to 34·9) −5·9% (−19·4 to 1·3) 48·6% (40·7 to 65·2) −9·1% (−13·0 to −4·4) Samoa 13·2 (0·1 to 85·8) 13·2 (0·1 to 95·2) 26·4 (0·3 to 172) 4·6 (0·1 to 25·7) 4·5 (0

·0% (−1·2 to 3·3) Papua New Guinea 1730 (155 to 7240) 1320 (108 to 5660) 3050 (268 to 12 000) 688 (339 to 2200) 636 (285 to 1880) 1320 (627 to 4030) 25·2% (19·0 to 34·9) −5·9% (−19·4 to 1·3) 48·6% (40·7 to 65·2) −9·1% (−13·0 to −4·4) Samoa 13·2 (0·1 to 85·8) 13·2 (0·1 to 95·2) 26·4 (0·3 to 172) 4·6 (0·1 to 25·7) 4·5 (0 ·1 to 26·8) 9·1 (0·1 to 55·5) 0·7% (−11·3 to 15·9) 8·1% (−5·0 to 16·3) 6·1% (−4·2 to 21·4) 4·6% (−6·7 to 17·0) Solomon Islands 42·8 (0·4 to 278) 37·0 (0·4 to 268) 79·8 (0·9 to 519) 15·1 (0·2 to 91·6) 13·0 (0·2 to 77·9) 28·1 (0·4 to 166) −0·1% (−12·2 to 14·5) 7·5% (−5·3 to 15·8) 5·9% (−4·0 to 20·8) 3·7% (−7·5 to 15·9) Tonga 1·1 (0·6 to 2·0) 1·5 (0·8 to 2·5) 2·6 (1·4 to 4·4) 0·5 (0·4 to 0·6) 0·7 (0·6 to 0·9) 1·2 (0·9 to 1·6) −0·2% (−7·0 to 5·8) 4·0% (0·5 to 8·1) 8·4% (3·7 to 16·3) 1·5% (−1·0 to 3·4) Vanuatu 20·7 (0·2 to 137) 17·7 (0·2 to 123) 38·3 (0·4 to 254) 7·0 (0·1 to 39·1) 6·4 (0·1 to 36·7) 13·4 (0·2 to 78·5) 3·3% (−9·0 to 18·4) 2·4% (−19·4 to 15·7) 6·8% (−3·5 to 22·4) 4·0% (−5·8 to 16·0) Southeast Asia 29 800 (24 100 to 38 400) 62 700 (52 500 to 81 000) 92 500 (78 300 to 116 000) 16 800 (13 300 to 22 800) 34 300 (30 700 to 38 900) 51 000 (44 700 to 60 300) −1·5% (−4·2 to 1·6) −2·2% (−3·8 to 0·5) 16·4% (15·0 to 18·0) −3·4% (−5·5 to −1·7) Cambodia 561 (108 to 1390) 485 (94·4 to 1230) 1050 (208 to 2520) 502 (303 to 893) 876 (546 to 1380) 1380 (884 to 1990) 2·3% (−3·8 to 9·6) −10·4% (−23·8 to −2·7) 36·8% (32·1 to 42·2) −15·2% (−18·6 to −11·4) Indonesia 6480 (5090 to 8730) 11 300 (8800 to 15 600) 17 700 (14 100 to 24 100) 2500 (2190 to 3060) 4270 (3850 to 4870) 6770 (6090 to 7680) 5·8% (3·1 to 64·4) 3·2% (1·8 to 4·8) 32·1% (30·8 to 34·4) 6·9% (6·3 to 7·6) Laos 223 (1·6 to 1370) 579 (6·0 to 3920) 802 (8·0 to 5270) 138 (0·3 to 1120) 388 (1·2 to 2780) 526 (1·5 to 4300) 25·4% (14·4 to 38·9) −6·6% (−17·3 to 1·9) 56·2% (40·6 to 67·6) 1·6% (−14·3 to 10·0) Malaysia 1200 (904 to 1540) 3180 (2060 to 3920) 4370 (3010 to 5230) 615 (328 to 844) 964 (744 to 1270) 1580 (1120 to 2070) −0·7% (−3·5 to 2·3) 0·2% (−2·1 to 1·4) 19·4% (16·7 to 22·2) −3·2% (−6·7 to −1·1) Maldives 0·9 (0·7 to 1·3) 0·3 (0·2 to 0·4) 1·2 (0·9 to 1·7) 0·6 (0·6 to 0·7) 0·1 (0·1 to 0·1) 0·8 (0·7 to 0·8) 4·2% (1·8 to 6·9) −2·3% (−4·9 to 0·3) 6·5% (4·8 to 8·4) 0·5% (−2·2 to 3·3) Mauritius 104 (71·7 to 140) 194 (145 to 269) 298 (233 to 382) 16·4 (15·3 to 17·5) 79·4 (73·9 to 85·3) 95·8 (90·2 to 102) 20·9% (18·9 to 21·8) 3·1% (0·3 to 5·8) 6·5% (6·0 to 7·0) 5·3% (4·4 to 6·2) Myanmar 4970 (2730 to 7240) 4740 (2970 t