Browse the corpus

Introduction Mid-life is a key stage for the development of obesity and cardiovascular disease risk.1 In England, 78% of men and 65% of women aged 45–75 years are overweight or obese,1 and 44% of adults aged 55–64 years do not meet recommended levels of physical activity.2 In the past 50 years, mass adoption of private motorised transport and the modification of built environments to facilitate car use has coincided with a decline in active travel and a rise in population prevalence of overweight and obesity. Laverty and colleagues3 reported that adults aged 50–65 years were 55% less likely to commute via public transport, 45% less likely to commute on foot, and 30% less likely to commute by bicycle than were 16–29-year olds. The commute to work has been identified by the UK National Institute for Health and Care Excellence (NICE) as a key intervention point.4 In England and Wales, 23·7 million working-age individuals commute regularly to a workplace, with 67% travelling by car.5 For many, a transition to more active modes might be possible, without requiring unacceptable time or financial costs.

l Institute for Health and Care Excellence (NICE) as a key intervention point.4 In England and Wales, 23·7 million working-age individuals commute regularly to a workplace, with 67% travelling by car.5 For many, a transition to more active modes might be possible, without requiring unacceptable time or financial costs. Previous studies3, 6, 7, 8, 9 have found a strong, independent, cross-sectional association between active or public transport commuting and reduced obesity risk. Compared with car commuters, individuals who used active and public transport had lower body-mass indexes (BMIs) and percentage body fat, and lower rates of diagnosed diabetes and hypertension.3, 6, 7, 9 A graded effect has also been found, whereby the magnitude of effect is greater across successively more active transport modes.9 However, a limitation of the evidence has been an overreliance on cross-sectional data, limiting causal inference. Martin and colleagues10 used longitudinal data from the British Household Panel Study to show that commuters who switched from car commuting to active or public modes experienced a significant, independent reduction in self-reported BMI. Equally those who transitioned from active to car commuting reported a significant increase in BMI. Mytton and colleagues11 used two waves of data from the Commuting in Cambridge study to demonstrate that maintenance of cycle commuting was associated with lower BMI when compared with maintenance of sedentary commuting. However corroborative results were not found for walking.11 The protective effects of cycle commuting have also been reported for schoolchildren.12 A Norwegian study13 showed that maintenance of active travel in pregnancy predicted lower gestational weight gain than switching to more sedentary modes. Much of the present evidence-base is hampered by reliance on self-reported height and weight, which are prone to bias.14 Besides changing of residential or employment locations, evidence is lacking on what socioeconomic, demographic, and health-related factors predict transitions from car to active commuting, or vice versa.

of the present evidence-base is hampered by reliance on self-reported height and weight, which are prone to bias.14 Besides changing of residential or employment locations, evidence is lacking on what socioeconomic, demographic, and health-related factors predict transitions from car to active commuting, or vice versa. Research in context Evidence before this study Studies have repeatedly shown that active commuting to work contributes to greater overall physical activity and is associated with reduced body-mass index (BMI), percentage body fat, and risk of reporting hypertension and type 2 diabetes diagnoses. Previous work has shown a graded effect of active commuting on BMI, wherein greater magnitudes of association are observed across progressively more active transportation modes. However, much of the existing evidence base is hampered by cross-sectional study designs and self-reported health outcome data. These limitations make causality hard to establish, and accuracy difficult to ensure. Previous studies using longitudinal data have contributed valuable evidence by showing that BMI decreases as individuals transition to, or maintain, active commuting. Added value of this study

Studies have repeatedly shown that active commuting to work contributes to greater overall physical activity and is associated with reduced body-mass index (BMI), percentage body fat, and risk of reporting hypertension and type 2 diabetes diagnoses. Previous work has shown a graded effect of active commuting on BMI, wherein greater magnitudes of association are observed across progressively more active transportation modes. However, much of the existing evidence base is hampered by cross-sectional study designs and self-reported health outcome data. These limitations make causality hard to establish, and accuracy difficult to ensure. Previous studies using longitudinal data have contributed valuable evidence by showing that BMI decreases as individuals transition to, or maintain, active commuting. Added value of this study This longitudinal study builds on these foundations by using objectively measured height and weight to derive an objective change in BMI outcome. The dataset, UK Biobank, allows for a focus on a lifecourse stage during which individuals are at particularly high risk for development of obesity and its behavioural risk factors: mid-life. The study shows that switching from more active (walking, cycling, or public transport) to more passive (car) commuting independently predicted a significant decrease in BMI of about 0·3 kg/m2. Conversely, switching from passive to more active commuting significantly and independently predicted a BMI increase of the same magnitude. Change in household income was found to be the key driver of commute mode transitions.

ive (car) commuting independently predicted a significant decrease in BMI of about 0·3 kg/m2. Conversely, switching from passive to more active commuting significantly and independently predicted a BMI increase of the same magnitude. Change in household income was found to be the key driver of commute mode transitions. Implications of all the available evidence Active commuting is a significant, independent determinant of bodyweight in mid-life. Public health policies that promote active travel to work, through encouragement of walking, cycling, and the use of public transport, could help prevent obesity in this critical period of the lifecourse (age 40–69 years). In this study, we used longitudinal data from UK Biobank, a large population-based study of UK adults in mid-life, to investigate associations between changing commute mode and objective measures of BMI, and to identify the determinants of transitions to more active modes of commuting. We aimed to determine which socioeconomic and demographic characteristics predicted switching to or from active commuting, investigate whether switching from car to active commuting (or the reverse) independently predicts change in objectively measured BMI, and to ascertain whether any association is attenuated by socioeconomic, demographic, or behavioural factors.

emographic characteristics predicted switching to or from active commuting, investigate whether switching from car to active commuting (or the reverse) independently predicts change in objectively measured BMI, and to ascertain whether any association is attenuated by socioeconomic, demographic, or behavioural factors. Methods Study design and data collection We used survey data from UK Biobank (project 5935) to longitudinally study adults aged 40–69 years, selected via National Health Service (NHS) patient registers and recruited to 22 regional assessment centres. Biobank collected baseline data nationwide between March, 2006, and July, 2010. The project also did repeat assessment at a single location (Stockport UK Biobank Coordinating Centre) between August, 2012, and June, 2013, for a subset of these participants.

ient registers and recruited to 22 regional assessment centres. Biobank collected baseline data nationwide between March, 2006, and July, 2010. The project also did repeat assessment at a single location (Stockport UK Biobank Coordinating Centre) between August, 2012, and June, 2013, for a subset of these participants. The sample of individuals who were present at both baseline and follow-up was refined to include only participants with complete data for all analytic variables at both timepoints. Four analytic samples were derived to address three objectives. For objective 1, we assessed individuals who had complete data for all hypothesised predictors and had either experienced a transition from car to active or public transport or conversely a transition from active or public transport to car commuting. For objective 2, we assessed individuals who experienced a transition from car to active or public transport or remained car commuters, and had complete data for all covariates. For objective 3, we assessed individuals who experienced a transition from active or public transport to car commuting or remained public or active transport users, and had complete data for all covariates. UK Biobank has approval from the North West Multi-centre Research Ethics Committee, the Patient Information Advisory Group, and the Community Health Index Advisory Group. Further details on the rationale, study design, survey methods, data collection, and ethical approval are available elsewhere.15, 16, 17, 18

The sample of individuals who were present at both baseline and follow-up was refined to include only participants with complete data for all analytic variables at both timepoints. Four analytic samples were derived to address three objectives. For objective 1, we assessed individuals who had complete data for all hypothesised predictors and had either experienced a transition from car to active or public transport or conversely a transition from active or public transport to car commuting. For objective 2, we assessed individuals who experienced a transition from car to active or public transport or remained car commuters, and had complete data for all covariates. For objective 3, we assessed individuals who experienced a transition from active or public transport to car commuting or remained public or active transport users, and had complete data for all covariates. UK Biobank has approval from the North West Multi-centre Research Ethics Committee, the Patient Information Advisory Group, and the Community Health Index Advisory Group. Further details on the rationale, study design, survey methods, data collection, and ethical approval are available elsewhere.15, 16, 17, 18 Procedures At both timepoints, participants were asked “what types of transport do you use to get to and from work?” and were able to select one or more of the following mode categories: car or motor vehicle, walk, public transport, or cycle. Responses were dichotomised to create a binary variable indicating whether the individual commuted solely by car, or by any other mode or mix of modes. This result was then used to derive two binary variables indicating whether the respondent had experienced one of the following transitions between baseline and repeat assessment: transition from car commuting to active or public transport commuting or transition from active or public transport to car commuting. These variables were used as outcome variables in the analyses for objective 1, and exposure variables for objectives 2 and 3.

llowing transitions between baseline and repeat assessment: transition from car commuting to active or public transport commuting or transition from active or public transport to car commuting. These variables were used as outcome variables in the analyses for objective 1, and exposure variables for objectives 2 and 3. Outcomes Change in BMI between baseline and follow-up was the primary outcome for study objectives 2 and 3. Anthropometric measurements were taken by trained staff using standard procedures detailed elsewhere.18 Height (measured using the Seca 202 stadiometer (Seca; Birmingham, UK) and weight (Tanita BC-418MA body composition analyser (Tanita; Amsterdam, Netherlands), was used to derive BMI via the standard formula. Change in BMI was calculated for each individual by subtracting BMI at baseline from BMI at follow-up.

ewhere.18 Height (measured using the Seca 202 stadiometer (Seca; Birmingham, UK) and weight (Tanita BC-418MA body composition analyser (Tanita; Amsterdam, Netherlands), was used to derive BMI via the standard formula. Change in BMI was calculated for each individual by subtracting BMI at baseline from BMI at follow-up. Covariates Factors hypothesised to confound the association between commute mode transition and BMI change were adjusted for in statistical analyses. They comprised both time-invariant factors (fixed characteristics or baseline measurements) and time-varying factors (changes between baseline and follow-up). Hypothesised time-invariant confounders were age at baseline, sex, ethnicity (white British, other white background, south Asian, Black Caribbean, Black African, Chinese, mixed ethnicity, other), baseline BMI, baseline highest educational qualification (university or college degree, further education [A level or equivalent], higher secondary education [ordinary level, GCSEs, or equivalent], secondary education [CSEs or equivalent], vocational qualifications [NVQ, Higher National Diploma, Higher National Certificate, or equivalent], professional qualifications, or none), baseline gross annual household income category (<£18 000, £18000–30 999, £31 000–51 000, £52 000–100 000, or >£100 000), manual occupation at baseline (usually or always vs rarely or never), baseline job involves standing or walking (usually or always vs rarely or never); and baseline days per week of at least 10 min moderate leisure physical activity. Hypothesised time-varying confounders between baseline and follow-up were change in income category (stable, decrease, or increase), self-rated general health (stable good health, good to poor health transition, poor to good health transition, or stable poor health), manual occupation status (stable, transition to non-manual work, or transition to manual work), days per week of moderate physical activity (stable, decrease, or increase), and occupational standing or walking level (stable, decrease, or increase).

lth transition, poor to good health transition, or stable poor health), manual occupation status (stable, transition to non-manual work, or transition to manual work), days per week of moderate physical activity (stable, decrease, or increase), and occupational standing or walking level (stable, decrease, or increase). Statistical analysis The analytic sample size was large enough to produce reliable estimates of BMI change. We used descriptive analysis to identify the prevalence of switching from active to sedentary modes of commuting and to describe the distribution of other key variables. We fitted separate bivariate logistic regression models to identify which socioeconomic, demographic, health, and behavioural factors predicted a transition to or from car commuting. To assess effects on BMI, we used two series of nested multivariate linear regression models to investigate the effects of switching from car commuting to active or public modes and to investigate the effects of switching from active or public commuting modes to car commuting. In each series of nested models, model 1 tested for a bivariate association between the commute transition exposure and the obesity outcome. Demographic and socioeconomic covariates were added for model 2 (baseline BMI, age, sex, ethnicity, baseline household income, household income change, and educational attainment). For the final model (model 3), health, physical activity, and occupational covariates were added (self-rated general health transitions, manual occupation transitions, days per week of leisure moderate physical activity, and changes between baseline and follow-up, occupational physical activity transitions). All analyses were done with Stata/SE, version 14.

physical activity, and occupational covariates were added (self-rated general health transitions, manual occupation transitions, days per week of leisure moderate physical activity, and changes between baseline and follow-up, occupational physical activity transitions). All analyses were done with Stata/SE, version 14. Role of the funding source The sponsors had no role in the design of the study; collection, analysis, and interpretation of the data; or writing of the report. The corresponding author had full access to the data and responsibility for the decision to submit for publication. Results 502 656 adults were surveyed at baseline. 20 346 (21%) participated in the repeat assessment (median follow-up 4·4 years [IQR 3·7–4·9]).18 5861 individuals had complete data for all analytic variables at both baseline and follow-up (table 1). For objective 1, 2993 individuals with complete data transitioned from car to active or public transport (table 2) and 1277 individuals with complete data transitioned from active or public transport to car commuting (table 2). For objective 2, 4126 individuals with compete covariate data transitioned from car to active or public transport or remained car commuters (table 3). For objective 3, 1735 individuals with compete covariate data transitioned from active or public transport to car commuting or remained public or active transport users (table 4).

ive 2, 4126 individuals with compete covariate data transitioned from car to active or public transport or remained car commuters (table 3). For objective 3, 1735 individuals with compete covariate data transitioned from active or public transport to car commuting or remained public or active transport users (table 4). 3646 baseline car commuters remained car commuters at follow-up (table 1). However, 480 (8%) individuals switched to active or public modes of commuting. Of these individuals, 44 (9%) had switched from car to exclusive walking or cycling, with 436 (91%) using public transport for part of their journey. Conversely, 416 (7%) individuals who commuted by active public modes at baseline had switched to car commuting at follow-up. Of these, 33 (8%) switched from exclusive walking or cycling with the rest switching from public transport. 1319 individuals used active or public modes at both baseline and follow-up. 1011 (17%) of 5861 had a decline in their household income category between baseline and follow-up, while a similar proportion reported an increase (18%, table 1). Although 4421 (75%) reported good general health at both timepoints, 616 (11%) had poor health at both timepoints and 490 (8%) went from good to poor general health.

) of 5861 had a decline in their household income category between baseline and follow-up, while a similar proportion reported an increase (18%, table 1). Although 4421 (75%) reported good general health at both timepoints, 616 (11%) had poor health at both timepoints and 490 (8%) went from good to poor general health. Only income was consistently associated with commuting mode transitions, for both sexes (table 2). Compared with individuals who remained in the same income category at both timepoints, respondents who experienced income loss were more likely to transition from car commuting to active or public modes (unadjusted OR 1·34, 95% CI 1·02–1·76; p=0·033). However, respondents who experienced income loss were also more likely to report a transition from active to public transport commuting to car travel compared with those who were active or public mode users at both timepoints (1·46, 1·04–2·05; p=0·0280). These results probably reflect changes in occupation, which might explain both income category and commute mode. Indeed, experiencing an income category increase was also predictive of transitioning from active or public modes to car use (1·62, 1·17–2·24; p=0·0040). The 480 individuals who switched from car commuting at baseline to active or public commuting modes at follow-up were compared with the 3646 individuals who remained car commuters at both timepoints (table 3). In the fully adjusted model, a transition to a more active commute was significantly and independently predictive of a 0·30 kg/m2 decrease in BMI (95% CI −0·47 to −0·13; p=0·0005).

ive or public commuting modes at follow-up were compared with the 3646 individuals who remained car commuters at both timepoints (table 3). In the fully adjusted model, a transition to a more active commute was significantly and independently predictive of a 0·30 kg/m2 decrease in BMI (95% CI −0·47 to −0·13; p=0·0005). The 416 individuals who switched from active or public commuting modes at baseline to car commuting at follow-up were compared with the 1319 individuals who reported commuting via active or public modes at both timepoints (table 4). In the fully adjusted model, experiencing a transition to car commuting was significantly and independently predictive of a 0·32 kg/m2 increase in BMI (95% CI 0·13 to 0·50; p=0·0008). Adjustment for hypothesised time varying and time invariant confounders did not attenuate the effects of commute mode transition.

le 4). In the fully adjusted model, experiencing a transition to car commuting was significantly and independently predictive of a 0·32 kg/m2 increase in BMI (95% CI 0·13 to 0·50; p=0·0008). Adjustment for hypothesised time varying and time invariant confounders did not attenuate the effects of commute mode transition. Discussion In our comparison of BMI changes in middle-aged adults who switched mode of commute with their counterparts who maintained their mode of commute, individuals who transitioned from car commuting at baseline to using active or public modes at follow-up had an average BMI decrease of about 0·3 kg/m2. This effect was not attenuated by adjustment for hypothesised demographic, socioeconomic, health, and behavioural confounders. The inverse effect was also found: individuals who transitioned from active or public modes at baseline to car commuting at follow-up typically had a BMI increase of about 0·3 kg/m2. This effect was also independent of fixed or changing demographic, socioeconomic, health, and behavioural factors. Of these factors, only income emerged as a consistent, independent predictor of commute mode transition. For the average man in the baseline sample (aged 52 years, 176·6 cm tall, weighing 85·1 kg) a BMI decrease of 0·3 kg/m2 is equivalent to a weight loss of approximately 1·0 kg (2·2 lbs). For the average woman in the baseline sample (aged 51 years, 163·9 cm tall, weighing 70·0 kg) a BMI decrease of 0·3 kg/m2 is equivalent to a weight loss of approximately 0·8 kg (1·8 lbs). Economic modelling undertaken to inform NICE obesity guidelines19 suggested that for overweight adults, weight management interventions costing £100 or less per head are cost-effective for the NHS if they result in a maintained weight loss of at least 1 kg.

equivalent to a weight loss of approximately 0·8 kg (1·8 lbs). Economic modelling undertaken to inform NICE obesity guidelines19 suggested that for overweight adults, weight management interventions costing £100 or less per head are cost-effective for the NHS if they result in a maintained weight loss of at least 1 kg. Most transitions reported here were between car use and public transport use, and vice versa. Previous cross-sectional studies have shown that compared with car commuting, public transport independently and significantly predicts lower BMI, with similar effect sizes to active modes.3, 6 However in previous work using cross-sectional data from UK Biobank’s baseline sample,9 although public transport commuting predicted lower BMI than car use, cycling and to a lesser extent walking to work was associated with lower BMI scores than public transport. Therefore, findings from the present study are probably an underestimation of the BMI decrease one would expect if it had been possible to model transitions between, for example, car commuting and cycle commuting. However this study adds strength to the argument that the incidental physical activity associated with the use of public transport, such as walking to and from transit stops, might play an important part in obesity prevention.

d been possible to model transitions between, for example, car commuting and cycle commuting. However this study adds strength to the argument that the incidental physical activity associated with the use of public transport, such as walking to and from transit stops, might play an important part in obesity prevention. These results support and corroborate the findings of Martin and colleagues10 in showing an independent, significant association between switching between sedentary and active commute modes and BMI change in the British Household Panel Survey 2004–07. Effect sizes are strikingly consistent: they found that switching from car to active or public transport commuting predicted a decrease in self-reported BMI of 0·32 kg/m2 (95% CI −0·60 to −0·05). They also found that the opposite transition predicted a BMI increase of 0·34 kg/m2 (0·05 to 0·64). Together these two studies provide strong evidence for an association between active commuting and BMI.

ctive or public transport commuting predicted a decrease in self-reported BMI of 0·32 kg/m2 (95% CI −0·60 to −0·05). They also found that the opposite transition predicted a BMI increase of 0·34 kg/m2 (0·05 to 0·64). Together these two studies provide strong evidence for an association between active commuting and BMI. Our study has strengths and limitations. UK Biobank is a high quality data resource that allows the use of objectively measured height and weight to provide unbiased BMI data. The comprehensive dataset also allows adjustment for a wide range of time-varying and time-invariant confounders. As randomised controlled trials are difficult to do in this area of research, longitudinal observational data might represent the best available evidence for policy development. However, the study is also subject to a range of limitations, many stemming from the relatively constrained sample size. First and foremost is the loss of nuance created by the need to combine active modes with public transport modes. This combination was attributable to the low prevalence of walking and cycling and the even lower incidence of transitions involving walkers and cyclists. Most respondents who switched from car commuting transitioned to public transport rather than to walking or cycling (and vice versa). The effect sizes reported in this study are therefore expected to be an underestimation of the likely BMI effects of transitions to walking or cycling. The necessary exclusion of commute distance from analyses is a limitation of this study, and a source of effect underestimation for long distance walkers or cyclists. The precise point at which a mode transition occurred is not known, and duration of exposure to a new commute mode is likely to be heterogeneous. As a result of these limitations, the precise effect sizes for the association between commute mode transitions and BMI change are subject to uncertainty.

alkers or cyclists. The precise point at which a mode transition occurred is not known, and duration of exposure to a new commute mode is likely to be heterogeneous. As a result of these limitations, the precise effect sizes for the association between commute mode transitions and BMI change are subject to uncertainty. Residual confounding by factors such as menopausal status, dietary energy intake, and physical activity might have occurred. Although we adjusted for leisure, occupational, and non-commute travel physical activity, these variables were self-reported and not comprehensive. Thus, although this study benefits from the inclusion of an objectively measured health outcome, the use of a self-reported exposure is a limitation. Social desirability bias might lead to under-reporting of car commuting. However, individuals’ propensity to misreport mode is likely to remain relatively fixed over time, strengthening the internal validity of the study.

on of an objectively measured health outcome, the use of a self-reported exposure is a limitation. Social desirability bias might lead to under-reporting of car commuting. However, individuals’ propensity to misreport mode is likely to remain relatively fixed over time, strengthening the internal validity of the study. The study is also subject to limitations stemming from attrition (mostly due to retirement) and missing data. Individuals who dropped out of the study could be systematically different from those who contributed to both waves of data collection. By definition, sample members with data at both timepoints were in the Stockport assessment centre catchment area. This may limit the generalisability of results to this geographical area. Only 21% of those invited by UK Biobank to take part in the repeat assessment did so, as described in the UK Biobank Repeat Assessment documentation.18 Furthermore, UK Biobank is not strictly representative of the UK mid-life population so results might not be fully generalisable.

of results to this geographical area. Only 21% of those invited by UK Biobank to take part in the repeat assessment did so, as described in the UK Biobank Repeat Assessment documentation.18 Furthermore, UK Biobank is not strictly representative of the UK mid-life population so results might not be fully generalisable. This study shows that individuals who switched from car commuting to public transport or active modes experienced a decrease in BMI. This decrease was independent of changes in the socioeconomic, demographic, health, and behavioural factors observed over the same period. These findings suggest that policies that enable and encourage the maintenance and uptake of commuting by more active modes such as public transportation, walking, or cycling could have an effect on obesity prevalence in this high-risk age group. Only 896 (15%) individuals in this study had a commute mode transition, suggesting untapped potential exists for interventions to facilitate uptake of active or public transport. Most individuals who switched from car commuting transitioned to public transport. The effects observed in this study are therefore primarily related to mass transit and the benefits gained from the incidental physical activity associated with its use. Thus, this study is likely to underestimate the effects on BMI of walking or cycling to work. Efforts to increase active travel to work through widening of access to mass transit systems and integrating them with opportunities for walking and cycling might represent an effective policy response to the obesity epidemic.

Thus, this study is likely to underestimate the effects on BMI of walking or cycling to work. Efforts to increase active travel to work through widening of access to mass transit systems and integrating them with opportunities for walking and cycling might represent an effective policy response to the obesity epidemic. Acknowledgments This research has been conducted using the UK Biobank Resource. UK Biobank is a registered charity that receives funding from the Wellcome Trust, the UK Medical Research Council MRC, the UK Department of Health, the Scottish Government, the Welsh Assembly, the British Heart Foundation, Diabetes UK, and the Northwest Regional Development Agency. This study was funded by a UK Medical Research Council Strategic Skills Postdoctoral Fellowship in Population Health, awarded to EF. SC is supported by a National Institute of Health Research Senior Fellowship. EW is funded by the ESRC International Centre for Lifecourse Studies. UK Medical Research Council. Skills Postdoctoral Fellowship in Population Health, awarded to EF. Contributors EF, SC, and EW conceived of the study and planned the analytic approach. EF did the analysis. EF, SC, and EW interpreted the results. EF drafted the paper. SC and EW commented on and edited further drafts. EF produced the final manuscript. SC and EW approved the final manuscript. Declaration of interests We declare no competing interests. Table 1 Descriptive analysis

Contributors EF, SC, and EW conceived of the study and planned the analytic approach. EF did the analysis. EF, SC, and EW interpreted the results. EF drafted the paper. SC and EW commented on and edited further drafts. EF produced the final manuscript. SC and EW approved the final manuscript. Declaration of interests We declare no competing interests. Table 1 Descriptive analysis Respondents (n=5861) Baseline age (years) 51 (6·25) Baseline BMI (kg/m2) 26·67 (4·47) Days per week of ≥10 min moderate physical activity 3·22 (2·27) Car commuter at t0 and t1 3646 (62%) Commuted by active or public commute modes at t0 and t1 1319 (23%) Transitioned from car to active or public modes, t0 to t1 480 (8%) Transitioned from active or public modes to car, t0 to t1 416 (7%) Sex Male 2977 (51%) Female 2884 (49%) Ethnicity White British 5293 (90%) Other white background 388 (7%) South Asian 47 (1%) Black Caribbean 18 (<1%) Black African 22 (<1%) Chinese 21 (<1%) Mixed background 28 (<1%) Other ethnic background 44 (1%) Gross annual household income <£18 000 406 (7%) £18 000–£30 999 1158 (20%) £31 000–£51 999 1892 (32%) £52 000–£100 000 1932 (33%) >£100 000 473 (8%) Gross annual household income category change, t0 to t1 Stable 3793 (65%) Decrease 1011 (17%) Increase 1057 (18%) Highest educational qualification at baseline College or university degree 2919 (50%) A levels or equivalent 803 (14%) O levels/GCSEs or equivalent 1083 (18%) CSEs or equivalent 287 (5%) NVQ, HND, HNC, or equivalent 338 (6%) Other professional qualifications 229 (4%) None of the above qualifications 202 (3%) Self-rated health transition between t0 and t1 Stable good health 4421 (75%) Good to poor health 490 (8%) Poor to good health 334 (6%) Stable poor health 616 (11%) Manual work status Non-manual work 4209 (72%) Manual work 1652 (28%) Change in manual work status between t0 and t1 Stable 5193 (89%) Transition to non-manual work 359 (6%) Transition to manual work 309 (5%) Change in days per week of ≥10min moderate activity, t0 to t1 Stable 1670 (28%) Decrease 2012 (34%) Increase 2179 (37%) Job involves standing or walking Never/rarely 2456 (42%) Sometimes 1878 (32%) Usually/always 1527 (26%) Change in occupational standing or walking levels, t0 to t1 Stable 4226 (72%) Decrease 832 (14%) Increase 803 (14%) Data are mean (SD) or n (%). t0=baseline. t1=repeat assessment.

rease 2012 (34%) Increase 2179 (37%) Job involves standing or walking Never/rarely 2456 (42%) Sometimes 1878 (32%) Usually/always 1527 (26%) Change in occupational standing or walking levels, t0 to t1 Stable 4226 (72%) Decrease 832 (14%) Increase 803 (14%) Data are mean (SD) or n (%). t0=baseline. t1=repeat assessment. Table 2 Separate bivariate logistic regression models assessing associations between demographic, socioeconomic, health, and behavioural factors for individuals transitioning from car commuting to active or public mode commuting between baseline (t0) and follow-up (t1; n=2993) or vice versa (n=1277)

rease 2012 (34%) Increase 2179 (37%) Job involves standing or walking Never/rarely 2456 (42%) Sometimes 1878 (32%) Usually/always 1527 (26%) Change in occupational standing or walking levels, t0 to t1 Stable 4226 (72%) Decrease 832 (14%) Increase 803 (14%) Data are mean (SD) or n (%). t0=baseline. t1=repeat assessment. Table 2 Separate bivariate logistic regression models assessing associations between demographic, socioeconomic, health, and behavioural factors for individuals transitioning from car commuting to active or public mode commuting between baseline (t0) and follow-up (t1; n=2993) or vice versa (n=1277) From car to active or public mode commuting (n=2993) From active or public mode to car commuting (n=1277) Unadjusted OR (95% CI) p value Unadjusted OR (95% CI) p value Baseline BMI 0·98 (0·96–1·01) 0·17 1·04 (1·01 to 1·07) 0·0050 Baseline age (years) 0·99 (0·98–1·01) 0·57 0·99 (0·97 to 1·01) 0·34 Sex Male 1 (reference) 1 (reference) Female 1·09 (0·88–1·36) 0·42 0·97 (0·75 to 1·25) 0·82 Highest educational qualification at baseline College or university degree 1 (reference) 1 (reference) A levels or equivalent 0·88 (0·64–1·21) 0·42 1·64 (1·14 to 2·48) 0·0080 O levels/GCSEs or equivalent 0·64 (0·47–0·88) 0·0060 1·74 (1·23 to 2·47) 0·0020 CSEs or equivalent 1·10 (0·70–1·75) 0·68 1·96 (0·96 to 4·03) 0·070 NVQ, HND, HNC, or equivalent 0·38 (0·21–0·72) 0·0030 1·25 (0·68 to 2·31) 0·47 Other professional qualifications 0·50 (0·27–0·94) 0·033 0·82 (0·31 to 2·19) 0·69 None of the above qualifications 0·56 (0·27–1·16) 0·12 0·72 (0·24 to 2·10) 0·55 Gross annual household income <£18 000 1 (reference) 1 (reference) £18 000–£30 999 0·90 (0·54–1·51) 0·69 1·48 (0·85 to 2·59) 0·17 £31 000–£51 999 0·87 (0·53–1·42) 0·58 1·46 (0·86 to 2·50) 0·16 £52 000–£100 000 0·70 (0·43–1·15) 0·16 1·64 (0·96 to 2·78) 0·070 >£100 000 0·98 (0·56–1·74) 0·96 1·45 (0·74 to 2·84) 0·28 Gross annual household income category change, t0 to t1 Stable 1 (reference) 1 (reference) Decrease 1·34 (1·02–1·76) 0·033 1·46 (1·04 to 2·05) 0·0280 Increase 1·08 (0·81–1·44) 0·60 1·62 (1·17 to 2·24) 0·0040 Self-rated health transition between t0 and t1 Stable good health 1 (reference) 1 (reference) Good to poor health 1·15 (0·79–1·67) 0·48 1·26 (0·80 to 1·98) 0·32 Poor to good health 1·24 (0·78–1·96) 0·36 0·84 (0·48 to 1·48) 0·55 Stable poor health 0·77 (0·51–1·16) 0·21 1·50 (0·95 to 2·37) 0·080 Days per week of ≥10 min moderate physical activity 0·99 (0·95–1·04) 0·82 0·93 (0·88 to 0·99) 0·0110 Walking for pleasure 0·98 (0·91–1·06) 0·64 0·99 (0·90 to 1·09) 0·87 Baseline manual work status Non-manual work 1 (reference) 1 (reference) Manual work 0·80 (0·62–1·02) 0·070 1·20 (0·89 to 1·61) 0·24 Change in occupational standing or walking levels, t0 to t1 Stable 1 (reference) 1 (reference) Decrease 0·89 (0·65–1·22) 0·48 1·11

r pleasure 0·98 (0·91–1·06) 0·64 0·99 (0·90 to 1·09) 0·87 Baseline manual work status Non-manual work 1 (reference) 1 (reference) Manual work 0·80 (0·62–1·02) 0·070 1·20 (0·89 to 1·61) 0·24 Change in occupational standing or walking levels, t0 to t1 Stable 1 (reference) 1 (reference) Decrease 0·89 (0·65–1·22) 0·48 1·11 (0·76 to 1·62) 0·59 Increase 1·00 (0·73–1·37) 0·99 1·33 (0·93 to 1·91) 0·12 Table 3 Nested multivariate linear regression models testing whether experiencing a transition from car commuting to active or public mode commuting between baseline and follow independently predicted body-mass index change (n=4126)

r pleasure 0·98 (0·91–1·06) 0·64 0·99 (0·90 to 1·09) 0·87 Baseline manual work status Non-manual work 1 (reference) 1 (reference) Manual work 0·80 (0·62–1·02) 0·070 1·20 (0·89 to 1·61) 0·24 Change in occupational standing or walking levels, t0 to t1 Stable 1 (reference) 1 (reference) Decrease 0·89 (0·65–1·22) 0·48 1·11 (0·76 to 1·62) 0·59 Increase 1·00 (0·73–1·37) 0·99 1·33 (0·93 to 1·91) 0·12 Table 3 Nested multivariate linear regression models testing whether experiencing a transition from car commuting to active or public mode commuting between baseline and follow independently predicted body-mass index change (n=4126) Model 1 Model 2 Model 3 Coefficient (95% CI) p value Coefficient (95% CI) p value Coefficient (95% CI) p value Stable car user 0 0 0 Experienced transition from car to active or public modes, t0 to t1 −0·31 (−0·48 to −0·13) 0·0005 −0·32 (−0·49 to −0·15) 0·0002 −0·30 (−0·47 to −0·13) 0·0005 Baseline BMI −0·07 (−0·09 to −0·06) <0·0001 −0·08 (−0·09 to −0·07) <0·0001 Age (years) −0·01 (−0·02 to −0·00) 0·013 −0·01 (−0·02 to −0·00) 0·0093 Sex Male 0 0 Female −0·03 (−0·14 to 0·08) 0·60 −0·03 (−0·14 to 0·08) 0·59 Ethnicity White British 0 0 Other white background 0·06 (−0·16 to 0·28) 0·58 0·03 (−0·19 to 0·25) 0·79 South Asian −0·12 (−0·70 to 0·47) 0·70 −0·19 (−0·76 to 0·39) 0·53 Black Caribbean 0·60 (−0·46 to 1·66) 0·27 0·50 (−0·54 to 1·54) 0·35 Black African −0·46 (−1·47 to 0·55) 0·37 −0·26 (−1·26 to 0·73) 0·61 Chinese −0·32 (−1·26 to 0·62) 0·51 −0·15 (−1·07 to 0·77) 0·75 Mixed background −0·81 (−1·66 to 0·04) 0·060 −0·78 (−1·62 to 0·06) 0·070 Other ethnic background −0·34 (−0·98 to 0·29) 0·29 −0·37 (−1·00 to 0·25) 0·24 Gross annual household income <£18 000 0 0 £18 000–£30 999 −0·16 (−0·42 to 0·10) 0·24 −0·20 (−0·46 to 0·06) 0·13 £31 000–£51 999 −0·21 (−0·46 to 0·04) 0·11 −0·23 (−0·49 to 0·02) 0·070 £52 000–£100 000 −0·18 (−0·44 to 0·08) 0·18 −0·22 (−0·48 to 0·05) 0·11 >£100 000 −0·11 (−0·42 to 0·21) 0·51 −0·15 (−0·47 to 0·16) 0·34 Gross annual household income category change, t0 to t1 Stable 0 0 Decrease −0·05 (−0·20 to 0·10) 0·53 −0·06 (−0·21 to 0·09) 0·44 Increase −0·10 (−0·25 to 0·05) 0·18 −0·11 (−0·26 to 0·03) 0·13 Highest educational qualification at baseline College or university degree 0 0 A levels or equivalent 0·17 (0·00 to 0·34) 0·051 0·19 (0·02 to 0·36) 0·025 O levels/GCSEs or equivalent 0·05 (−0·10 to 0·20) 0·51 0·07 (−0·08 to 0·22) 0·35 CSEs or equivalent 0·20 (−0·06 to 0·45) 0·13 0·19 (−0·06 to 0·44) 0·15 NVQ, HND, HNC, or equivalent 0·06 (−0·18 to 0·29) 0·65 0·11 (−0·12 to 0·35) 0·36 Other professional qualifications 0·34 (0·07 to 0·62) 0·014 0·35 (0·08 to 0·62) 0·012 None of the above qualifications 0·04 (−0·26 to 0·35) 0·79 0·05 (−0·25 to 0·35) 0·74 Self-rated health transition between t0 and t1 Stable good health 0 Good to poor health 0·73 (0·53 to

(−0·18 to 0·29) 0·65 0·11 (−0·12 to 0·35) 0·36 Other professional qualifications 0·34 (0·07 to 0·62) 0·014 0·35 (0·08 to 0·62) 0·012 None of the above qualifications 0·04 (−0·26 to 0·35) 0·79 0·05 (−0·25 to 0·35) 0·74 Self-rated health transition between t0 and t1 Stable good health 0 Good to poor health 0·73 (0·53 to 0·92) <0·0001 Poor to good health −0·66 (−0·90 to −0·42) <0·0001 Stable poor health 0·46 (0·28 to 0·63) <0·0001 Manual work status at baseline Non-manual work 0 Manual work −0·07 (−0·25 to 0·12) 0·47 Change in manual work status between t0 and t1 Stable 0 Transition to non-manual work 0·19 (−0·06 to 0·44) 0·14 Transition to manual work 0·08 (−0·17 to 0·33) 0·52 Days per week of ≥10 min moderate physical activity 0·00 (−0·03 to 0·03) 0·98 Change in days per week of ≥10 min moderate activity, t0 to t1 Stable 0 Decrease 0·16 (0·01 to 0·30) 0·031 Increase −0·18 (−0·32 to −0·04) 0·011 Job involves standing or walking at baseline Never/rarely 0 Sometimes 0·02 (−0·13 to 0·16) 0·84 Usually/always −0·05 (−0·23 to 0·14) 0·61 Change in occupational standing or walking levels, t0 to t1 Stable 0 Decrease 0·19 (0·02 to 0·36) 0·032 Increase −0·18 (−0·35 to −0·01) 0·036 Table 4 Nested multivariate linear regression models testing whether experiencing a transition from active or public mode commuting to car commuting between baseline and follow independently predicted body-mass index change (n=1735)

to t1 Stable 0 Decrease 0·19 (0·02 to 0·36) 0·032 Increase −0·18 (−0·35 to −0·01) 0·036 Table 4 Nested multivariate linear regression models testing whether experiencing a transition from active or public mode commuting to car commuting between baseline and follow independently predicted body-mass index change (n=1735) Model 1 Model 2 Model 3 Coefficient (95% CI) p value Coefficient (95% CI) p value Coefficient (95% CI) p value Stable active or public modes user 0 0 0 Experienced transition from active or public modes to car, t0 to t1 0·31 (0·13 to 0·49) 0·0009 0·36 (0·17 to 0·54) 0·0001 0·32 (0·13 to 0·50) 0·0008 Baseline BMI −0·06 (−0·08 to −0·04) <0·0001 −0·07 (−0·08 to −0·05) <0·0001 Age (years) 0·01 (0·00 to 0·02) 0·168 0·01 (0·00 to 0·02) 0·206 Sex Male 0 0 Female −0·03 (−0·19 to 0·13) 0·735 −0·01 (−0·16 to 0·15) 0·936 Ethnicity White British 0 0 Other white background −0·07 (−0·38 to 0·23) 0·636 −0·10 (−0·40 to 0·20) 0·520 South Asian 0·30 (−0·67 to 1·28) 0·542 0·26 (−0·72 to 1·23) 0·606 Black Caribbean 0·69 (−0·54 to 1·91) 0·270 0·64 (−0·58 to 1·86) 0·303 Black African −1·02 (−2·04 to 0·01) 0·052 −1·09 (−2·11 to −0·070) 0·036 Chinese 0·44 (−0·78 to 1·66) 0·482 0·48 (−0·74 to 1·69) 0·443 Mixed background −0·71 (−1·68 to 0·27) 0·156 −0·66 (−1·64 to 0·31) 0·181 Other ethnic background 0·18 (−0·72 to 1·08) 0·696 0·07 (−0·83 to 0·97) 0·876 Gross annual household income <£18 000 0 0 £18 000–£30 999 −0·12 (−0·42 to 0·18) 0·426 −0·07 (−0·37 to 0·23) 0·651 £31 000–£51 999 −0·10 (−0·40 to 0·20) 0·520 −0·06 (−0·37 to 0·24) 0·673 £52 000–£100 000 −0·09 (−0·40 to 0·22) 0·581 −0·03 (−0·35 to 0·28) 0·849 >£100 000 0·12 (−0·29 to 0·53) 0·574 0·19 (−0·22 to 0·60) 0·367 Gross annual household income category change, t0 to t1 Stable 0 0 Decrease −0·30 (−0·52 to −0·08) 0·0071 −0·30 (−0·52 to −0·08) 0·0082 Increase −0·05 (−0·26 to 0·17) 0·661 −0·07 (−0·28 to 0·15) 0·540 Highest educational qualification at baseline College or university degree 0 0 A levels or equivalent 0·15 (−0·09 to 0·38) 0·221 0·12 (−0·11 to 0·36) 0·310 O levels/GCSEs or equivalent 0·20 (−0·03 to 0·43) 0·084 0·18 (−0·05 to 0·41) 0·125 CSEs or equivalent 0·64 (0·21 to 1·07) 0·0033 0·57 (0·14 to 1·00) 0·0094 NVQ, HND, HNC, or equivalent 0·14 (−0·24 to 0·53) 0·458 0·03 (−0·36 to 0·42) 0·881 Other professional qualifications 0·43 (−0·05 to 0·91) 0·080 0·44 (−0·04 to 0·92) 0·072 None of the above qualifications 0·17 (−0·33 to 0·67) 0·505 0·07 (−0·44 to 0·57) 0·801 Self-rated health transition between t0 and t1 Stable good he

NVQ, HND, HNC, or equivalent 0·14 (−0·24 to 0·53) 0·458 0·03 (−0·36 to 0·42) 0·881 Other professional qualifications 0·43 (−0·05 to 0·91) 0·080 0·44 (−0·04 to 0·92) 0·072 None of the above qualifications 0·17 (−0·33 to 0·67) 0·505 0·07 (−0·44 to 0·57) 0·801 Self-rated health transition between t0 and t1 Stable good he alth 0 Good to poor health 0·55 (0·27 to 0·83) 0·0001 Poor to good health −0·46 (−0·79 to −0·14) 0·0055 Stable poor health 0·17 (−0·12 to 0·46) 0·244 Manual work status at baseline Non-manual work 0 Manual work 0·19 (−0·08 to 0·47) 0·171 Change in manual work status between t0 and t1 Stable 0 Transition to non-manual work −0·10 (−0·49 to 0·29) 0·623 Transition to manual work −0·39 (−0·79 to 0·00) 0·049 Days per week of ≥10 min moderate physical activity −0·01 (−0·05 to 0·03) 0·596 Change in days per week of ≥10 min moderate activity, t0 to t1 Stable 0 Decrease 0·10 (−0·10 to 0·29) 0·333 Increase −0·05 (−0·25 to 0·16) 0·660 Job involves standing or walking at baseline Never/rarely 0 Sometimes 0·06 (−0·15 to 0·27) 0·573 Usually/always 0·03 (−0·26 to 0·31) 0·849 Change in occupational standing or walking levels, t0 to t1 Stable 0 Decrease 0·11 (−0·15 to 0·36) 0·398 Increase 0·00 (−0·24 to 0·23) 0·982

of the estimates apply to countries outside our focus. Evidence from post-communist countries on health outcomes of politico-economic transitions remains inconclusive, although available data suggest that adverse health outcomes could result from rising unemployment and stress levels associated with mass privatisation. Added value of this study

Introduction Individuals at high risk of serious complication after an influenza infection have historically been the target for seasonal influenza annual vaccination programmes worldwide.1, 2, 3 However, because of the indirect effects of vaccination, termed herd protection, the vaccination of groups who are important for transmission of infection is often also cost-effective.4, 5 A large proportion of this transmission is attributable to children and adolescents because of the relatively high number of contacts they have with others,6, 7, 8 the fraction of these contacts that involve touching,6, 7, 8 and their susceptibility to influenza infection.9, 10 Some countries have broadened their seasonal influenza vaccine recommendations to immunise healthy children and adolescents every year, for example in the USA11 and more recently in the UK in 2012.12, 13

8 the fraction of these contacts that involve touching,6, 7, 8 and their susceptibility to influenza infection.9, 10 Some countries have broadened their seasonal influenza vaccine recommendations to immunise healthy children and adolescents every year, for example in the USA11 and more recently in the UK in 2012.12, 13 If paediatric immunisation programmes gain high coverage early enough in the influenza season to interrupt transmission to risk groups, any existing risk group-based vaccine programme becomes less cost-effective. If a national paediatric programme renders elderly and risk group vaccination programmes not cost-effective, then removing annual vaccination from these target groups will allow a large annual saving. Previous analysis in the context of the UK suggests that substantial uncertainty exists regarding the cost-effectiveness of an elderly low-risk programme in the presence of a moderate vaccine uptake of 50% in healthy children12, 13 and adolescents, with a third of simulations finding the elderly vaccine not to be cost-effective.

nalysis in the context of the UK suggests that substantial uncertainty exists regarding the cost-effectiveness of an elderly low-risk programme in the presence of a moderate vaccine uptake of 50% in healthy children12, 13 and adolescents, with a third of simulations finding the elderly vaccine not to be cost-effective. To reduce seasonal influenza-associated serious disease in elderly people and in individuals clinically at risk from influenza-related disease, England and Wales are introducing a publicly funded universal funded paediatric vaccination programme,12, 13 using live attenuated influenza vaccine (LAIV). The LAIV programme began in 2013 and offered the intranasal vaccines to all healthy children aged 2–3 years and to healthy children aged 4 years and those in primary grades 1–6 (children commencing the school year aged 5–10 years) who were enrolled in a pilot study done in seven regions. In 2014–15, the programme extended to all children aged 2–4 years, and the ongoing pilot studies expanded to include some regions offering vaccines to secondary school-aged adolescents in school grades 7–8 (those commencing the school year aged 11–12 years). In 2015–16, the programme offered vaccines to primary grades 1–2 (those commencing the school year aged 5–6 years) and continued to offer vaccines to children aged 2–4 years and children in the previous pilot areas. From 2016 onwards, the programme extends year on year to children in the rest of the primary school grades (aged 5–10 years when commencing the school year) and secondary school adolescents (aged 11–15 years when commencing the school year).14 Case-control studies have been implemented over this period to ascertain the direct effectiveness across each season.15, 16

n year to children in the rest of the primary school grades (aged 5–10 years when commencing the school year) and secondary school adolescents (aged 11–15 years when commencing the school year).14 Case-control studies have been implemented over this period to ascertain the direct effectiveness across each season.15, 16 Research in context Evidence before this study Using a transmission model, our previous study showed that a paediatric vaccination programme can effectively control seasonal influenza because of the substantial herd protection conferred to unvaccinated children and adults. The decision to implement a paediatric vaccine programme in England was initiated after the evidence presented in a later study that combined the previous mathematical model with a cost-effectiveness analysis. Although this study showed paediatric vaccination to be very cost-effective, the study only superficially assessed the effect of this programme on existing target group vaccination. Added value of this study In this Article, we assess the epidemiological effect and cost-effectiveness of the existing elderly and high-risk group vaccination in the presence of paediatric vaccination. Our study suggests that a mass paediatric vaccination programme will not affect the existing recommendation of risk group vaccination in England. However, the continued cost-effectiveness of a low-risk elderly vaccination programme is uncertain. Implications of all the available evidence

In this Article, we assess the epidemiological effect and cost-effectiveness of the existing elderly and high-risk group vaccination in the presence of paediatric vaccination. Our study suggests that a mass paediatric vaccination programme will not affect the existing recommendation of risk group vaccination in England. However, the continued cost-effectiveness of a low-risk elderly vaccination programme is uncertain. Implications of all the available evidence This study highlights the importance of a cost-effective paediatric vaccination scheme in curtailing influenza transmission to the most vulnerable risk groups. However, important uncertainties exist surrounding the cost-effectiveness of elderly programmes in the UK that might warrant reconsideration of this strategy in the coming years. We suggest that other countries introducing paediatric vaccination should reconsider their full vaccination schedule in view of these results.

owever, important uncertainties exist surrounding the cost-effectiveness of elderly programmes in the UK that might warrant reconsideration of this strategy in the coming years. We suggest that other countries introducing paediatric vaccination should reconsider their full vaccination schedule in view of these results. Using England as a case study, we assess the likely effect and cost-effectiveness of the existing elderly and risk group vaccination programme under the new policy of mass paediatric vaccination. Specifically, we assess whether maintaining a low-risk elderly and high-risk seasonal influenza vaccination programme remains cost-effective in the presence of preschool and school-based vaccination with varying vaccine coverages achieved at different speeds. Although the focus of our study assesses the situation after the decision in England in 2012 to introduce preschool-age and school-age influenza vaccination, we extend this result to general policy considerations for other countries where paediatric programmes are currently under review.

s achieved at different speeds. Although the focus of our study assesses the situation after the decision in England in 2012 to introduce preschool-age and school-age influenza vaccination, we extend this result to general policy considerations for other countries where paediatric programmes are currently under review. Methods Epidemiological effect of paediatric vaccination For this cost-effectiveness analysis, to predict the direct and indirect effects of seasonal influenza vaccine programmes, we use a previously described transmission model of seasonal influenza in England and Wales that is calibrated to 14 seasons of data.12, 13 Using this calibrated model, we assessed the influenza incidence in two groups: the high-risk population (individuals older than 6 months who have a diagnosed clinical disorder that puts them at risk of complications after influenza infection); and the low-risk elderly population (individuals older than 64 years who are not categorised as high risk). We assess the influenza incidence with and without low-risk elderly and high-risk vaccination, in the presence of three different paediatric programmes: (1) preschool age (2–4 years) only; (2) preschool and primary school age (2–10 years); and (3) preschool, primary, and secondary school age (2–16 years). Because children are known to be epidemiological drivers for seasonal influenza transmission, the speed at which children are vaccinated will probably affect the disease burden of the rest of the population. Therefore, we considered three administration speeds: (1) slow (uptake achieved between Jan 1 and Jan 31); (2) linear (uptake achieved between Oct 1 and Jan 31); and (3) fast (uptake achieved between Oct 1 and Oct 31). We also assessed the effect of different paediatric vaccine coverages on the effectiveness of the low-risk elderly and high-risk vaccine programmes.

peeds: (1) slow (uptake achieved between Jan 1 and Jan 31); (2) linear (uptake achieved between Oct 1 and Jan 31); and (3) fast (uptake achieved between Oct 1 and Oct 31). We also assessed the effect of different paediatric vaccine coverages on the effectiveness of the low-risk elderly and high-risk vaccine programmes. Transmission model and LAIV assumptions We use an age-specific and risk-specific mathematical model that captures seasonal influenza transmission in England and Wales. The model is calibrated to the number of influenza-like illness consultations and the frequency of virological confirmations from 1995 to 2009. The calibrated model is parameterised with data on vaccination coverage, vaccine uptake speeds, and vaccine effectiveness data for elderly and high-risk inactivated influenza vaccination (IIV) in the absence of any paediatric vaccination. We calibrated the model using a Bayesian evidence synthesis approach, which captures uncertainty in the model parameters and is able to generate a distribution of model outcomes consistent with available data. The model captures the dynamics of A/H1N1, A/H3N2, and B strains separately. In each year, IIV is either matched or unmatched to the circulating strain. IIV efficacy was assumed to be 70% for people younger than 65 years and 46% for people aged 65 years or older for matched years, and 42% for people younger than 65 years and 28% for people aged 65 years or older for unmatched years. More details on the model and model parameters have been published previously.12

train. IIV efficacy was assumed to be 70% for people younger than 65 years and 46% for people aged 65 years or older for matched years, and 42% for people younger than 65 years and 28% for people aged 65 years or older for unmatched years. More details on the model and model parameters have been published previously.12 After the calibration, we used the model to predict the distribution of elderly and high-risk influenza cases averted in the presence of the new LAIV programme. In our base case scenario, we assumed LAIV efficacy across all strains was equal to IIV efficacy. Unless otherwise specified, our model was parameterised using the observed vaccine coverage and the speed of the uptake reported during the paediatric programme rollout for children aged 2–4 years and the pilot programme for children aged 4–16 years.17, 18

med LAIV efficacy across all strains was equal to IIV efficacy. Unless otherwise specified, our model was parameterised using the observed vaccine coverage and the speed of the uptake reported during the paediatric programme rollout for children aged 2–4 years and the pilot programme for children aged 4–16 years.17, 18 Incremental cost-effectiveness of the elderly vaccination programme Extending the economic framework previously used for decision making on the paediatric programme in England,12, 13 we assessed the net benefits accrued from continuing both the low-risk elderly and high-risk annual influenza vaccine programmes in England in the presence of a paediatric programme. We present the results in terms of the incremental cost for each quality-adjusted life-year gained (£ per QALY). For our base case analysis, we use 70% coverage for all children aged 2–16 years (consistent with uptakes in Scotland and Wales), together with the reported uptake speeds achieved in England for preschool-age children and school-age children (71% vs 40% achieved coverage by Oct 31, 89% vs 87% by Nov 30, and 94% vs 100% by Dec 31).18 For additional analyses, we also assessed the net benefits achieved under a range of feasible paediatric vaccination coverages (status quo, 50%, and 90%), together with a fast administration of children and adolescents by the end of October.

hieved coverage by Oct 31, 89% vs 87% by Nov 30, and 94% vs 100% by Dec 31).18 For additional analyses, we also assessed the net benefits achieved under a range of feasible paediatric vaccination coverages (status quo, 50%, and 90%), together with a fast administration of children and adolescents by the end of October. Officially, for an intervention to be deemed cost-effective in England, its incremental cost-effectiveness ratio (ICER) must typically have been less than £20 000 per QALY, although interventions with ICERs less than £30 000 per QALY were also considered.19 The mean ICER and the probability that the ICER falls below these thresholds (ie, the probability of cost-effectiveness) is also considered in the decision to fund a programme. Methodological research suggests that this threshold should be reduced to £13 000 per QALY20 and there is currently debate about whether the threshold will be updated.21 In practice, a threshold of about £15 000 per QALY would probably be deemed cost-effective. With these issues in mind, we present the mean ICER, and cumulative probabilities for three thresholds: £15 000 per QALY, £20 000 per QALY, and £30 000 per QALY.

is currently debate about whether the threshold will be updated.21 In practice, a threshold of about £15 000 per QALY would probably be deemed cost-effective. With these issues in mind, we present the mean ICER, and cumulative probabilities for three thresholds: £15 000 per QALY, £20 000 per QALY, and £30 000 per QALY. Economic evaluation assumptions To calculate the cost-effectiveness of the elderly programme and high-risk programmes, we integrated the transmission model into an economic evaluation. The economic evaluation tracks the number of general practitioner consultations, hospital admissions, and deaths for each year for the three strains. QALYs lost are assumed for febrile cases to have a mean of 7·49 × 10−3, and for cases admitted to hospital, to be normally distributed with a mean of 0·018 (SD 0·0018).13 Vaccine costs associated with vaccine price reimbursement and administration were triangularly distributed (£11·00, £15·55, £20·00). General practitioner and hospital treatment costs are assumed to be normally distributed with mean prices £37·00 (SD 8·40) and £839·00 (192·10). Further details of the specific health economic values used, including the health burden of each of the associated health outcomes has been previously published.13 No discounting was applied because the economic evaluation results report the uncertainty in the cost per QALY gained over a single year. Statistical analysis The analysis was run in R, using R Studio with the R package fluEvidenceSynthesis. Plots were drawn using Mathematica version 10.3.0.0.

Economic evaluation assumptions To calculate the cost-effectiveness of the elderly programme and high-risk programmes, we integrated the transmission model into an economic evaluation. The economic evaluation tracks the number of general practitioner consultations, hospital admissions, and deaths for each year for the three strains. QALYs lost are assumed for febrile cases to have a mean of 7·49 × 10−3, and for cases admitted to hospital, to be normally distributed with a mean of 0·018 (SD 0·0018).13 Vaccine costs associated with vaccine price reimbursement and administration were triangularly distributed (£11·00, £15·55, £20·00). General practitioner and hospital treatment costs are assumed to be normally distributed with mean prices £37·00 (SD 8·40) and £839·00 (192·10). Further details of the specific health economic values used, including the health burden of each of the associated health outcomes has been previously published.13 No discounting was applied because the economic evaluation results report the uncertainty in the cost per QALY gained over a single year. Statistical analysis The analysis was run in R, using R Studio with the R package fluEvidenceSynthesis. Plots were drawn using Mathematica version 10.3.0.0. 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. KEA, MB, JP-G, EvL, RP, and MR had access to the raw data. The corresponding author had full access to all of the data in the study and the final responsibility for the decision to submit for publication.

e study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. KEA, MB, JP-G, EvL, RP, and MR had access to the raw data. The corresponding author had full access to all of the data in the study and the final responsibility for the decision to submit for publication. Results As a universal paediatric programme expands to older ages (ie, from preschool age [2–4 years] to primary age [children commencing the school year aged 5–10 years] to secondary age [adolescents commencing the school year aged 11–15 years]), the number of cases of influenza averted by an elderly or high-risk vaccination programme diminishes (figure 1). Importantly, however, as the paediatric programme expands, and there are concurrent drops in influenza incidence in the elderly and high-risk populations, the speed at which the paediatric programme is implemented becomes important. A rapid implementation each year of any vaccination programme of preschool, primary school, and secondary school children, even in the absence of an elderly vaccine programme, would avert a similar number of cases as the same paediatric programme implemented gradually over the influenza season in the presence of an elderly programme.

ntation each year of any vaccination programme of preschool, primary school, and secondary school children, even in the absence of an elderly vaccine programme, would avert a similar number of cases as the same paediatric programme implemented gradually over the influenza season in the presence of an elderly programme. Because of the high morbidity and mortality in the high-risk group, vaccination of these individuals remains cost-effective in the presence of all paediatric strategies with an ICER of less than £15 000 per QALY gained for all simulations (figure 2). By contrast, the presence of paediatric vaccination produces a qualitative shift in the ICER for low-risk elderly vaccination. Under the 70% vaccine coverage base case, we find that although low-risk elderly vaccination is likely to maintain an ICER of less than £30 000 per QALY (83% of simulations), the ICER itself has increased by £9982 per QALY relative to no paediatric programme. Compared with no paediatric programme, the low-risk elderly programme has a much higher chance of falling between £20 000 per QALY and £30 000 per QALY (29%, compared with 12%) or more than £30 000 per QALY (17%, compared with 1%; figure 2). Conversely, with a cost-effectiveness threshold of £15 000 per QALY, the elderly programme is no longer cost-effective, with a mean ICER of £22 000 per QALY, and 80% of simulations providing ICER estimates of greater than £15 000 per QALY.

pared with 12%) or more than £30 000 per QALY (17%, compared with 1%; figure 2). Conversely, with a cost-effectiveness threshold of £15 000 per QALY, the elderly programme is no longer cost-effective, with a mean ICER of £22 000 per QALY, and 80% of simulations providing ICER estimates of greater than £15 000 per QALY. To assess the effect of paediatric uptake speed, we calculated the cost-effectiveness of elderly vaccination under a possible scenario in which the target coverage (70% vaccine uptake) was reached by the beginning of November (figure 2) compared with the present scenario, in which school-based administration increases gradually from October through December; achievement of the target coverage earlier (ie, by the beginning of November) resulted in a reduction in the cost-effectiveness of elderly vaccination and an increase in the ICER by £479 per QALY. Our model suggests that changing the cost-effectiveness threshold across a reasonable range determines whether the low-risk elderly vaccination remains cost-effective. For example, even under optimistic conditions of fast delivery by the beginning of November and a coverage of 90% for the 2–16 years paediatric programme, the probability that elderly vaccination is cost-effective at £30 000 per QALY is still 68%. However, if the cost-effectiveness threshold falls to £15 000 per QALY under likely base case conditions, the elderly programme will cease to be cost-effective, with only 5% of simulations falling below the threshold (figure 2).

he probability that elderly vaccination is cost-effective at £30 000 per QALY is still 68%. However, if the cost-effectiveness threshold falls to £15 000 per QALY under likely base case conditions, the elderly programme will cease to be cost-effective, with only 5% of simulations falling below the threshold (figure 2). We also assessed the effect of a reduced whole-season direct effectiveness of LAIV relative to the IIV that is currently given to elderly and high-risk individuals (figure 3). Under base case assumptions, the probability that the elderly programme is cost-effective increases as the whole-season LAIV direct effectiveness decreases. Under the conservative threshold of £15 000 per QALY, reducing LAIV effectiveness to half of IIV effectiveness increases the probability that the elderly programme is cost-effective from 42% to 65%.

obability that the elderly programme is cost-effective increases as the whole-season LAIV direct effectiveness decreases. Under the conservative threshold of £15 000 per QALY, reducing LAIV effectiveness to half of IIV effectiveness increases the probability that the elderly programme is cost-effective from 42% to 65%. Discussion We used a previously described calibrated mathematical model of seasonal influenza transmission in England and Wales to assess the likely epidemiological and economic effect of a paediatric programme, which is currently being rolled out across England, on the existing seasonal influenza vaccine programme. We found that under reasonable assumptions of vaccine coverage achieved in the pilot schemes, uncertainty exists about the continuing cost-effectiveness of the low-risk elderly vaccination programme that is available to all individuals with no underlying chronic disorders. The uncertainty surrounding these results stems primarily from the cost-effectiveness threshold assumed. We found that vaccinating high-risk individuals was always cost-effective.

cost-effectiveness of the low-risk elderly vaccination programme that is available to all individuals with no underlying chronic disorders. The uncertainty surrounding these results stems primarily from the cost-effectiveness threshold assumed. We found that vaccinating high-risk individuals was always cost-effective. Our analysis shows the potentially large effect of additional vaccination programmes on the cost-effectiveness of existing schemes. Particularly, the analysis highlights the importance of the assessment of existing influenza vaccination policies in the presence of increased herd protection after paediatric vaccination. Although our results suggest that vaccinating high-risk individuals will always be cost-effective over all plausible outcomes of the paediatric vaccination programme, the decision to maintain low-risk elderly vaccination is not so straightforward. As the new paediatric programme in England and Wales increases its scope from preschool-age children only to also include school-age children, with uptake consistent with other school-based programmes, significant uncertainty will probably arise concerning the cost-effectiveness of a low-risk elderly vaccination programme. Our analysis suggests three important questions for policy makers in countries wishing to introduce a paediatric programme (panel).

ldren, with uptake consistent with other school-based programmes, significant uncertainty will probably arise concerning the cost-effectiveness of a low-risk elderly vaccination programme. Our analysis suggests three important questions for policy makers in countries wishing to introduce a paediatric programme (panel). In England, the LAIV school-based pilot programme aimed at children aged 5–10 years reached a coverage of 52% in the first season of introduction and 57% in the second.18 However, national school-based administration of vaccines usually reaches much higher coverage. For example, at present, the first dose uptake for the human papillomavirus vaccine in England for girls aged 12–13 years is 89%,22 and for the 3-in-1 Tetanus-Diphtheria-Polio in Scotland for students aged 15–16 years is 88%.22, 23 Although the human papillomavirus and 3-in-1 vaccines are usually administered through the entire school year, influenza vaccines must achieve the target coverage in a short timeframe, typically before the Christmas holidays. As such, once the LAIV programme is established, a school-based administration of LAIV will probably achieve much higher coverage than the coverage that exists at present, but not as high as the coverage achieved by other school-administered vaccines. Unlike the timings for preschool, elderly, and high-risk vaccine administration, which are patient-led, the timing for a school-based vaccine administration would be determined by the regional National Health Service (NHS) commissioner, and is limited by the capacity of the providers, the number of schools, and the available days in the school term. If maximum uptakes are to be reached before November, additional resources would probably need to be diverted to enable sufficient local vaccination teams. Allocation of extra resources to regional NHS groups responsible for school-based vaccine administration might result in an increase in either the speed at which maximum coverage is met or the total vaccine uptake. Our results suggest that rapid administration of paediatric vaccines will provide substantial protection to both children and adults. This result is consistent with research7, 8 suggesting that influenza transmission is driven in most seasons by the young.

d at which maximum coverage is met or the total vaccine uptake. Our results suggest that rapid administration of paediatric vaccines will provide substantial protection to both children and adults. This result is consistent with research7, 8 suggesting that influenza transmission is driven in most seasons by the young. In this analysis, we assess national vaccination policies; however, differences in vaccine uptake, speed of administration, and the population age distribution across NHS regions will probably result in heterogeneity in herd protection levels reached. If the same number of LAIV doses were purchased, administration of them by the end of October would prevent the need for the entire low-risk elderly vaccine programme (figure 1). Because children and adolescents are known to be drivers of influenza transmission6, 7, 8 and, as such, are targeted for vaccination, cost-effective vaccination depends not only on the target population, but also on the speed at which the population is targeted.24

e entire low-risk elderly vaccine programme (figure 1). Because children and adolescents are known to be drivers of influenza transmission6, 7, 8 and, as such, are targeted for vaccination, cost-effective vaccination depends not only on the target population, but also on the speed at which the population is targeted.24 The US Advisory Committee on Immunization Practices temporarily withdrew their recommendation for LAIV administration on the basis of evidence from the USA for inefficacy of the vaccine, by contrast with other settings such as the UK.25 Calculation of the season-wide direct effectiveness of LAIV is problematic for several reasons: first, the direct effectiveness is known to change with age of recipient, number of vaccine doses (eg, two doses in the USA vs one in the UK), and endpoint measured (eg, virologically confirmed or syndromic presentation only); and second, seasonal fluctuations exist in the efficacy of the vaccine due to strain composition of the vaccine, frequency of circulating strains, and vaccine manufacture. Our sensitivity analysis assesses the likely effect of variation in the season-wide direct effectiveness. Because of the wide variation in empirical estimates for direct vaccine effectiveness, particularly between A and B strains, the extent to which LAIV efficacy differs from IIV remains unclear.

vaccine manufacture. Our sensitivity analysis assesses the likely effect of variation in the season-wide direct effectiveness. Because of the wide variation in empirical estimates for direct vaccine effectiveness, particularly between A and B strains, the extent to which LAIV efficacy differs from IIV remains unclear. Quantifying the population-level effect of a paediatric vaccination programme relies on capturing the social mixing patterns that facilitate disease transmission. Although this study uses empirical data on contact patterns from the same time period as the clinical data to which the model is calibrated, the methodology has some important caveats. For example, how the risk of influenza infection varies with the number of social contacts relative to their type or duration (eg, familial relationships in the same household, workplace meetings, etc) is not generally not well understood; additionally, how behaviour changes as a result of infection, and the implications for disease transmission, are also uncertain. Through consideration of the empirical-derived contact patterns that are used in the model as prior information, which updates to reflect other sources of data during the model calibration procedure, the dependency of the model on these data can be reduced. The accuracy of the model predictions are also contingent on the disease burden data to which the model is calibrated. Although the modelling approach is able to convey uncertainty in the predictions, robust age-specific influenza surveillance information is key for accurate model predictions.

hese data can be reduced. The accuracy of the model predictions are also contingent on the disease burden data to which the model is calibrated. Although the modelling approach is able to convey uncertainty in the predictions, robust age-specific influenza surveillance information is key for accurate model predictions. With substantial uncertainty around the cost-effectiveness of an elderly vaccination in the presence of a paediatric programme in England, other developed countries with a large fraction of the population older than 65 years might wish to reassess their funded elderly programme. In 2014, 9·5 million individuals were older than 65 years in England, 17·6% of the total population and the same number of those aged 2–16 years.26 Assuming an elderly vaccine uptake of 70% and a dose and administration cost of about £16 per vaccine, discontinuing elderly vaccination would save more than £106 million annually. Most developed countries have an ageing population as a result of a decreasing birth rate and an increasing life expectancy. These demographic shifts will not only increase the scale of an age-targeted programme, but they will also affect their cost-effectiveness through changes in age-specific social mixing patterns that affect the level of herd protection accrued. Frequent reassessment of policy decisions is essential to ensure these changes are accounted for.

fts will not only increase the scale of an age-targeted programme, but they will also affect their cost-effectiveness through changes in age-specific social mixing patterns that affect the level of herd protection accrued. Frequent reassessment of policy decisions is essential to ensure these changes are accounted for. The evidence needed to withdraw a funded vaccine programme might be different to that needed to initiate one. Should a decision be under consideration to remove a long-standing vaccination programme as a result of improved public health measures elsewhere, it might be wise to offer a scaled-back programme rather than completely remove an established policy. How such a scaled-back programme would be implemented depends on the details of the at-risk groups. For example, 55% of the high-risk population are themselves older than 65 years. Alternatively, extenuating circumstances might exist that take precedent over low probabilities of cost-effectiveness. Moreover, our results suggest that in countries that currently have no funded influenza vaccination programme in place, it is necessary to consider a suite of age-group and risk-group based strategies concurrently to optimise any national influenza programme.

take precedent over low probabilities of cost-effectiveness. Moreover, our results suggest that in countries that currently have no funded influenza vaccination programme in place, it is necessary to consider a suite of age-group and risk-group based strategies concurrently to optimise any national influenza programme. Acknowledgments KEA, EvL, and MB acknowledge funding from the National Institute for Health Research through the Health Protection Research Unit programme in partnership with Public Health England (KEA and MB in Immunisation at the London School of Hygiene & Tropical Medicine and EvL in Respiratory Infections at Imperial College London). EvL also acknowledges funding from the UK Medical Research Council (Project MR/J008761/1). DH received funding from a Medical Research Council PhD Studentship (administered through CoMPLEX University College London). The views expressed are those of the authors and not necessarily those of the UK National Health Service, the UK National Institute for Health Research, the UK Medical Research Council, the UK Department of Health, or Public Health England. Contributors KEA, MB, JP-G, and EvL conceived the study. DH did the data analysis. KEA wrote the manuscript, which was revised by MB, RP, MR, EvL, and JP-G. Declaration of interests KEA reports personal fees from Sanofi Pasteur, outside the submitted work. MR reports grants from GlaxoSmithKline and Pfizer, outside the submitted work. All other authors declare no competing interests.

Contributors KEA, MB, JP-G, and EvL conceived the study. DH did the data analysis. KEA wrote the manuscript, which was revised by MB, RP, MR, EvL, and JP-G. Declaration of interests KEA reports personal fees from Sanofi Pasteur, outside the submitted work. MR reports grants from GlaxoSmithKline and Pfizer, outside the submitted work. All other authors declare no competing interests. Figure 1 Effect of low-risk elderly and high-risk vaccination programmes in the presence of paediatric vaccination administered at different speeds through the influenza season Preschool are children aged 2–4 years, primary are children aged 5–10 years, and secondary are adolescents aged 11–16 years. The paediatric vaccination uptake speeds are associated with the accumulation of the fixed paediatric vaccine coverage across October, November, December, and January. Figure 2 Cost-effectiveness of low-risk elderly and high-risk vaccination programmes in the presence of paediatric vaccination

Preschool are children aged 2–4 years, primary are children aged 5–10 years, and secondary are adolescents aged 11–16 years. The paediatric vaccination uptake speeds are associated with the accumulation of the fixed paediatric vaccine coverage across October, November, December, and January. Figure 2 Cost-effectiveness of low-risk elderly and high-risk vaccination programmes in the presence of paediatric vaccination Preschool are children aged 2–4 years, primary are children aged 5–10 years, and secondary are adolescents aged 11–16 years. (A) All paediatric vaccination coverage is administered at speeds consistent with those reported during the full rollout of preschool or the pilot rollout of school-age children. (B) All paediatric vaccination coverage is administered by the end of October (fast uptake). Cost-effectiveness regions are coloured for readability: more than £30 000 per QALY (orange; not cost-effective), £20 000–30 000 per QALY (pink; cost-effective under current protocol), £15 000–20 000 per QALY (green; very cost-effective under current protocol), and less than £15 000 per QALY (blue; cost-effective under proposed protocol). The dark grey area corresponds to the incremental cost-effectiveness of a high-risk vaccine programme in the presence of the respective paediatric vaccine programme. The white areas correspond to the incremental cost-effectiveness of an elderly vaccine programme in the presence of the respective paediatric programme. Each grey and white probability distribution area is scaled to have a fixed maximum height, whereas each distribution represents an area of one unit. The mean (solid), 75% (dashed), and 90% (dotted) quantiles are shown to indicate in which cost-effectiveness region the mean of the distribution, 75%, or 90% of its simulations lie. QALY=quality-adjusted life-year.

area is scaled to have a fixed maximum height, whereas each distribution represents an area of one unit. The mean (solid), 75% (dashed), and 90% (dotted) quantiles are shown to indicate in which cost-effectiveness region the mean of the distribution, 75%, or 90% of its simulations lie. QALY=quality-adjusted life-year. Figure 3 Effect of the whole-season direct effectiveness of live attenuated influenza vaccine on the probability that the elderly vaccination programme is cost-effective Preschool are children aged 2–4 years, primary are children aged 5–10 years, and secondary are adolescents aged 11–16 years. Paediatric vaccination coverage is set at 70% and is reached at speeds consistent with those observed during the full rollout of preschool and the pilot rollout of school-age children. Three different incremental cost-effectiveness ratio thresholds are considered: £30 000 per QALY (red; cost-effective), £20 000 per QALY (blue; very cost-effective under current protocol), and £15 000 per QALY (green; cost-effective under proposed protocol). A relative direct effectiveness of 1 corresponds to the LAIV having the same whole-season direct effectiveness as the IIV (base case value), and a relative direct effectiveness of 0 corresponds to the LAIV having no effect on influenza epidemiology and is therefore equivalent to the cost-effectiveness of the elderly programme with no paediatric coverage. LAIV=live attenuated influenza vaccine. QALY=quality-adjusted life-year. IIV=inactivated influenza vaccine.

a relative direct effectiveness of 0 corresponds to the LAIV having no effect on influenza epidemiology and is therefore equivalent to the cost-effectiveness of the elderly programme with no paediatric coverage. LAIV=live attenuated influenza vaccine. QALY=quality-adjusted life-year. IIV=inactivated influenza vaccine. Panel Policy considerations for countries wishing to introduce a paediatric influenza vaccination programme We reassess the cost-effectiveness of the elderly vaccination policy in England because it was one of the first countries to introduce a publicly funded programme of paediatric influenza vaccination. With results suggesting that this is a highly cost-effective health policy, other countries will probably also follow a similar route. We would therefore recommend that decision makers consider the following three questions before adapting and renegotiating their influenza vaccination policies: 1 Are they willing to pay the same per vaccine dose for an elderly influenza vaccination programme in the presence of a national paediatric programme? In our base case scenario in the English context, the mean incremental cost-effectiveness increases from £12 552 per quality-adjusted life-year (QALY) to £29 145 per QALY after a successful preschool-based and school-based programme is introduced with high coverage.

e in the presence of a national paediatric programme? In our base case scenario in the English context, the mean incremental cost-effectiveness increases from £12 552 per quality-adjusted life-year (QALY) to £29 145 per QALY after a successful preschool-based and school-based programme is introduced with high coverage. 2 How quickly would the vaccines be available and administered via a preschool-based and school-based programme? Our results show that swift paediatric vaccination can negate the requirement for elderly vaccination, whereas slow uptake wastes vaccine doses by protecting children too late to have any effect on the reduction of transmission. 3 Would a focus on the achievement of a high coverage across all paediatric groups be more cost-effective than maintaining a historically low elderly coverage? When combined with previous results, we conclude that the per dose cost-effectiveness of a paediatric (2–16 years) vaccine is higher than that for an elderly vaccine, such that expansion of the paediatric programme as much as possible will always be economically advisable.

Introduction The period from 1990 to 1995 in the newly independent post-Soviet states was marked by an estimated extra 7 million premature deaths, 4 million in Russia alone.1 Many of these deaths were due to external causes and cardiovascular diseases.2 The group most affected by this rapidly increasing mortality was working-age men.3 There is extensive work documenting the importance of alcohol, unstable employment, and stress4, 5, 6, 7 as proximal determinants. However, the upstream determinants are less well understood, yet crucially important to inform policy makers in future periods of rapid political and economic transition.8 One of the central pillars of the post-communist transition has been privatisation.9 Stuckler and colleagues10 reported a cross-national association between extremely fast and extensive privatisation (so-called mass privatisation) with higher working-age male mortality, suggesting that unemployment was a primary mechanism linking privatisation and premature deaths. While it is well known that state enterprises hoarded an excessive and often ineffective labour force during the Soviet era, resisting massive layoffs during the post-communist crisis probably indirectly affected public health. Being employed, at least part-time, or even nominally, provided people with minimal security and gave them the feeling of being in control of their lives. Unemployment and related stress led to a hitherto unseen drop in life expectancy.11 Research in context Evidence before this study

Stuckler and colleagues10 reported a cross-national association between extremely fast and extensive privatisation (so-called mass privatisation) with higher working-age male mortality, suggesting that unemployment was a primary mechanism linking privatisation and premature deaths. While it is well known that state enterprises hoarded an excessive and often ineffective labour force during the Soviet era, resisting massive layoffs during the post-communist crisis probably indirectly affected public health. Being employed, at least part-time, or even nominally, provided people with minimal security and gave them the feeling of being in control of their lives. Unemployment and related stress led to a hitherto unseen drop in life expectancy.11 Research in context Evidence before this study We undertook an extensive analysis of the scientific literature on the effect of economic reforms on public health, morbidity, and mortality in post-communist Russia, and relevant electronic databases. We searched PubMed and Embase for reports published between 1990 and 2013, published in any language, with the following search terms: “mortality in Russia”, “working age mortality”, “male mortality”, “mortality and transition”, “privatization”, “alcohol”. We also searched our own extensive library of literature on mortality in the former communist countries of Europe. These data were highly heterogeneous and had limited utility as they were not suitable for linking the speed of privatisation in various settlements to individuals' mortality. Meta-analysis of the effects of privatisation on the health of employees and populations does not permit drawing a pooled estimate, as most of the estimates apply to countries outside our focus. Evidence from post-communist countries on health outcomes of politico-economic transitions remains inconclusive, although available data suggest that adverse health outcomes could result from rising unemployment and stress levels associated with mass privatisation.

of the estimates apply to countries outside our focus. Evidence from post-communist countries on health outcomes of politico-economic transitions remains inconclusive, although available data suggest that adverse health outcomes could result from rising unemployment and stress levels associated with mass privatisation. Added value of this study Employing an innovative design, we were able to quantify and compare inequalities in all-cause mortality, caused by policy intervention, during the 1990s in the urban population of Russian mono-industrial towns. We clarify the uncertainty associated with the link between the pace of privatisation in the former Soviet Union and the increasing mortality. This study is unique in using individual-level data collected specifically to study the health effects of privatisation and thus fills gaps in the scientific literature. Methodological contribution includes the use of the propensity score matching method to isolate the effect of privatisation by closely matching the settlements for a prospective survey prior to collecting the data via survey. Indirect demographic techniques were used to show how mortality levels in Russia changed over time. We used the method of establishing a convenience cohort study, based on the Brass indirect method that surveys random population samples to collect data on deaths of respondents' relatives, so as to estimate key population mortality parameters. Most studies focusing on this period examined macro data, while high-quality data on the level of individuals remain scarce. Even though the design of the study was dictated by the need to test the privatisation thesis, a large new empirical base provides rich data for individual-level health risk analysis.

parameters. Most studies focusing on this period examined macro data, while high-quality data on the level of individuals remain scarce. Even though the design of the study was dictated by the need to test the privatisation thesis, a large new empirical base provides rich data for individual-level health risk analysis. Implications of all the available evidence Our findings contribute to an emerging specialty at the intersection of comparative political economy and public health, and provide an evidence base for the public and scholarly debate on the effect of privatisation on public health. Our multi-level work contributes to the methodological literature, offering a concrete example of how to study the effects of structural socioeconomic changes and individual-level factors on population health. This new research tradition extends the social determinants of morbidity and mortality research programme by linking it to specific political and economic policies and processes. Findings are also relevant for policy makers considering similar interventions in post-communist societies and beyond.

on population health. This new research tradition extends the social determinants of morbidity and mortality research programme by linking it to specific political and economic policies and processes. Findings are also relevant for policy makers considering similar interventions in post-communist societies and beyond. Another possible explanation for the effects of privatisation on health could be the fact that before reforms, large state enterprises offered a wide range of social benefits to employees and their families. This was especially the case for city-forming enterprises that bore substantial responsibilities for the day-to-day wellbeing of their employees by providing company housing, health care, catering, day care, and holiday recreation.12, 13, 14 When these enterprises were privatised, such provisions ceased. Usually, in transition economies these responsibilities were taken over by municipal and local administrative divisions, but due to serious budget deficits, these organisations were unable to provide adequate quality and scope of social services.15 The loss of health services can affect directly population health. However, the loss of social services, as well as the uncertainty created by privatisation, could increase levels of stress and associated risky behaviours like excessive drinking, which is a well documented cause of increased post-communist male mortality.16 It is estimated that as much as 30% of overall male mortality in Russia is accounted for by excessive drinking, and even more in working ages.4, 17 King and colleagues18 recorded a cross-national association between mass privatisation, reduced health-care resources, unemployment and stress-related outcomes like alcohol consumption, suicides, ischaemic heart disease, and homicide.

in Russia is accounted for by excessive drinking, and even more in working ages.4, 17 King and colleagues18 recorded a cross-national association between mass privatisation, reduced health-care resources, unemployment and stress-related outcomes like alcohol consumption, suicides, ischaemic heart disease, and homicide. Finally, extremely fast and extensive privatisation could be linked to poor health outcomes by damaging firm performance and state capacity. Findings of cross-national time-series analysis, cross-sectional firm-level regressions, and case studies15, 19 show that at the micro-level, mass privatisated firms experienced reduced innovation, employed fewer employees, lowered economic output, and paid fewer taxes. At the macro-level, mass privatisation was associated with substantially lower economic growth and poorer state performance.20, 21 While national-level studies can plausibly test macro-mechanisms, critics have noted potential limitations such as omitted variable bias and ecological fallacy (ie, when the factors associated with group-level mortality rates fail to be associated with individual-level mortality).22 The aim of this study was to assess the effect of rapid privatisation at the individual level with data collected via survey within the framework of the European Research Council (ERC) Advanced Grant project, The Impact of Privatization on the Mortality Crisis in Eastern Europe (PrivMort).23

ividual-level mortality).22 The aim of this study was to assess the effect of rapid privatisation at the individual level with data collected via survey within the framework of the European Research Council (ERC) Advanced Grant project, The Impact of Privatization on the Mortality Crisis in Eastern Europe (PrivMort).23 Methods Study design The PrivMort project23 was designed to retrieve demographic and socioeconomic characteristics of individuals indirectly via surviving relatives. The method, developed by William Brass, was originally designed to estimate rates of vital events (births and deaths in countries with low literacy and numeracy). Because the information about relatives was not collected directly, it is often referred to as indirect estimation or the Brass technique after its developer. This information provides estimates of mortality rates that are acceptably representative of the underlying population.24, 25 This approach is recommended by the UN for inclusion in censuses in such situations in which alternative official sources are not available26 and has been used successfully for analysing adult mortality differentials in Russia.27 There were no changes made to the previously published protocol.23 The full protocol describes the data collected in all towns that participated in the survey, while this study analyses the data from a smaller survey sample of towns. The remainder are five multi-industrial towns matched with five mono-industrial towns. They are not well matched with the other towns and were selected to test a different set of hypotheses that will be reported elsewhere. These are therefore not studied in this article.

the data from a smaller survey sample of towns. The remainder are five multi-industrial towns matched with five mono-industrial towns. They are not well matched with the other towns and were selected to test a different set of hypotheses that will be reported elsewhere. These are therefore not studied in this article. Study populations Here we report analysis of data from 20 Russian towns whose economy is dominated by a single company. We selected mono-towns or towns of city-forming enterprise (monogoroda) because such settlements allow us to better isolate the effects of privatisation. Whereas some city-forming enterprises were fully privatised within 1 or 2 years, in others more gradual privatisation strategies were adopted.

ed by a single company. We selected mono-towns or towns of city-forming enterprise (monogoroda) because such settlements allow us to better isolate the effects of privatisation. Whereas some city-forming enterprises were fully privatised within 1 or 2 years, in others more gradual privatisation strategies were adopted. Following convention, we defined mono-industrial towns as settlements in which a single major enterprise employed more than 7·5% of total population, and the second largest enterprise had to be at least three times smaller in terms of its share in total employment in 1991. Every town had a population of between 5000 and 100 000 in 1991. An extensive review of sources was undertaken to compile enterprise-level data.23 We then classified mono-towns into three groups: towns where the major enterprise had been privatised rapidly, transferring 90% or more of its shares in any 2 consecutive years between 1992 and 1998; towns that adopted a more gradual approach in which less than 50% of shares were privatised in 2 consecutive years between 1992 and 1998, a period that provides sufficient time to account for any lagged effects of privatisation on mortality; and towns with a medium rate of privatisation (defined by privatisation of state shares of between 50% and 90%). We excluded towns with medium-pace privatisation to maximise the contrast in the speed of privatisation between the groups of firms. Future analysis using official gross mortality data and a larger sample of towns will allow us to perform an ecological analysis to assess whether there is a dose–response relationship between privatisation and mortality.

privatisation to maximise the contrast in the speed of privatisation between the groups of firms. Future analysis using official gross mortality data and a larger sample of towns will allow us to perform an ecological analysis to assess whether there is a dose–response relationship between privatisation and mortality. We chose a set of 20 mono-towns in west Russia, closely matched by initial conditions, so that ten towns that experienced fast privatisation (treatment group) were matched to ten that experienced gradual privatisation (control group). Interviews were performed between November, 2014, and March, 2015, by the Russian Agency for Public Opinion Research (VCIOM). The appendix lists the matched mono-industrial towns (appendix p 3). We used standard propensity score matching, logistic regression with no replacement, using the nearest neighbour approach. Every town in the treatment group was matched to a town in the control group with the closest score, so that a unit was selected only once, identifying the top ten pairs with closest propensity scores. We estimated conditional probabilities based on eight potential predictors of mortality, all measured for the pre-transition year, 1991, with the exception of wages, which were not available before 1992. We also used other matching algorithms such as trying alternative measures of the covariates (eg, number of hospital beds rather than number of physicians) and different distance measures (eg, Mahalanobis). The alternative methods explored produced town lists only marginally differing (ie, 1–2 towns changing), leading to the conclusion that the selected matching method is robust. The appendix shows these eight covariates across the matched towns (appendix p 4).

Our study is, to our knowledge, the largest investigation into factors linking social isolation and loneliness to an increased mortality risk. We did a mediation analysis and found that the association between social isolation and mortality reduced by 64% after taking into account differences in lifestyle, socioeconomic factors, and mental health problems between socially isolated and non-isolated individuals. These risk factors explained the association between loneliness and mortality. Implications of all the available evidence Isolation and loneliness are markers of many risk factors, such as socioeconomic adversity, unhealthy lifestyles, and lowered mental wellbeing. Policies and public health interventions that tackle these risk factors in general could potentially reduce excess mortality among the isolated and the lonely. This study was done under generic approval from the National Health Service National Research Ethics Service (June 17, 2011; Ref 11/NW/0382). Participants provided electronic informed consent for the baseline assessments and the register linkage.

hysicians) and different distance measures (eg, Mahalanobis). The alternative methods explored produced town lists only marginally differing (ie, 1–2 towns changing), leading to the conclusion that the selected matching method is robust. The appendix shows these eight covariates across the matched towns (appendix p 4). Study participants In every settlement, 20 to 45 starting points were identified using a grid. Interviewers were instructed to launch the random walk procedure at a randomly assigned address within each cell selected, inviting face-to-face interviews at every fourth household. The respondents were asked to provide information about vital status, sociodemographic and socioeconomic characteristics and health-related behaviours of their parents, two eldest siblings (if eligible), and first husbands or long-term partners. Ethics approval was obtained from the University of Cambridge Department of Sociology ethics committee and ERC ethics advisers. The data were anonymised to prevent any potential identification of individual respondents. Statistical analysis We calculated indirect age-standardised mortality rates (SMR) in rapidly and slowly privatised towns and then, in multivariate analyses, calculated Poisson proportional incidence rate ratios for men and women separately, with corresponding 95% confidence intervals (CIs) and robust standard errors that control for heteroscedasticity, to estimate associations of relatives' characteristics and speed of privatisation with the risk of death from all causes. Regression models were estimated with Stata (version 12.0).

en and women separately, with corresponding 95% confidence intervals (CIs) and robust standard errors that control for heteroscedasticity, to estimate associations of relatives' characteristics and speed of privatisation with the risk of death from all causes. Regression models were estimated with Stata (version 12.0). The multivariate analysis focuses on the effect of the speed of privatisation on mortality of those of working age, who experienced the greatest increase in mortality during the post-communist transition.28 Relatives younger than 20 years or older than 69 years in 1992, those who died before the beginning of the reforms (in 1991 or earlier) or did not have full information on education, marital status, professional occupation, material deprivation, alcohol drinking and smoking habits were excluded from the analysis. We also excluded those who were not residing in the 20 matched towns during privatisation, that is, for most of the 1990s. Work-related data were collected only on relatives who had not retired by 1992. Role of the funding source The funder had no role in study design, data collection, analyses, and reporting. AA, DI, LK, MMu, AG, GS, MB, DSte, and PH had access to raw data. The corresponding author had full access to all of the data and the final responsibility to submit for publication.

The multivariate analysis focuses on the effect of the speed of privatisation on mortality of those of working age, who experienced the greatest increase in mortality during the post-communist transition.28 Relatives younger than 20 years or older than 69 years in 1992, those who died before the beginning of the reforms (in 1991 or earlier) or did not have full information on education, marital status, professional occupation, material deprivation, alcohol drinking and smoking habits were excluded from the analysis. We also excluded those who were not residing in the 20 matched towns during privatisation, that is, for most of the 1990s. Work-related data were collected only on relatives who had not retired by 1992. Role of the funding source The funder had no role in study design, data collection, analyses, and reporting. AA, DI, LK, MMu, AG, GS, MB, DSte, and PH had access to raw data. The corresponding author had full access to all of the data and the final responsibility to submit for publication. Results Between November, 2014, and March, 2015, 21 494 households were identified in 20 towns. In 647 households, nobody could be contacted (after five visits), a further 5473 declined to be interviewed, and 732 were unable to participate due to physical or mental impairment. Overall, 14 642 interviews were completed. After data cleaning, a sample was reduced to 13 932 valid interviews, with a 65% response rate. Information about 38 339 relatives (21 634 men and 16 705 women) was collected. After exclusion criteria were applied, 19 167 were eligible for inclusion (12 086 men and 7081 women). Table 1 presents the availability of data for individuals in towns of each type and by sex. The distribution of people by vital status, education, occupation, and other characteristics in both types of town was similar.Table 1 Independent variables for fast and slow privatised towns

clusion (12 086 men and 7081 women). Table 1 presents the availability of data for individuals in towns of each type and by sex. The distribution of people by vital status, education, occupation, and other characteristics in both types of town was similar.Table 1 Independent variables for fast and slow privatised towns Fast-privatised towns Slow-privatised towns Men Women Men Women People 5018 (62%) 3031 (38%) 7068 (64%) 4050 (36%) Age in 1992 44·7 (12·3) 50·2 (11·8) 44·7 (12·3) 49·9 (11·9) Vital status at the end of 1998 Alive 4484 (89%) 2842 (94%) 6412 (91%) 3833 (95%) Deceased 534 (11%) 189 (6%) 656 (9%) 217 (5%) Education Elementary or incomplete secondary 1182 (24%) 906 (30%) 1546 (22%) 1302 (32%) Complete academic and vocational secondary 1926 (38%) 1032 (34%) 2952 (42%) 1290 (32%) Vocational higher education or incomplete higher 1367 (27%) 821 (27%) 1912 (28%) 1074 (27%) Complete academic higher education 543 (11%) 272 (9%) 658 (9%) 384 (10%) Occupation in 1990s Military 65 (1%) 4 (<1%) 100 (1%) 4 (<1%) Managerial 232 (5%) 77 (3%) 352 (5%) 98 (2%) High professional 587 (12%) 600 (20%) 647 (9%) 813 (20%) Low professional/routine non-manual 191 (4%) 485 (16%) 273 (4%) 675 (17%) Skilled manual 3039 (61%) 698 (23%) 4340 (61%) 765 (19%) Unskilled manual 268 (5%) 193 (6%) 409 (6%) 376 (9%) Unemployed in 1990s 636 (13%) 974 (32%) 947 (13%) 1319 (33%) Marital status at the date of interview/death Partnered 4464 (89%) 2554 (84%) 6416 (91%) 3447 (85%) Single 64 (1%) 91 (3%) 70 (1%) 105 (3%) Separated 490 (10%) 386 (13%) 582 (8%) 498 (12%) Smoking status Never smoked 1215 (24%) 2827 (93%) 1837 (26%) 3802 (94%) Currently/was a regular smoker 2879 (57%) 112 (4%) 3851 (55%) 142 (4%) Used to smoke but quit 924 (18%) 92 (3%) 1380 (20%) 106 (3%) Alcohol consumption Almost every day or several times a week 731 (15%) 23 (1%) 874 (12%) 35 (1%) About 2–4 times a month or up to once a month 2876 (57%) 1165 (38%) 3882 (55%) 1251 (31%) A couple of times a year 342 (7%) 427 (14%) 517 (7%) 561 (14%) Used to drink but quit 552 (11%) 67 (2%) 911 (13%) 80 (2%) Never 517 (10%) 1349 (45%) 884 (13%) 2213 (52%) Material deprivation Often or sometimes 119 (2%) 62 (2%) 202 (3%) 104 (3%) Rarely or never 4899 (98%) 2969 (98%) 6866 (97%) 3956 (97%) Data are n (%) or mean (SD).

imes a year 342 (7%) 427 (14%) 517 (7%) 561 (14%) Used to drink but quit 552 (11%) 67 (2%) 911 (13%) 80 (2%) Never 517 (10%) 1349 (45%) 884 (13%) 2213 (52%) Material deprivation Often or sometimes 119 (2%) 62 (2%) 202 (3%) 104 (3%) Rarely or never 4899 (98%) 2969 (98%) 6866 (97%) 3956 (97%) Data are n (%) or mean (SD). Fast privatisation was strongly associated with higher working-age male mortality rates both between 1992 and 1998 (age-standardised mortality ratio in men aged 20–69 years in fast vs slow privatised towns: 1·13, SMR 0·83 [95% CI 0·77–0·88] vs 0·73 [0·69–0·77], respectively) and from 1999 to 2006 (1·15, 0·91, 0·86–0·97 vs 0·79, 0·75–0·84). These preliminary results suggest that mortality rates were significantly higher in fast-privatised mono-industrial towns than in slow-privatised mono-industrial towns both in 1992–1998 and 1998–2006.

[95% CI 0·77–0·88] vs 0·73 [0·69–0·77], respectively) and from 1999 to 2006 (1·15, 0·91, 0·86–0·97 vs 0·79, 0·75–0·84). These preliminary results suggest that mortality rates were significantly higher in fast-privatised mono-industrial towns than in slow-privatised mono-industrial towns both in 1992–1998 and 1998–2006. All towns except one in the sample initiated privatisation between 1992 and 1998, thus enabling us to test the privatisation hypothesis. Models 1 in table 2 and table 3 report age-adjusted incidence rate ratios, while models 2 and 3 include controls for sociodemographic and socioeconomic characteristics and health-related behaviours. As the towns were closely matched with respect to population size, dependency ratio, level of health provision, housing provision, pollution, initial mortality rate, alcohol poisoning, and level of income, these town-level variables were not included in the models. We controlled for individual-level education, occupation, material deprivation in the 1990s, marital status, and frequency of drinking and smoking.Table 2 Age-adjusted incidence rate ratios of death from Poisson models in men aged 20–69 years between 1992 and 1998

e town-level variables were not included in the models. We controlled for individual-level education, occupation, material deprivation in the 1990s, marital status, and frequency of drinking and smoking.Table 2 Age-adjusted incidence rate ratios of death from Poisson models in men aged 20–69 years between 1992 and 1998 Model 1 Model 2 Model 3 Speed of privatisation (ref: slow speed) 1·15 (1·03–1·29) 1·17 (1·05–1·31) 1·13 (1·01–1·26) Education (ref: elementary) Complete academic and vocational secondary .. 0·99 (0·86–1·14) 1·04 (0·91–1·19) Vocational higher education or incomplete higher .. 0·83 (0·70–0·98) 0·93 (0·78–1·10) Complete academic higher education .. 0·70 (0·52–0·93) 0·82 (0·62–1·10) Occupation (ref: unskilled manual) Military .. 0·92 (0·47–1·79) 0·95 (0·49–1·97) Managerial .. 0·79 (0·43–1·19) 0·92 (0·62–1·38) High professional .. 0·73 (0·49–1·03) 0·80 (0·57–1·13) Low professional/routine non-manual .. 1·20 (0·73–1·75) 1·29 (0·88–1·88) Skilled manual .. 0·86 (0·60–1·09) 0·91 (0·72–1·16) Was not working in the 90s .. 1·32 (0·94–1·70) 1·40 (1·08–1·81) Material deprivation (ref: rarely or never) Often or sometimes .. .. 1·21 (0·91–1·61) Marital status (ref: partnered) Single .. .. 0·88 (0·48–1·63) Separated .. .. 1·19 (0·99–1·44) Alcohol consumption (ref: a couple of times a year) Almost every day or several times a week .. .. 1·58 (1·21–2·05) About 2–4 times a month or up to once a month .. .. 1·20 (0·94–1·53) Used to drink but quit .. .. 0·62 (0·45–0·85) Never .. .. 1·24 (0·93–1·67) Smoking (ref: never smoked) Used to smoke but quit .. .. 0·85 (0·69–1·05) Currently/was a regular smoker .. .. 1·67 (1·44–1·95) Data are incidence rate ratio (95% CI). Data are for 12 086 men and 1190 events. ref=reference.

. .. 1·20 (0·94–1·53) Used to drink but quit .. .. 0·62 (0·45–0·85) Never .. .. 1·24 (0·93–1·67) Smoking (ref: never smoked) Used to smoke but quit .. .. 0·85 (0·69–1·05) Currently/was a regular smoker .. .. 1·67 (1·44–1·95) Data are incidence rate ratio (95% CI). Data are for 12 086 men and 1190 events. ref=reference. Table 3 Age-adjusted incidence rate ratios of death from Poisson models among female cohorts aged 20–69 between 1992 and 1998

. .. 1·20 (0·94–1·53) Used to drink but quit .. .. 0·62 (0·45–0·85) Never .. .. 1·24 (0·93–1·67) Smoking (ref: never smoked) Used to smoke but quit .. .. 0·85 (0·69–1·05) Currently/was a regular smoker .. .. 1·67 (1·44–1·95) Data are incidence rate ratio (95% CI). Data are for 12 086 men and 1190 events. ref=reference. Table 3 Age-adjusted incidence rate ratios of death from Poisson models among female cohorts aged 20–69 between 1992 and 1998 Model 1 Model 2 Model 3 Speed of privatisation ref: slow speed) 1·16 (0·95–1·40) 1·17 (0·97–1·42) 1·18 (0·97–1·43) Education (ref: elementary) Complete academic and vocational secondary .. 0·93 (0·73–1·19) 0·93 (0·73–1·19) Vocational higher education or incomplete higher .. 1·08 (0·80–1·45) 1·08 (0·80–1·45) Complete academic higher education .. 0·64 (0·37–1·10) 0·64 (0·37–1·10) Occupation (ref: unskilled manual) Military .. 0·00 (0·00–0·00) 0·00 (0·00–0·00) Managerial .. 1·09 (0·44–2·72) 1·11 (0·45–2·78) High professional .. 0·88 (0·50–1·54) 0·89 (0·51–1·56) Low professional/routine non-manual .. 1·04 (0·63–1·73) 1·04 (0·63–1·73) Skilled manual .. 0·70 (0·42–1·17) 0·71 (0·42–1·18) Was not working in the 90s .. 1·56 (1·02–2·37) 1·57 (1·04–1·39) Material deprivation (ref: rarely or never) Often or sometimes .. .. 1·35 (0·78–2·33) Marital status (ref: partnered) Single .. .. 1·18 (0·68–2·05) Separated .. .. 1·03 (0·73–1·45) Alcohol consumption (ref: a couple of times a year) Almost every day or several times a week .. .. 1·78 (0·59–5·35) About 2–4 times a month or up to once a month .. .. 0·90 (0·66–1·23) Used to drink but quit .. .. 0·70 (0·31–1·60) Never .. .. 0·91 (0·69–1·21) Smoking (ref: never smoked) Used to smoke but quit .. .. 1·11 (0·57–2·19) Currently/was a regular smoker .. .. 0·99 (0·47–2·11) Data are incidence rate ratio (95% CI). Data are for 7081 women and 406 events. ref=reference.

.. .. 0·90 (0·66–1·23) Used to drink but quit .. .. 0·70 (0·31–1·60) Never .. .. 0·91 (0·69–1·21) Smoking (ref: never smoked) Used to smoke but quit .. .. 1·11 (0·57–2·19) Currently/was a regular smoker .. .. 0·99 (0·47–2·11) Data are incidence rate ratio (95% CI). Data are for 7081 women and 406 events. ref=reference. The incidence rate ratio for fast privatisation in the basic (age-adjusted) model 1 was 1·15 (95% CI 1·03–1·29) times higher than for the slow group for male relatives (table 2). In model 2 (main model), with controls for education and occupation, the incidence rate ratio was 1·17 (95% CI 1·05–1·31) times higher for men. In the fully adjusted models 3, which additionally include variables measuring material deprivation, alcohol consumption, smoking and marital status, the association with privatisation is attenuated but remains important with mortality 1·13 times higher (95% CI 1·01–1·26) in fast-privatised settlements, a similar estimate to that obtained in the comparison of SMRs. The model-predicted estimates of the mortality rates in fast privatisation and slow privatisation towns are 0·014 and 0·016, respectively. All control variables showed effects that were in the expected directions. We re-fitted the models 1–3 for both sexes on a dataset with imputed missing values using multivariable imputation via chained equations and the resulting coefficients were in line with those presented in table 4.Table 4 Age-adjusted incidence rate ratios of death from Poisson models (robustness check)

cted directions. We re-fitted the models 1–3 for both sexes on a dataset with imputed missing values using multivariable imputation via chained equations and the resulting coefficients were in line with those presented in table 4.Table 4 Age-adjusted incidence rate ratios of death from Poisson models (robustness check) Incidence rate ratio N N (events) Men 1992–1998, main model 1·17 (1·05–1·31) 12 086 1190 1992–1998, main model, hierarchical 1·17 (1·01–1·36) 12 086 1190 1992–1998, main model, for older cohorts* 1·15 (1·03–1·30) 7768 1061 5 years after privatisation, main model 1·21 (1·08–1·37) 11 819 1062 1992–2006, main model 1·16 (1·08–1·23) 12 086 3288 1999–2006, main model 1·18 (1·08–1·30) 10 315 1714 Women 1992–1998, main model 1·17 (0·97–1·42) 7081 406 1992–1998, main model, hierarchical 1·20 (0·92–1·56) 7081 406 1992–1998, main model, for older cohorts* 1·18 (0·97–1·43) 5709 394 Five years after privatisation, main model 1·24 (1·01–1·51) 6928 382 1992–2006, main model 1·14 (1·03–1·26) 7081 1431 1999–2006, main model 1·13 (0·97–1·32) 5794 623 Data are incidence rate ratio (95% CI) or n. The coefficients give the hazard ratio of death in relation to an absolute change in the independent variable (privatisation speed). Relative hazard ratios are presented with 95% CIs based on robust (heteroscedasticity-corrected) standard errors. We report age-adjusted hazard ratios of death from Poisson models among cohorts aged 20–69 years, with covariates: individuals' occupation, educational level. * Aged 40–69 years.

Incidence rate ratio N N (events) Men 1992–1998, main model 1·17 (1·05–1·31) 12 086 1190 1992–1998, main model, hierarchical 1·17 (1·01–1·36) 12 086 1190 1992–1998, main model, for older cohorts* 1·15 (1·03–1·30) 7768 1061 5 years after privatisation, main model 1·21 (1·08–1·37) 11 819 1062 1992–2006, main model 1·16 (1·08–1·23) 12 086 3288 1999–2006, main model 1·18 (1·08–1·30) 10 315 1714 Women 1992–1998, main model 1·17 (0·97–1·42) 7081 406 1992–1998, main model, hierarchical 1·20 (0·92–1·56) 7081 406 1992–1998, main model, for older cohorts* 1·18 (0·97–1·43) 5709 394 Five years after privatisation, main model 1·24 (1·01–1·51) 6928 382 1992–2006, main model 1·14 (1·03–1·26) 7081 1431 1999–2006, main model 1·13 (0·97–1·32) 5794 623 Data are incidence rate ratio (95% CI) or n. The coefficients give the hazard ratio of death in relation to an absolute change in the independent variable (privatisation speed). Relative hazard ratios are presented with 95% CIs based on robust (heteroscedasticity-corrected) standard errors. We report age-adjusted hazard ratios of death from Poisson models among cohorts aged 20–69 years, with covariates: individuals' occupation, educational level. * Aged 40–69 years. The findings in men changed little when we fitted our model with different, more conservative specifications. While the short-to-medium term effect of the rapid privatisation was captured in the main model, we explored long-term effect of privatisation on health and mortality in models based on the data for longer (1992–2006) and later (1999–2006) periods. In these models the effect of privatisation was significant in all specifications for men, and significant for women in a model covering the longer period (table 4). The magnitude of the privatisation effect was slightly larger than for 1992 to 1998 and significance levels were not changed. When we restricted the sample to cohorts of 40 years and older, point estimates and significance levels were unaffected (table 4).

cant for women in a model covering the longer period (table 4). The magnitude of the privatisation effect was slightly larger than for 1992 to 1998 and significance levels were not changed. When we restricted the sample to cohorts of 40 years and older, point estimates and significance levels were unaffected (table 4). We also fitted a Poisson model in which we took into account the actual year of privatisation. Although most towns experienced privatisation in 1993, some privatisation started in 1992. In this model we specified the start of exposure as the year of privatisation, and the end of exposure 5 years after the start of privatisation (table 4). Taking into account individuals' occupation and educational level, the incidence rate ratio of living in a town with rapid privatisation in this specification for men was 1·21 (95% CI 1·08–1·37).

start of exposure as the year of privatisation, and the end of exposure 5 years after the start of privatisation (table 4). Taking into account individuals' occupation and educational level, the incidence rate ratio of living in a town with rapid privatisation in this specification for men was 1·21 (95% CI 1·08–1·37). The findings were also robust when we used multi-level regressions with town-level random effects. Incidence rate ratios for males from this model (1·17, 95% CI 1·01–1·36) were very similar to the findings of the default models for men (table 4). The next two models analysed the mortality during the pre-intervention period: from 1985 to 1991. In these placebo analysis models (appendix p 10) mortality differentials between towns where the city-forming enterprise underwent slow or fast privatisation disappeared. As a final placebo analysis, we compared the relatives from fast and gradual privatised towns who did not live in those towns for a large part of the 1990s. There was no difference in the mortality rate between those groups (appendix p 10). Results for women showed weaker effects, with sizes varying from 13 to 24 in most models but rarely reaching significance. Only in the model accounting for the actual year of privatisation was the female incidence rate ratio large and significant (1·24, 95% CI 1·01–1·51; table 4).

The findings were also robust when we used multi-level regressions with town-level random effects. Incidence rate ratios for males from this model (1·17, 95% CI 1·01–1·36) were very similar to the findings of the default models for men (table 4). The next two models analysed the mortality during the pre-intervention period: from 1985 to 1991. In these placebo analysis models (appendix p 10) mortality differentials between towns where the city-forming enterprise underwent slow or fast privatisation disappeared. As a final placebo analysis, we compared the relatives from fast and gradual privatised towns who did not live in those towns for a large part of the 1990s. There was no difference in the mortality rate between those groups (appendix p 10). Results for women showed weaker effects, with sizes varying from 13 to 24 in most models but rarely reaching significance. Only in the model accounting for the actual year of privatisation was the female incidence rate ratio large and significant (1·24, 95% CI 1·01–1·51; table 4). Discussion The PrivMort project, to our knowledge, is the largest multi-level indirect retrospective cohort study to date done in the post-communist countries. We found clear differences in mortality in working-age men between towns where the city-forming enterprise underwent slow rather than fast privatisation. Working-age men in fast-privatised towns experienced 13% to 21% higher mortality than working-age men in slow-privatised towns, depending on the specification or sample restrictions applied. Findings in women were broadly consistent but not significant, mainly due to the smaller number of deaths in women.

rivatisation. Working-age men in fast-privatised towns experienced 13% to 21% higher mortality than working-age men in slow-privatised towns, depending on the specification or sample restrictions applied. Findings in women were broadly consistent but not significant, mainly due to the smaller number of deaths in women. Our results are consistent with a previous cross-national study by Stuckler and colleagues,10 which noted that rapid privatisation was associated with a rise in male death rates. The magnitude of association in their findings,10 a 12·8% (95% CI 7·9–17·7) rise linked with introduction of mass privatisation was lower than in the present study (17%, 95% CI 5–31). This finding is unsurprising in view of the different operationalisation of privatisation used in the study by Stuckler and colleagues.10 The greater estimate in this study is to be expected because the national-level finding estimates mortality in all towns and cities, both gradual and fast privatisers. Additionally, we were better able to isolate the effect of privatisation. Whereas in the previous study10 the percentage change in mortality ratio was adjusted for the European Bank for Reconstruction and Development (EBRD) price liberalisation index, EBRD trade liberalisation index, the democratisation index, dummy for military or ethnic conflict, the percentage of urban population, dependency ratio, per capita log gross domestic product, and education level, here we adjusted for all of these characteristics and a wider set of variables. The first five covariates were accounted for by restricting the analysis to the urban population of a single country. The next two covariates were accounted for by matching towns on dependency ratio and average wage; additionally, we matched on other socioeconomic and demographic variables. Unlike the previous work,10 we were able to adjust for education on the level of individuals, not country level, which makes our estimates more precise.

xt two covariates were accounted for by matching towns on dependency ratio and average wage; additionally, we matched on other socioeconomic and demographic variables. Unlike the previous work,10 we were able to adjust for education on the level of individuals, not country level, which makes our estimates more precise. Previous researchers have suggested that women have been able to cope with the post-Soviet transitions better than men.29 Although Russian women faced a double burden from participating in the labour market and taking care of their families,30, 31 paradoxically this greater responsibility seems to have given their lives more meaning. It has also been suggested that, faced with unemployment, Russian men turned to cheap and easily available alcohol to seek compensation for their lost provider's status,32 while heavy alcohol consumption was stigmatised among women.33 The indirect estimation technique we used adequately assesses overall mortality levels and differences in mortality between different socioeconomic subgroups, as findings of several large-scale studies in Russia have previously shown.16, 34, 35, 36 In fact, the present study has several advantages common to indirect techniques of estimating mortality, mainly in terms of cost and time efficiency and ability to retrieve information about people usually unreachable by direct techniques. Additionally, the PrivMort study retrieved a broad range of socioeconomic and demographic information not otherwise available.

ges common to indirect techniques of estimating mortality, mainly in terms of cost and time efficiency and ability to retrieve information about people usually unreachable by direct techniques. Additionally, the PrivMort study retrieved a broad range of socioeconomic and demographic information not otherwise available. However, our study has limitations. First, the sample was not representative of the entire population of Russia because we focused only on the urban population of industrial mono-enterprise towns in the European part of the country. Moreover, the fact that national statistics average across the whole population of Russia—including the Muslim population with lower reported alcohol consumption37—makes it difficult to compare these figures with the sample of 20 towns that have a mainly Russian population. The total number of Muslims is estimated to be lower than 2·8 million, or less than 2% of Russia's total population.38 Second, potential selection bias might have arisen from differentials in family-wise mortality levels: representatives from high-mortality families were less likely to have participated in the survey because a higher proportion would not survive until the interview date, making such families less likely to be captured by the sampling, and pass the screening criteria for the interviews, which require respondents' relatives to reside in the settlements in question. However, research by Murphy and colleagues27 suggests that this is not likely to bias the mortality estimates. Third, as those who emigrated away from the 20 settlements were excluded from the study, future investigations should look into the migrant differentials at settlement level. Possible bias might arise from out-migration from fast privatisation towns; however, this could both overestimate and underestimate the effect of the speed of privatisation. If healthy people left a fast privatisation town, this could introduce an upward bias, if families who lost their fathers (ie, the main earners) left the town, this would contribute to underestimation of the effect of privatisation on mortality.

h overestimate and underestimate the effect of the speed of privatisation. If healthy people left a fast privatisation town, this could introduce an upward bias, if families who lost their fathers (ie, the main earners) left the town, this would contribute to underestimation of the effect of privatisation on mortality. The fourth concern relates to temporal inconsistency of the recorded individual characteristics and mortality outcomes. While individual characteristics such as education, marital status, alcohol consumption, and smoking were recoded at the time of death for decedents (between 1992 and 1998), for surviving relatives these characteristics were collected at time of the interview and at the time of death for those who died after 1998, which might be as late as 2015. However, because we were only studying relatives aged 20 years and older, and, in view of low transition rates to higher education for adults in Russia, education status is unlikely to change. Marital status and alcohol and tobacco use might have changed between the exposure period and the date of interview, hence the result of the full models 3 must be treated with caution. Finally, there were potential problems in accurately recalling events that happened 20–25 years ago; this can potentially bias the estimates of frequency of drinking and smoking, education, and frequency of communication for relatives. Anticipating this, interviewers were instructed to introduce auxiliary sentences and memory cards to facilitate accurate recall and obtain accurate estimates of characteristics and behaviours of their relatives. However, for all five of these biases, there was no reason to think they would vary with the speed of privatisation in the settlements.

wers were instructed to introduce auxiliary sentences and memory cards to facilitate accurate recall and obtain accurate estimates of characteristics and behaviours of their relatives. However, for all five of these biases, there was no reason to think they would vary with the speed of privatisation in the settlements. The method of isolating the effect of privatisation by closely matching the settlements before collecting the data also entails certain limitations. We deployed propensity score matching on baseline covariates that are most important to health outcomes, but one could argue that fast-privatised towns are systematically different from slow-privatised towns in some other ways. A systematic literature review of the causes of privatisation15 showed that the main determinants of privatisation were political, not underlying economic weakness or social conditions. The political choice of the speed of privatisation was made on the level of enterprises or local and regional authorities, thus it can be claimed that it was probably orthogonal to mortality outcomes.

than one” (1 point for no participation in social activities at least weekly). Thus, individuals could score a total of 0–3; an individual was defined as socially isolated if he or she scored 2 or 3; those who scored 0 or 1 were classified as not isolated. Similar scales have been used previously in other UK studies.12 Loneliness was assessed with two questions: “Do you often feel lonely?” (no=0, yes=1) and “How often are you able to confide in someone close to you?” (0=almost daily to once every few months; 1=never or almost never). An individual was defined as lonely if he or she scored 2, and not lonely if he or she scored 0 or 1. Similar questions are included in scales such as the revised UCLA Loneliness Scale.13 Follow-up for all deaths irrespective of cause started at inclusion in the UK Biobank study (from national death registers) and ended on Aug 14, 2015, or upon death, for all participants. The cause-specific-mortality International Classification of Diseases (ICD) codes were as follows: neoplasms (C00–D48), diseases of the circulatory system (I05–I89), and other diseases (all remaining ICD-10 codes).

owed that the main determinants of privatisation were political, not underlying economic weakness or social conditions. The political choice of the speed of privatisation was made on the level of enterprises or local and regional authorities, thus it can be claimed that it was probably orthogonal to mortality outcomes. Clearly, there are several factors contributing to excess mortality in Russia, acting at different points along several causal pathways, with alcohol playing a key role in several, including that related to speed of privatisation. In a study by Walberg and colleagues5 that reported an association between labour turnover in Russian regions, itself linked with privatisation, researchers noted that much of the excess mortality was alcohol-related. Treisman39 invoked the fall in alcohol prices from 1990 to 1994 as the primary determinant of increased mortality during the 1990s. Treisman's measure of price was based on the cost of vodka in the capital city of each of Russia's regions. There were no data available for alcohol prices at the settlement level that would allow us to directly test this hypothesis. Another possible component of several causal pathways involves social protection. Increased mortality could be related to variations in the provision of social benefits at the settlement level. However, local expenditures are strongly dependent on the health of city-forming enterprises.15 Similarly, mortality could be related to sector or firm-specific characteristics, possibly operating through lower wages. However, without data for these variables, we cannot test all possible causal pathways.

tlement level. However, local expenditures are strongly dependent on the health of city-forming enterprises.15 Similarly, mortality could be related to sector or firm-specific characteristics, possibly operating through lower wages. However, without data for these variables, we cannot test all possible causal pathways. It is essential to recognise that the health effects of transition were complex and defy mono-causal explanation. We do not claim that mass privatisation was the only cause of increased mortality. In fact, it would be remarkable if one single policy or factor explained all the variation in the post-communist mortality crisis. We believe our findings provide strong evidence for the hypothesis that rapid privatisation contributed to raised working-age male mortality. Importantly, the findings of this study complement the epidemiological and public health explanations linking alcohol consumption and psychological stress to the Russian mortality crisis of 1992 to 1998,40 identifying rapid privatisation as an important underlying cause. Regression coefficients for fast privatisation were consistently positive and significant for men, but usually not significant for women. Robustness checks did not significantly change the magnitude or the direction of the covariates of mortality. This study adds to previous, less conclusive findings on the effects of economic governance on all-cause mortality risk in working-age men, and makes a valuable methodological and empirical contribution to the topic. Supplementary Material Supplementary appendix

It is essential to recognise that the health effects of transition were complex and defy mono-causal explanation. We do not claim that mass privatisation was the only cause of increased mortality. In fact, it would be remarkable if one single policy or factor explained all the variation in the post-communist mortality crisis. We believe our findings provide strong evidence for the hypothesis that rapid privatisation contributed to raised working-age male mortality. Importantly, the findings of this study complement the epidemiological and public health explanations linking alcohol consumption and psychological stress to the Russian mortality crisis of 1992 to 1998,40 identifying rapid privatisation as an important underlying cause. Regression coefficients for fast privatisation were consistently positive and significant for men, but usually not significant for women. Robustness checks did not significantly change the magnitude or the direction of the covariates of mortality. This study adds to previous, less conclusive findings on the effects of economic governance on all-cause mortality risk in working-age men, and makes a valuable methodological and empirical contribution to the topic. Supplementary Material Supplementary appendix Acknowledgments All authors acknowledge financial support from the European Research Council (ERC). DStu is funded by a Wellcome Trust Investigator Award.

It is essential to recognise that the health effects of transition were complex and defy mono-causal explanation. We do not claim that mass privatisation was the only cause of increased mortality. In fact, it would be remarkable if one single policy or factor explained all the variation in the post-communist mortality crisis. We believe our findings provide strong evidence for the hypothesis that rapid privatisation contributed to raised working-age male mortality. Importantly, the findings of this study complement the epidemiological and public health explanations linking alcohol consumption and psychological stress to the Russian mortality crisis of 1992 to 1998,40 identifying rapid privatisation as an important underlying cause. Regression coefficients for fast privatisation were consistently positive and significant for men, but usually not significant for women. Robustness checks did not significantly change the magnitude or the direction of the covariates of mortality. This study adds to previous, less conclusive findings on the effects of economic governance on all-cause mortality risk in working-age men, and makes a valuable methodological and empirical contribution to the topic. Supplementary Material Supplementary appendix Acknowledgments All authors acknowledge financial support from the European Research Council (ERC). DStu is funded by a Wellcome Trust Investigator Award. Contributors AA compiled the datasets, designed and did the empirical analysis, coordinated the early stages of the project, and drafted the report. LK led the project, was the author of the grand design of the study, oversaw the statistical analysis, developed the core ideas and contributed to the draft. MMc oversaw the grand design of the study, discussed core ideas, and amended the report. MMu helped to design the statistical analysis, oversaw the grand design of the study, facilitated the interpretation of results, and commented on drafts. MB oversaw the grand design, facilitated the interpretation of results and helped to draft the report. MMa jointly led the project, oversaw the grand design, and discussed the core ideas. DStu oversaw the grand design, facilitated the interpretation of results and helped to draft the report. DI provided background information, oversaw the raw data collection via survey, facilitated the interpretation of results, and contributed to the report. MF provided background information, contributed to the empirical analysis and facilitated the interpretation of results. AG facilitated interpretation of findings, did multiple imputations to the dataset, and contributed to the report. IK compiled settlement-level data for empirical analysis, provided other background statistical information, and facilitated the interpretation of results. VP and IS facilitated the interpretation of results, and commented on drafts. GS, DSte, and PH facilitated the interpretation of findings. All authors saw and approved the final version of the report.

r empirical analysis, provided other background statistical information, and facilitated the interpretation of results. VP and IS facilitated the interpretation of results, and commented on drafts. GS, DSte, and PH facilitated the interpretation of findings. All authors saw and approved the final version of the report. Declaration of interests We declare no competing interests.

ank study (from national death registers) and ended on Aug 14, 2015, or upon death, for all participants. The cause-specific-mortality International Classification of Diseases (ICD) codes were as follows: neoplasms (C00–D48), diseases of the circulatory system (I05–I89), and other diseases (all remaining ICD-10 codes). Details of the assessments of participants' variables are publicly available.14 Briefly, participants completed several touch-screen computer-based questionnaires, and then had a face-to-face interview with a trained researcher. The information collected included basic demographics (sex and age), ethnic origin (white vs other), socioeconomic factors (educational attainment, household income, and postcode of residence with the corresponding Townsend deprivation index score), and chronic diseases (diabetes, cardiovascular disease, cancer, and other long-standing illness, disability, or infirmity). The Townsend deprivation index is an integrated neighbourhood-level measure of unemployment, non-car ownership, non-home ownership, and household overcrowding across the UK.15

vation index score), and chronic diseases (diabetes, cardiovascular disease, cancer, and other long-standing illness, disability, or infirmity). The Townsend deprivation index is an integrated neighbourhood-level measure of unemployment, non-car ownership, non-home ownership, and household overcrowding across the UK.15 To assess biological factors, trained data collectors measured height and weight in all participants during clinic attendance using standard operating procedures, and the body-mass index (BMI) was subsequently calculated. Procedures for measuring systolic and diastolic blood pressure and handgrip strength are reported in the UK Biobank protocol, which is available online.11 Behavioural factors, including cigarette smoking (current smoker [yes or no]; ex-smoker [yes or no]), physical activity (moderate and vigorous), and alcohol intake frequency (at least three times a week vs twice a week or less) were self-reported on a questionnaire.

Isolation and loneliness are markers of many risk factors, such as socioeconomic adversity, unhealthy lifestyles, and lowered mental wellbeing. Policies and public health interventions that tackle these risk factors in general could potentially reduce excess mortality among the isolated and the lonely. This study was done under generic approval from the National Health Service National Research Ethics Service (June 17, 2011; Ref 11/NW/0382). Participants provided electronic informed consent for the baseline assessments and the register linkage. Procedures The social isolation scale used by the UK Biobank was constructed from three questions: (1) “Including yourself, how many people are living together in your household? Include those who usually live in the house such as students living away from home during term time, partners in the armed forces or professions such as pilots” (1 point for living alone); (2) “How often do you visit friends or family or have them visit you?” (1 point for friends and family visit less than once a month); and (3) “Which of the following [leisure/social activities] do you engage in once a week or more often? You may select more than one” (1 point for no participation in social activities at least weekly). Thus, individuals could score a total of 0–3; an individual was defined as socially isolated if he or she scored 2 or 3; those who scored 0 or 1 were classified as not isolated. Similar scales have been used previously in other UK studies.12

the UK Biobank protocol, which is available online.11 Behavioural factors, including cigarette smoking (current smoker [yes or no]; ex-smoker [yes or no]), physical activity (moderate and vigorous), and alcohol intake frequency (at least three times a week vs twice a week or less) were self-reported on a questionnaire. Psychological factors comprised current depressive symptoms and general cognitive capacity. Depressive symptoms were measured using the frequency of four items from the Patient Health Questionnaire (PHQ):16, 17 (1) depressed mood, (2) disinterest or absence of enthusiasm, (3) tenseness or restlessness, and (4) tiredness or lethargy in the previous 2 weeks. General cognitive capacity (numeric memory, verbal–numerical reasoning, reaction time, and visual memory) was assessed by use of a touch-screen application.18 Self-rated health was assessed using the following question answered on a four-point scale (1=poor; 4=excellent): “In general, how would you rate your overall health?”

Introduction Socially isolated and lonely individuals have a higher mortality risk than people with social contacts.1, 2, 3, 4, 5 Several factors might contribute to these associations.6 According to one hypothesis, losing social connections and feeling lonely could be associated with depressive mood and cognitive decline,7 with accompanying downstream biological changes such as increased cortisol secretion, deterioration in immune function, and weight gain.6 Social isolation could also be associated with unhealthy lifestyle factors, such as increased smoking, increased alcohol consumption, and physical inactivity.8 Similarly, socioeconomic adversity is associated with an increased likelihood of social isolation,9 and thus might explain the reported associations. However, few extensive prospective data exist on which to test these hypotheses and assess the associations in different groups such as old and young individuals, low and high socioeconomic groups, and those with and without chronic disease. All these factors might confound the association of social isolation and loneliness with mortality.

e prospective data exist on which to test these hypotheses and assess the associations in different groups such as old and young individuals, low and high socioeconomic groups, and those with and without chronic disease. All these factors might confound the association of social isolation and loneliness with mortality. A better understanding of the factors underlying the associations between social isolation (ie, having no or few contacts with others), loneliness (ie, feeling lonely or unable to share one's thoughts), and mortality might facilitate the design of interventions to reduce excess health risk in socially isolated, lonely people. We used data from the UK Biobank study to quantify the extent to which the associations of social isolation and loneliness with mortality are related to biological, behavioural, socioeconomic, and psychological risk factors. Methods Study design and participants We analysed baseline data and mortality follow-up data from the UK Biobank study.10 UK National Health Service registers maintain records of all individuals legally registered as resident in the UK. With the help of these records, invitations were sent to individuals aged 40–69 years living within a sensible travelling distance of the 22 assessment centres across Great Britain in 2007–10.10 For the UK Biobank project, baseline questionnaires and physical measures (eg, standard anthropometry and spirometry) were collected and blood and urine samples were stored, as described elsewhere.11 502 656 individuals were recruited (5% of the eligible population) in the UK Biobank.

across Great Britain in 2007–10.10 For the UK Biobank project, baseline questionnaires and physical measures (eg, standard anthropometry and spirometry) were collected and blood and urine samples were stored, as described elsewhere.11 502 656 individuals were recruited (5% of the eligible population) in the UK Biobank. Research in context Evidence before this study Social isolation and loneliness are associated with increased health problems and excess mortality risk. We searched PubMed for studies published in English up to May 31, 2016, using the following combinations of search terms (in title): (A) social support AND mortalit*; (B) social relations AND mortalit*; (C) social networks AND mortalit*; (D) social isolation AND mortalit*; and (E) loneliness AND mortalit*. Search combination A yielded 37 publications, B eight, C 14, D 11, and E 14. Findings from these studies suggest that there is an association between social isolation and mortality and between loneliness and mortality. Added value of this study Our study is, to our knowledge, the largest investigation into factors linking social isolation and loneliness to an increased mortality risk. We did a mediation analysis and found that the association between social isolation and mortality reduced by 64% after taking into account differences in lifestyle, socioeconomic factors, and mental health problems between socially isolated and non-isolated individuals. These risk factors explained the association between loneliness and mortality. Implications of all the available evidence

l cognitive capacity (numeric memory, verbal–numerical reasoning, reaction time, and visual memory) was assessed by use of a touch-screen application.18 Self-rated health was assessed using the following question answered on a four-point scale (1=poor; 4=excellent): “In general, how would you rate your overall health?” Statistical analysis We did analyses first using data from those participants who did not have any missing data (complete case analyses) and then using imputed datasets. We examined the associations of social isolation and loneliness with all-cause mortality using Cox proportional hazard models with age as a timescale. Associations with cause-specific mortality were examined using competing-risks survival regression (based on Fine and Gray's proportional sub-hazards model), which is the appropriate method for estimation of competing actual risks.19 All the models were adjusted for age, sex, and ethnic origin, with additional adjustment for chronic disease. To measure the robustness of these associations, we did additional subgroup analyses separately for men and women, three age groups (37–52 years, 53–60 years, and 61–73 years), different ethnic groups (white vs non-white), and participants with and without chronic disease at baseline. Subgroup analyses by sex, age, ethnic origin, and chronic disease were chosen because these factors represent potential confounders for the association between social relations and mortality. Men, individuals belonging to ethnic minorities, elderly people, and those with long-standing illness tend to have fewer social relations and also are at increased risk of mortality.4, 6 Similar three-level age categorisations have been used in a previous study based on UK Biobank data.20

ocial relations and mortality. Men, individuals belonging to ethnic minorities, elderly people, and those with long-standing illness tend to have fewer social relations and also are at increased risk of mortality.4, 6 Similar three-level age categorisations have been used in a previous study based on UK Biobank data.20 To assess the extent to which baseline biological, behavioural, socioeconomic, psychological, and health-related risk factors explained the associations of social isolation and loneliness with mortality, we calculated the percentage of excess risk mediated (PERM) for the following five groups of explanatory variables: (1) biological (BMI, diastolic and systolic blood pressure, and handgrip strength); (2) behavioural (smoking, alcohol consumption, and physical activity); (3) socioeconomic (Townsend deprivation index, education, and household income); (4) psychological factors (depressive symptoms and cognitive performance); and (5) self-rated health. For each risk-factor group, we estimated the percentage of PERM as:21 PERM=[Hazard 0​ratio (age, sex ethnicity, andchronic disease adjusted)-hazardratio (age, sex, ethnicity, chronicdisease, and risk factor adjusted)Hazard 0​ratio (age, sex ethnicity, andchronic disease adjusted)-1]×100 Sex, ethnic origin, chronic disease, smoking status, education, and high alcohol consumption were treated as categorical and the other risk factors as continuous variables in the analyses. Finally, all the risk factors were included in the same model simultaneously (final model).

To assess the extent to which baseline biological, behavioural, socioeconomic, psychological, and health-related risk factors explained the associations of social isolation and loneliness with mortality, we calculated the percentage of excess risk mediated (PERM) for the following five groups of explanatory variables: (1) biological (BMI, diastolic and systolic blood pressure, and handgrip strength); (2) behavioural (smoking, alcohol consumption, and physical activity); (3) socioeconomic (Townsend deprivation index, education, and household income); (4) psychological factors (depressive symptoms and cognitive performance); and (5) self-rated health. For each risk-factor group, we estimated the percentage of PERM as:21 PERM=[Hazard 0​ratio (age, sex ethnicity, andchronic disease adjusted)-hazardratio (age, sex, ethnicity, chronicdisease, and risk factor adjusted)Hazard 0​ratio (age, sex ethnicity, andchronic disease adjusted)-1]×100 Sex, ethnic origin, chronic disease, smoking status, education, and high alcohol consumption were treated as categorical and the other risk factors as continuous variables in the analyses. Finally, all the risk factors were included in the same model simultaneously (final model). To assess the extent to which the associations followed a dose–response pattern, we did dose–response analyses using sum scores from the individual items of the isolation and loneliness measures. We analysed isolation and loneliness separately to assess whether there was a pattern across the continuous score as a predictor of mortality.

nt to which the associations followed a dose–response pattern, we did dose–response analyses using sum scores from the individual items of the isolation and loneliness measures. We analysed isolation and loneliness separately to assess whether there was a pattern across the continuous score as a predictor of mortality. We accounted for missing data by multiple imputation by chained equations, which generated five imputed datasets.22 The imputation model included age, sex, social isolation, loneliness, all confounding and mediating variables, the Nelson-Aalen estimate of cumulative hazard, and survival status.23 We fitted Cox proportional hazards models within each imputed dataset and combined them in accordance with Rubin's rules. To test whether reverse-causation bias (ie, the effect of chronic disease on social isolation) affected our results, we did a sensitivity analysis examining the association between social isolation and all-cause mortality (imputed data, but chronic disease missingness not imputed) after adjustment for all covariates in those without chronic disease at baseline. We used Stata (version 13.1) for all analyses. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. ME and CH had full access to all the data in the study and ME and MK had final responsibility for the decision to submit for publication.

To test whether reverse-causation bias (ie, the effect of chronic disease on social isolation) affected our results, we did a sensitivity analysis examining the association between social isolation and all-cause mortality (imputed data, but chronic disease missingness not imputed) after adjustment for all covariates in those without chronic disease at baseline. We used Stata (version 13.1) for all analyses. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. ME and CH had full access to all the data in the study and ME and MK had final responsibility for the decision to submit for publication. Results 466 901 (93%) of the 502 656 individuals recruited to the UK Biobank provided complete data on social isolation, loneliness, and mortality, and were included in the present analysis. There were statistically significant differences in all the study variables between those who were and those who were not included, although the absolute between-group differences were small (appendix p 2). The mean age of study participants was 56·5 years (SD 8·1; range 40–69), 55% were women, and 95% were white (table). 42 548 (9%) of 466 901 participants were categorised as socially isolated and 29 442 (6%) as lonely.Table Descriptive statistics of the study participants

een-group differences were small (appendix p 2). The mean age of study participants was 56·5 years (SD 8·1; range 40–69), 55% were women, and 95% were white (table). 42 548 (9%) of 466 901 participants were categorised as socially isolated and 29 442 (6%) as lonely.Table Descriptive statistics of the study participants Participants (n=466 901) Age (years) 56·5 (8·1) Sex Women 254 919 (55%) Men 211 982 (45%) Ethnic origin Non-white 21 482 (5%) White 444 118 (95%) Data missing 1301 (<1%) Townsend index score* –1·37 (3·05) Education No secondary education 77 329 (17%) Secondary education 232 222 (50%) University degree 153 810 (33%) Data missing 3540 (1%) Annual household income Less than £31 000 193 196 (41%) At least £31 000 212 753 (46%) Data missing 60 952 (13%) Chronic illness No 224 947 (48%) Yes 229 595 (49%) Data missing 12 359 (3%) Social isolation No 424 353 (91%) Yes 42 548 (9%) Loneliness No 437 459 (94%) Yes 29 442 (6%) Body-mass index (kg/m2) 27·4 (4·8) Diastolic blood pressure (mm Hg) 82·2 (10·1) Systolic blood pressure (mm Hg) 137·8 (18·6) Handgrip strength (kg) 30·7 (11·0) Smoker No 416 921 (89%) Yes 48 542 (10%) Data missing 1438 (<1%) Ex-smoker No 303 113 (65%) Yes 162 350 (35%) Data missing 1438 (<1%) Alcohol intake frequency Twice or less per week 261 310 (56%) At least three times per week 205 351 (44%) Data missing 240 (<1%) Physical activity† Moderate (range 0–7)‡ 3·7 (2·3) Vigorous (range 0–7)§ 1·9 (1·9) Cognitive performance (range 0–13)18 6·2 (2·1) Patient Health Questionnaire Depressed mood (range 1–4) 1·3 (0·6) Disinterest or no enthusiasm (range 1–4) 1·3 (0·6) Tenseness or restlessness (range 1–4) 1·3 (0·6) Tiredness or lethargy (range 1–4) 1·7 (0·8) Self-rated health (range 1–4) 1·9 (0·7) Data are mean (SD) or number (%). Some percentages do not add up to 100 because of rounding.

onnaire Depressed mood (range 1–4) 1·3 (0·6) Disinterest or no enthusiasm (range 1–4) 1·3 (0·6) Tenseness or restlessness (range 1–4) 1·3 (0·6) Tiredness or lethargy (range 1–4) 1·7 (0·8) Self-rated health (range 1–4) 1·9 (0·7) Data are mean (SD) or number (%). Some percentages do not add up to 100 because of rounding. * A standardised measure of deprivation, including area-level unemployment (as a percentage of those aged 16 years and older who are economically active [ie, not retired or living in care]), non-car ownership (as a percentage of all households), non-home ownership (as a percentage of all households), and household overcrowding. † Number of days per week of physical activity lasting more than 10 min. ‡ Activities that needed moderate effort, resulting in slight shortness of breath. § Activities that caused sweating or heavy breathing, such as cycling, aerobics, or heavy lifting. During a mean follow-up of 6·5 years (SD 0·8) and 3·0 million person-years at risk, 11 593 individuals died. The most common causes of death were neoplasms (6758 deaths) and diseases of the circulatory system (2032 deaths). Other causes (2803 deaths) included diseases of the respiratory and digestive systems and external causes.

ean follow-up of 6·5 years (SD 0·8) and 3·0 million person-years at risk, 11 593 individuals died. The most common causes of death were neoplasms (6758 deaths) and diseases of the circulatory system (2032 deaths). Other causes (2803 deaths) included diseases of the respiratory and digestive systems and external causes. In the complete case analysis, the hazard ratio (adjusted for age, sex, ethnic origin, and chronic disease) for the risk of death from any cause among socially isolated people compared with their non-isolated counterparts was 1·74 (95% CI 1·65–1·83). This association was consistent across sex and age groups, ethnic groups, and participants with and without chronic diseases at baseline, although it was weaker in women than in men (appendix p 6). The results of the cause-specific mortality analyses followed a similar pattern to those for all-cause mortality. Socially isolated individuals had an increased risk of death from neoplasms (minimally adjusted sub-hazard ratio 2·06, 95% CI 1·92–2·20), diseases of the circulatory system (1·68, 1·59–1·77), and other causes (1·57, 1·48–1·66; figure 1).Figure 1 Proportions of the social isolation–mortality association attributable to biological, behavioural, and psychological factors HR=hazard ratio. PERM=percentage of excess risk mediated. SHR=sub-hazard ratio. *Adjusted for age, sex, ethnic origin, and chronic disease.

In the complete case analysis, the hazard ratio (adjusted for age, sex, ethnic origin, and chronic disease) for the risk of death from any cause among socially isolated people compared with their non-isolated counterparts was 1·74 (95% CI 1·65–1·83). This association was consistent across sex and age groups, ethnic groups, and participants with and without chronic diseases at baseline, although it was weaker in women than in men (appendix p 6). The results of the cause-specific mortality analyses followed a similar pattern to those for all-cause mortality. Socially isolated individuals had an increased risk of death from neoplasms (minimally adjusted sub-hazard ratio 2·06, 95% CI 1·92–2·20), diseases of the circulatory system (1·68, 1·59–1·77), and other causes (1·57, 1·48–1·66; figure 1).Figure 1 Proportions of the social isolation–mortality association attributable to biological, behavioural, and psychological factors HR=hazard ratio. PERM=percentage of excess risk mediated. SHR=sub-hazard ratio. *Adjusted for age, sex, ethnic origin, and chronic disease. In the multivariable analyses, the hazard ratio for social isolation compared with no social isolation was 1·73 (95% CI 1·65–1·82) after adjustment for age, sex, ethnic origin, and chronic disease (ie, minimally adjusted), and decreased by 10% after adjustment for biological risk factors, 34% after adjustment for behavioural risk factors, 35% after adjustment for socioeconomic factors, 18% after adjustment for depressive symptoms, 4% after adjustment for cognitive performance, and 32% after adjustment for self-rated health (figure 1). The overall attenuation after adjustment for all these factors (ie, fully adjusted) was 64% (hazard ratio 1·26, 95% CI 1·20–1·33). Similar patterns were found in all the cause-specific mortality groups. Socioeconomic factors (PERM 32–41%), behavioural factors (33–36%), and self-rated health (31–37%) were the strongest explanatory variables, but the fully adjusted sub-hazard ratios remained statistically significant and ranged from 1·32 to 1·22 (figure 1).

patterns were found in all the cause-specific mortality groups. Socioeconomic factors (PERM 32–41%), behavioural factors (33–36%), and self-rated health (31–37%) were the strongest explanatory variables, but the fully adjusted sub-hazard ratios remained statistically significant and ranged from 1·32 to 1·22 (figure 1). In the complete case analysis, the minimally adjusted hazard ratio for the risk of death from any cause in lonely individuals compared with those who were not lonely was 1·37 (95% CI 1·28–1·46). This association was consistent across sex and age groups, ethnic groups, and participants with and without chronic disorders at baseline (appendix p 7). Lonely individuals had an increased risk of death from neoplasms (minimally adjusted sub-hazard ratio 1·75, 95% CI 1·61–1·91), diseases of the circulatory system (1·30, 1·23–1·39), and other causes (1·24, 1·15–1·34; figure 2). Serial adjustments led to complete attenuation of the association (fully adjusted hazard ratio 0·99, 95% CI 0·93–1·06; figure 2). A similar pattern emerged in all the cause-specific mortality groups. Depressive symptoms (PERM 60–77%) and self-rated health (61–84%) seemed to be the strongest mediators, in addition to socioeconomic (41–50%) and behavioural factors (32–53%; figure 2). Social isolation and loneliness were associated with higher levels of depressive symptoms, smoking, and high alcohol intake frequency, all of which are well established correlates of social isolation and loneliness (appendix p 3).Figure 2 Proportions of the loneliness–mortality association attributable to biological, behavioural, and psychological factors

associated with higher levels of depressive symptoms, smoking, and high alcohol intake frequency, all of which are well established correlates of social isolation and loneliness (appendix p 3).Figure 2 Proportions of the loneliness–mortality association attributable to biological, behavioural, and psychological factors HR=hazard ratio. PERM=percentage of excess risk mediated. SHR=sub-hazard ratio. *Adjusted for age, sex, ethnic origin, and chronic disease. The multivariable, minimally adjusted complete-case analysis showed similar associations between all-cause mortality and social isolation (hazard ratio 1·29, 95% CI 1·14–1·47) and loneliness (0·93, 0·77–1·11). The dose–response analyses revealed a dose–response pattern in the associations of social isolation and loneliness with all-cause mortality (appendix p 5). When both isolation and loneliness were tested in the same model, only social isolation predicted all-cause mortality (hazard ratio 1·27, 95% CI 1·20–1·33). In the sensitivity analysis testing for reverse-causation bias, the association between social isolation and all-cause mortality was still apparent (hazard ratio 1·22, 95% CI 1·08–1·38). The frequencies of complete and imputed variables are reported in the appendix (p 4; imputed data sample size 499 238).

The dose–response analyses revealed a dose–response pattern in the associations of social isolation and loneliness with all-cause mortality (appendix p 5). When both isolation and loneliness were tested in the same model, only social isolation predicted all-cause mortality (hazard ratio 1·27, 95% CI 1·20–1·33). In the sensitivity analysis testing for reverse-causation bias, the association between social isolation and all-cause mortality was still apparent (hazard ratio 1·22, 95% CI 1·08–1·38). The frequencies of complete and imputed variables are reported in the appendix (p 4; imputed data sample size 499 238). Discussion In this UK Biobank study, social isolation was associated with increased mortality in the total cohort as well as in the subgroups of men and women, younger and older individuals, initially healthy and unhealthy people, and ethnic subgroups. We also found a similar relation for cause-specific mortality, including deaths from neoplasms and diseases of the circulatory system. Risk factors explained 64% of the association between social isolation and mortality, leaving over a third independent of socioeconomic factors, health-related behaviours, depressive symptoms, biological factors, and cognitive capacity. Loneliness was also associated with increased mortality, but, unlike social isolation, differences in risk-factor levels, especially depressive symptoms, between lonely individuals and others explained its association with all-cause and cause-specific mortality.

ive symptoms, biological factors, and cognitive capacity. Loneliness was also associated with increased mortality, but, unlike social isolation, differences in risk-factor levels, especially depressive symptoms, between lonely individuals and others explained its association with all-cause and cause-specific mortality. To our knowledge, this is the first large-scale study on the contribution of biological, behavioural, socioeconomic, and psychological risk factors to associations between social isolation and all-cause and cause-specific mortality. Our research complements findings from meta-analyses,3, 4, 5 which reported an association between social isolation and increased all-cause mortality. The novel aspects in our analysis include the identification of explanatory factors with greater precision than previously, and the fact that we examined not only all-cause mortality, but also deaths from neoplasms and circulatory diseases.4 The strength of the social isolation–mortality association broadly reflected what has been previously reported, with a hazard ratio of 1·73 in the present study compared with 1·3 in previous meta-analyses.3, 4, 5 This similarity was also the case for loneliness, with a hazard ratio of 1·38 in the present study compared with a minimally adjusted relative risk of 1·3 in a previous meta-analysis.4 However, unlike in some previous studies, our findings suggest that the association between loneliness and mortality is fully attributable to the worse risk-factor levels in the exposed group. Of the risk factors, increased depressive symptoms explained 66% and socioeconomic factors 44% of the excess mortality in lonely individuals. This result broadly confirms the less precise estimates reported in a smaller study in elderly people.24

fully attributable to the worse risk-factor levels in the exposed group. Of the risk factors, increased depressive symptoms explained 66% and socioeconomic factors 44% of the excess mortality in lonely individuals. This result broadly confirms the less precise estimates reported in a smaller study in elderly people.24 Findings from previous studies3, 4 have suggested that socioeconomic factors explain part of the association between social isolation and disease. In support of these findings, we show that 35% of the relation between mortality and social isolation was attributable to education, neighbourhood deprivation, and household income (ie, socioeconomic factors). With regard to loneliness, the contribution of socioeconomic factors was 44%. Contrary to previous findings,6 the present results do not imply that biological factors (eg, obesity, high diastolic and systolic blood pressure, and low handgrip strength) are major contributing factors. However, because our analyses were adjusted for prevalent chronic diseases at baseline, the estimated contribution of biological risk factors did not include variation in these diseases.

ological factors (eg, obesity, high diastolic and systolic blood pressure, and low handgrip strength) are major contributing factors. However, because our analyses were adjusted for prevalent chronic diseases at baseline, the estimated contribution of biological risk factors did not include variation in these diseases. The association between mortality and social isolation seemed stronger than the association between mortality and loneliness. These two factors measure different aspects of social relations and thus also have slightly different associations with health outcomes and mortality. Whereas isolation measures the scarcity of contact with other people and related health resources, loneliness is a perception of detachment associated potentially with emotional states such as depressive symptoms. People can feel lonely even if they are married or living with someone, and that feeling might be less closely related to an absence of practical support than to actual isolation.25

lated health resources, loneliness is a perception of detachment associated potentially with emotional states such as depressive symptoms. People can feel lonely even if they are married or living with someone, and that feeling might be less closely related to an absence of practical support than to actual isolation.25 Some methodological issues should be taken into account when interpreting our findings. Studies in this specialty have typically relied on small samples that are vulnerable to chance findings. Meta-analyses of these studies are additionally limited by heterogeneity in study populations, the diversity of measures used, and differences in the levels of statistical adjustment. The UK Biobank provided an opportunity to investigate the associations of social isolation and loneliness with mortality and their links to risk factors in a large sample, substantially reducing the risk of random error. However, the response rate was low, and selection bias should be taken into account especially when making inferences about population prevalence figures. However, prevalence and incidence were not the focus of the present study. We used a multi-item assessment of social isolation, which is referred to in previous studies as the best predictive validity for such a measurement strategy in relation to mortality.4 Assessing more than one type of social-relationship measurement (structural and functional) might better capture the many effects of social relationships.5, 26

essment of social isolation, which is referred to in previous studies as the best predictive validity for such a measurement strategy in relation to mortality.4 Assessing more than one type of social-relationship measurement (structural and functional) might better capture the many effects of social relationships.5, 26 We measured only simple forms of the complex phenomenon of social networks and interaction, although similar results have been reported in studies using more advanced network analysis.27 Missing data on the covariates reduced the sample size at each successive stage of the regression model adjustment, although the results were similar when the analyses were repeated with only participants who provided complete data on all the variables. The possibility of residual confounding cannot be completely ruled out in observational studies such as ours, although the association between social isolation and mortality remained even after adjustment for a wide range of potential confounders. Similarly, reverse causality can affect the results of observational research. For example, in our study, chronic disease might have affected the risk of both social isolation and mortality. However, we noted an association between social isolation and mortality even among those with no prevalent chronic disease at baseline and after adjustment for a range of health-related covariates. Finally, the sample comprised participants aged between 40 years and 69 years; hence, the findings cannot be extrapolated beyond this age range.

ssociation between social isolation and mortality even among those with no prevalent chronic disease at baseline and after adjustment for a range of health-related covariates. Finally, the sample comprised participants aged between 40 years and 69 years; hence, the findings cannot be extrapolated beyond this age range. In conclusion, data from the UK Biobank suggest that social isolation is associated with overall excess mortality and death attributable to neoplasms and circulatory diseases. Most of the excess mortality among socially isolated and lonely people could be attributed to adverse socioeconomic conditions, an unhealthy lifestyle, and lower mental wellbeing. Public health policies addressing these issues might reduce this excess. Such policies have been designed to increase longevity in the general population. The results of the present study suggest that isolated and lonely people in particular would benefit from successful implementation of targeted policies. Future studies should assess the potential benefits, harms, and cost-effectiveness of interventions and policies aimed at tackling risk factors in lonely and isolated people. Supplementary Material Supplementary appendix Acknowledgments This study was funded by the Academy of Finland (265977), NordForsk, and the UK Medical Research Council (K013351). This research was done using the UK Biobank resource.

In conclusion, data from the UK Biobank suggest that social isolation is associated with overall excess mortality and death attributable to neoplasms and circulatory diseases. Most of the excess mortality among socially isolated and lonely people could be attributed to adverse socioeconomic conditions, an unhealthy lifestyle, and lower mental wellbeing. Public health policies addressing these issues might reduce this excess. Such policies have been designed to increase longevity in the general population. The results of the present study suggest that isolated and lonely people in particular would benefit from successful implementation of targeted policies. Future studies should assess the potential benefits, harms, and cost-effectiveness of interventions and policies aimed at tackling risk factors in lonely and isolated people. Supplementary Material Supplementary appendix Acknowledgments This study was funded by the Academy of Finland (265977), NordForsk, and the UK Medical Research Council (K013351). This research was done using the UK Biobank resource. Contributors ME and CH designed the original hypothesis and CH ran all the analyses in collaboration with ME. ME wrote the first draft of the report, apart from the methods section, which was written by CH. All authors interpreted the results, revised the text, and approved the final draft of the report. Declaration of interests We declare no competing interests.

Introduction The substantial expansion in life expectancy and population ageing during the 20th century is continuing into the 21st century. Life expectancy at age 65 years among the 27 countries of the European Union has increased from 17·8 years in 2002, to 20·0 years in 2014.1 Rapid ageing of populations in developed countries is set to continue; however, evidence about trends in morbidity and disability prevalence in the past few decades is inconsistent.2, 3 Policy makers, service planners, and clinicians need reliable forecasts of future trends in life expectancy and the burden of disease and disability. Current projections involve simple extrapolations that fail to consider the combined effect that trends in disease incidence, particularly cardiovascular disease and dementia, will have on the health status of older people. In the UK, concerns exist regarding potential increases in age-related disability. Between 1991 and 2011, findings from the Cognitive Function and Ageing Study (CFAS)4 showed that although total life expectancy and disability-free life expectancy increased, the proportion of life without disability decreased.

r people. In the UK, concerns exist regarding potential increases in age-related disability. Between 1991 and 2011, findings from the Cognitive Function and Ageing Study (CFAS)4 showed that although total life expectancy and disability-free life expectancy increased, the proportion of life without disability decreased. Trends in life expectancy and disability are shaped primarily by trends in the burden of cardiovascular disease and dementia.5, 6, 7 Both conditions are important underlying causes of age-related disability, particularly in middle-income and high-income countries.7 Cardiovascular disease morbidity and mortality have fallen greatly in the past few decades.8, 9 The associated prolongation of life expectancy has enlarged the pool of individuals surviving to old age and hence susceptible to dementia. Furthermore, because dementia and cardiovascular disease share behavioural and biomedical risk factors,5, 10 reduction in vascular risk might also reduce age-specific dementia incidence.11, 12, 13 On the basis of these two opposing effects, forecasting of the projected prevalence of disability requires simultaneous modelling of both conditions. Research in context Evidence before this study

Trends in life expectancy and disability are shaped primarily by trends in the burden of cardiovascular disease and dementia.5, 6, 7 Both conditions are important underlying causes of age-related disability, particularly in middle-income and high-income countries.7 Cardiovascular disease morbidity and mortality have fallen greatly in the past few decades.8, 9 The associated prolongation of life expectancy has enlarged the pool of individuals surviving to old age and hence susceptible to dementia. Furthermore, because dementia and cardiovascular disease share behavioural and biomedical risk factors,5, 10 reduction in vascular risk might also reduce age-specific dementia incidence.11, 12, 13 On the basis of these two opposing effects, forecasting of the projected prevalence of disability requires simultaneous modelling of both conditions. Research in context Evidence before this study Between Oct 1, and Oct 30, 2016, we searched PubMed for studies forecasting future trends in disability or dementia in the UK, with the search terms “disability”, “dementia”, “longevity”, “life expectancy”, “forecasting”, “simulation”, “model” and synonyms of “United Kingdom”. The appendix (pp 12, 13) lists our complete PubMed search strings and shows results of our systematic review of the literature. We did additional hand searches with lists of references retrieved from relevant papers. We identified only two studies forecasting total life expectancy at age 65 years, neither of which investigated disability or disability-free life expectancies, and two studies reporting a future number of cases with disability in England and Wales. None of these studies modelled future trends in disability and life expectancy while explicitly including interactions between trends in cardiovascular disease and dementia.

stigated disability or disability-free life expectancies, and two studies reporting a future number of cases with disability in England and Wales. None of these studies modelled future trends in disability and life expectancy while explicitly including interactions between trends in cardiovascular disease and dementia. Added value of this study To our knowledge, this is the first study to model future trends in disability in the UK using empirical longitudinal data for England and Wales while also taking into account interactions over time between cardiovascular disease, dementia, and disability. Our findings show that people in England and Wales will live longer but, on average, a quarter of the extra years gained after age 65 years will involve disability. The overall burden of disability will grow primarily as a consequence of population ageing rather than an increase in the prevalence of disability. These predictions have profound individual and societal implications. Implications of all available evidence Changes in vascular risk factors are considered to be the primary drivers of trends in cardiovascular disease and dementia incidence; therefore, future forecasts of disability need to take into account the interaction of these conditions over time. Simulation modelling offers a platform to gain new insights to inform these projections and highlight opportunities for further refinement. Subsequent research should identify which prevention strategies might provide the biggest health and economic benefits in the future.

the interaction of these conditions over time. Simulation modelling offers a platform to gain new insights to inform these projections and highlight opportunities for further refinement. Subsequent research should identify which prevention strategies might provide the biggest health and economic benefits in the future. Previous studies have not considered the complex synergies of life expectancy, cardiovascular disease, and dementia, nor the contribution of these chronic conditions to disability over time.14 We therefore aimed to forecast trends in disability and life expectancy in England and Wales up to 2025, simultaneously accounting for time-varying trends in morbidity and mortality. Methods Model design We developed and validated the IMPACT-Better Ageing Model (IMPACT-BAM)—a discrete-time probabilistic Markov model that follows the progression of a healthy population (aged 35–100 years) of England and Wales from 2006 to 2025 into eight different states characterised by the presence or absence of cardiovascular disease, cognitive impairment, and moderate-to-severe functional impairment (moderate-to-severe disability), and two states for death from cardiovascular disease and non-cardiovascular disease causes (appendix p 14). Movements of the population between these ten states are governed by 1 year age-specific, sex-specific, and year-specific probabilities of transition. IMPACT-BAM is a population model such that for each year in the simulation, a new cohort of 35-year-olds enters the system through the disease-free state.

ppendix p 14). Movements of the population between these ten states are governed by 1 year age-specific, sex-specific, and year-specific probabilities of transition. IMPACT-BAM is a population model such that for each year in the simulation, a new cohort of 35-year-olds enters the system through the disease-free state. Data sources We combined age-specific and sex-specific population estimates from the Office for National Statistics with prevalence data from the English Longitudinal Study of Ageing (ELSA)15 to populate all the states in the model at the start of the simulation. We used projections from the Office for National Statistics until 2025 to create the input population vector of 35-year-olds assumed to be disease-free at entry. Data for calculation of probabilities of transition were also from ELSA.15

ng (ELSA)15 to populate all the states in the model at the start of the simulation. We used projections from the Office for National Statistics until 2025 to create the input population vector of 35-year-olds assumed to be disease-free at entry. Data for calculation of probabilities of transition were also from ELSA.15 Health states We defined cardiovascular disease as a diagnosis of cardiovascular disease; myocardial infarction; or stroke or angina, or both. We defined cognitive impairment without dementia as impairment in two or more domains of cognitive function tests applied to ELSA participants (such as orientation to time, immediate and delayed memory, verbal fluency, and numeracy function). For individuals who were unable to take the cognitive function tests, the Informant Questionnaire on Cognitive Decline (IQCODE) was administered to a proxy informant (usually an immediate family member).16 A score higher than 3·6 on the IQCODE was used to identify cognitive impairment without dementia. We defined functional impairment or disability as the inability to independently do one or more activities of daily living, which included getting in or out of bed, walking across a room, bathing or showering, using the toilet, dressing, cutting food, and eating. This definition of disability captures numbers of individuals who have difficulty maintaining self-care independence and require supportive care.

r more activities of daily living, which included getting in or out of bed, walking across a room, bathing or showering, using the toilet, dressing, cutting food, and eating. This definition of disability captures numbers of individuals who have difficulty maintaining self-care independence and require supportive care. We defined dementia on the basis of the coexistence of cognitive impairment and functional impairment, or a report of a doctor diagnosis of dementia by the participant or caregiver. Transitions to the two death states (cardiovascular disease or non-cardiovascular disease causes) were possible from any health state (appendix p 14). We distinguished four disability states: cardiovacular disease-related disability; dementia-related disability; cardiovacular disease-related and dementia-related disability; and other disease-related disability, defined as other forms of disability not linked to cardiovascular disease or dementia. Model assumptions and outputs We assumed that population trends in cardiovascular disease incidence would parallel the rate of decline in cardiovascular disease mortality up to 2025, as observed in ELSA for the period 2002–13.15 We further assumed that dementia incidence would follow a 2·7% relative annual decline, as also observed in ELSA.

outputs We assumed that population trends in cardiovascular disease incidence would parallel the rate of decline in cardiovascular disease mortality up to 2025, as observed in ELSA for the period 2002–13.15 We further assumed that dementia incidence would follow a 2·7% relative annual decline, as also observed in ELSA. IMPACT-BAM was used to calculate future trends in the prevalence of disability (both age-specific and age-standardised using the population in 2015), life expectancy, disabled life expectancy, and disability-free life expectancy according to the Sullivan method.17 The model was implemented as a statistical package in R software. Appendix pp 13–23 provide detailed information about data sources, definitions of health states, and methods. Sensitivity analyses In view of uncertainty regarding trends in dementia incidence, we tested two alternative assumptions for the trend in future dementia incidence: constant incidence (no annual decline) and an annual decline of 4% in dementia incidence (appendix pp 24, 25). To explore the effect of parameter uncertainty on model outputs, we did a probabilistic sensitivity analysis using Monte Carlo simulation. The procedure entailed iterative sampling from specified distributions for the input parameters used in the model and recalculation of the outputs. We did 1000 iterations to estimate 95% uncertainty intervals (UIs) for the output variables. Appendix p 25 provides additional details of distributions and statistical functions.

ocedure entailed iterative sampling from specified distributions for the input parameters used in the model and recalculation of the outputs. We did 1000 iterations to estimate 95% uncertainty intervals (UIs) for the output variables. Appendix p 25 provides additional details of distributions and statistical functions. Model validation To validate cardiovascular disease and non-cardiovascular disease deaths, we compared our model estimates with observed mortality data from the Office of National Statistics for the period 2006–12. For cardiovascular disease prevalence, we compared our model estimates with those published by the Health Survey for England 2011. For disability prevalence, we compared 2014 prevalence estimates from the most recent wave in ELSA (not used in the design of the model) with our estimates. For dementia prevalence, we compared our model estimates with those reported in the CFAS II.3 Finally, we compared our estimates of life expectancy at age 65 years for the period 2006–12 with those reported by the European Health and Life Expectancy Information System (EHLEIS) and the Office of National Statistics (appendix pp 27–31). 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 final responsibility for the decision to submit for publication.

Model validation To validate cardiovascular disease and non-cardiovascular disease deaths, we compared our model estimates with observed mortality data from the Office of National Statistics for the period 2006–12. For cardiovascular disease prevalence, we compared our model estimates with those published by the Health Survey for England 2011. For disability prevalence, we compared 2014 prevalence estimates from the most recent wave in ELSA (not used in the design of the model) with our estimates. For dementia prevalence, we compared our model estimates with those reported in the CFAS II.3 Finally, we compared our estimates of life expectancy at age 65 years for the period 2006–12 with those reported by the European Health and Life Expectancy Information System (EHLEIS) and the Office of National Statistics (appendix pp 27–31). 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 final responsibility for the decision to submit for publication. Results Our projections indicate that the number of people in England and Wales aged 65 years and older will increase by roughly 19·4% (95% UI 17·7–20·9), from 10·4 million (10·37–10·41 million) in 2015, to 12·4 million (12·23–12·57 million) in 2025. Notably, the number of people aged 85 years and older will increase by 38·9% (95% UI 31·9–45·5), from 1·4 million (1·38–1·45 million) in 2015, to 2·0 million (1·87–2·07 million) in 2025.

by roughly 19·4% (95% UI 17·7–20·9), from 10·4 million (10·37–10·41 million) in 2015, to 12·4 million (12·23–12·57 million) in 2025. Notably, the number of people aged 85 years and older will increase by 38·9% (95% UI 31·9–45·5), from 1·4 million (1·38–1·45 million) in 2015, to 2·0 million (1·87–2·07 million) in 2025. The age-specific disability prevalence between 2002 and 2013 in ELSA ranged from 14% in people aged 65–69 years to 57% in those aged 90 years and older (appendix p 26). Additionally, we estimated that 53% of all people aged 65 years or older with disability had cardiovascular disease or cognitive impairment (appendix pp 26, 27). We predict that, between 2015 and 2025, the number of people living with disability will increase by about 2·3% per annum (table 1). By 2025, roughly 2·8 million people will be living with disability, an additional 560 000 cases compared with 2015 (a 25·0% overall increase; table 1). The number of men living with disabilities will increase by roughly 3% per year to reach 1·24 million cases in 2025, whereas the number of women living with disabilities will increase by 1·7% per year to reach 1·58 million cases in 2025 (table 1, figure 1).Figure 1 Projected number of cases (A) and prevalence (B) of disability in men and women aged 65 years or older from 2015 to 2025 in England and Wales Shaded areas represent 95% uncertainty intervals. Prevalence of disability is standardised to 2015 population estimates for England and Wales. Table 1 Projected number of disability cases (in thousands) by sex and age in 2015 and 2025 in England and Wales

The age-specific disability prevalence between 2002 and 2013 in ELSA ranged from 14% in people aged 65–69 years to 57% in those aged 90 years and older (appendix p 26). Additionally, we estimated that 53% of all people aged 65 years or older with disability had cardiovascular disease or cognitive impairment (appendix pp 26, 27). We predict that, between 2015 and 2025, the number of people living with disability will increase by about 2·3% per annum (table 1). By 2025, roughly 2·8 million people will be living with disability, an additional 560 000 cases compared with 2015 (a 25·0% overall increase; table 1). The number of men living with disabilities will increase by roughly 3% per year to reach 1·24 million cases in 2025, whereas the number of women living with disabilities will increase by 1·7% per year to reach 1·58 million cases in 2025 (table 1, figure 1).Figure 1 Projected number of cases (A) and prevalence (B) of disability in men and women aged 65 years or older from 2015 to 2025 in England and Wales Shaded areas represent 95% uncertainty intervals. Prevalence of disability is standardised to 2015 population estimates for England and Wales. Table 1 Projected number of disability cases (in thousands) by sex and age in 2015 and 2025 in England and Wales 2015 2025 Annual relative change (%) Relative change between 2015 and 2025 (%) All people (age, years) ≥65 2251 (2235–2268) 2811 (2727–2890) 2·3% (2·0–2·5) 25·0% (21·3–28·2) 65–84 1692 (1679–1706) 2010 (1969–2049) 1·7% (1·5–1·9) 18·9% (16·6–20·9) ≥85 559 (552–567) 800 (750–852) 3·7% (3·0–4·3) 43·2% (34·2–52·1) Men (age, years) ≥65 922 (911–933) 1236 (1187–1279) 3·0% (2·5–3·3) 34·0% (28·6–38·5) 65–84 745 (735–755) 914 (888–939) 2·1% (1·8–2·3) 22·6% (19·2–25·5) ≥85 177 (173–181) 322 (297–349) 6·1% (5·3–6·9) 81·6% (67·2–95·3) Women (age, years) ≥65 1329 (1316–1343) 1578 (1508–1639) 1·7% (1·3–2·1) 18·7% (13·7–23·2) 65–84 947 (936–959) 1098 (1068–1126) 1·5% (1·2–1·7) 16·0% (12·9–18·8) ≥85 382 (376–388) 480 (434–521) 2·3% (1·3–3·2) 25·4% (14·2–36·4) Data in parentheses are 95% uncertainty intervals.

2 (297–349) 6·1% (5·3–6·9) 81·6% (67·2–95·3) Women (age, years) ≥65 1329 (1316–1343) 1578 (1508–1639) 1·7% (1·3–2·1) 18·7% (13·7–23·2) 65–84 947 (936–959) 1098 (1068–1126) 1·5% (1·2–1·7) 16·0% (12·9–18·8) ≥85 382 (376–388) 480 (434–521) 2·3% (1·3–3·2) 25·4% (14·2–36·4) Data in parentheses are 95% uncertainty intervals. Between 2015 and 2025, other disease-related disability will remain the most frequent type of disability among people aged 65–84 years, whereas dementia-related and cardiovascular disease-related disability will persist as the most common types of disability among people aged 85 years or older (table 2). The numbers of other disease-related disability cases among people aged 65–84 years will increase in the next decade by about 31%, and the number of dementia-related disability cases by about 40%; however, the cases of cardiovascular disease-related disability will decline by about 17% (table 2). In people aged 85 years or older, the numbers of other disease-related disability, dementia-related disability, and cardiovascular disease-related disability will increase by 84%, 63%, and 6%, respectively (table 2).Table 2 Projected number of disease-related disability cases (in thousands) by age in 2015 and 2025 in England and Wales

5 years or older, the numbers of other disease-related disability, dementia-related disability, and cardiovascular disease-related disability will increase by 84%, 63%, and 6%, respectively (table 2).Table 2 Projected number of disease-related disability cases (in thousands) by age in 2015 and 2025 in England and Wales 2015 2025 Relative change between 2015 and 2025 (%) ≥65 years old CVD-related disability 588 (576 to 599) 527 (505 to 547) −10·3% (−13·6 to 7·3) Dementia-related disability 468 (447 to 491) 699 (654 to 745) 49·1% (41·5 to 56·2) Dementia and CVD-related disability 177 (171 to 183) 191 (178 to 203) 7·7% (1·3 to 14·0) Other disease-related disability 1018 (995 to 1041) 1395 (1355 to 1440) 37·1% (34·3 to 39·8) 65 to 84 years old CVD-related disability 419 (410 to 428) 348 (335 to 363) −16·9% (−19·8 to 3·9) Dementia-related disability 289 (269 to 309) 405 (374 to 436) 40·3% (34·7 to 46·3) Dementia and CVD-related disability 84 (79 to 88) 79 (73 to 85) −5·3% (−10·8 to 0·7) Other disease-related disability 900 (879 to 921) 1177 (1145 to 1213) 30·9% (28·5 to 33·1) ≥85 years old CVD -related disability 168 (163 to 174) 179 (166 to 190) 6·0% (−0·8 to 12·6) Dementia-related disability 179 (171 to 188) 293 (267 to 320) 63·1% (50·7 to 76·5) Dementia and CVD-related disability 93 (89 to 97) 111 (102 to 120) 19·3% (9·8 to 29·0) Other disease-related disability 118 (111 to 126) 217 (200 to 235) 84·2% (74·6 to 94·2) Data in parentheses are 95% uncertainty intervals. CVD=cardiovascular disease.

disability 179 (171 to 188) 293 (267 to 320) 63·1% (50·7 to 76·5) Dementia and CVD-related disability 93 (89 to 97) 111 (102 to 120) 19·3% (9·8 to 29·0) Other disease-related disability 118 (111 to 126) 217 (200 to 235) 84·2% (74·6 to 94·2) Data in parentheses are 95% uncertainty intervals. CVD=cardiovascular disease. The age-standardised prevalence of disability in the population aged 65 years or older will remain broadly constant to 2025 in both men and women (figure 1). A modest increase of 1·6% in prevalence is predicted in the oldest men in the next decade (table 3); however, trends in age-standardised prevalence will differ by type of disability: prevalence of cardiovascular disease-related disability will decline by about 30%, whereas prevalence of dementia-related and other disease-related disability will increase by about 14% (appendix pp 4–6).Table 3 Projected percentage of disability cases by sex and age in 2015 and 2025 in England and Wales 2015 2025 All people (age, years) ≥65 21·7% (21·5–21·8) 21·6% (21·3–21·8) ≥75 28·8% (28·5–29·0) 29·0% (28·5–29·3) ≥85 39·5% (38·9–40·0) 39·8% (39·0–40·4) Men (age, years) ≥65 19·6% (19·3–19·8) 19·8% (19·4–20·1) ≥75 25·6% (25·2–26·0) 26·4% (25·9–26·9) ≥85 34·8% (34·0–35·7) 36·4% (35·4–37·3) Women (age, years) ≥65 23·4% (23·1–23·6) 23·2% (22·9–23·6) ≥75 31·1% (30·7–31·4) 31·0% (30·4–31·5) ≥85 42·1% (41·4–42·8) 42·2% (41·2–43·0) Data in parentheses are 95% uncertainty intervals. Percentages are age-standardised to the 2015 population of England and Wales.

(25·9–26·9) ≥85 34·8% (34·0–35·7) 36·4% (35·4–37·3) Women (age, years) ≥65 23·4% (23·1–23·6) 23·2% (22·9–23·6) ≥75 31·1% (30·7–31·4) 31·0% (30·4–31·5) ≥85 42·1% (41·4–42·8) 42·2% (41·2–43·0) Data in parentheses are 95% uncertainty intervals. Percentages are age-standardised to the 2015 population of England and Wales. Overall in people aged 65 years, total life expectancy, disability-free life expectancy, and life expectancy with disability will increase in the entire population between 2015 and 2025 (table 4), and there will be increases in life expectancy in all age groups (figure 2). Appendix p 31 compares the life expectancy estimates in our study with those from other studies. Disability-free life expectancy at age 65 years will increase by 1·0 years, from 15·4 years to 16·4 years; however, life expectancy with disability will grow more in relative terms (about 15% increase from 2015; table 4). For both health expectancies, bigger increases are predicted for men than for women (table 4). Overall, the proportion of the life expectancy lived with disabilities at age 65 years will increase from 21·4% to 24·0% in men and 24·9% to 25·8% in women (table 4).Figure 2 Projected life expectancy trends between 2015 and 2025 in England and Wales decomposed by years lived with and without disability from different age points Error bars show 95% uncertainty intervals. Table 4 Projected estimates of life expectancy in people aged 65 years in 2015 and 2025 in England and Wales

Overall in people aged 65 years, total life expectancy, disability-free life expectancy, and life expectancy with disability will increase in the entire population between 2015 and 2025 (table 4), and there will be increases in life expectancy in all age groups (figure 2). Appendix p 31 compares the life expectancy estimates in our study with those from other studies. Disability-free life expectancy at age 65 years will increase by 1·0 years, from 15·4 years to 16·4 years; however, life expectancy with disability will grow more in relative terms (about 15% increase from 2015; table 4). For both health expectancies, bigger increases are predicted for men than for women (table 4). Overall, the proportion of the life expectancy lived with disabilities at age 65 years will increase from 21·4% to 24·0% in men and 24·9% to 25·8% in women (table 4).Figure 2 Projected life expectancy trends between 2015 and 2025 in England and Wales decomposed by years lived with and without disability from different age points Error bars show 95% uncertainty intervals. Table 4 Projected estimates of life expectancy in people aged 65 years in 2015 and 2025 in England and Wales 2015 2025 Difference All people Total 20·1 (19·9 to 20·3) 21·8 (20·2 to 23·6) 1·7 (0·1 to 3·6) Without disability 15·4 (15·3 to 15·5) 16·4 (15·5 to 17·3) 1·0 (0·1 to 1·9) With disability 4·7 (4·6 to 4·8) 5·4 (4·7 to 6·4) 0·7 (0·0 to 1·7) Proportion lived with disability 23·4% (23·1 to 23·6) 24·9% (23·2 to 27·0) 1·5% (−0·1 to 3·7) Men Total 19·0 (18·7 to 19·3) 21·7 (19·9 to 23·9) 2·7 (0·9 to 4·9) Without disability 14·9 (14·7 to 15·1) 16·5 (15·4 to 17·6) 1·6 (0·5 to 2·7) With disability 4·1 (3·9 to 4·2) 5·2 (4·4 to 6·3) 1·1 (0·4 to 2·2) Proportion lived with disability 21·4% (21·0 to 21·7) 24·0% (22·2 to 26·4) 2·6% (0·8 to 5·0) Women Total 21·0 (20·8 to 21·2) 22·1 (19·7 to 24·7) 1·1 (−1·3 to 3·8) Without disability 15·8 (15·7 to 15·9) 16·4 (15·1 to 17·7) 0·6 (−0·7 to 1·9) With disability 5·2 (5·1 to 5·3) 5·7 (4·6 to 7·1) 0·5 (−0·6 to 1·9) Proportion lived with disability 24·9% (24·5 to 25·2) 25·8% (23·5 to 28·9) 0·9% (−1·4 to 4·1) Data in parentheses are 95% uncertainty intervals.

9·7 to 24·7) 1·1 (−1·3 to 3·8) Without disability 15·8 (15·7 to 15·9) 16·4 (15·1 to 17·7) 0·6 (−0·7 to 1·9) With disability 5·2 (5·1 to 5·3) 5·7 (4·6 to 7·1) 0·5 (−0·6 to 1·9) Proportion lived with disability 24·9% (24·5 to 25·2) 25·8% (23·5 to 28·9) 0·9% (−1·4 to 4·1) Data in parentheses are 95% uncertainty intervals. Results of our sensitivity analysis showed that estimates of numbers of people with disability remained almost unchanged despite different calendar trends in incidence of dementia (appendix p 8). Furthermore, the proportion of life expectancy lived with disability will remain virtually unchanged from the baseline scenario for both men and woman (appendix p 11). However, the two alternative assumptions regarding the trend in future dementia incidence do affect the numbers in the disease-specific disability (appendix pp 7–11). If dementia incidence remains unchanged over the next decade, the cases of dementia-related disability will increase compared to our main prediction. This increase will be counterbalanced by a decrease in the number of cases of other types of disability, including cardiovascular disease-related disability. Conversely, a faster annual decline in dementia incidence would result in fewer cases of dementia-related disability, but an increase in the numbers of other types of disability.

ll be counterbalanced by a decrease in the number of cases of other types of disability, including cardiovascular disease-related disability. Conversely, a faster annual decline in dementia incidence would result in fewer cases of dementia-related disability, but an increase in the numbers of other types of disability. We validated key model outputs against empirical observations. The model provided a good match with the Office for National Statistics estimates of the number of cardiovascular and non-cardiovascular deaths (appendix pp 28, 29). Our estimates of cardiovascular disease prevalence in 2011 were similar to those reported by the Health Survey England, particularly for men (appendix p 29). Our findings for disability prevalence in 2014 for women were very similar to those reported by ELSA wave 7; however, our model estimates a slightly lower prevalence of disability in men (appendix p 30). Our age-specific estimates of dementia in 2011 were similar to those reported in CFAS II (appendix p 30). Our estimates of life expectancy at 65 years for 2006–15 were similar to those reported by the Office of National Statistics and EHLEIS (appendix pp 30, 31).

y lower prevalence of disability in men (appendix p 30). Our age-specific estimates of dementia in 2011 were similar to those reported in CFAS II (appendix p 30). Our estimates of life expectancy at 65 years for 2006–15 were similar to those reported by the Office of National Statistics and EHLEIS (appendix pp 30, 31). Discussion Our life expectancy and disability forecasts are based on a Markov model, which, for the first time, synthesises the combined effects of present trends in incidence of cardiovascular disease, dementia disability, and mortality. Our findings show that the number of people aged 65 years or older with care needs in England and Wales could reach 2·8 million by 2025. This 25% increase will mainly reflect population ageing rather than an increase in the prevalence of disability. Lifespans will increase further, but a quarter of life expectancy at age 65 years will involve disability.

people aged 65 years or older with care needs in England and Wales could reach 2·8 million by 2025. This 25% increase will mainly reflect population ageing rather than an increase in the prevalence of disability. Lifespans will increase further, but a quarter of life expectancy at age 65 years will involve disability. Other forms of disability not linked to cardiovascular disease or dementia will persist between 2015 and 2025 as the most frequent type of disability among people aged 65–84 years. Our findings are consistent with those from a 2016 analysis of the most important contributors to global disability burden for this age group.18 For people older than 85 years, future disability levels will be influenced mainly by the joint evolution of the burdens of dementia and cardiovascular disease over time. Evidence suggests that declines in cardiovascular disease incidence and mortality in the past few decades are set to continue.8, 9 Encouragingly, progressive declines in the incidence of dementia have been reported in the past 20 years in Europe and the USA;11, 12, 13 however, the size of the decline varied across these study populations. To account for this uncertainty, we modelled the 2·7% annual decline observed in ELSA15 and showed that the projected burden of disability was robust to two alternative assumptions about the future dementia incidence trend, 0% and 4%.

SA;11, 12, 13 however, the size of the decline varied across these study populations. To account for this uncertainty, we modelled the 2·7% annual decline observed in ELSA15 and showed that the projected burden of disability was robust to two alternative assumptions about the future dementia incidence trend, 0% and 4%. Our projections of life expectancy at 65 years are broadly similar to those from the Office of National Statistics and other studies;19, 20 the largest differences are for women, for whom our projections are lower. In developed countries, women have tended to live longer, but to have worse health than men.21 Excess mortality in men has been attributed to higher prevalence of life-threatening chronic conditions such as cardiovascular disease.22 Therefore, the higher increment in life expectancy for men than for women between 2015 and 2025, could be explained by life expectancy in men being more sensitive to the declining trends in cardiovascular disease incidence and mortality observed in ELSA over the past decade and projected forward in our model. Although women have poorer health than men, attributable to a higher prevalence of dementia and functional limitations, our results suggest the present sex difference in disabled life expectancy will diminish because men will have a relatively larger increase in the burden of disability.

projected forward in our model. Although women have poorer health than men, attributable to a higher prevalence of dementia and functional limitations, our results suggest the present sex difference in disabled life expectancy will diminish because men will have a relatively larger increase in the burden of disability. Two previous studies have predicted future numbers of disability cases in England and Wales. Jagger and colleagues23 used a macro-simulation model of cohort transitions between non-disabled, disabled, and death states from 1992 to 2026. The investigators estimated transition probabilities from CFAS I conditional on the prevalence of several chronic conditions and assuming their prevalence to be constant over the forecasting period. Comas-Herrera and colleagues24 applied age-specific and sex-specific prevalence of cognitive impairment and disability observed in 2002 (also based on CFAS I) to Government Actuary Department's population projections in 2031, again assuming that dementia incidence will be constant over time. Comparisons of these studies with ours are indirect because the definitions of disability differ. Correspondingly, both previous studies predicted a smaller future burden of disability than IMPACT-BAM. This divergence arises largely because estimates of disability prevalence in the base year of respective models (CFAS I: 1992, ELSA: 2002) depend on the number of activities included in the activities of daily living instrument. The ELSA instrument included six activities, compared with three in CFAS I.25 The greater the number of items of activities of daily living, the higher the probability of detection of disabilities.

odels (CFAS I: 1992, ELSA: 2002) depend on the number of activities included in the activities of daily living instrument. The ELSA instrument included six activities, compared with three in CFAS I.25 The greater the number of items of activities of daily living, the higher the probability of detection of disabilities. Credible predictions of future prevalence of disability and life expectancy need to consider the complex population dynamics of lifespan and the major disability determinants. IMPACT-BAM responds to these requirements by modelling two major diseases with shared and strongly interrelated determinants, both of which are associated with substantial functional impairment.5, 10 Few studies have attempted to estimate future disability in England and Wales on the basis of dementia trends.23, 24 Moreover, none has explicitly considered transitions between cardiovascular disease, dementia, and functional impairment states while simultaneously taking into account changing time trends in dementia incidence, cardiovascular disease incidence, cardiovascular disease mortality, and non-cardiovascular disease mortality. Furthermore, our model parameters derive from the best available population-based longitudinal evidence in the UK, and cross-validation of our projections used external data sources and estimates.

ia incidence, cardiovascular disease incidence, cardiovascular disease mortality, and non-cardiovascular disease mortality. Furthermore, our model parameters derive from the best available population-based longitudinal evidence in the UK, and cross-validation of our projections used external data sources and estimates. Prediction modelling has limitations. First, IMPACT-BAM aggregates disability caused by conditions such as musculoskeletal disorders, mental disorders, diabetes, chronic respiratory diseases, and other chronic diseases in a single other disease-related disability state. Transitions from this state to other states in the model are not conditioned to the prevalence of these specific chronic conditions. However, because ELSA is representative of the English population, the estimated transition probabilities to cardiovascular disease, dementia, or death from this combined state are likely to represent a weighted average of risk across all comorbidities. Although outside the scope of our study, consideration of disability on a spectrum of severity has important implications for health-care planning and policy, because each level of severity generates corresponding needs in care and assistance. Our binary definition of disability is relevant to future demand of long-term care because it provides a transparent basis for estimation of the trend in numbers of individuals who would require supportive care. Finally, the cohort of 35-year-olds populating the model every year was obtained from official population projections.26 These principal projections are based on assumptions regarding future levels of fertility, migration, and mortality, which might add uncertainty to our estimates. However, population projections have proved to be relatively robust to mortality assumptions, whereas fertility and migrant variant assumptions only affect the projected numbers of children and young adults.26

regarding future levels of fertility, migration, and mortality, which might add uncertainty to our estimates. However, population projections have proved to be relatively robust to mortality assumptions, whereas fertility and migrant variant assumptions only affect the projected numbers of children and young adults.26 The societal, economic, and public health implications of our forecast are substantial. In particular, our findings draw attention to the scale of societal costs associated with disability in the coming decade. Public and private expenditure on long-term care will need to increase considerably by 2025, in view of the predicted 25% rise in the number of people who will have age-related disability. This situation has serious implications for a cash-strapped and overburdened National Health Service, and an under-resourced social care system.

ivate expenditure on long-term care will need to increase considerably by 2025, in view of the predicted 25% rise in the number of people who will have age-related disability. This situation has serious implications for a cash-strapped and overburdened National Health Service, and an under-resourced social care system. In the context of the rapid and continuing rise in the number of older dependent people in the UK, the government needs to give urgent consideration to options for more cost-effective health and social care provision in all its forms. First, national capacity for institutional care needs to increase. Second, informal and home care require stronger policy support, for example by means of tax allowances or cash benefits. Affected individuals and their families pay an estimated 40% of the national cost of long-term care from income and savings.27 Notably, the disadvantage of older people on lower incomes unable to live independently will increase if the shortage of caregivers and the precarious state of institutional and domiciliary care provision is not addressed.28

families pay an estimated 40% of the national cost of long-term care from income and savings.27 Notably, the disadvantage of older people on lower incomes unable to live independently will increase if the shortage of caregivers and the precarious state of institutional and domiciliary care provision is not addressed.28 Cardiovascular disease and dementia share risk factors—namely, poor diet, smoking, high alcohol consumption, hypertension, diabetes, and physical inactivity. Effective prevention strategies are therefore strongly advocated by UK health charities, the National Institute of Health and Care Excellence, and WHO.29, 30 This shared-determinants approach is an important but neglected strategy for reducing the future burden of disability at the population level. Immediate substantial investment in such prevention policies could be substantially cost-saving.31 Policy simulation modelling based on interventions aimed at the drivers of disability is needed to inform this debate. Potential prevention strategies include healthy food interventions, fiscal measures for obesity control, and disease-specific health-care-based interventions.

cies could be substantially cost-saving.31 Policy simulation modelling based on interventions aimed at the drivers of disability is needed to inform this debate. Potential prevention strategies include healthy food interventions, fiscal measures for obesity control, and disease-specific health-care-based interventions. In conclusion, our evidence-based forecasting model, incorporating the expected continuing declines in cardiovascular disease and dementia incidence, predicts that the number of people living with disability will increase over the next decade. This increase mainly reflects the expected continuing rise in life expectancy and resulting upward shift in the age distribution of the population. The rising burden of age-related disability accompanying population ageing poses a substantial societal challenge and emphasises the urgent need for policy development that includes effective prevention interventions. Supplementary Material Supplementary appendix Acknowledgments This study is funded by the British Heart Foundation (grant number RG/13/2/30098). Contributors All authors made a substantial contribution to study conception and design. EJB and MO'F developed the original idea. MG-C and PB designed and implemented the model, with inputs from MO'F, SA-A, MJS, and EJB. SA-A, MJS, and MG-C analysed and prepared the data. All authors contributed to interpreting the results, drafting the manuscript and the revisions. Declaration of interests We declare no competing interests.

Introduction Trauma-related hospital admission following serious injuries and poisonings, whether self-inflicted, perpetrated by other people, or occurring accidentally, can be highly distressing experiences for children and their families. Furthermore, these episodes predict short-term and long-term increases in risk of suicide and other causes of premature death,1, 2, 3, 4 and they also absorb scarce health-care resources disproportionately.4 To inform the coordinated development of multiagency initiatives to reduce adverse outcome risks among psychosocially vulnerable young people, it is therefore necessary to better understand the long-term trajectories of children who require hospital admission following major injuries or poisonings, to help to ensure their safe progression through to healthy adult maturity. These trauma-related hospital admissions have been described as salient so-called teachable moments, providing an opportunity to deliver considered interventions that could potentially prevent future harmful behaviours.5

injuries or poisonings, to help to ensure their safe progression through to healthy adult maturity. These trauma-related hospital admissions have been described as salient so-called teachable moments, providing an opportunity to deliver considered interventions that could potentially prevent future harmful behaviours.5 A paper published in 2015 reported on a large longitudinal investigation of hospital admissions for violent, drug-related or alcohol-related, self-inflicted, and accidental injury among adolescents aged 10–19 years in England.3 The study indicated strong links between this exposure and risks of death and emergency readmission within 10 years of the index hospitalisation episode. Two previous investigations reporting results from large routinely collected datasets in New Zealand have examined subsequent risks of assaultive injury,6 self-injury, and suicide,7 without specifying an age range for experiencing trauma-related hospital admission. We examined such hospital admissions that occurred specifically during childhood, to establish the longer-term trajectories of this particularly vulnerable group of young people. Research in context Evidence before this study

A paper published in 2015 reported on a large longitudinal investigation of hospital admissions for violent, drug-related or alcohol-related, self-inflicted, and accidental injury among adolescents aged 10–19 years in England.3 The study indicated strong links between this exposure and risks of death and emergency readmission within 10 years of the index hospitalisation episode. Two previous investigations reporting results from large routinely collected datasets in New Zealand have examined subsequent risks of assaultive injury,6 self-injury, and suicide,7 without specifying an age range for experiencing trauma-related hospital admission. We examined such hospital admissions that occurred specifically during childhood, to establish the longer-term trajectories of this particularly vulnerable group of young people. Research in context Evidence before this study We searched for article titles published in English in MEDLINE and Embase up until Feb 24, 2017, that included the following combination of terms: “child” or “youth” or “adolesc” AND “hospitali” or “admitted” or “admission” AND “trauma” or “self-harm” or “self-inflict” or “self-injur” or “self-poison” or “suicid” or “violen” or “assault” or “accident” or “death” or “die” or “mortality”. We discovered that the existing evidence-base for this topic was limited. Previous national registry studies have examined links between history of trauma-related hospital admission and future risks of self-harm and suicide, assaultive injury, emergency readmission, and premature death. These studies have consistently shown strong associations, but several important research questions remain unanswered, including the long-term trajectories of people who experienced trauma-related hospital admissions during childhood, assessment of self-harm versus violent offending risks in the same study cohort, and cumulative risk estimation.

ntly shown strong associations, but several important research questions remain unanswered, including the long-term trajectories of people who experienced trauma-related hospital admissions during childhood, assessment of self-harm versus violent offending risks in the same study cohort, and cumulative risk estimation. Added value of this study We did a national cohort study of more than 1 million people. We identified all trauma-related hospital admissions during childhood through to the 15th birthday and examined adverse outcomes between mid-adolescence and age 35 years. We compared risks for internalised versus externalised violence, and we generated measures of sex-specific absolute risks in this population by deriving estimates of cumulative incidence values that accounted for competing risks. Around one in seven men admitted to hospital during childhood following self-harm and one in four due to interpersonal violence will be convicted for committing a violent crime by age 35 years. About one in five women admitted to hospital after an episode of self-harm or interpersonal violence during childhood will be admitted to hospital again following self-harm between their 15th and 35th birthdays. Implications of all the available evidence

We did a national cohort study of more than 1 million people. We identified all trauma-related hospital admissions during childhood through to the 15th birthday and examined adverse outcomes between mid-adolescence and age 35 years. We compared risks for internalised versus externalised violence, and we generated measures of sex-specific absolute risks in this population by deriving estimates of cumulative incidence values that accounted for competing risks. Around one in seven men admitted to hospital during childhood following self-harm and one in four due to interpersonal violence will be convicted for committing a violent crime by age 35 years. About one in five women admitted to hospital after an episode of self-harm or interpersonal violence during childhood will be admitted to hospital again following self-harm between their 15th and 35th birthdays. Implications of all the available evidence The psychosocial wellbeing of individuals who experienced trauma-related hospital admission while growing up warrants careful consideration in the development of comprehensive strategies to address internalised and externalised violence in young people. National clinical guidelines for provision of psychosocial assessment target single problems such as self-harm, but they could be usefully broadened to encompass other adverse events such as hospital admission of children following episodes of interpersonal violence or serious accidents. Trauma-related hospital admission of a child might present an important opportunity to implement a family-oriented intervention in the hospital setting, with the proactive purpose of reducing future risk of harmful or self-destructive behaviours. Particularly, close monitoring and robust support are indicated for young women who were admitted to hospital as children on more than one occasion following trauma and for those who were admitted post trauma for multiple reasons during their childhood.

of reducing future risk of harmful or self-destructive behaviours. Particularly, close monitoring and robust support are indicated for young women who were admitted to hospital as children on more than one occasion following trauma and for those who were admitted post trauma for multiple reasons during their childhood. Published reports have tended to focus on a single adverse outcome, such as youth suicide,8 whereas we set out to harness the potential of national Danish registers to examine the longer-term trajectories of affected children. We investigated self-harm and violent criminality as adverse outcomes because self-directed and externalised violence are associated harmful behaviours that share common causal factors. A Swedish national registry study from 2006 reported a five times increasd risk of violent crime conviction among people with a history of hospital-treated self-harm, with an independent doubled risk after adjustment for psychiatric comorbidity and environmental factors.9 The combined societal costs of these two related deleterious behaviours are immense,10 prompting calls for concerted action to tackle them in unison.11

ion among people with a history of hospital-treated self-harm, with an independent doubled risk after adjustment for psychiatric comorbidity and environmental factors.9 The combined societal costs of these two related deleterious behaviours are immense,10 prompting calls for concerted action to tackle them in unison.11 The aims of this national cohort study were: (1) to estimate the relative and absolute risks of self-harm and violent criminality among youths and young adults who experienced hospital admission due to injuries and poisonings during their childhood; (2) to compare these estimates by sex and by cause of hospital admission—self-harm, interpersonal violence, or accident; (3) to assess confounding by parental socioeconomic status (SES); and (4) to evaluate effect modification by frequency of and multiple reasons for trauma-related hospital admission during childhood. We anticipated observing especially increased risks among individuals admitted to hospital as children following interpersonal violence or self-harm versus the reference category of people who had no experience of trauma-related hospital admissions during childhood. We also examined exposure to hospital admission following accidental injury or poisoning as an additional comparison. We acknowledge that a sizeable proportion of admissions following accidents might not have been anywhere near as traumatising to the affected children as those that followed internalised or externalised violence.

amined exposure to hospital admission following accidental injury or poisoning as an additional comparison. We acknowledge that a sizeable proportion of admissions following accidents might not have been anywhere near as traumatising to the affected children as those that followed internalised or externalised violence. Methods Study design and participants Since 1968 the Civil Registration System has registered all Danish residents12 by capturing date and place of birth, sex, and continuously updated vital status information. Its mandatory unique personal identification numbering system enables complete and accurate linkage to health-related administrative registers as well as similarly comprehensive parent-offspring linkage. The study cohort consisted of all people born in Denmark from Jan 1, 1977, to Dec 31, 1997, who still resided in Denmark at their 15th birthday, and whose parents were both Danish-born, thereby accounting for increased risks of self-harm and violent criminality linked with first and second generation immigrant status.13 Furthermore, the study registers provided less complete information about experience of trauma-related hospital admissions during childhood for first-generation immigrants, because only those episodes experienced during residence in Denmark are captured in the nationwide administrative registers.

on immigrant status.13 Furthermore, the study registers provided less complete information about experience of trauma-related hospital admissions during childhood for first-generation immigrants, because only those episodes experienced during residence in Denmark are captured in the nationwide administrative registers. In this register-based study, consent to participate from cohort members was not needed. Cohort members were followed up between their 15th and 35th birthdays. Follow-up ended at the first occurrence of the adverse outcome of interest, emigration, death, or the final observation date of Dec 31, 2012, whichever date was earliest. Self-harm and violent criminal offending risks were assessed from the 15th birthday and onwards in the 1977–97 birth cohort. Therefore, the earliest outcome ascertainment date was Jan 1, 1992, and the latest was Dec 31, 2012, meaning that only those cohort members born during 1977 could provide complete follow-up information through to their 35th birthdays. Approval to conduct the study was given by the Danish Data Protection Agency, and data access was granted by the State Serum Institute and by Statistics Denmark.

In this register-based study, consent to participate from cohort members was not needed. Cohort members were followed up between their 15th and 35th birthdays. Follow-up ended at the first occurrence of the adverse outcome of interest, emigration, death, or the final observation date of Dec 31, 2012, whichever date was earliest. Self-harm and violent criminal offending risks were assessed from the 15th birthday and onwards in the 1977–97 birth cohort. Therefore, the earliest outcome ascertainment date was Jan 1, 1992, and the latest was Dec 31, 2012, meaning that only those cohort members born during 1977 could provide complete follow-up information through to their 35th birthdays. Approval to conduct the study was given by the Danish Data Protection Agency, and data access was granted by the State Serum Institute and by Statistics Denmark. Procedures Classification of trauma-related hospital admissions during childhood Hospital admissions for injuries or poisonings between cohort members' births and their 15th birthdays were identified on the basis of “reason for contact” coding recorded in the National Patient Register14 and according to the 8th15 and 10th16 revisions of the International Classification of Diseases (ICD) as recorded in the National Patient Register and Psychiatric Central Research Register,17 as follows: (1) self-harm (“reason for contact”=4; complex ICD-based algorithm published previously);18 (2) interpersonal violence (“reason for contact”=3; ICD-8: E960-E969; ICD-10: X85-Y09); and (3) accident (“reason for contact”=2; ICD-8: E800-E949; ICD-10: V01-X59). In delineating individuals who were admitted to hospital following self-harm during childhood, we only included episodes in which the child had reached their tenth birthday on the admission date. Emergency room and ambulatory care visits could not be included in any of the exposure classifications because they were only captured in the two hospital registers from 1995. Thus, the stringent exposure classification included only infants and children who were admitted to hospital, thereby providing a proxy for more serious episodes that would usually constitute a more traumatic experience for affected children.

ions because they were only captured in the two hospital registers from 1995. Thus, the stringent exposure classification included only infants and children who were admitted to hospital, thereby providing a proxy for more serious episodes that would usually constitute a more traumatic experience for affected children. Ascertainment of adverse outcomes between the 15th and 35th birthdays To ascertain admissions following self-harm occurring beyond the 15th birthday, we used the aforementioned previously reported complex algorithm18 that entailed linkage to and extraction of data from the Psychiatric Central Research Register and the National Patient Register. The study cohort was also linked to the National Crime Register19 to identify all violent crime convictions, including homicide, assault, robbery, aggravated burglary or arson, possessing a weapon in a public place, violent threats, extortion, human trafficking, abduction, kidnapping, rioting and other public order offenses, terrorism, and sexual offences.

e National Crime Register19 to identify all violent crime convictions, including homicide, assault, robbery, aggravated burglary or arson, possessing a weapon in a public place, violent threats, extortion, human trafficking, abduction, kidnapping, rioting and other public order offenses, terrorism, and sexual offences. Adjustment for parental SES This potentially important confounding influence was measured for the entire study population at cohort members' 15th birthdays according to income quintile, highest educational attainment level (primary school, high school or vocational training, or higher education) and employment status (employed, unemployed, or outside the workforce for other reasons). As a sensitivity analysis, we compared models adjusted for parental SES measured at ages 5 years, 10 years, and 15 years versus those adjusted for parental SES measured at age 15 years only. These data were extracted from the Integrated Database for Labour Market Research.20

or outside the workforce for other reasons). As a sensitivity analysis, we compared models adjusted for parental SES measured at ages 5 years, 10 years, and 15 years versus those adjusted for parental SES measured at age 15 years only. These data were extracted from the Integrated Database for Labour Market Research.20 Statistical analysis Data were analysed with SAS statistical software version 9.4 (SAS Institute Inc, Cary, NC, USA). Sex- specific analyses are reported throughout this report due to substantial differences that we have previously shown between male and female incidence rates for self-harm and violent criminal offending among Danish youths and young adults.21 Incidence rate ratios (IRRs) were estimated by fitting log-linear Poisson regression models22 adjusted for age group and calendar year as time-dependent variables. Parental SES was adjusted for using time-fixed variables. The outcomes, self-harm and violent crime conviction, were examined separately according to time to first event since the 15th birthday. Thus, the estimated IRRs pertain to these first events only as repeat events were not assessed. The reference category for all IRRs reported was people who did not experience a hospital admission for injury or poisoning between their birth and 15th birthday. p values and 95% CIs were calculated from likelihood ratio tests. From competing risks survival analysis,23 cumulative incidence (absolute risk) was calculated as the probability of experiencing the specific outcome of interest, taking into account emigration or death. As with the IRR estimates, these cumulative incidence values were calculated separately for each childhood hospital admission exposure type: self-harm, interpersonal violence, or accident.

risk) was calculated as the probability of experiencing the specific outcome of interest, taking into account emigration or death. As with the IRR estimates, these cumulative incidence values were calculated separately for each childhood hospital admission exposure type: self-harm, interpersonal violence, or accident. Role of the funding source The funder of the study had no role in the study design, data collection, data analysis, and data interpretation, or writing of the report. The corresponding author had full access to the data and had final responsibility to submit for publication.

risk) was calculated as the probability of experiencing the specific outcome of interest, taking into account emigration or death. As with the IRR estimates, these cumulative incidence values were calculated separately for each childhood hospital admission exposure type: self-harm, interpersonal violence, or accident. Role of the funding source The funder of the study had no role in the study design, data collection, data analysis, and data interpretation, or writing of the report. The corresponding author had full access to the data and had final responsibility to submit for publication. Results 1 087 672 Danish people were enrolled in this study. Table 1 shows the sex-specific prevalence values for trauma-related hospital admissions between cohort members' births and their 15th birthdays, with estimates reported separately by reason for hospital admission. The prevalence of any trauma-related hospital admission was 10% (105 753 per 1 087 672; men: 64 454 [11%]; women: 44 299 [8%]) and for both sexes, accident was by far the most prevalent of the categories assessed (men: 59 011 [11%]; women: 40 756 [8%]. All estimated prevalence values were substantially greater for men than for women, as indicated by the male to female prevalence rate ratios (PRRs), except for hospital admission following self-harm (PRR 0·58 [95% CI 0·55–0·61]). The rarest reason for hospital admission was interpersonal violence, and this category also had the largest observed male:female PRR (1·79 [95% CI 1·59–2·02]).Table 1 Prevalence and male to female prevalence rate ratios by reason for trauma-related hospital admission before the 15th birthday

R 0·58 [95% CI 0·55–0·61]). The rarest reason for hospital admission was interpersonal violence, and this category also had the largest observed male:female PRR (1·79 [95% CI 1·59–2·02]).Table 1 Prevalence and male to female prevalence rate ratios by reason for trauma-related hospital admission before the 15th birthday Prevalence in men (n=557 976) Prevalence in women (n=529 696) Male: female prevalence Self-harm 2371 (<1%) 3896 (1%) 0·58 (0·55–0·61) Interpersonal violence 781 (<1%) 414 (<1%) 1·79 (1·59–2·02) Accident 59 011 (11 %) 40 756 (8%) 1·37 (1·36–1·39) Any trauma-related hospital admission 61 454 (11%) 44 299 (8%) 1·32 (1·30–1·33) Data are n (%) or rate ratio (95% CI).

women (n=529 696) Male: female prevalence Self-harm 2371 (<1%) 3896 (1%) 0·58 (0·55–0·61) Interpersonal violence 781 (<1%) 414 (<1%) 1·79 (1·59–2·02) Accident 59 011 (11 %) 40 756 (8%) 1·37 (1·36–1·39) Any trauma-related hospital admission 61 454 (11%) 44 299 (8%) 1·32 (1·30–1·33) Data are n (%) or rate ratio (95% CI). Table 2 shows the IRRs for adverse outcomes at ages 15–35 years linked with trauma-related hospital admission at least once during childhood. In men and women, there was a significant increase in risk for both adverse outcomes. Essentially the same patterns of increased risk were observed in both sexes, although the IRRs were consistently and significantly larger in women (self-harm: IRR 1·94 [95% CI 1·85–2·02]; violent criminality: 2·16 [1·97–2·36]) than in men (self-harm: 1·61 [1·53–1·69]; violent criminality: 1·58 [1·53–1·63]). The observed associations were only slightly attenuated following adjustment for parental SES measured at cohort members' 15th birthdays. Additional adjustment for parental SES measured at ages 5 years, 10 years, and 15 years, compared with adjustment for parental SES measured at age 15 years only, made almost no difference to the degree of attenuation observed (data not shown).Table 2 Incidence rate ratios (IRRs) for self-harm and violent criminality at ages 15–35 years, linked with any trauma-related hospital admission before 15th birthday

compared with adjustment for parental SES measured at age 15 years only, made almost no difference to the degree of attenuation observed (data not shown).Table 2 Incidence rate ratios (IRRs) for self-harm and violent criminality at ages 15–35 years, linked with any trauma-related hospital admission before 15th birthday People with adverse outcome (n) Incidence rate per 10 000 person-years IRR 1*(95% CI) IRR 2†(95% CI) Men Self-harm Any trauma-related hospital admission 1806 30·9 1·61 (1·53–1·69) 1·50 (1·42–1·58) No trauma-related hospital admission (ref) 8871 18·5 1·00 1·00 Violent crime conviction Any trauma-related hospital admission 4306 76·0 1·58 (1·53–1·63) 1·49 (1·44–1·54) No trauma-related hospital admission (ref) 21 662 46·0 1·00 1·00 Women Self-harm Any trauma-related hospital admission 2343 58·8 1·94 (1·85–2·02) 1·85 (1·77–1·94) No trauma-related hospital admission (ref) 13 028 28·4 1·00 1·00 Violent crime conviction Any trauma-related hospital admission 579 14·1 2·16 (1·97–2·36) 1·91 (1·73–2·10) No trauma-related hospital admission (ref) 2815 6·0 1·00 1·00 * Adjusted for age band and calendar year period. † Adjusted for age band and calendar year period, and also for parental socioeconomic status measured at the 15th birthday.

People with adverse outcome (n) Incidence rate per 10 000 person-years IRR 1*(95% CI) IRR 2†(95% CI) Men Self-harm Any trauma-related hospital admission 1806 30·9 1·61 (1·53–1·69) 1·50 (1·42–1·58) No trauma-related hospital admission (ref) 8871 18·5 1·00 1·00 Violent crime conviction Any trauma-related hospital admission 4306 76·0 1·58 (1·53–1·63) 1·49 (1·44–1·54) No trauma-related hospital admission (ref) 21 662 46·0 1·00 1·00 Women Self-harm Any trauma-related hospital admission 2343 58·8 1·94 (1·85–2·02) 1·85 (1·77–1·94) No trauma-related hospital admission (ref) 13 028 28·4 1·00 1·00 Violent crime conviction Any trauma-related hospital admission 579 14·1 2·16 (1·97–2·36) 1·91 (1·73–2·10) No trauma-related hospital admission (ref) 2815 6·0 1·00 1·00 * Adjusted for age band and calendar year period. † Adjusted for age band and calendar year period, and also for parental socioeconomic status measured at the 15th birthday. The IRRs presented in table 3 are reported according to reason for hospital admission during childhood (self-harm, interpersonal violence, or accident) in relation to the two adverse outcomes at ages 15–35 years: self-harm and violent criminality. Hospital admission during childhood following episodes of self-harm or interpersonal violence was strongly associated with later risks of self-harm and violent offending (table 3). Much weaker links were seen between childhood hospital admission following an accident and the two adverse outcomes through older adolescence and young adulthood (table 3). Again, the patterns of increased risk as indicated by the IRRs were similar for both sexes, although stronger associations were noted in women than in men. The increase in violent offending risk was almost ten-fold among women who experienced hospital admission following interpersonal violence during childhood (IRR 9·85 [95% CI 6·67–13·92]).Table 3 Incidence rate ratios (IRRs) and cumulative incidence values (%) for self-harm and violent offending at ages 15–35 by reason for trauma-related hospital admission before the 15th birthday

ng women who experienced hospital admission following interpersonal violence during childhood (IRR 9·85 [95% CI 6·67–13·92]).Table 3 Incidence rate ratios (IRRs) and cumulative incidence values (%) for self-harm and violent offending at ages 15–35 by reason for trauma-related hospital admission before the 15th birthday People with adverse outcome (n) Incidence rate per 10 000 person-years IRR*(95% CI) Cumulative incidence†(95% CI) Men Self-harm‡ Self-harm§ 110 50·8 2·24 (1·84–2·69) 7·5% (5·7–9·7) Interpersonal violence 54 65·6 3·14 (2·37–4·05) 9·4% (7·1–12·0) Accident 1688 30·1 1·54 (1·46–1·62) 5·0% (4·6–5·3) No trauma-related hospital admission (ref) 8871 18·5 1·00 3·2% (3·1–3·2) Violent crime conviction Self-harm 230 110·3 1·92 (1·69–2·19) 13·7% (11·7–15·8) Interpersonal violence 149 196·5 3·82 (3·24–4·47) 25·0% (21·2–28·9) Accident 4012 73·8 1·50 (1·45–1·56) 10·4% (10·0–10·8) No trauma-related hospital admission (ref) 21 662 46·0 1·00 6·8% (6·7–6·9) Women Self-harm‡ Self-harm§ 657 248·0 6·60 (6·10–7·14) 21·4% (19·8–23·1) Interpersonal violence 54 152·2 4·16 (3·14–5·37) 18·3% (13·5–23·6) Accident 1789 47·9 1·47 (1·40–1·55) 6·5% (6·1–7·0) No trauma-related hospitalisation (ref) 13 028 28·4 1·00 4·0% (4·0–4·1) Violent crime conviction Self-harm 163 55·4 6·41 (5·45–7·48) 6·4% (5·2–7·8) Interpersonal violence 29 77·8 9·85 (6·67–13·92) 8·4% (5·7–11·6) Accident 430 11·2 1·58 (1·42–1·74) 1·5% (1·3–1·6) No trauma-related hospital admission (ref) 2815 6·0 1·00 0·9% (0·9–0·9) * Adjusted for age band and calendar year period.

·0–4·1) Violent crime conviction Self-harm 163 55·4 6·41 (5·45–7·48) 6·4% (5·2–7·8) Interpersonal violence 29 77·8 9·85 (6·67–13·92) 8·4% (5·7–11·6) Accident 430 11·2 1·58 (1·42–1·74) 1·5% (1·3–1·6) No trauma-related hospital admission (ref) 2815 6·0 1·00 0·9% (0·9–0·9) * Adjusted for age band and calendar year period. † Measures the probability or risk of experiencing the outcome of interest by the 35th birthday. ‡ Self harm as an outcome (15–35 years). § Self harm as a reason for hospital admission (before age 15 years).

·0–4·1) Violent crime conviction Self-harm 163 55·4 6·41 (5·45–7·48) 6·4% (5·2–7·8) Interpersonal violence 29 77·8 9·85 (6·67–13·92) 8·4% (5·7–11·6) Accident 430 11·2 1·58 (1·42–1·74) 1·5% (1·3–1·6) No trauma-related hospital admission (ref) 2815 6·0 1·00 0·9% (0·9–0·9) * Adjusted for age band and calendar year period. † Measures the probability or risk of experiencing the outcome of interest by the 35th birthday. ‡ Self harm as an outcome (15–35 years). § Self harm as a reason for hospital admission (before age 15 years). The cumulative incidence percentage values shown in table 2 indicate that: (1) for young adult men, the highest absolute risk noted was for violent offending among individuals admitted to hospital for interpersonal violence injury during childhood (cumulative incidence 25·0% [95% CI 21·2–28·9]); (2) for young adult women, absolute risk was highest for subsequent self-harm repetition among those admitted to hospital following self-harm in childhood (cumulative incidence 21·4% [95% CI 19·8–23·1]). Figure 1 shows the substantial sex differences in cumulative incidence (absolute risk) across the 15–35 years age range. Among men admitted to hospital during childhood following interpersonal violence, about a fifth will have committed a violent crime by their 25th birthday and a quarter will have done so by age 35 years (figure 1C). Among women admitted to hospital during childhood after self-harm, almost a fifth will be admitted to hospital again due to self-harm repetition by age 25 years (figure 1B).Figure 1 Cumulative incidence (%) for self-harm (men, A; women, B) and violent crime conviction (men, C; women, D) at ages 15–35 years by reason for trauma-related hospital admission before the 15th birthday

-harm, almost a fifth will be admitted to hospital again due to self-harm repetition by age 25 years (figure 1B).Figure 1 Cumulative incidence (%) for self-harm (men, A; women, B) and violent crime conviction (men, C; women, D) at ages 15–35 years by reason for trauma-related hospital admission before the 15th birthday Figure 2 shows the IRRs for later self-harm and violent offending reported separately according to the number of times a cohort member experienced a trauma-related hospital admission during childhood versus the reference category of zero trauma-related hospital admissions. The prevalence of multiple hospital admissions was low in the study cohort (men: admitted twice, n=5528 [1%]; admitted 3 times or more, n=742 [<1%]; women: twice, n=3434 [1%]; 3 times or more, n=427 [<1%]. For both adverse outcomes, we noted an incremental increase in the observed IRR with rising frequency of trauma-related hospital admissions experienced in childhood, and these dose–response relations were much stronger in women than in men: three or more trauma-related hospital admissions (vs none) linked with later self-harm risk (women: IRR 7·40 [95% CI 5·87–9·17]; men: 2·54 [1·78–3·50]); three or more trauma-related hospital admissions (vs none) linked with later violent criminality risk (women: IRR 11·02 [95% CI 7·41–15·65]; men: 2·57 [2·05–3·16]). Similarly, figure 3 shows IRRs for the two adverse outcomes according to the number of trauma-related hospital admission types during childhood (self-harm, interpersonal violence, or accident). Being admitted to hospital for more than one reason, with comparable exposure prevalence of 0·13% (n=742) in men versus 0·14% (n=765) in women, was associated with substantially greater risks of later self-harm and violent offending compared with one reason only. As with hospital admission frequency, the risk gradients were again far steeper in women than in men: two or three trauma-related hospital admission types (vs none) linked with later self-harm risk (women: IRR 9·18 [95% CI 7·81–10·71]; men: 3·73 [2·75–4·92]); two or three trauma-related hospital admission types (vs none) linked with later violent criminality risk (women: IRR 9·99 [95% CI 7·27– 13·31]; men: 2·80 [2·24–3·44]).Figure 2 Incidence rate ratios (IRRs)* for self-harm and violent offending at ages 15–35 years according to frequency of trauma-related hospital admissions before the 15th birthday

hospital admission types (vs none) linked with later violent criminality risk (women: IRR 9·99 [95% CI 7·27– 13·31]; men: 2·80 [2·24–3·44]).Figure 2 Incidence rate ratios (IRRs)* for self-harm and violent offending at ages 15–35 years according to frequency of trauma-related hospital admissions before the 15th birthday *Reference category for IRR estimates: cohort members who did not experience trauma-related hospital admission before their 15th birthday; IRRs adjusted for age band and calendar year period. Figure 3 Incidence rate ratios (IRRs)* for self-harm and violent offending at 15–35 years according to number of trauma-related hospitalisation types (self-harm, interpersonal violence, or accident) before the 15th birthday *Reference category for IRR estimates: cohort members who did not experience trauma-related hospital admission before their 15th birthday; IRRs adjusted for age band and calendar year period.

Figure 3 Incidence rate ratios (IRRs)* for self-harm and violent offending at 15–35 years according to number of trauma-related hospitalisation types (self-harm, interpersonal violence, or accident) before the 15th birthday *Reference category for IRR estimates: cohort members who did not experience trauma-related hospital admission before their 15th birthday; IRRs adjusted for age band and calendar year period. Discussion Incidence rates and cumulative incidence values for self-harm and violent offending were significantly raised among youths and young adults who experienced a trauma-related hospital admission at least once during their childhood. Confounding by parental SES explained very little of these increased risks. In men and women, individuals who were admitted to hospital during childhood following self-harm or interpersonal violence had substantially increased risks for later self-harm and violent criminal offending at ages 15–35 years. Around one in seven men admitted to hospital during childhood following self-harm and one in four due to interpersonal violence will be convicted for committing a violent crime by age 35 years. About one in five women admitted to hospital after an episode of self-harm or interpersonal violence during childhood will be admitted to hospital again following self-harm between their 15th and 35th birthdays. As we expected, the increases in risk linked with hospital admission as a consequence of an accident were far smaller than they were following self-harm or interpersonal violence, probably reflecting a lesser degree of trauma on average and perhaps also lower prevalence of environmental and genetic risk factors among individuals who experienced these episodes. More frequent trauma-related hospital admissions during childhood, and experiencing hospital admissions for multiple types of trauma at such an early age, conferred substantial increases in self-harm and violent criminality risks through late adolescence and young adulthood. Especially steep gradients of this nature were noted among female cohort members.

admissions during childhood, and experiencing hospital admissions for multiple types of trauma at such an early age, conferred substantial increases in self-harm and violent criminality risks through late adolescence and young adulthood. Especially steep gradients of this nature were noted among female cohort members. Our review of the existing published literature indicated that this topic has not been extensively researched, a conclusion that has also been reached by other investigators.8 The findings generated from previous studies do, however, largely concur with what we observed. For example, studies done in New Zealand7 and Sweden8 also showed that history of hospital admission for injury caused by interpersonal violence, as well as previous hospital-treated episodes of self-harm, predict future increased risk of suicidality. Earlier research has reported on subsequent risk of assaultive injury in individuals with a history of hospital admission following self-injury and assault,6 but our cohort study is the first to examine links between hospital-treated self-harm occurring before mid-adolescence and later risk of perpetrating violent criminal offence. Consideration of sex-specific absolute risk via cumulative incidence estimates that accounted for competing risks is another distinctive feature of our investigation. Little robust epidemiological evidence exists regarding confounding by familial SES, but the apparent little influence of this phenomenon that we observed concurs with findings reported by a previous national registry study examining links between hospital admission for injury and later suicide risk among Swedish youths.8

tle robust epidemiological evidence exists regarding confounding by familial SES, but the apparent little influence of this phenomenon that we observed concurs with findings reported by a previous national registry study examining links between hospital admission for injury and later suicide risk among Swedish youths.8 We believe that the findings generated from this study are likely to be generalisable internationally. Our national cohort study had some major strengths including comprehensive record linkage between multiple registers, absence of recall bias and other sources of information bias, capacity to account for all deaths and emigrations throughout follow-up, nationwide coverage of the registry datasets, and abundant statistical power and precision to examine fairly rare adverse events in a study cohort consisting of more than 1 million people. Thus, for example, our study was entirely free of the linkage errors and resulting selection biases that were reported by a previous study of this topic done using Hospital Episode Statistics in England.3 A specific strength of our register-based cohort study was that we could delineate childhood trauma-related hospital admissions to a clinically significant level of severity, because the exposure classification included only those individuals who were actually admitted to hospital. We were also unable to examine emergency department contacts following trauma without admission in this study. However, this should not be regarded as an omission, because such an approach would have identified a much larger set of relatively minor accidental injuries to children, most of whom would be unlikely to have increased risk of experiencing the longer-term adverse outcomes we examined beyond their 15th birthdays.

tudy. However, this should not be regarded as an omission, because such an approach would have identified a much larger set of relatively minor accidental injuries to children, most of whom would be unlikely to have increased risk of experiencing the longer-term adverse outcomes we examined beyond their 15th birthdays. A widely recognised generic limitation of registry studies is residual confounding.24 Thus, adjustment for parental SES took account of only three parameters, and there are many other important determinants of risk, including child abuse and neglect, bullying by peers, and household dysfunction,25 that were not measured systematically in the administrative registers that we had access to. Studies of this particular topic share a common limitation that a sizeable proportion of violence and self-harm are undetected at hospital admission and might frequently be misclassified as accidents.4 A limitation specific to our cohort study was that we could not examine trauma-related hospital admissions that were of undetermined cause, which was possible in studies done using routinely collected datasets in other countries such as New Zealand6, 7 and the USA.4 This “reason for hospitalisation” category was unavailable to us because of anomalies in registration procedures and absence of consistency in ICD 8th and 10th revision coding usage in the hospital registers across the whole of the study's observation period. Individuals admitted for injuries of undetermined cause have been reported to show substantially increased risk for subsequent suicidal behaviour.7 The exclusion of this group might have attenuated some of the relative risk estimates we observed, specifically in table 1 and in Figure 2, Figure 3, albeit only to a marginal degree because this is likely to be an exposure subgroup with rare prevalence. Finally, in this study we did not examine the potential influences of psychopathology among cohort members and also their parents' mental disorders. The latter could have important roles in the cause of trauma-related hospital admission during childhood26 as well as self-harm and interpersonal violence during older adolescence and early adulthood.27 A comprehensive investigation of these potential causal mechanisms would require a complex study design and a sophisticated analytical approach, ideally with linkage to both primary and secondary care data sources to enable capture of all diagnosed mental illnesses among cohort members and their parents.

y adulthood.27 A comprehensive investigation of these potential causal mechanisms would require a complex study design and a sophisticated analytical approach, ideally with linkage to both primary and secondary care data sources to enable capture of all diagnosed mental illnesses among cohort members and their parents. Trauma-related hospital admission during childhood could be a useful marker for myriad forms of familial dysfunction and distress that in later years promote emotional dysregulation and impulsive, self-destructive reactions to stress and adversity in adulthood.28 In developing multiagency strategies to tackle internalised and externalised violence among susceptible youths and young adults, the wellbeing of individuals who experienced hospital admission during childhood for self-harm or interpersonal violence merits especially close attention. High school and college-oriented programmes might be particularly beneficial in this regard.29 National clinical guidelines typically address single problems such as self-harm,30 but might need to be broadened to encompass other types of adverse events to encourage family-level psychosocial assessments following hospital admission of children due to serious episodes of interpersonal violence3 and accidental injury or poisoning caused by apparent parental neglect. Childhood hospital admissions following trauma might therefore represent crucially important opportunities for hospital-based family-oriented interventions aimed at reducing future risk of accidental, self-inflicted and assaultive injuries, and poisonings.

dental injury or poisoning caused by apparent parental neglect. Childhood hospital admissions following trauma might therefore represent crucially important opportunities for hospital-based family-oriented interventions aimed at reducing future risk of accidental, self-inflicted and assaultive injuries, and poisonings. To conclude, young people who harm themselves or who are aggressive or violent towards others have often experienced trauma during their childhoods, and therefore they should be treated sympathetically rather than as problematic and undeserving. Among those individuals who experienced trauma-related hospital admission during their childhood, risks of adverse outcome are increased, irrespective of parental socioeconomic status. Therefore, a more psychosocial approach to meet the needs of these vulnerable young people and their families is indicated for successful prevention of future episodes of harmful behaviour. The novelty of this study lies in its examination of two associated destructive behaviours for an under-researched exposure-outcome association that spans two adjacent periods in the life-course that are crucially important in an individual's development through to healthy adult maturity. Reporting of sex-specific absolute and relative risks also breaks new ground for this research topic—an approach that maximises the relevance and use of the findings for clinicians, public health professionals, and policy makers Acknowledgments This study was supported by a European Research Council starting grant awarded to RTW (ref no 335905).

To conclude, young people who harm themselves or who are aggressive or violent towards others have often experienced trauma during their childhoods, and therefore they should be treated sympathetically rather than as problematic and undeserving. Among those individuals who experienced trauma-related hospital admission during their childhood, risks of adverse outcome are increased, irrespective of parental socioeconomic status. Therefore, a more psychosocial approach to meet the needs of these vulnerable young people and their families is indicated for successful prevention of future episodes of harmful behaviour. The novelty of this study lies in its examination of two associated destructive behaviours for an under-researched exposure-outcome association that spans two adjacent periods in the life-course that are crucially important in an individual's development through to healthy adult maturity. Reporting of sex-specific absolute and relative risks also breaks new ground for this research topic—an approach that maximises the relevance and use of the findings for clinicians, public health professionals, and policy makers Acknowledgments This study was supported by a European Research Council starting grant awarded to RTW (ref no 335905). Contributors SA, MJC, PLHM, CBP, and RTW designed and conceived the study. All authors collected, analysed, and interpreted the data, and reviewed the report. RTW drafted the report. SA did the statistical analysis. RTW obtained the funding. RTW and SA provided administrative and technical support.

Acknowledgments This study was supported by a European Research Council starting grant awarded to RTW (ref no 335905). Contributors SA, MJC, PLHM, CBP, and RTW designed and conceived the study. All authors collected, analysed, and interpreted the data, and reviewed the report. RTW drafted the report. SA did the statistical analysis. RTW obtained the funding. RTW and SA provided administrative and technical support. Declaration of interests We declare no competing interests.

Introduction Respiratory syncytial virus (RSV) is a highly seasonal respiratory virus (the season runs from from late autumn to early spring).1 Exposure to RSV does not lead to long-lasting protection and hence people can have many infections over their lifetime.2 Infection mainly leads to mild disease, but in very young children (aged <6 months), elderly people, and immunocompromised patients it can result in serious disease or death.3 Currently, the only effective preventive strategy against RSV is passive immunisation with palivizumab, a humanised monoclonal RSV-specific antibody. Because of its high price, this antibody is only used in the highest-risk groups of individuals during the RSV season (November to February)—usually young children who are born prematurely and have other respiratory or cardiac conditions.4, 5 However, around 60 RSV vaccine and monoclonal antibody candidates are in development, 16 of which are in clinical trials,6, 7 although trial results in adults aged 60 years and older for the most advanced vaccine candidate (Resolve RSV-F vaccine) have not shown efficacy.8 Besides older adults, other potential candidates are pregnant women (to protect newborn babies through passive immunity), newborn babies (through passive immunisation with antibodies), and infants. An RSV vaccine could possibly be licensed in the next 5–10 years.9 Additionally, at least one extended, half-life monoclonal antibody designed to protect infants from birth, along with at least three maternal vaccines, are in clinical trials.6

newborn babies (through passive immunisation with antibodies), and infants. An RSV vaccine could possibly be licensed in the next 5–10 years.9 Additionally, at least one extended, half-life monoclonal antibody designed to protect infants from birth, along with at least three maternal vaccines, are in clinical trials.6 Decision makers will need to understand the potential health and economic effects of the different vaccine and antibody options to select strategies that maximise the effect of health-care resources. Although the exact characteristics of future maternal or infant vaccines or prophylactic antibodies for newborn babies are unknown, understanding the burden of RSV disease and the drivers of vaccine effects and value can help to inform decisions about prioritisation of vaccination or antibody strategies, and protocols for clinical trials.7 Such analyses can also identify and help to ensure that data are obtained in advance about the key drivers of cost-effectiveness. Research in context Evidence before this study

Decision makers will need to understand the potential health and economic effects of the different vaccine and antibody options to select strategies that maximise the effect of health-care resources. Although the exact characteristics of future maternal or infant vaccines or prophylactic antibodies for newborn babies are unknown, understanding the burden of RSV disease and the drivers of vaccine effects and value can help to inform decisions about prioritisation of vaccination or antibody strategies, and protocols for clinical trials.7 Such analyses can also identify and help to ensure that data are obtained in advance about the key drivers of cost-effectiveness. Research in context Evidence before this study Respiratory syncytial virus (RSV) disease is the primary contributor to childhood lower respiratory tract infections. More than 60 biological candidates for RSV prophylaxis (vaccines and prophylactic monoclonal and polyclonal antibodies) are undergoing development, of which more than 25% have progressed to human trials, and one or more is likely to be licenced in the next 5–10 years. The candidates target different patient populations, and the optimum prophylactic strategy is yet to be determined. We did a search of the scientific literature, based on expert opinions. Despite some previous studies separately assessing the incidence of RSV-attributable clinical disease, and the economic impact of vaccination, as yet there have been no studies that combine this information, and few published studies can be used by decision-making bodies to assess the cost-effectiveness of different RSV vaccination strategies.

dies separately assessing the incidence of RSV-attributable clinical disease, and the economic impact of vaccination, as yet there have been no studies that combine this information, and few published studies can be used by decision-making bodies to assess the cost-effectiveness of different RSV vaccination strategies. Added value of this study We used data from laboratory reports and on health-care attendances for acute respiratory illness to estimate disease burden and health-care costs associated with RSV in England (UK). The estimates agreed with those from previous studies, while providing greater insight into the timing of outbreaks and ages most affected. We present the first quantitative analysis to highlight how the month of birth affects RSV-attributable health-care outcomes in a temperate climate. We then assessed the effect and cost-effectiveness of various vaccine and antibody strategies in pregnant women and young children. We showed that children born immediately before the RSV season, which runs from late autumn to early spring, have a two-fold higher risk of primary-care attendance and a four-fold higher risk of being admitted to hospital than children born after the season. Implications of all the available evidence Given the difference in these risks between children born before and after RSV season, the most cost-effective strategies, and ones that have the potential to avert the most severe disease and deaths, are those that protect children born just before the RSV season, such as maternal vaccination or long-lasting prophylactic monoclonal antibodies.

ese risks between children born before and after RSV season, the most cost-effective strategies, and ones that have the potential to avert the most severe disease and deaths, are those that protect children born just before the RSV season, such as maternal vaccination or long-lasting prophylactic monoclonal antibodies. So far, few studies are available to inform about the potential cost-effectiveness of different RSV vaccination strategies10 and the need for further cost-effectiveness information has been identified as a priority by WHO's Strategic Advisory Group of Experts on Immunisation.11 To help to address this need, we present a detailed analysis of the disease burden of RSV and the associated health-care costs in England (UK). We then used England as an example of a high-income country in the temperate zone that is considering RSV vaccination in the future. This allowed us to illustrate general principles and to explore the potential effect and cost-effectiveness of different vaccine and antibody strategies to protect young children in high income, temperate climates with a similar epidemiology to England.

temperate zone that is considering RSV vaccination in the future. This allowed us to illustrate general principles and to explore the potential effect and cost-effectiveness of different vaccine and antibody strategies to protect young children in high income, temperate climates with a similar epidemiology to England. Methods Disease burden estimation Most people who present to health-care services with respiratory symptoms are not routinely tested for RSV, so the incidence of primary care attendances and hospital admissions for RSV has to be inferred. We used a statistical regression model12 to ecologically link clinical attendances for acute respiratory infection to organisms detected in routine clinical microbiological testing, based on temporal trends in both datasets, similar to our work estimating the burden of seasonal influenza.12 Similar methods have been previously used to explore the burden of seasonal organisms, including RSV, influenza virus, and rotavirus.13, 14 We synthesised information from general practice attendances and hospital admissions for acute respiratory symptoms and positive laboratory reports for respiratory pathogens with data from the scientific literature15 to explore the detailed age distribution of these clinical attendances in children younger than 5 years. Full details of our methods are in the appendix (p 1).

nces and hospital admissions for acute respiratory symptoms and positive laboratory reports for respiratory pathogens with data from the scientific literature15 to explore the detailed age distribution of these clinical attendances in children younger than 5 years. Full details of our methods are in the appendix (p 1). Economic model We used a static cohort model (ie, a model that does not account for the indirect or herd effects of vaccination) to explore the potential direct effect of paediatric vaccination or long-lasting monoclonal antibody use on its recipient (appendix p 5). We used the results of the model to estimate the net cost and cost-effectiveness of the interventions. We estimated the maximum cost-effective price (MCEP) per fully protected individual that could be paid for both the purchase and the administration costs of a course of vaccines or prophylactic antibodies (including any required booster doses), so as not to exceed the threshold of £20 000 per quality-adjusted life-year (QALY) gained, which is commonly used as a measure of cost-effectiveness in England.16 This value is close to the UK's gross domestic product per capita, which has been suggested17 as a possible threshold to use for an intervention to be deemed very cost-effective. The maximum price payable for each fully vaccinated individual for a range of assumptions on vaccine efficacy is in the appendix (p 10). Further details including cost-related and health-related quality-of-life parameters are in table 1, and the appendix (p 9).Table 1 Base case parameters for the cost-effectiveness model and variations used in the sensitivity analysis

d individual for a range of assumptions on vaccine efficacy is in the appendix (p 10). Further details including cost-related and health-related quality-of-life parameters are in table 1, and the appendix (p 9).Table 1 Base case parameters for the cost-effectiveness model and variations used in the sensitivity analysis Base case value Lower limit Upper limit Vaccine efficacy* 70% 50% 100% Age of infant vaccine administration 3 months 2 months 4 months Duration of maternal antibody protection 3 months 2 months 4 months Duration of neonatal prophylactic antibody protection 6 months 4 months 8 months Multiplier for deaths 1·0 0 2·0 Multiplier for QALYs 1·0 0·8 1·2 Multiplier for costs 1·0 0·8 1·2 Discounting 3·5% 1·5% 3·5% QALY=quality-adjusted life-year. * Because there were no available data to inform on vaccine efficacy, we chose a mid-range efficacy value. Interventions We considered vaccination strategies that targeted either infants, pregnant women, or neonates. We assumed that neonates would be protected either through passive immunisation via maternal vaccination, anticipated to give 3 months' protection, or through an extended half-life monoclonal antibody administered to newborn infants and providing passive immunisation, anticipated to give 6 months' protection. We also considered the strategy of vaccinating neonates born in certain months of the year only (appendix p 7). Assumptions behind all vaccination strategies are in the appendix (p 7).

-life monoclonal antibody administered to newborn infants and providing passive immunisation, anticipated to give 6 months' protection. We also considered the strategy of vaccinating neonates born in certain months of the year only (appendix p 7). Assumptions behind all vaccination strategies are in the appendix (p 7). Statistical analysis We ran a probabilistic sensitivity analysis, varying the number of RSV-attributable cases, costs, and QALYs (table 2, appendix). 95% uncertainty intervals (UIs) are the result of 10 000 simulations. Additionally, to determine the sensitivity of our cost-effectiveness estimates to different model variables, we ran a sensitivity analysis, sequentially altering model parameters from the baseline values. We did all analyses with R version 3.2.2.Table 2 Estimated number of annual general practitioner (GP) consultations, admissions to hospital, and deaths in hospital attributable to respiratory syncytial virus in children younger than 5 years

vity analysis, sequentially altering model parameters from the baseline values. We did all analyses with R version 3.2.2.Table 2 Estimated number of annual general practitioner (GP) consultations, admissions to hospital, and deaths in hospital attributable to respiratory syncytial virus in children younger than 5 years Age less than 6 months Age 6 months to less than 5 years Age less than 5 years GP consultations 64 570 (63 700–65 430) 288 000 (284 230–291 770) 352 570 (348 700–356 440) Incidence per 100 population 21·42 (21·13–21·70) 10·92 (10·77–11·06) 11·99 (11·86–12·12) Admissions to hospital 13 250 (13 200–13 310) 13 150 (13 020–13 280) 26 400 (26 270–26 550) Incidence per 100 population 4·4 (4·38–4·41) 0·5 (0·49–0·50) 0·9 (0·89–0·90) GP consultations leading to hospital admissions 20·53% 4·57% 13·00% Deaths in hospital 7·49 (7·19–7·79) 17·39 (16·92–17·86) 24·88 (24·32–25·44) Incidence per 100 population 0·00248 (0·00238–0·00258) 0·00066 (0·00064–0·00068) 0·00085 (0·00083–0·00087) Data are n (95% CI), incidence per 100 population (95% CI), or n (%). 95% CIs were used to fit a normal distribution in the probabilistic sensitivity analysis. 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 (DC) had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Age less than 6 months Age 6 months to less than 5 years Age less than 5 years GP consultations 64 570 (63 700–65 430) 288 000 (284 230–291 770) 352 570 (348 700–356 440) Incidence per 100 population 21·42 (21·13–21·70) 10·92 (10·77–11·06) 11·99 (11·86–12·12) Admissions to hospital 13 250 (13 200–13 310) 13 150 (13 020–13 280) 26 400 (26 270–26 550) Incidence per 100 population 4·4 (4·38–4·41) 0·5 (0·49–0·50) 0·9 (0·89–0·90) GP consultations leading to hospital admissions 20·53% 4·57% 13·00% Deaths in hospital 7·49 (7·19–7·79) 17·39 (16·92–17·86) 24·88 (24·32–25·44) Incidence per 100 population 0·00248 (0·00238–0·00258) 0·00066 (0·00064–0·00068) 0·00085 (0·00083–0·00087) Data are n (95% CI), incidence per 100 population (95% CI), or n (%). 95% CIs were used to fit a normal distribution in the probabilistic sensitivity analysis. 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 (DC) had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Variations in baseline parameters are in table 1. Results from the regression model suggested that 352 570 (16%) of 2 217 400 acute respiratory general practitioner (GP) consultations and 26 400 (22%) of 122 100 admissions to hospital for acute respiratory conditions are attributable to RSV in children younger than 5 years (table 2). We estimated that RSV is responsible for around 12 primary care consultations (95% CI 11·86–12·12) and 0·9 admissions to hospital annually per 100 children younger than 5 years (95% CI 0·89–0·90), with the major burden occurring in infants younger than 6 months (table 2, figure 1). In children younger than 6 months, RSV accounted for more than half of all admissions to hospital for acute respiratory conditions and for more than 70% of those admissions occurring between October and January (table 2). We estimated that there were around 25 deaths (95% CI 24·32–25·44) in children younger than 5 years (table 2). Although proportionally, the number of RSV-attributable outcomes was higher in children younger than 6 months, in absolute terms most of the burden occurred in children aged 6 months to 5 years (table 2).Figure 1 Outcomes attributable to respiratory syncytial virus by month of birth

dren younger than 5 years (table 2). Although proportionally, the number of RSV-attributable outcomes was higher in children younger than 6 months, in absolute terms most of the burden occurred in children aged 6 months to 5 years (table 2).Figure 1 Outcomes attributable to respiratory syncytial virus by month of birth (A) General practitioner (GP) consultations, (B) admissions to hospital. RSV is extremely seasonal, with peaks of incidence in December and January that predominantly affected children younger than 6 months (appendix p 3). Children born in winter had more RSV-attributable GP consultations and admissions to hospital (figure 1), and a higher proportion of their primary care outcomes occurred when they were younger than 6 months (dark red shading in figure 1), similar to that previously reported for RSV-attributable laboratory reports.1 The incidence of RSV-attributable GP consultations in the first year of life fluctuated between 13·9 per 100 children born in March up to 27·4 per 100 children born in November (and similarly from 1·55 per 100 births to 6·47 per 100 births for admissions to hospital).

reported for RSV-attributable laboratory reports.1 The incidence of RSV-attributable GP consultations in the first year of life fluctuated between 13·9 per 100 children born in March up to 27·4 per 100 children born in November (and similarly from 1·55 per 100 births to 6·47 per 100 births for admissions to hospital). Health-care costs for RSV in children younger than 5 years are £54 million annually (95% UI 50 million–57 million) or £87·58 annually per child (82–93). Most of this cost (£37 million) resulted from RSV-attributable admissions to hospital (including admissions to intensive care, which are assumed to be more likely in preterm infants), and was split approximately equally between children younger than 6 months (£19·1 million [95% UI 18·7 million–19·4 million]) and children aged between 6 months and 5 years (£18·4 million [16·3 million–20·4 million]). RSV-attributable GP consultations cost the health-care service £16 million annually (95% UI 14 million–19 million), with £13 million (82%) of the total cost (£16 million) for older children and £3 million (18%) for children younger than 6 months. Almost 70% (3526 of 5221) of QALYs lost stemmed from events associated with GP consultations and almost 75% (3841 of 5221) of QALY losses were caused by outcomes in older children.

illion), with £13 million (82%) of the total cost (£16 million) for older children and £3 million (18%) for children younger than 6 months. Almost 70% (3526 of 5221) of QALYs lost stemmed from events associated with GP consultations and almost 75% (3841 of 5221) of QALY losses were caused by outcomes in older children. Results of vaccination are shown for a vaccine with 100% efficacy to present the maximum possible effect; however, the proportional effect from different strategies was the same regardless of the vaccine efficacy (figure 2). Most of the cost-savings from any strategy resulted from averted admissions to hospital and intensive care (62% using the infant strategy and 86–88% when protecting newborn babies; figure 2), despite most of the averted cases being in primary care (95%, 83%, and 80% for infant, newborn, and maternal strategies, respectively). QALY gains from all strategies were driven largely by averting GP consultations (73% of gains for an infant strategy and 47% and 43% for newborn and maternal strategies, respectively; figure 2). For all strategies, just over 10% of the QALYs gained were from avoidance of RSV-attributable deaths in hospital. With the same vaccine efficacy, a newborn strategy averted many more RSV-attributable outcomes in babies younger than 6 months than a programme for infants aged 3 months and older.Figure 2 Cases of respiratory syncytial virus averted and costs or QALYs saved for different vaccination strategies with complete vaccine efficacy

he same vaccine efficacy, a newborn strategy averted many more RSV-attributable outcomes in babies younger than 6 months than a programme for infants aged 3 months and older.Figure 2 Cases of respiratory syncytial virus averted and costs or QALYs saved for different vaccination strategies with complete vaccine efficacy Data given per 100 annual births for (A) general practitioner (GP) consultations averted,(B) hospital admissions averted, (C) deaths in hospital averted, (D) health-care costs saved, (E) QALYs saved, and (F) maximum cost-effective price (MCEP) of vaccination strategy. M=maternal immunisation strategy. N=newborn passive immunisation strategy. C=infant strategy, N+C=newborn and infant strategies. ICU=intensive-care unit. QALY= quality-adjusted life-year. In the base case, the maximum price payable per fully protected person that should be paid for infant, newborn, and maternal vaccination strategies without seasonal restrictions was £192 (95% UI 168–219), £81 (76–86), and £54 (51–57), respectively. The MCEP for a strategy that combined a newborn and infant programme was £246 (95% UI 219–275). However, if a newborn programme was already in place (hence reducing disease burden and thus the benefit of any further immunisation strategies), then the MCEP of an infant vaccine would drop from £192 to £165 (143–190). Likewise, if an infant programme was already in place, then the MCEP for a newborn strategy decreased from £81 to £54 (51–57).

amme was already in place (hence reducing disease burden and thus the benefit of any further immunisation strategies), then the MCEP of an infant vaccine would drop from £192 to £165 (143–190). Likewise, if an infant programme was already in place, then the MCEP for a newborn strategy decreased from £81 to £54 (51–57). Regardless of the actual cost of an immunisation strategy or its efficacy, because of the extreme seasonality of RSV, and its propensity to infect very young children, a strategy to protect newborn babies is most cost-effective if it is only administered during certain months of the year. In the UK, the most cost-effective strategy was to protect only neonates born in November (before the start of the RSV season; MCEP of £220 [95% UI 208–232] per fully protected newborn infant). We noted that nine of the top ten most cost-effective strategies involved restricting prophylaxis to neonates born in only 4 months (or fewer) of the year, and who were born before the peak in RSV incidence (figure 3).Figure 3 Ten most cost-effective periods over which to offer newborn vaccination or prophylactic antibodies against respiratory syncytial virus Red shading shows the birth months of the children that are most cost-effective to protect via a newborn strategy. Prices are per fully protected child for vaccination or prophylactic antibodies. All of the ten most cost-effective programmes involved protecting newborn infants for 5 months or less of the year, the top nine protected for 4 months or less, and the top eight recommended protecting children born in November.

strategy. Prices are per fully protected child for vaccination or prophylactic antibodies. All of the ten most cost-effective programmes involved protecting newborn infants for 5 months or less of the year, the top nine protected for 4 months or less, and the top eight recommended protecting children born in November. The model was most sensitive to vaccine efficacy, and for the maternal or newborn strategy, to the duration of vaccine protection (figure 4). It was difficult to predict at this stage the potential effect that an RSV vaccine or antibody might have on infection transmission and hence indirect (herd) benefits, so the results presented here provide a conservative lower bound of the maximum price to pay per protected person, in the absence of consideration of herd effects. To estimate the maximum benefit that might be conferred through indirect effects, we made the assumption that introduction of a vaccine would completely eliminate disease transmission and hence disease in all children younger than 5 years (including those too young to receive the vaccine). Under this assumption, the maximum price for a full course and administration of a vaccine was £246 (220–276).Figure 4 Sensitivity to model parameters of cost-effectiveness calculations for respiratory syncytial virus vaccination

e in all children younger than 5 years (including those too young to receive the vaccine). Under this assumption, the maximum price for a full course and administration of a vaccine was £246 (220–276).Figure 4 Sensitivity to model parameters of cost-effectiveness calculations for respiratory syncytial virus vaccination (A) Infant strategy. (B) Newborn infant strategy. (C) Maternal strategy. Bars show by how much the maximum cost-effective price changes from its base case level when model parameters are varied. Changing the discounting strategy, or excluding deaths from the model, had little effect on the maximum cost-effective price (MCEP) estimates. Similarly, modifying the costs or quality-adjusted life-years (QALYs) associated with each health-care outcome had little effect on model estimates.

model parameters are varied. Changing the discounting strategy, or excluding deaths from the model, had little effect on the maximum cost-effective price (MCEP) estimates. Similarly, modifying the costs or quality-adjusted life-years (QALYs) associated with each health-care outcome had little effect on model estimates. Discussion There is a large and costly RSV disease burden in children younger than 5 years, especially infants younger than 6 months (particularly for admissions to hospital) and in the winter months. Indeed, in children younger than 5 years RSV is responsible for nearly twice as many GP consultations and nearly five times as many admissions to hospital as influenza, for which paediatric vaccination has been found to be cost-effective.18 RSV accounts for more than 75% of infants admitted to hospitals for respiratory conditions between the beginning of November and the end of January, consistent with other studies showing RSV to be a leading cause of infant admissions to hospital.19 We have shown how RSV-attributable health-care outcomes vary based on month of birth, with children born just before the start of the RSV season having double the risk of an RSV-attributable GP consultation and more than a four-fold higher risk of an RSV-attributable admission to hospital in their first year of life than those born after the RSV season. The general trends and conclusions for England are likely to be similar in high-income temperate countries with similar epidemiology, widespread health services and existing, well supported vaccination programmes.

n RSV-attributable admission to hospital in their first year of life than those born after the RSV season. The general trends and conclusions for England are likely to be similar in high-income temperate countries with similar epidemiology, widespread health services and existing, well supported vaccination programmes. Our estimates for RSV-attributable admissions to hospital are in line with other UK estimates,13, 20, 21 and many reported in western Europe,22, 23 but lower than estimates reported for Spain (a full comparison is in the appendix p 8). Our findings about the marked seasonality of RSV disease agree with those from a study in England1 showing that infants born just before or during the RSV season had a much higher risk of laboratory-confirmed RSV in their first year than those born just after the RSV season.1 Around 20% of infants visited GPs for RSV-attributable consultations in the first 6 months of life, and a fifth of these were admitted to hospital. Although the burden of RSV decreased as infants matured, nearly half of all children aged 6 months to 5 years visited GPs for RSV-attributable respiratory illnesses, with 5% resulting in admissions to hospital. These estimates agree with those from a study13 based on a restricted regression analysis of UK data that only incorporated laboratory reports for influenza and RSV rather than all respiratory pathogens.

nths to 5 years visited GPs for RSV-attributable respiratory illnesses, with 5% resulting in admissions to hospital. These estimates agree with those from a study13 based on a restricted regression analysis of UK data that only incorporated laboratory reports for influenza and RSV rather than all respiratory pathogens. One of the limitations of our study was that our disease estimates were based on statistical models, similar to those used to understand the burden of other respiratory12, 13 and diarrhoeal diseases.14 Therefore the normal caveats apply, such as difference in sensitivity between tests, reporting bias, testing practices, and unattributable changes over time. However, our RSV model was based on previous work testing nine models incorporating adjustments suggested by others24, 25 on six different age groups and selected the best-fitting model.12 Our assumptions of the effects that RSV infection has on quality of life, although from a previous cost-effectiveness analysis,10 were based on expert opinion rather than data. Indeed, RSV-averted deaths through vaccination might have occurred in children who had a lower quality of life or shorter life expectancy than average because of other comorbidities, and this differential might make vaccination less cost-effective. Future cost-effectiveness studies could benefit from better understanding of the effect of RSV disease on quality of life in young children, more detailed information on RSV incidence by month of birth in children younger than 1 year, and more detailed information about disease in preterm infants. Our burden estimates were based on data from 2001 to 2008, and although they were in agreement with other estimates, all were done before the introduction of paediatric vaccination for influenza; thus, future studies are required to consider the implications of this policy change on RSV burden.

ease in preterm infants. Our burden estimates were based on data from 2001 to 2008, and although they were in agreement with other estimates, all were done before the introduction of paediatric vaccination for influenza; thus, future studies are required to consider the implications of this policy change on RSV burden. The exact effect of an RSV vaccine or monoclonal antibody depends heavily on the age at which it can be given, and the age profile of RSV disease burden in very young infants. We based our analysis on several existing studies which showed RSV-associated admissions to hospital peaked at around 2 months of age, and decreased thereafter.15, 26, 27 However, RSV probably induces more severe disease in younger age groups28 and therefore the age profile of milder disease might be different. Understanding both the age-distribution and seasonality of RSV disease is key to selecting the best preventive strategy; hence, further direct active surveillance is needed to get better estimates. Additionally, although the burden of RSV in low-income and middle-income countries is substantial,22 further work is needed to assess the effect of interventions in these settings because of differences in seasonality of disease, access to care, resources available to pay for interventions, and population comorbidities.

dditionally, although the burden of RSV in low-income and middle-income countries is substantial,22 further work is needed to assess the effect of interventions in these settings because of differences in seasonality of disease, access to care, resources available to pay for interventions, and population comorbidities. Since neither the mechanism of action or the efficacy of RSV-immunisation strategies for either newborn babies or infants are known, a transmission model was not used in this analysis, therefore herd effects that might protect infants too young to be vaccinated, other unvaccinated children, and older individuals could not be assessed. If an RSV vaccine can prevent transmission as well as disease, the vaccine is likely to be even more cost-effective than our analysis suggests, and the results of future clinical trials will be essential to determine vaccine efficacy for each strategy. Using an assumption of complete disease elimination in children younger than 5 years, we concluded that the maximum price payable for the full purchase and administration of an RSV-immunisation programme would be £244. Herd effects might thus render a year-round vaccination strategy more cost-effective than a seasonal one, since a seasonal strategy is unlikely to elicit these effects. Hence, once suitable data on vaccine mechanisms become available the cost-effectiveness should be reassessed using a dynamic model. Additionally, because of the uncertainties described above, and the uncertainty in vaccine price, we did not use the traditional approach of comparing the cost-effectiveness of different options incrementally to each other, since this would require knowing the cost of each option. Once further details of the vaccines become available, a full incremental cost-effectiveness analysis of all options together would be helpful.

not use the traditional approach of comparing the cost-effectiveness of different options incrementally to each other, since this would require knowing the cost of each option. Once further details of the vaccines become available, a full incremental cost-effectiveness analysis of all options together would be helpful. We did not consider the vaccination of children older than early infancy. Such a strategy would not directly protect infants at the age of highest disease incidence, but might have a large effect on disease by protecting the younger group through herd (indirect) protection. The most cost-effective strategy assessed was a seasonal strategy that protected children who are born just before the RSV season from birth for the first few months of life. The exact month of vaccination should be determined on the basis of the epidemiology of each country. Such a seasonal vaccination strategy would probably only be feasible for a single-dose immunisation strategy, either given to the mother before birth, or to a child in the first few weeks of life. This suggests that efforts focused on developing an efficacious maternal vaccine, or a birth dose of a long-lasting monoclonal antibody, and on investigating the potential for vertical protection are well placed. Single-dose prophylactic antibodies have completed phase 1 trials in adults29 and are in phase 1b and 2a clinical trials in infants.6

focused on developing an efficacious maternal vaccine, or a birth dose of a long-lasting monoclonal antibody, and on investigating the potential for vertical protection are well placed. Single-dose prophylactic antibodies have completed phase 1 trials in adults29 and are in phase 1b and 2a clinical trials in infants.6 Under certain conditions, protecting older infants would be more cost-effective than protecting neonates. However, these conditions only hold under optimistic assumptions about an infant vaccine—ie, that it will confer full protection from age 6 months to 5 years compared with the rapidly waning protection from a newborn dose of monoclonal antibody or maternal immunisation. Additionally, we showed that most QALY gains from vaccination were attributable to avoiding GP consultations rather than hospital admissions and deaths. This result drives the greater relative economic value of infant strategies compared with newborn or maternal strategies, even though maternal strategies might prevent more severe RSV cases. Although, as Black argues30 the goal of immunisation programmes is primarily to prevent severe disease and death.

ons and deaths. This result drives the greater relative economic value of infant strategies compared with newborn or maternal strategies, even though maternal strategies might prevent more severe RSV cases. Although, as Black argues30 the goal of immunisation programmes is primarily to prevent severe disease and death. Severe RSV infection early in life might be linked to later development of chronic conditions such as wheezing and asthma.31 Such long-term chronic conditions can be influential in cost-effectiveness analyses because of their long-term implications. However the relation between RSV infection and long-term outcomes is uncertain and has only been most clearly described for preterm infants.32 This additional complication was not included in our analysis but should be considered, particularly when more is known about the likely groups indicated for vaccination and the parameters of the relation between RSV and its sequelae.

outcomes is uncertain and has only been most clearly described for preterm infants.32 This additional complication was not included in our analysis but should be considered, particularly when more is known about the likely groups indicated for vaccination and the parameters of the relation between RSV and its sequelae. RSV burden is substantial in children under 5 years, particularly in young infants. Passive or active immunisation directed at pregnant women, neonates, or infants could reduce this burden and would be good value for money if priced appropriately. There is potential for an RSV vaccine that protects infants and young children to be cost-effective because of the high disease burden in these groups. A maternal or newborn vaccination strategy is likely to avert the most severe disease and deaths, especially if it can be targeted at protecting infants born during the RSV season between late autumn and early spring. Vaccination of older children with a long-lasting vaccine might avert more health-care costs and episodes of mild disease. Our conclusions are based on ecological analyses of syndromic and laboratory data with economic modelling using a range of characteristics of potential prophylactic interventions. They will need to be validated when results from clinical trials and post-licensure studies become available. Supplementary Material Supplementary appendix

RSV burden is substantial in children under 5 years, particularly in young infants. Passive or active immunisation directed at pregnant women, neonates, or infants could reduce this burden and would be good value for money if priced appropriately. There is potential for an RSV vaccine that protects infants and young children to be cost-effective because of the high disease burden in these groups. A maternal or newborn vaccination strategy is likely to avert the most severe disease and deaths, especially if it can be targeted at protecting infants born during the RSV season between late autumn and early spring. Vaccination of older children with a long-lasting vaccine might avert more health-care costs and episodes of mild disease. Our conclusions are based on ecological analyses of syndromic and laboratory data with economic modelling using a range of characteristics of potential prophylactic interventions. They will need to be validated when results from clinical trials and post-licensure studies become available. Supplementary Material Supplementary appendix Acknowledgments AJvH and MJ were supported by the UK National Institute for Health Research Health Protection Research (NIHR) Unit in Immunisation at the London School of Hygiene and Tropical Medicine (London, UK) in partnership with Public Health England (PHE). The views expressed are those of the authors and not necessarily those of the UK National Health Service, the NIHR, the UK Department of Health, or PHE.

lth Protection Research (NIHR) Unit in Immunisation at the London School of Hygiene and Tropical Medicine (London, UK) in partnership with Public Health England (PHE). The views expressed are those of the authors and not necessarily those of the UK National Health Service, the NIHR, the UK Department of Health, or PHE. Contributors DC analysed and interpreted the data, searched the literature, designed the figures, and wrote the manuscript. AJvH searched the literature, interpreted the data, and wrote the manuscript. ATN initiated the concept and wrote the manuscript. MJ initiated the concept, interpreted the data, and wrote the manuscript. AJP provided epidemiological and immunological advice on respiratory syncytial virus and reviewed the manuscript.

t. AJvH searched the literature, interpreted the data, and wrote the manuscript. ATN initiated the concept and wrote the manuscript. MJ initiated the concept, interpreted the data, and wrote the manuscript. AJP provided epidemiological and immunological advice on respiratory syncytial virus and reviewed the manuscript. Declaration of interests MJ and AJP are members of the Respiratory Syncytial Virus Consortium in Europe (RESCEU). RESCEU has received funding from the Innovative Medicines Initiative 2 Joint Undertaking under grant agreement 116019. This Joint Undertaking receives support from the European Union's Horizon 2020 research and innovation programme and the European Federation of Pharmaceutical Industries and Associations. Neither MJ nor his research group has received any forms of pecuniary or other support from the pharmaceutical industry. AJP's department received unrestricted educational grants from Pfizer, GlaxoSmithKline (GSK), and Astra Zeneca in July, 2016, and from Gilead, Merck Sharpe Dohme, GSK, and Astra Zeneca in June, 2017, for a 3-day course on infection and immunity in children. He is Chair of the UK Department of Health's Joint Committee on Vaccination and Immunisation and the European Medicines Agency Scientific Advisory Group in vaccines and a member of WHO's SAGE. All other authors declare no competing interests.

Introduction In 2015, the UK Scientific Advisory Committee on Nutrition published a report1 on the evidence for the association between consumption of carbohydrates and health. The report clarified the role of sugar for development of dental caries and identified sugar-sweetened beverages (SSBs) as a specific risk factor for weight gain and type 2 diabetes, recommending that SSB consumption should be minimised. Both Public Health England2 and the UK House of Parliament's Health Committee3 subsequently advised a tax on SSBs and, in March, 2016, the budget statement4 included proposals for a soft drinks industry levy. Taxes on SSBs have been previously introduced in Mexico, France, Hungary, and elsewhere;5 however, the UK would be the first to introduce a three-tiered levy. The levy is presented as an incentive for the industry to reformulate existing products to remove sugar, reduce portion sizes, and promote new or existing low-sugar alternatives. The levy is due to be introduced in 2018, subject to parliament passing the legislation in 2017, with revenue hypothecated for an increase of physical activity and breakfast clubs in schools.4

try to reformulate existing products to remove sugar, reduce portion sizes, and promote new or existing low-sugar alternatives. The levy is due to be introduced in 2018, subject to parliament passing the legislation in 2017, with revenue hypothecated for an increase of physical activity and breakfast clubs in schools.4 Although the UK Government has expressed a desire that the levy is not passed on to purchasers through price rises, this request cannot be mandated and the industry response is unknown. Other outcomes could include reformulation to reduce sugar content or changes in marketing to encourage purchasers to switch to low-sugar products or small portion sizes. Different responses will have different effects on consumption patterns for soft drinks and hence determine the health effects of the levy.6 The aim of this study is to appraise the health effect of various discrete industry responses so that legislation for the soft drinks levy can be designed to maximise health gain. Research in context Evidence before this study

Although the UK Government has expressed a desire that the levy is not passed on to purchasers through price rises, this request cannot be mandated and the industry response is unknown. Other outcomes could include reformulation to reduce sugar content or changes in marketing to encourage purchasers to switch to low-sugar products or small portion sizes. Different responses will have different effects on consumption patterns for soft drinks and hence determine the health effects of the levy.6 The aim of this study is to appraise the health effect of various discrete industry responses so that legislation for the soft drinks levy can be designed to maximise health gain. Research in context Evidence before this study The UK Government announced a soft drinks industry levy in March, 2016. Multiple observational and modelling studies have analysed the effect of soft drink taxes in other international settings; however, the UK would be first to introduce a tiered industry levy (high tax for drinks with >8 g of sugar per 100 mL, moderate tax for 5–8 g, and no tax for <5 g) rather than a sales tax, as has been applied elsewhere. To our knowledge, no analyses of its potential impact have been done and no international precedent exists from which to predict the potential response of soft drink manufacturers to the levy. Added value of this study

The UK Government announced a soft drinks industry levy in March, 2016. Multiple observational and modelling studies have analysed the effect of soft drink taxes in other international settings; however, the UK would be first to introduce a tiered industry levy (high tax for drinks with >8 g of sugar per 100 mL, moderate tax for 5–8 g, and no tax for <5 g) rather than a sales tax, as has been applied elsewhere. To our knowledge, no analyses of its potential impact have been done and no international precedent exists from which to predict the potential response of soft drink manufacturers to the levy. Added value of this study This study, to our knowledge, is the first to estimate the health impact of the UK soft drinks industry levy. It focuses on obesity, diabetes, and oral health, for which evidence of a causal link between soft drink consumption and health is strongest. Previous evidence has suggested that soft drink taxes lead to price rises and subsequent reductions in purchases of targeted drinks. This study goes further and estimates the effects of six scenarios to illustrate the relative health impacts of three possible industry responses to the levy: reformulation, price rises, and changes to product market share. Implications of all the available evidence

This study, to our knowledge, is the first to estimate the health impact of the UK soft drinks industry levy. It focuses on obesity, diabetes, and oral health, for which evidence of a causal link between soft drink consumption and health is strongest. Previous evidence has suggested that soft drink taxes lead to price rises and subsequent reductions in purchases of targeted drinks. This study goes further and estimates the effects of six scenarios to illustrate the relative health impacts of three possible industry responses to the levy: reformulation, price rises, and changes to product market share. Implications of all the available evidence Each of the three responses modelled could lead to important health gains, with industry likely to react to the levy using a combination of all three. This study extends previous analyses of the effect of soft drink taxes to show the benefits of reformulation stimulated by the tiered levy. Our analyses show that substantial health benefits could occur if the levy stimulates reformulation. Further important health benefits from price changes will be mitigated if industry spread the price increase across their entire portfolio. Increases in market share for mid-sugar and low-sugar drinks could have substantial health benefits, but only if the market share comes at the expense of high-sugar drinks rather than people shifting from low-sugar to mid-sugar drinks.

ll be mitigated if industry spread the price increase across their entire portfolio. Increases in market share for mid-sugar and low-sugar drinks could have substantial health benefits, but only if the market share comes at the expense of high-sugar drinks rather than people shifting from low-sugar to mid-sugar drinks. Methods Scenarios We developed a comparative risk assessment model to estimate the effects of SSB reformulation, price changes, and changes to SSB market share on obesity, dental caries, and type 2 diabetes in the UK (figure). The baseline for the model was the 2014 UK population and we took all data for the model from the closest year for which data were freely available (appendix). We identified and modelled three possible industry responses. First, reformulation to reduce sugar concentration; second, a rise in price; and third, activities to change the relative market share of high-sugar, mid-sugar, and low-sugar drinks. For each of these responses, the magnitude of the response is uncertain. Informed by evidence where available and expert opinion, for each response we identified better-case and worse-case scenarios for reduction of sugar consumption, resulting in six scenarios (table 1).Figure Conceptual model DMFT=decayed, missing, or filled teeth. RCT=randomised controlled trial. SSB=sugar-sweetened beverage. Table 1 Simulated scenarios

Methods Scenarios We developed a comparative risk assessment model to estimate the effects of SSB reformulation, price changes, and changes to SSB market share on obesity, dental caries, and type 2 diabetes in the UK (figure). The baseline for the model was the 2014 UK population and we took all data for the model from the closest year for which data were freely available (appendix). We identified and modelled three possible industry responses. First, reformulation to reduce sugar concentration; second, a rise in price; and third, activities to change the relative market share of high-sugar, mid-sugar, and low-sugar drinks. For each of these responses, the magnitude of the response is uncertain. Informed by evidence where available and expert opinion, for each response we identified better-case and worse-case scenarios for reduction of sugar consumption, resulting in six scenarios (table 1).Figure Conceptual model DMFT=decayed, missing, or filled teeth. RCT=randomised controlled trial. SSB=sugar-sweetened beverage. Table 1 Simulated scenarios Better case for sugar reduction Worse case for sugar reduction Reformulation Scenario 1: high-sugar drinks reduce sugar content by 30% and mid-sugar drinks by 15% Scenario 2: mid-sugar and high-sugar drinks both reduce sugar content by 5% Price change Scenario 3: increase in price of high-sugar and mid-sugar drinks such that 50% of levy is passed on to consumers with a maximum 20% price rise Scenario 4: increase in price of all packaged drinks* by the same percentage such that 50% of the tax is borne by customers Change to SSB market share Scenario 5: breakdown in sales of soft drinks shifts from 58% to 64% for low-sugar drinks, 6% to 12% for mid-sugar drinks, and 36% to 24% for high-sugar drinks Scenario 6: breakdown in sales of soft drinks shifts to 55% for low-sugar drinks, 12% for mid-sugar drinks, and 33% for high-sugar drinks Low-sugar drinks is less than 5 g of sugar per 100 mL, medium-sugar drinks is 5–8 g of sugar per 100 mL, and high-sugar drinks is more than 8 g of sugar per 100 mL. SSB=sugar-sweetened beverage.

breakdown in sales of soft drinks shifts to 55% for low-sugar drinks, 12% for mid-sugar drinks, and 33% for high-sugar drinks Low-sugar drinks is less than 5 g of sugar per 100 mL, medium-sugar drinks is 5–8 g of sugar per 100 mL, and high-sugar drinks is more than 8 g of sugar per 100 mL. SSB=sugar-sweetened beverage. * Including low-sugar or zero-sugar drinks, bottled water, fruit juice, and sweetened milk drinks, and not including tea, coffee, unsweetened milk, and alcohol. We adopt the government definitions of high-sugar drinks as those with more than 8 g of sugar per 100 mL, mid-sugar drinks as those with 5–8 g of sugar per 100 mL, and low-sugar drinks as those with less than 5 g of sugar per 100 mL. Soft drinks are defined as all drinks with added sugar or sweetener; SSBs are drinks with added sugar, excluding milk-based drinks, tea, and coffee; concentrated SSBs are defined as SSBs that are intended to be diluted with water, and regular SSBs are intended to be drunk as sold. Small producers will be excluded from the levy.4 We searched all soft drinks sold through the Tesco website and extracted the names of manufacturers. We used the Companies House website to identify manufacturers fulfilling the UK Government definition of a small company9 and identified 13 small companies, which together contributed 0·6% of total UK SSB sales. Therefore, we did not adjust our analyses to account for these companies.

racted the names of manufacturers. We used the Companies House website to identify manufacturers fulfilling the UK Government definition of a small company9 and identified 13 small companies, which together contributed 0·6% of total UK SSB sales. Therefore, we did not adjust our analyses to account for these companies. The better-case reformulation scenario (scenario 1) assumed that industry would reduce the sugar concentration of high-sugar drinks by 30% and mid-sugar drinks by 15%. This assumption is based on the reformulation of Sprite and Lipton Ice Tea, which have both reduced their sugar concentration by 30% since 2013.10, 11 In the worse-case scenario (scenario 2), we assumed a 5% reduction in sugar concentration of both high-sugar and mid-sugar drinks. This assumption was based on Coca-Cola's pledge made to the Public Health Responsibility Deal of a 5% reduction in calories across their sparkling drink range between 2012 and 2014; they achieved a 5·3% reduction.11 Under both these scenarios, the volume consumed is assumed to remain constant.

gar and mid-sugar drinks. This assumption was based on Coca-Cola's pledge made to the Public Health Responsibility Deal of a 5% reduction in calories across their sparkling drink range between 2012 and 2014; they achieved a 5·3% reduction.11 Under both these scenarios, the volume consumed is assumed to remain constant. To derive the price change scenarios, we used estimates from the Office for Budget Responsibility that the levy will be 18 pence per L on mid-sugar drinks and 24 pence per L on high-sugar drinks.12 Low-sugar drinks will not be taxed. Previous sugary drink taxes have been passed on at rates of between 50% and 100%,13, 14, 15 and if the tax was entirely passed through to consumers, high-sugar concentrated drinks would, on average, increase in price by 75% and high-sugar regular drinks would increase by 31% (table 2). Such price rises are markedly greater than in other examples of SSB taxes (most countries have adopted smaller tax rates and therefore, despite high pass-on rates, only result in a 5–15% price rise)17 and are larger than the 20% often cited as being necessary to affect substantial behavioural change and improve health.18Table 2 Baseline price16 and change in price for the taxed drinks categories with different rates of tax pass-throughs and as modelled in scenario 3

ass-on rates, only result in a 5–15% price rise)17 and are larger than the 20% often cited as being necessary to affect substantial behavioural change and improve health.18Table 2 Baseline price16 and change in price for the taxed drinks categories with different rates of tax pass-throughs and as modelled in scenario 3 Baseline price (pence per L) Price with 100% pass-through (pence per L) Price with 50% pass-through (pence per L) Scenario 3 modelled price (pence per L) Concentrated high sugar 32·1 56·1 (+75%) 44·1 (+37%) 38·6 (+20%) Concentrated mid sugar 40·1 58·1 (+45%) 49·1 (+22%) 48·1 (+20%) Regular high sugar 77·6 101·6 (+31%) 89·6 (+15%) 89·6 (+15%) Regular mid sugar 99·0 117·0 (+18%) 108·0 (+9%) 108·0 (+9%) Percentages in parentheses indicate percentage change in baseline price.

ncentrated high sugar 32·1 56·1 (+75%) 44·1 (+37%) 38·6 (+20%) Concentrated mid sugar 40·1 58·1 (+45%) 49·1 (+22%) 48·1 (+20%) Regular high sugar 77·6 101·6 (+31%) 89·6 (+15%) 89·6 (+15%) Regular mid sugar 99·0 117·0 (+18%) 108·0 (+9%) 108·0 (+9%) Percentages in parentheses indicate percentage change in baseline price. We therefore assumed that 50% of the price increase would be passed on to purchasers and that companies would not increase prices by more than 20% (table 2). The better case for price change (scenario 3) assumed that the tax is passed on only through SSBs. However, major soft drink manufacturers produce various beverages. Therefore, in a worse case for price change (scenario 4), we assume that the levy is passed on evenly across all soft drinks (both diet beverages and SSBs), fruit juice, and bottled water, resulting in a 6% price rise. We modelled passing on 100% of the price increase to consumers as a sensitivity analysis. As stated by the UK Government, we applied tax rates to concentrated drinks given their price per litre as drunk, assuming a ratio of concentrate to water of one to four, as used by the British Soft Drinks Association (BSDA).10, 19 The appendix gives details of how we modelled the price change.

a sensitivity analysis. As stated by the UK Government, we applied tax rates to concentrated drinks given their price per litre as drunk, assuming a ratio of concentrate to water of one to four, as used by the British Soft Drinks Association (BSDA).10, 19 The appendix gives details of how we modelled the price change. A change in SSB market shares might result from changes in product marketing, changing product size, or the introduction of new mid-sugar and low-sugar products. For example, the BSDA reports a 70% increase in expenditure on advertising of low-calorie or zero-calorie brands and growth in the sales of small pack sizes,10 and new mid-sugar products have emerged, such as Coca-Cola Life, which has 30% less sugar than does full-sugar Coca-Cola.20 Few data exist to inform the extent to which these activities drive changes in purchasing behaviour. However, the soft drinks industry has pledged to reduce energy intake from soft drinks by 20% from 2015 levels by 2020.10 To achieve this target, we calculate that the market share of high-sugar drinks would need to fall by 12% alongside a 6% increase for each of mid-sugar and low-sugar drinks, as shown in scenario 5, our better case for sugar reduction. The worse case (scenario 6) acknowledges that increased marketing of new mid-sugar drinks might lead consumers to switch to this category from low-sugar drinks. We assume that mid-sugar drinks double their market share alongside equal reductions in the market share of high-sugar and low-sugar drinks.

r sugar reduction. The worse case (scenario 6) acknowledges that increased marketing of new mid-sugar drinks might lead consumers to switch to this category from low-sugar drinks. We assume that mid-sugar drinks double their market share alongside equal reductions in the market share of high-sugar and low-sugar drinks. Health impact modelling We developed a comparative risk assessment model to estimate the effect of the changes to SSB purchasing on incidence of dental caries and type 2 diabetes and prevalence of obesity. Comparative risk assessment modelling requires identification of risk factor-disease pairs. In this case, the risk factor is SSB consumption and the diseases are dental caries, type 2 diabetes, and obesity. A two-step process then estimates the impact of the risk factor on the diseases. First, changes in the risk factor between baseline (current behaviour) and scenarios are estimated. Second, changes in the diseases as a result of changes in the risk factor are calculated using population impact fractions and applied to baseline levels of disease in the population. Such methods are common to the field of comparative risk assessment modelling21 and are based on model parameters representing baseline risk factor and disease status and the epidemiological relations between risk factors and diseases, which are assumed to be causal.

ied to baseline levels of disease in the population. Such methods are common to the field of comparative risk assessment modelling21 and are based on model parameters representing baseline risk factor and disease status and the epidemiological relations between risk factors and diseases, which are assumed to be causal. We sought parameters describing the direct relation between SSB consumption and health outcomes where possible from meta-analyses of randomised controlled trials where available or cohort studies (table 3).22, 23, 24, 25, 26, 27, 28 We modelled the relation between SSB consumption and diabetes and bodyweight as a function of SSB consumption. The reformulation scenarios (scenarios 1 and 2) assumed that SSB consumption stays constant, but the amount of sugar in the drinks reduces. To estimate the effect of these scenarios on obesity and diabetes, we derived estimates of equivalised SSB consumption, which rises and falls in direct proportion to volume of SSB consumed and average SSB sugar concentration. We standardised against the average sugar concentrations in drinks in the baseline scenario. For example, in the baseline scenario, the average sugar concentration of SSBs was 9·2 g per 100 mL and the average consumption was 213 mL per day. A reduction in the average sugar concentration to 8·2 g per 100 mL at the same level of consumption would have an equivalised SSB consumption of 190 mL per day. This equivalised SSB consumption arises because a reduction in consumption of SSBs from 213 mL per day to 190 mL per day at constant sugar concentrations would result in the same reduction of sugar as a reduction of sugar concentration from 9·2 g per 100 mL to 8·2 g per 100 mL at the same level of consumption. We used equivalised SSB consumption as an input for diabetes and obesity modelling in all scenarios.Table 3 Model input parameters and data sources

concentrations would result in the same reduction of sugar as a reduction of sugar concentration from 9·2 g per 100 mL to 8·2 g per 100 mL at the same level of consumption. We used equivalised SSB consumption as an input for diabetes and obesity modelling in all scenarios.Table 3 Model input parameters and data sources Parameter Data source Bodyweight Increase in weight of 0·09 kg (95% CI −0·11 to 0·29) in adults and 0·45 kg (0·24–0·66) in children per additional 100 mL SSB consumed per day Meta-analysis of randomised controlled trials of SSB consumption and bodyweight; two studies22, 23 identified and combined for adults and two24, 25 for children Diabetes Relative risk of incident diabetes of 1·42 (95% CI 1·19–1·69) per additional 250 mL serving per day for adults and children Imamura et al26 Dental caries Increase in number of decayed, missing, or filled teeth of 0·008 (95% CI 0·002–0·014) per person per year for every additional 10 g of sugar consumed per day Bernabé et al27 SSB=sugar-sweetened beverage. Uncertainty intervals reflect the uncertainty in baseline sugar drink sales and consumption, disease burden, sensitivity to price changes, and associations between sugar or sugary drink consumption and health outcomes. We estimated them using 5000 iterations of a Monte Carlo analysis, with model parameters drawn from the published or estimated uncertainty of each parameter (appendix).

nk sales and consumption, disease burden, sensitivity to price changes, and associations between sugar or sugary drink consumption and health outcomes. We estimated them using 5000 iterations of a Monte Carlo analysis, with model parameters drawn from the published or estimated uncertainty of each parameter (appendix). We applied all results to the 2014 UK population29 and made separate estimates for each outcome by sex and age group using age-specific and sex-specific estimates of baseline SSB consumption and disease burden. Further details of the health impact model are in the appendix, including a sensitivity analysis in which the direct effect on weight of SSB consumption is replaced with an energy balance equation for comparison with other work that has used this method.30 Role of the funding source There was no specific funding source for this study. The corresponding author had full access to all the data and had final responsibility for the decision to submit for publication.

We applied all results to the 2014 UK population29 and made separate estimates for each outcome by sex and age group using age-specific and sex-specific estimates of baseline SSB consumption and disease burden. Further details of the health impact model are in the appendix, including a sensitivity analysis in which the direct effect on weight of SSB consumption is replaced with an energy balance equation for comparison with other work that has used this method.30 Role of the funding source There was no specific funding source for this study. The corresponding author had full access to all the data and had final responsibility for the decision to submit for publication. Results The better case for reformulation (scenario 1) resulted in a fall in mean sugar content of SSBs equivalent to a reduction of 58·5 mL (95% uncertainty interval [95% UI 54·5–62·6; 10 kcal [9–10]) of SSBs per person per day (table 4). This reduction is the largest among scenarios modelled. All simulated scenarios led to a fall in equivalised SSB consumption except for the worse case for market share, which resulted in a small increase. The largest falls for both sexes were among 11–18-year-olds who consume the largest volume of SSBs.Table 4 Reduction in equivalised* volume of sugar-sweetened beverage consumed with each scenario

led to a fall in equivalised SSB consumption except for the worse case for market share, which resulted in a small increase. The largest falls for both sexes were among 11–18-year-olds who consume the largest volume of SSBs.Table 4 Reduction in equivalised* volume of sugar-sweetened beverage consumed with each scenario Reformulation Price change Change in market share Scenario 1 Scenario 2 Scenario 3 Scenario 4 Scenario 5 Scenario 6 Male sex Boys aged 4–10 years 61·7 11·2 34·5 12·4 38·6 −3·8 Boys aged 11–18 years 137·6 25·0 77·0 27·7 86·0 −8·6 Men aged 19–64 years 71·0 12·9 39·7 14·3 44·4 −4·4 Men aged ≥65 years 24·0 4·4 13·4 4·8 15·0 −1·5 Female sex Girls aged 4–10 years 51·9 9·5 29·1 10·4 32·5 −3·2 Girls aged 11–18 years 93·2 17·0 52·1 18·7 58·3 −5·8 Women aged 19–64 years 49·7 9·0 27·8 10·0 31·1 −3·1 Women aged ≥65 years 23·5 4·3 13·2 4·7 14·7 −1·5 Total Total (95% UI) 58·5 (54·5 to 62·6) 10·7 (10·0 to 11·4) 32·7 (30·3 to 35·3) 11·8 (10·9 to 12·7) 36·6 (34·9 to 38·3) −3·6 (−3·8 to −3·4) Data are in mL per person per day. UI=uncertainty interval. * Where equivalisation results in the same sugar intake for each equivalised unit of sugar-sweetened beverage.

Reformulation Price change Change in market share Scenario 1 Scenario 2 Scenario 3 Scenario 4 Scenario 5 Scenario 6 Male sex Boys aged 4–10 years 61·7 11·2 34·5 12·4 38·6 −3·8 Boys aged 11–18 years 137·6 25·0 77·0 27·7 86·0 −8·6 Men aged 19–64 years 71·0 12·9 39·7 14·3 44·4 −4·4 Men aged ≥65 years 24·0 4·4 13·4 4·8 15·0 −1·5 Female sex Girls aged 4–10 years 51·9 9·5 29·1 10·4 32·5 −3·2 Girls aged 11–18 years 93·2 17·0 52·1 18·7 58·3 −5·8 Women aged 19–64 years 49·7 9·0 27·8 10·0 31·1 −3·1 Women aged ≥65 years 23·5 4·3 13·2 4·7 14·7 −1·5 Total Total (95% UI) 58·5 (54·5 to 62·6) 10·7 (10·0 to 11·4) 32·7 (30·3 to 35·3) 11·8 (10·9 to 12·7) 36·6 (34·9 to 38·3) −3·6 (−3·8 to −3·4) Data are in mL per person per day. UI=uncertainty interval. * Where equivalisation results in the same sugar intake for each equivalised unit of sugar-sweetened beverage. The reduction in obesity prevalence resulting from each scenario is estimated to be greatest after scenario 1 (better case for reformulation; table 5), leading to an estimated reduction of 144 383 (95% UI 5102–306 743) of 15 470 813 individuals with obesity, 0·9% of the obese population. This figure is compared with the better cases for price change (scenario 3), which reduces the obese population by 81 594 (3588–182 669; 0·5%), and with change in market share (scenario 5), which reduces the obese population by 91 042 (4289–204 903; 0·6%). Results varied by age, with larger reductions in the number of children with obesity than in that of adults in scenario 1. The relative reduction in obesity prevalence was predicted to be greater in male individuals than in female individuals because male individuals consume a greater volume of SSBs (appendix). Effect size estimates were significantly increased when an energy balance equation was used (appendix).Table 5 Reduction in the number of obese individuals with each scenario

was predicted to be greater in male individuals than in female individuals because male individuals consume a greater volume of SSBs (appendix). Effect size estimates were significantly increased when an energy balance equation was used (appendix).Table 5 Reduction in the number of obese individuals with each scenario Reformulation Price change Change in market share Scenario 1 Scenario 2 Scenario 3 Scenario 4 Scenario 5 Scenario 6 Male sex Boys aged 4–10 years 29 227 (10·4%) 5524 (2·0%) 16 689 (6·0%) 6095 (2·2%) 18 592 (6·6%) −1911 (−0·7%) Boys aged 11–18 years 31 793 (6·0%) 5907 (1·1%) 17 987 (3·4%) 6521 (1·2%) 20 066 (3·8%) −2033 (−0·4%) Men aged 19–64 years 25 582 (0·5%) 4663 (0·1%) 14 324 (0·3%) 5149 (0·1%) 16 005 (0·3%) −1596 (0·0%) Men aged ≥65 years 3002 (0·2%) 547 (0·0%) 1680 (0·1%) 603 (0·0%) 1877 (0·1%) −187 (0·0%) Female sex Girls aged 4–10 years 16 455 (8·9%) 3097 (1·7%) 9374 (5·0%) 3418 (1·8%) 10 447 (5·6%) −1070 (−0·6%) Girls aged 11–18 years 17 581 (4·8%) 3257 (0·9%) 9930 (2·7%) 3595 (1·0%) 11 081 (3·0%) −1120 (−0·3%) Women aged 19–64 years 17 328 (0·3%) 3157 (0·1%) 9700 (0·2%) 3487 (0·1%) 10 839 (0·2%) −1081 (0·0%) Women aged ≥65 years 3415 (0·2%) 622 (0·0%) 1911 (0·1%) 697 (0·0%) 2135 (0·1%) −213 (0·0%) Total Total; 95% UI 144 383 (0·9%); 5102 to 30 6743 26 774 (0·2%); 1276 to 63 806 81 594 (0·5%); 3588 to 182 669 29 555 (0·2%); 1379 to 69 804 91 042 (0·6%); 4289 to 204 903 −9211 (−0·1%); −22  776 to −485 Data are n (%). UI=uncertainty interval.

ars 3415 (0·2%) 622 (0·0%) 1911 (0·1%) 697 (0·0%) 2135 (0·1%) −213 (0·0%) Total Total; 95% UI 144 383 (0·9%); 5102 to 30 6743 26 774 (0·2%); 1276 to 63 806 81 594 (0·5%); 3588 to 182 669 29 555 (0·2%); 1379 to 69 804 91 042 (0·6%); 4289 to 204 903 −9211 (−0·1%); −22  776 to −485 Data are n (%). UI=uncertainty interval. Across the scenarios modelled, the pattern of results seen with obesity is repeated for type 2 diabetes. Scenario 1 (better case for reformulation) resulted in an estimated 19 094 (95% UI 6920–32 678; incidence reduction of 31·1 per 100 000 person-years) fewer new cases of diabetes per year and scenario 6 (worse case for change in market share) led to an increase of 1238 (455–2359; incidence increase of 2·0 per 100 000 person-years) cases per year (table 6). However, by contrast with the obesity results, adults aged older than 65 years saw the largest absolute reduction in diabetes incidence, reflecting the positive association between age and disease burden.Table 6 Reduction in the number of cases of diabetes per year with each scenario

on-years) cases per year (table 6). However, by contrast with the obesity results, adults aged older than 65 years saw the largest absolute reduction in diabetes incidence, reflecting the positive association between age and disease burden.Table 6 Reduction in the number of cases of diabetes per year with each scenario Reformulation Price Change in market share Scenario 1 Scenario 2 Scenario 3 Scenario 4 Scenario 5 Scenario 6 Male sex Boys aged 4–10 years 71 (2·6) 13 (0·5) 40 (1·5) 15 (0·5) 45 (1·6) −5 (−0·2) Boys aged 11–18 years 224 (7·5) 44 (1·5) 131 (4·3) 49 (1·6) 145 (4·8) −15 (−0·5) Men aged 19–64 years 8364 (43·5) 1585 (8·2) 4783 (24·9) 1749 (9·1) 5327 (27·6) −549 (−2·8) Men aged ≥65 years 2539 (49·4) 469 (9·1) 1431 (27·8) 517 (10·1) 1598 (31·1) −160 (−3·1) Female sex Girls aged 4–10 years 52 (2·0) 10 (0·4) 29 (1·1) 11 (0·4) 33 (1·2) −3 (−0·1) Girls aged 11–18 years 223 (7·8) 43 (1·5) 128 (4·5) 47 (1·7) 143 (4·9) −15 (−0·5) Women aged 19–64 years 5192 (26·7) 972 (5·0) 2950 (15·1) 1073 (5·5) 3289 (16·9) −336 (−1·7) Women aged ≥65 years 2429 (38·8) 448 (7·2) 1369 (21·8) 495 (7·9) 1528 (24·4) −1549 (−2·5) Total Total; 95% UI 19 094 (31·1); 6920 to 32 678 3584 (5·8); 1289 to 6466 10 861 (17·7); 3899 to 18 964 3955 (6·4); 1420 to 7085 1528 (19·7); 4414 to 21 785 −1238 (−2·0); −2359 to −455 Data in parentheses are reductions in incidence per 100 000 person-years. UI=uncertainty interval.

495 (7·9) 1528 (24·4) −1549 (−2·5) Total Total; 95% UI 19 094 (31·1); 6920 to 32 678 3584 (5·8); 1289 to 6466 10 861 (17·7); 3899 to 18 964 3955 (6·4); 1420 to 7085 1528 (19·7); 4414 to 21 785 −1238 (−2·0); −2359 to −455 Data in parentheses are reductions in incidence per 100 000 person-years. UI=uncertainty interval. All scenarios except for scenario 6 led to a fall in the numbers of teeth affected with dental caries (measured by the number of decayed, missing, or filled teeth [DMFT]; table 7). The better case for reformulation (scenario 1) had the largest effect size, reducing the annual incidence of DMFT by 269 375 (95% UI 82 211–470 928; incidence reduction of 4·4 per 1000 person-years). As with results for obesity and diabetes, the better case for change in market share (scenario 5) and price change (scenario 3) scenarios had the next largest effects respectively. Those aged 11–18 years were expected to have the greatest relative benefit because of their higher baseline SSB consumption than those aged older or younger than this age group (appendix).Table 7 Reduction in number of decayed, missing, or filled teeth per year with each scenario

the next largest effects respectively. Those aged 11–18 years were expected to have the greatest relative benefit because of their higher baseline SSB consumption than those aged older or younger than this age group (appendix).Table 7 Reduction in number of decayed, missing, or filled teeth per year with each scenario Reformulation Price Change in market share Scenario 1 Scenario 2 Scenario 3 Scenario 4 Scenario 5 Scenario 6 Male sex Boys aged 4–10 years 12 735 (4·6) 2318 (0·8) 7022 (2·6) 3376 (1·2) 8096 (2·9) −741 (−0·3) Boys aged 11–18 years 31 040 (10·3) 5650 (1·9) 17 268 (5·7) 7015 (2·3) 19 967 (6·6) −1922 (−0·6) Men aged 19–64 years 102 477 (5·3) 18 654 (1·0) 56 909 (3·0) 23 776 (1·2) 65 792 (3·4) −6282 (−0·3) Men aged ≥65 years 9239 (1·8) 1682 (0·3) 5081 (1·0) 2554 (0·5) 5924 (1·2) −563 (−0·1) Female sex Girls aged 4–10 years 10 225 (3·9) 1861 (0·7) 5645 (2·2) 2640 (1·0) 6496 (2·5) −593 (−0·2) Girls aged 11–18 years 19 977 (7·0) 3637 (1·2) 11 099 (3·9) 4610 (1·6) 12 808 (4·5) −1216 (−0·4) Women aged 19–64 years 72 625 (3·7) 13 220 (0·7) 40 270 (2·1) 17 240 (0·9) 46 516 (2·4) −4397 (−0·2) Women aged ≥65 years 11 056 (1·8) 2013 (0·3) 6086 (1·0) 3029 (0·5) 7119 (1·1) −688 (−0·1) Total Total; 95% UI 269 375 (4·4); 82 211 to 470 928 49 036 (0·8); 14 929 to 85 630 149 378 (2·4); 45 231 to 262 013 64 240 (1·1); 19 643 to 112 371 172 718 (2·8); 47 919 to 294 499 −16 401 (−0·3); −28 037 to −4604 Data in parentheses are reductions in incidence per 1000 person-years. UI=uncertainty interval.

g for smoking cessation and tobacco cessation treatment to Medicaid recipients NR 113 137 NR 0% (−6 to 6) NA The state-wide health reform legislation in MA, USA, was not associated with significant changes in emergency department visits for lower RTIs NR=not reported. NA=not applicable. RTI=respiratory tract infection. * Presumptive eligibility: low-income pregnant women are presumed to be eligible for Medicaid, so they can receive care (including smoking cessation services) while their Medicaid applications are still pending. The unborn-child option: the state can consider a fetus a “targeted low-income child”, allowing coverage of prenatal care (including smoking cessation services) and delivery to low-income pregnant women, even if they cannot provide documentation of citizenship or residency. Table 3 Association between implementation of tobacco taxation and primary outcomes

88 (−0·1) Total Total; 95% UI 269 375 (4·4); 82 211 to 470 928 49 036 (0·8); 14 929 to 85 630 149 378 (2·4); 45 231 to 262 013 64 240 (1·1); 19 643 to 112 371 172 718 (2·8); 47 919 to 294 499 −16 401 (−0·3); −28 037 to −4604 Data in parentheses are reductions in incidence per 1000 person-years. UI=uncertainty interval. In our sensitivity analysis where 100% of the levy is passed on to consumers, equivalised SSB consumption would reduce by 71 mL (95% UI 66–77; 11 kcal [10–12]) per person per day. This reduction would lead to 174 818 (7536–367 647) fewer individuals with obesity, 23 046 (8419–39 965) fewer cases of diabetes per year, and 324 488 (89 073–553 840) fewer DMFT per year. Discussion The proposed UK soft drinks industry levy has the potential to reduce obesity prevalence, diabetes incidence, and dental caries incidence. The effect on health and the ranking of scenarios is sensitive to the manner in which industry responds to the levy and the uncertainty in the modelling. Our estimates suggest that the greatest benefits will result from reformulation, with less but still positive health effects after price changes and changes to SSB market share to increase the proportion of low-sugar drinks sold, although in the worse-case scenario for change in market share, the health effect was actually negative. Children will have the greatest relative health benefit in terms of obesity and caries, with absolute reductions in diabetes incidence rates increasing with age.

to increase the proportion of low-sugar drinks sold, although in the worse-case scenario for change in market share, the health effect was actually negative. Children will have the greatest relative health benefit in terms of obesity and caries, with absolute reductions in diabetes incidence rates increasing with age. The main strength of this study is the timely assessment of a planned government policy by simulating a set of discrete scenarios for how industry might respond to the levy to inform the detail of the legislation. Other strengths include modelling of multiple health outcomes, use of age-specific and sex-specific data, use of own-price and cross-price elasticities for high-sugar, mid-sugar, and low-sugar drinks, and use of equivalised SSB consumption to allow for changes in both sugar content and SSB volume.

l of the legislation. Other strengths include modelling of multiple health outcomes, use of age-specific and sex-specific data, use of own-price and cross-price elasticities for high-sugar, mid-sugar, and low-sugar drinks, and use of equivalised SSB consumption to allow for changes in both sugar content and SSB volume. Uncertainty intervals estimate the uncertainty arising from model parameters; however, the greatest uncertainty is how the soft drinks industry will respond to the levy. Given this uncertainty, our results should not be read as precise estimates of the impact of the levy, but instead should be used to compare the relative effects of different scenarios. Moreover, industry is likely to respond with a blended approach that combines elements of reformulation, price changes, and marketing. Although the results have wide and overlapping uncertainty intervals, much of the uncertainty is correlated between scenarios. In all of the iterations of our Monte Carlo analysis, the best-case reformulation scenario was associated with the best health outcomes, which suggests that the ranking of scenarios is robust. We have not estimated uncertainty in how much of the levy is passed on to consumers (although a 100% pass-on is modelled as a sensitivity analysis) and we did not use child-specific estimates of the effect of SSB consumption on diabetes and dental caries incidence because of an absence of data available. We have assumed that disease risk from SSBs is dependent on the quantity of sugar consumed (more sugar leads to higher risk). Although findings from studies24, 25 have shown benefits from a swap from SSBs to artificially sweetened beverages, we are not aware of any studies that have described the effect of a swap of SSBs with a high sugar content to SSBs with a low sugar content.

e quantity of sugar consumed (more sugar leads to higher risk). Although findings from studies24, 25 have shown benefits from a swap from SSBs to artificially sweetened beverages, we are not aware of any studies that have described the effect of a swap of SSBs with a high sugar content to SSBs with a low sugar content. We have not modelled a temporal component. The effects on DMFT could occur soon after the change in SSB consumption, and the trials22, 23, 24, 25 used to parameterise the relation between SSB consumption and weight suggest that falls in obesity would be expected within 6 months for adults and 12 months for children. The effects on type 2 diabetes could take longer to be realised than for obesity (median follow-up of observational studies26 used for this parameter ranged between 3·4 years and 21·1 years). We have also not modelled results for different subgroups. Individuals from different socioeconomic backgrounds, ages, and baseline consumption levels could respond differently to each industry response simulated. Finally, we have not modelled the long-term health benefits of falls in obesity, the possible educational role the levy might have in highlighting that SSBs cause disease,31 and the health effect of use of the revenue to improve school sport and nutrition.

could respond differently to each industry response simulated. Finally, we have not modelled the long-term health benefits of falls in obesity, the possible educational role the levy might have in highlighting that SSBs cause disease,31 and the health effect of use of the revenue to improve school sport and nutrition. This study is, to our knowledge, the first to appraise the potential health effects of the UK soft drinks industry levy. The results of our study vary from those of a report by Oxford Economics,32 which calculated the impact of a price change associated with a 100% pass-on of the levy to targeted products only, with no reformulation or market share (most similar to our scenario 3). The authors estimated that the levy would result in a 5 kcal per-person per-day fall in energy intake. Our sensitivity analysis of a 100% pass-on rate would result in a reduction of 11 kcal per person per day (before adjustment for BSDA sales figures). Two principal explanations exist for the difference. First, we estimate the average price before tax of dilutables as 22 pence per L, whereas Oxford Economics estimate it as £1·76 per L. This discrepancy is likely to be due to Oxford Economics applying the tax before dilution. Second, Oxford Economics used estimates of how consumers respond to price changes of diet and non-diet SSBs taken from our 2013 study estimating the effect on obesity of a 20% UK SSB tax.30 In our present study, we have calculated estimates separately for high-sugar, mid-sugar, and low-sugar drinks.

ax before dilution. Second, Oxford Economics used estimates of how consumers respond to price changes of diet and non-diet SSBs taken from our 2013 study estimating the effect on obesity of a 20% UK SSB tax.30 In our present study, we have calculated estimates separately for high-sugar, mid-sugar, and low-sugar drinks. Our 2013 study30 estimated that a 20% price rise would lead to a 1·3% fall in the number of adults with obesity in the UK, compared with our scenario 3 estimate of 0·5% (after an average price rise of 15%). Our 2013 study did not estimate price elasticities separately for high-sugar and mid-sugar drinks, did not quantify the effect of the tax on children, and did not adjust for BSDA sales figures. It also used an energy balance equation rather than quantifying the direct effect of SSBs on bodyweight, which estimates larger effects on bodyweight than does quantifying the direct effect of SSBs on bodyweight. Conversely, in this study, we used an estimate of the direct relation between bodyweight and SSBs, which might more accurately represent substitution and other compensatory mechanisms secondary to changes in sugar (and energy) consumed from SSBs than use of an energy balance equation. This analysis substantiates our 2013 findings of greater relative reductions in obesity among younger adults than among older adults. This finding is explained by teenagers and young adults drinking more SSBs than do old adults and trial data suggesting that SSBs have a greater effect on weight gain in children than in adults.23, 24, 25, 26

iates our 2013 findings of greater relative reductions in obesity among younger adults than among older adults. This finding is explained by teenagers and young adults drinking more SSBs than do old adults and trial data suggesting that SSBs have a greater effect on weight gain in children than in adults.23, 24, 25, 26 Considering reformulation, Ma and colleagues33 estimated that a 40% reduction in sugar across all SSBs in the UK would lead to 800 000 fewer individuals with obesity. This estimate is substantially higher than our estimate in scenario 1 (about 144 400), which assumed a 30% reduction in sugar content of high-sugar drinks and 15% reduction in that of mid-sugar drinks. This discrepancy is, in part, due to Ma and colleagues estimating that the average reduction in energy consumed would be approximately twice our estimate and then using an energy balance equation to estimate the effect of energy intake on weight. Of note, estimates for the reduction in diabetes were similar between our study and Ma and colleagues' study. We also recognise that uncertainties exist around all parameters in the model that will affect the comparison of results with those from other simulation studies.

estimate the effect of energy intake on weight. Of note, estimates for the reduction in diabetes were similar between our study and Ma and colleagues' study. We also recognise that uncertainties exist around all parameters in the model that will affect the comparison of results with those from other simulation studies. The UK soft drinks industry levy has the potential to lead to important improvements in population health, particularly among children. Policy makers should engage with stakeholders to encourage responses to the levy that will maximise the potential health benefits of the new policy. Our results show the need for ongoing monitoring of the implementation strategies adopted by industry alongside modelling to estimate the long-term health consequences of their actions. Our results suggest that, of the scenarios modelled, reformulation would lead to the largest health benefits. Price rises and changes to product market share might also lead to important improvements in health. However, effects would be attenuated if manufacturers chose to pass the tax on to purchasers across all drinks or other products in their portfolio rather than just those targeted by the levy. Moreover, negative health effects might arise if the increase in market share of mid-sugar drinks comes at the expense of low-sugar drinks. Conversely, further health benefits might be realised if manufacturers pass on more than 50% of the levy to consumers or choose to reformulate to a greater extent than that modelled (as announced by Tesco34 and Lucozade Ribena Suntory35).

ase in market share of mid-sugar drinks comes at the expense of low-sugar drinks. Conversely, further health benefits might be realised if manufacturers pass on more than 50% of the levy to consumers or choose to reformulate to a greater extent than that modelled (as announced by Tesco34 and Lucozade Ribena Suntory35). The UK soft drinks industry levy could have valuable population health benefits, but the magnitude of its health impact will depend on how industry responds. The detail of the levy is yet to be decided, but we show important health benefits that could be maximised by substantial product reformulation, with further health gains arising through raising the price of high-sugar and mid-sugar drinks and increasing the market share of low-sugar products. Supplementary Material Supplementary appendix Acknowledgments We thank Eduardo Bernabé (King's College London, London, UK) for his help in identifying and interpreting data for sugar-sweetened beverages and oral health. Contributors ADMB, OTM, MR, SAJ, and PS conceived the study. ADMB, OTM, AK, RT, TB, and PS designed the methods. ADMB, OTM, AK, AE, and PS identified data and ran the analyses. ADMB wrote the first draft of the manuscript. All authors designed the scenarios analysed, interpreted results, commented on the manuscript and made critical revisions, and approved the final version of the manuscript.

RT, TB, and PS designed the methods. ADMB, OTM, AK, AE, and PS identified data and ran the analyses. ADMB wrote the first draft of the manuscript. All authors designed the scenarios analysed, interpreted results, commented on the manuscript and made critical revisions, and approved the final version of the manuscript. Declaration of interests RT and AK have previously done work on sugar-sweetened beverage taxes funded by the Union of European Soft Drinks Associations. MR is chair of Sustain and the Children's Food Campaign, which have campaigned for sugar drink taxes in the UK. MR is funded by the British Heart Foundation, grant number 006/PSS/CORE/2016/OXFORD. ADMB and OTM are members of the Faculty of Public Health, which has a position statement supporting sugary drink taxes. ADMB is funded by the Wellcome Trust, grant number 102730/Z/13/Z. OTM is a member of the UK Health Forum, which has also supported a UK sugar drinks tax. OTM is supported by a Wellcome Trust Clinical Doctoral Fellowship. SAJ was the independent Chair of the Department of Health Public Health Responsibility Deal Food Network from 2010 to 2015. SAJ is funded by the National Institute for Health Research Collaboration for Leadership in Applied Health Research and Care Oxford. The views expressed are those of the authors and not necessarily those of the National Health Service, National Institute for Health Research, or the Department of Health. PS is funded by the British Heart Foundation, grant number FS/15/34/31656. TB is funded the Health Research Council of New Zealand (16/443). AE declares no competing interests.

Introduction Almost half of children worldwide are regularly exposed to second-hand smoke, and 28% of the 600 000 deaths each year related to second-hand smoke occur in children.1, 2 Maternal smoking and second-hand smoke exposure during pregnancy are detrimental to fetal growth and development, leading to adverse birth outcomes such as preterm birth, low birthweight, being small for gestational age, and perinatal and infant mortality.3, 4, 5, 6, 7, 8 Additionally, second-hand smoke exposure presents substantial health risks postnatally by increasing the risk of asthma and respiratory tract infections.1, 9 Protection of children from the adverse health implications of second-hand smoke during important phases of development and the subsequent disease burden carried on into adulthood is crucial. The WHO Framework Convention on Tobacco Control (FCTC) aims to reduce tobacco consumption and second-hand smoke exposure through national tobacco control programmes.2 In 2008, six MPOWER measures were introduced to guide FCTC implementation (panel).2, 10 With tobacco use increasingly becoming a problem of developing countries already experiencing the largest burden of early-life morbidity and mortality, the absence of tobacco regulation is set to be a big driver of between-country inequality in child health outcomes.11 However, evaluations of the effectiveness of tobacco control interventions have generally excluded children, focusing instead on smoking rates and adult health outcomes.12, 13, 14Panel MPOWER policies2 Monitor tobacco use

cco regulation is set to be a big driver of between-country inequality in child health outcomes.11 However, evaluations of the effectiveness of tobacco control interventions have generally excluded children, focusing instead on smoking rates and adult health outcomes.12, 13, 14Panel MPOWER policies2 Monitor tobacco use Eligible policies include those that enforce accurate measurement of the extent of the tobacco epidemic and of the interventions to control it. Protect people from smoke Eligible policies include legislation to create smoke-free public environments (both indoors and outdoors). Offer help to quit tobacco use Eligible policies include tobacco cessation advice or interventions offered through health-care services, free telephone quit lines, and providing access to free or low-cost cessation medicines. Warn about the dangers of tobacco Eligible policies include health warnings on tobacco products, plain packaging of tobacco products, and mass media campaigns to educate the public about the dangers of tobacco. Enforce bans on tobacco advertising, promotion and sponsorship See WHO Framework Convention on Tobacco Control (FCTC) guidelines for implementation of Article 13, which provides a non-exhaustive list of advertising, promotion, and sponsorship within the terms of the FCTC.11 Raise taxes on tobacco Eligible policies include increasing percentage excise tax share in final tobacco. Research in context Evidence before this study

See WHO Framework Convention on Tobacco Control (FCTC) guidelines for implementation of Article 13, which provides a non-exhaustive list of advertising, promotion, and sponsorship within the terms of the FCTC.11 Raise taxes on tobacco Eligible policies include increasing percentage excise tax share in final tobacco. Research in context Evidence before this study Tobacco smoke exposure is the world's leading cause of preventable morbidity and premature mortality. Children cannot control their tobacco smoke exposure and therefore need protection through tobacco control measures. In a previous systematic review, we investigated the associations between smoke-free legislation and perinatal and child health outcomes. We searched 14 online medical research databases, the WHO International Clinical Trials Registry Platform, hand-searched references and citations, and consulted a panel of experts in the field to identify published and unpublished literature in any language from January, 1975, to May, 2013, on the associations between smoke-free legislation and our outcomes of interest. The primary outcomes were preterm birth, low birthweight, and hospital attendance for asthma. We identified 11 studies showing that smoke-free legislation was associated with significant reductions in preterm birth and severe asthma exacerbations. Studies have since addressed various knowledge gaps identified in our previous review, including assessments of the effect of smoke-free legislation on respiratory tract infections, the most important contributor to the global burden of paediatric morbidity and mortality associated with tobacco smoke exposure. The increased number of studies now available was also anticipated to allow investigation of another knowledge gap: exploration of a potential dose–response association between the comprehensiveness of smoke-free laws and their effect on child health. Furthermore, we sought to substantially broaden the focus of our study by evaluating the early-life health effect of the entire range of WHO-recommended tobacco control policies (ie, MPOWER). Following a prespecified and peer-reviewed protocol, we did a comprehensive literature search for experimental and quasi-experimental studies assessing associations between implementation of MPOWER policies and key perinatal and childhood outcomes associated with tobacco smoke exposure.

o control policies (ie, MPOWER). Following a prespecified and peer-reviewed protocol, we did a comprehensive literature search for experimental and quasi-experimental studies assessing associations between implementation of MPOWER policies and key perinatal and childhood outcomes associated with tobacco smoke exposure. Added value of this study To our knowledge, this is the first systematic review examining the association between the full spectrum of MPOWER policies and perinatal and child health. Our findings add value to the existing evidence base by identifying a link between smoke-free legislation and a substantial reduction in severe paediatric respiratory tract infections, providing consistent evidence that comprehensive smoke-free laws are associated with broad health effects, and collating evidence supporting the potential for other MPOWER measures to benefit child health. We also identified several key knowledge gaps, including a shortage of studies in low-income and middle-income countries, and of studies assessing MPOWER measures other than smoke-free legislation, tobacco tax increases, and smoking cessation services. Implications of all the available evidence

To our knowledge, this is the first systematic review examining the association between the full spectrum of MPOWER policies and perinatal and child health. Our findings add value to the existing evidence base by identifying a link between smoke-free legislation and a substantial reduction in severe paediatric respiratory tract infections, providing consistent evidence that comprehensive smoke-free laws are associated with broad health effects, and collating evidence supporting the potential for other MPOWER measures to benefit child health. We also identified several key knowledge gaps, including a shortage of studies in low-income and middle-income countries, and of studies assessing MPOWER measures other than smoke-free legislation, tobacco tax increases, and smoking cessation services. Implications of all the available evidence With most of the world's population currently not covered by comprehensive tobacco control policies, there is great potential for global public health gains by protecting unborn babies and children from tobacco smoke exposure. Future efforts should focus on increasing the uptake of comprehensive MPOWER policies worldwide to protect the health of children, while developing and evaluating new and ongoing tobacco control policy initiatives around the world.

ublic health gains by protecting unborn babies and children from tobacco smoke exposure. Future efforts should focus on increasing the uptake of comprehensive MPOWER policies worldwide to protect the health of children, while developing and evaluating new and ongoing tobacco control policy initiatives around the world. In a previous systematic review,15 we partly addressed this gap in the literature by synthesising available evidence on the effect of smoke-free legislation (ie, “P” in MPOWER, for ”Protect people from tobacco smoke”) on perinatal and child health. By combining data from 11 studies, we found smoke-free legislation to be associated with substantial reductions in preterm birth and hospital admissions for asthma among children. Studies have since addressed various knowledge gaps identified in our review, including assessments of the effect of smoke-free legislation on respiratory tract infections and on general practitioner (GP) consultations.16, 17, 18, 19 The increased number of studies now available was also anticipated to allow investigation of another knowledge gap: exploration of a potential dose–response association between the comprehensiveness of smoke-free laws and their effect on child health. In addition to addressing this association, we sought to substantially broaden the focus of our systematic review by systematically evaluating the early-life health effect of the entire range of MPOWER measures. This analysis has implications for the Sustainable Development Goal 3 (SDG 3) aims to strengthen FCTC implementation and reduce child mortality. As such, findings from this study can guide policy making for prioritisation of the most effective tobacco control policies to protect child health, especially in parts of the world where MPOWER implementation is lagging behind, while identifying the key remaining knowledge gaps that need to be addressed.

ld mortality. As such, findings from this study can guide policy making for prioritisation of the most effective tobacco control policies to protect child health, especially in parts of the world where MPOWER implementation is lagging behind, while identifying the key remaining knowledge gaps that need to be addressed. Methods Search strategy and selection criteria This systematic review and meta-analysis was done according to a peer-reviewed protocol that is published20 and registered with PROSPERO (CRD42015023448). We followed the PRISMA checklist when reporting our findings.21 Ethical approval was not required for this study. Studies were eligible for inclusion if they investigated the association between one or more MPOWER tobacco control policies and health outcomes among fetuses, neonates, or children (ie, the majority of the study population aged <12 years).

Methods Search strategy and selection criteria This systematic review and meta-analysis was done according to a peer-reviewed protocol that is published20 and registered with PROSPERO (CRD42015023448). We followed the PRISMA checklist when reporting our findings.21 Ethical approval was not required for this study. Studies were eligible for inclusion if they investigated the association between one or more MPOWER tobacco control policies and health outcomes among fetuses, neonates, or children (ie, the majority of the study population aged <12 years). We searched for published studies in the following databases: Cochrane Central Register of Controlled Trials (CENTRAL), MEDLINE, Embase, PsycINFO, Cumulative Index to Nursing and Allied Health Literature (CINAHL), WHO Global Health Library (in addition to MEDLINE, covering African Index Medicus [AIM], LILACS, Index Medicus for the Eastern Mediterranean Region [IMEMR], Index Medicus for South-East Asia Region [IMSEAR], Western Pacific Region Index Medicus [WPRIM], WHO Library Database [WHOLIS], and Scientific Electronic Library Online [SciELO]), IndMED, ISI Web of Science, KoreaMed, EconLit, Paediatric Economic Database Evaluation (PEDE), Google Scholar, and the ProQuest database of PhD dissertations. We searched the WHO International Clinical Trials Registry Platform (ICTRP) for unpublished studies.

WHOLIS], and Scientific Electronic Library Online [SciELO]), IndMED, ISI Web of Science, KoreaMed, EconLit, Paediatric Economic Database Evaluation (PEDE), Google Scholar, and the ProQuest database of PhD dissertations. We searched the WHO International Clinical Trials Registry Platform (ICTRP) for unpublished studies. The appendix (p 1) contains an overview of the search strategies for each database. We did not apply any language restrictions, and searched the full time period available for each database. Searches were updated on June 22, 2017. To identify any additional relevant studies, we hand-searched reference lists of, and citations to, included studies and relevant review papers, and consulted experts in the field (appendix p 2).

anguage restrictions, and searched the full time period available for each database. Searches were updated on June 22, 2017. To identify any additional relevant studies, we hand-searched reference lists of, and citations to, included studies and relevant review papers, and consulted experts in the field (appendix p 2). We focused on studies that evaluated governmental public health interventions that could be classified according to the MPOWER acronym (panel), with the exception of “M” since “Monitoring tobacco use and prevention policies” itself was not expected to affect health outcomes. We followed the methodological approach recommended by the Cochrane Effective Practice and Organization of Care (EPOC) group to select studies with the most robust designs for our primary analyses: randomised controlled trials (including cluster randomised controlled trials), controlled clinical trials (including cluster controlled clinical trials), interrupted time series studies (including difference-in-difference designs, which were categorised as controlled interrupted time series studies),22 and controlled before-and-after studies. To assess the robustness of our findings, we also included non-EPOC study designs in sensitivity analyses: uncontrolled before-and-after studies, prospective or retrospective cohort studies, and case-control and nested case-control studies. Primary and secondary outcomes were selected on the basis of their established associations with maternal smoking during pregnancy and prenatal or childhood second-hand smoke exposure,23, 24 and their relative contributions to the global burden of adverse child health.1, 25 Our primary outcomes of interest were perinatal mortality, preterm birth, asthma exacerbations requiring hospital attendance, and respiratory tract infections requiring hospital attendance. Secondary outcomes of interest were stillbirth, early neonatal mortality, neonatal mortality, late neonatal mortality, post-neonatal mortality, infant mortality, child mortality, extremely low birthweight, very low birthweight, low birthweight, birthweight (continuous scale), very small for gestational age, small for gestational age, extremely preterm birth, very preterm birth, gestational age (continuous scale), congenital anomalies, asthma, wheezing, respiratory tract infections, upper respiratory tract infections, lower respiratory tract infections, otitis media with effusion, and chronic cough.

ll for gestational age, small for gestational age, extremely preterm birth, very preterm birth, gestational age (continuous scale), congenital anomalies, asthma, wheezing, respiratory tract infections, upper respiratory tract infections, lower respiratory tract infections, otitis media with effusion, and chronic cough. Studies were excluded if they only measured smoking prevalence, smoking behaviour, second-hand smoke exposure, surrogate outcomes, or economic outcomes. Studies that reported outcomes for both adults and children were included if paediatric subgroup data were available. Data analysis Two reviewers (TF and AK) independently assessed all search results by title and abstract, and by full text for potential eligible studies identified. Any disagreements were resolved through joint discussion or via an adjudicator (JVB). Relevant data were extracted with a customised data extraction form (appendix, pp 3–5). Study authors were contacted for clarification where necessary and to obtain relevant data that were missing from the reports. A risk-of-bias assessment form was created on the basis of EPOC criteria for interrupted time series and controlled before-and-after studies.26 The Effective Public Health Practice Project (EPHPP) tool was adapted to assess the risk of bias of observational studies.27 Two reviewers (TF and AK) independently extracted data and assessed risk of bias, with disagreements resolved through discussion or arbitration (JVB).

s and controlled before-and-after studies.26 The Effective Public Health Practice Project (EPHPP) tool was adapted to assess the risk of bias of observational studies.27 Two reviewers (TF and AK) independently extracted data and assessed risk of bias, with disagreements resolved through discussion or arbitration (JVB). Point estimates and corresponding 95% CIs for effect sizes or association measures were extracted. For dichotomous outcomes, risk ratios (RRs) were extracted. Where RRs were not available, we calculated RRs from odds ratios (ORs) using the following formula, where PEER is the patient-expected event rate in the control group: RR=OR/(1-PEER)+(PEER × OR) When PEER was not available in interrupted time series studies we used the overall event rate across the study population as an approximation. For outcomes that could occur more than once (eg, hospital attendances for asthma and respiratory tract infections), we used incidence rate ratios (IRRs).

Point estimates and corresponding 95% CIs for effect sizes or association measures were extracted. For dichotomous outcomes, risk ratios (RRs) were extracted. Where RRs were not available, we calculated RRs from odds ratios (ORs) using the following formula, where PEER is the patient-expected event rate in the control group: RR=OR/(1-PEER)+(PEER × OR) When PEER was not available in interrupted time series studies we used the overall event rate across the study population as an approximation. For outcomes that could occur more than once (eg, hospital attendances for asthma and respiratory tract infections), we used incidence rate ratios (IRRs). Aggregated effect estimates were calculated to assess the association between each tobacco control policy and individual health outcomes, where feasible. Relative risk differences were extracted or calculated from absolute risk differences and were pooled in random-effects meta-analyses given anticipated heterogeneity. Step changes (ie, immediate risk changes) following introduction of an intervention were pooled in separate analyses from slope changes (ie, gradual risk changes). Heterogeneity was assessed by the I2 statistic. For the meta-analyses, we selected the effect estimate of the most comprehensive intervention within each MPOWER category from each study. In case of overlapping populations between studies, we selected one study according to the following hierarchy: the lowest risk of bias, the most comprehensive intervention, or the largest study population. We also extracted data on changes in smoking behaviour and second-hand smoke exposure if reported. The comprehensiveness of smoke-free legislation was assessed by counting the number of locations that were made completely smoke-free, out of eight prespecified options as suggested by WHO.2 Policies that were completely smoke-free in all eight locations were considered to be comprehensive.

smoke exposure if reported. The comprehensiveness of smoke-free legislation was assessed by counting the number of locations that were made completely smoke-free, out of eight prespecified options as suggested by WHO.2 Policies that were completely smoke-free in all eight locations were considered to be comprehensive. We did sensitivity analyses to explore the robustness of our findings by reanalysing the data for the primary outcomes with the addition of non-EPOC studies, and by restricting analyses to studies with low risk of bias and moderate risk of bias. Where possible, we did subgroup analyses according to the comprehensiveness of each intervention. Where possible, the effect of each intervention was reported according to socioeconomic status, alongside its overall effect. We assessed risk of bias across studies using funnel plots when ten or more studies were included in a meta-analysis. Role of the funding source The funders 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 assessed risk of bias across studies using funnel plots when ten or more studies were included in a meta-analysis. Role of the funding source The funders 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 We identified 25 478 citations from bibliographic databases and an additional 20 from other sources. After removal of duplicates, 12 392 unique citations were screened by title and abstract, and 65 full texts were sourced. Of these, 41 EPOC studies16, 17, 18, 19, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64 and three non-EPOC studies65, 66, 67 fit the inclusion criteria (figure 1; appendix, pp 6, 7). The EPOC studies included data from more than 57 million births, and from 4·6 million GP diagnoses and 2·7 million hospital admissions for respiratory conditions.Figure 1 PRISMA flow diagram EPOC= Effective Practice and Organization of Care (a Cochrane Review Group). MPOWER=WHO's recommended tobacco control policies (see panel).

Results We identified 25 478 citations from bibliographic databases and an additional 20 from other sources. After removal of duplicates, 12 392 unique citations were screened by title and abstract, and 65 full texts were sourced. Of these, 41 EPOC studies16, 17, 18, 19, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64 and three non-EPOC studies65, 66, 67 fit the inclusion criteria (figure 1; appendix, pp 6, 7). The EPOC studies included data from more than 57 million births, and from 4·6 million GP diagnoses and 2·7 million hospital admissions for respiratory conditions.Figure 1 PRISMA flow diagram EPOC= Effective Practice and Organization of Care (a Cochrane Review Group). MPOWER=WHO's recommended tobacco control policies (see panel). The appendix (pp 8–20) details the main characteristics of the EPOC studies. Among these, 26 were interrupted time series studies,16, 17, 18, 19, 31, 32, 37, 38, 40, 42, 43, 45, 46, 48, 50, 51, 52, 54, 55, 57, 58, 59, 61, 62, 63, 64 14 were controlled interrupted time series studies,28, 29, 33, 34, 35, 36, 39, 41, 44, 47, 49, 53, 56, 60 and one had a regression discontinuity design,30 a quasi-experimental design bearing close resemblance to interrupted time series methodology.68 The three non-EPOC studies were uncontrolled before-and-after studies (appendix, p 21).65, 66, 67 Model characteristics of individual studies can be found in the appendix (pp 22–31). The EPOC studies were done in 14 countries across North America (24 studies)18, 28, 29, 31, 33, 35, 36, 38, 39, 41, 42, 43, 45, 46, 47, 49, 50, 53, 54, 56, 57, 59, 60, 61 and Europe (16 studies),16, 17, 30, 32, 34, 37, 40, 44, 48, 51, 52, 55, 58, 62, 63, 64 with one study from Hong Kong, China.19 Several US studies assessed the same outcomes in partially overlapping study populations.18, 29, 31, 36, 38, 43, 45, 46, 49, 53, 56, 59, 61

2, 43, 45, 46, 47, 49, 50, 53, 54, 56, 57, 59, 60, 61 and Europe (16 studies),16, 17, 30, 32, 34, 37, 40, 44, 48, 51, 52, 55, 58, 62, 63, 64 with one study from Hong Kong, China.19 Several US studies assessed the same outcomes in partially overlapping study populations.18, 29, 31, 36, 38, 43, 45, 46, 49, 53, 56, 59, 61 Risk of bias of individual studies is reported in detail in the appendix (pp 32, 33). For the EPOC studies, risk of bias was low in 23 studies,16, 17, 18, 28, 30, 31, 32, 33, 36, 37, 39, 40, 44, 45, 46, 47, 48, 51, 52, 53, 58, 63, 64 moderate in 16,19, 34, 35, 41, 42, 43, 49, 50, 54, 55, 56, 57, 59, 60, 61, 62 and high in two.29, 38 For the non-EPOC studies, risk of bias was high for two studies66, 67 and unclear for one.65

risk of bias was low in 23 studies,16, 17, 18, 28, 30, 31, 32, 33, 36, 37, 39, 40, 44, 45, 46, 47, 48, 51, 52, 53, 58, 63, 64 moderate in 16,19, 34, 35, 41, 42, 43, 49, 50, 54, 55, 56, 57, 59, 60, 61, 62 and high in two.29, 38 For the non-EPOC studies, risk of bias was high for two studies66, 67 and unclear for one.65 28 studies assessed the association between smoke-free legislation and one or more primary outcomes (ie, perinatal mortality, preterm birth, asthma exacerbations requiring hospital attendance, and respiratory tract infections requiring hospital attendance), five assessed the association between tobacco taxation and primary outcomes, and two assessed the association between policies providing smoking cessation services and our primary outcomes (Table 1, Table 2, Table 3); four studies assessed a combination of these interventions, and ten studies only assessed secondary outcomes. A meta-analysis was only possible for studies on smoke-free legislation because studies on tax increases and smoking cessation services had variable outcome reporting and overlapping study populations.Table 1 Association between implementation of smoke-free legislation and primary outcomes

assessed secondary outcomes. A meta-analysis was only possible for studies on smoke-free legislation because studies on tax increases and smoking cessation services had variable outcome reporting and overlapping study populations.Table 1 Association between implementation of smoke-free legislation and primary outcomes Details of intervention Population at risk (n) Events (n) Slope before intervention (% change in events per year) Direct change in events (step change, %; 95% CI) Sustained change in events per year (slope change, %; 95% CI) Summary of findings Perinatal mortality Peelen (2016)58* First smoke-free law: workplaces and public transport except for restaurants and bars† (allowing designated smoking areas) Second smoke-free law: expansion of first smoke-free law to include restaurants and bars‡(allowing designated smoking areas) 1 980 727 13 027 NA because of non-linear time trend First smoke-free law: −1·99% (−8·95 to 5·96) Second smoke-free law: −5·96% (−12·93 to 1·99) NA National smoke-free workplaces and public transport, and smoke-free restaurants and bars, were not associated with significant changes in perinatal mortality Preterm birth Bakolis (2016)30 Public places and workplaces (including restaurants and bars) 1 800 906 126 527 NR Analysis of a 1, 2, 3, or 5 month time window around the intervention cutoff date (July 1, 2007): ±1 month, −4·67% (−16·00 to −0·93); ±2 months, −8·42% (−15·05 to −1·86); ±3 months, −5·60% (−10·31 to −0·93); ±5 months, −3·73% (−7·48 to −0·93) NA National comprehensive smoke-free legislation was associated with an immediate 4–9% decrease in preterm births Bartholomew (2016)31 Comprehensive (workplaces, restaurants, and bars) Restrictive (workplaces and restaurants, no restriction in bars) Moderate (workplaces, partial restriction in restaurants, and no restriction in bars) Limited (partial restriction in workplaces, any restriction in restaurants, and no restriction in bars) 293 715 32 250 NR Comprehensive: −0·015%§ (−0·022 to −0·008) Restrictive: 0·003%§ (−0·005 to 0·011) Moderate: 0·004%§ (−0·002 to 0·010) Limited: 0·001%§ (−0·006 to 0·007) NA County-wide, comprehensive smoke-free legislation was associated with a 0·015 percentage point decrease in preterm births, whereas less restrictive laws were not associated with changes in incidence of preterm births Bharadwaj (2014)34 Restaurants and bars (in addition to existing smoke-free laws in public places and workplaces) 822 (intervention group), 3185 (control group) 46 (interventio

·015 percentage point decrease in preterm births, whereas less restrictive laws were not associated with changes in incidence of preterm births Bharadwaj (2014)34 Restaurants and bars (in addition to existing smoke-free laws in public places and workplaces) 822 (intervention group), 3185 (control group) 46 (interventio n group), 189 (control group) NR −2·55%§ (−5·52 to 0·42) NA National smoke-free restaurants and bars were not associated with significant changes in preterm births among women working in restaurants and bars Cox (2013)37 Public places and workplaces (excluding catering industry); restaurants (in addition to existing smoke-free laws in public places and workplaces); and bars serving food (in addition to existing smoke-free laws in public places and workplaces, including restaurants) 606 877 36 663 NR Public places and workplaces: single smoke-free law¶, −0·59% (−2·63 to 1·49); final model‖, no significant changes Restaurants (in addition to public places and workplaces): single smoke-free law¶, −2·28% (−4·73 to −0·15); final model‖, −3·18% (−5·38 to −0·94) Bars serving food (in addition to restaurants and public places and workplaces): single smoke-free law¶, −1·24% (−3·05 to 0·60); final model‖, no significant changes Public places and workplaces: single smoke-free law†, −1·95% (−3·50 to −0·37); final model‡, no significant changes Restaurants (in addition to public places and workplaces): single smoke-free law†, −1·42% (−2·87 to 0·05); final model‡, no significant changes Bars serving food (in addition to restaurants and public places and workplaces): single smoke-free law†, −2·10% (−4·82 to 0·69); final model‡, −3·50% (−6·35 to −0·57) National smoke-free public places and workplaces were not associated with significant changes in preterm births; expansion of national smoke-free legislation to include restaurants was associated with an immediate 3·2% reduction in preterm births; and expansion of national smoke-free legislation to include bars was associated with a gradual 4% per year decrease in preterm births Hade (2011)43 Public places and workplaces (including restaurants and bars) 583 530 NR NR No significant changes** No significant changes** State-wide, smoke-free public places and workplaces were not associated with significant changes in preterm birth Hajdu (2017)44 Public places and workplaces (including restaurants and bars) 18 755 NR NR –1·9%§ (–4·3 to 0·5) NA National smoke-free legislation was not associated with significant changes in preterm birt

State-wide, smoke-free public places and workplaces were not associated with significant changes in preterm birth Hajdu (2017)44 Public places and workplaces (including restaurants and bars) 18 755 NR NR –1·9%§ (–4·3 to 0·5) NA National smoke-free legislation was not associated with significant changes in preterm birt h among female restaurant and bar workers compared with women working in places other than restaurants and bars Hankins (2016)45 Workplaces, restaurants, and bars NR NR NR Workplaces: 0·07%§ (−0·11 to 0·25) Restaurants: 0·09%§ (−0·13 to 0·31) Bars: −0·29%§ (−0·49 to −0·09) NA State-wide or county smoke-free workplaces and restaurants were not associated with significant changes in preterm births; state-wide or county smoke-free bars were associated with an immediate 0·3 percentage point decrease in preterm births Hawkins (2014)46 100% smoke-free workplaces and restaurants 16 198 654 1 555 071 NR 0·72%§ (−0·11 to 1·55) NA State-wide smoke-free workplaces and restaurants were not associated with significant changes in preterm births Mackay (2012)52 Public places and workplaces (including restaurants and bars) 709 756 41 998 NR Crude: −11·07% (−15·15 to −6·79) Adjusted: −11·72% (−15·87 to −7·35) Crude: 2·28% (−0·03 to 4·66) Adjusted: 3·83% (1·42 to 6·30) National smoke-free public places and workplaces were associated with an immediate 12% decrease in preterm births, and a subsequent gradual 4% increase per year Markowitz (2013)53 Workplaces with complete smoke-free law Workplaces with smoking restrictions (requiring designated smoking areas) Restaurants with complete smoke-free laws Restaurants with smoking restrictions (requiring designated smoking areas) Maternal age <20 years: 54 132Maternal age 20–24 years: 101 723Maternal age 25–34 years: 183 763Maternal age >34 years: 53 109 Maternal age <20 years: 5413Maternal age 20–24 years: 7120Maternal age 25–34 years: 11 026Maternal age >34 years: 3718 NR Workplaces with complete smoke-free laws: NRWorkplaces with smoking restrictions: NRRestaurants with complete smoke-free laws: maternal age <20 years, 0·7%§ (−3·5 to 4·9);20–24 years, −0·2%§ (−1·5 to 1·1); 25–34 years, −0·3%§ (−0·8 to 0·2); >34 years, −0·6%§ (−1·9 to 0·7) Restaurants with smoking restrictions: maternal age <20 years, −0·6%§ (−3·8 to 2·6); 20–24 years, −0·1%§ (−1·1 to 0·9); 25–34 years, −0·8%§ (−1·2 to −0·4); >34 years, −0·3%§ (−1·7 to 1·1) NA State-wide complete smoke-free laws were not associated with significant changes in preterm births, but state-

·6%§ (−1·9 to 0·7) Restaurants with smoking restrictions: maternal age <20 years, −0·6%§ (−3·8 to 2·6); 20–24 years, −0·1%§ (−1·1 to 0·9); 25–34 years, −0·8%§ (−1·2 to −0·4); >34 years, −0·3%§ (−1·7 to 1·1) NA State-wide complete smoke-free laws were not associated with significant changes in preterm births, but state- wide restaurant smoking restrictions were associated with a 0·8 percentage point decrease in preterm births among women aged 25–34 years McKinnon (2015)54 Public places and workplaces (including restaurants and bars) 470 199 19 321 NR Crude: −6% (−10 to −1) Adjusted: −5% (−10 to 0) NA State-wide smoke-free legislation was associated with a 5% decrease in preterm births 9 months after its implementation Page (2012)56 Public places and workplaces (including restaurants and bars) 6717 (intervention group), 32 293 (control group) 515 (intervention group), 2767 (control group) NR Crude: −20·6% (−34·7 to −3·4) Adjusted: −23·1% (−40·1 to −1·3) NA City-wide smoke-free public places and workplaces were associated with a 23% decrease in preterm births Peelen (2016)58* First smoke-free law†: workplaces and public transport except for restaurants and bars (allowing designated smoking areas) Second smoke-free law‡: expansion of first smoke-free law to include restaurants and bars (allowing designated smoking areas) 1 972 163 116 043 NA because of non-linear time trend First smoke-free law: 0·94% (−1·89 to 3·77) Second smoke-free law: −0·94% (−3·78 to 2·83) NA National smoke-free workplaces and public transport, and smoke-free restaurants and bars, were not associated with significant changes in preterm births Simón (2017)62 First smoke-free law: complete smoke-free workplaces and partial smoke-free restaurants and bars Second smoke-free law: public places and workplaces (including restaurants and bars) 5 302 374 416 595 NR First smoke-free law: 4·6% (2·9 to 6·2) Second smoke-free law: −4·5% (–6·1 to −2·9) NA National partial smoke-free legislation was associated with a 5% increase in preterm births; the subsequent national comprehensive smoke-free legislation was associated with a 5% decrease in preterm births Vicedo-Cabrera (2016)63 Public places and workplaces (including restaurants and bars), with several exceptions in the hospitality sector†† 446 492 24 482 NR −3·56% (−9·29 to 2·53) NA Federal smoke-free legislation was not associated with a significant change in preterm births Asthma exacerbations requiring hospital attendance Ciaccio (2016)36 Public places and workplac

s (including restaurants and bars), with several exceptions in the hospitality sector†† 446 492 24 482 NR −3·56% (−9·29 to 2·53) NA Federal smoke-free legislation was not associated with a significant change in preterm births Asthma exacerbations requiring hospital attendance Ciaccio (2016)36 Public places and workplac es (including restaurants and bars) 13 246 809 335 588 NR –17% (–18 to −15) NA State or local smoke-free legislation was associated with an immediate 17% decrease in emergency department visits for asthma Croghan (2015)38 Public places and workplaces (including restaurants and bars) NR 1531 1·1% (0·2 to 2·0) −24·9% (−40·5 to −5·3) −1·5% (−2·9 to −0·1) National smoke-free legislation was associated with an immediate 25% decrease in emergency department visits for children with asthma, and a subsequent gradual 1·5% decrease per year Galán (2017)40 First smoke-free law: complete smoke-free workplaces and partial smoke-free restaurants and bars Second smoke-free law: public places and workplaces (including restaurants and bars) NR NR NR First smoke-free law: 25·0% (–2·6 to 60·4) Second smoke-free law: −11·0% (–28·6 to 11·1) NA Partial and comprehensive national smoke-free legislation were not associated with significant immediate changes in asthma-related hospital admissions via emergency departments Gaudreau (2013)42 Public places and workplaces (including restaurants and bars), allowing designated smoking areas NR 3050 NR 11% (−37 to 95) 0% (−2 to 2) Provincial smoke-free public places and workplaces were not associated with significant changes in hospital admissions for paediatric asthma Hawkins (2016)18 State or local 100% smoke-free workplaces or restaurants, or both NR 128 807 NR State: −3% (−8 to 2) Local: 2% (−6 to 11) NA State or local smoke-free workplaces or restaurants were not associated with significant changes in emergency department visits for paediatric asthma Landers (2014)49 100% smoke-free workplaces, restaurants, and bars‡‡ NR NR Mean rate across all states and years: 9·02 per 10 000 per quarter (SD 9·66; range 0·00–144·47) Any state law: 0·12%§ (−0·38 to 0·62) Any county law: −1·32%§ (−2·64 to 0·00) Interaction term of state law and county law: 0·51%§ (−1·04 to 2·06) NA County-level smoke-free laws were associated with a one percentage point decrease in discharge rates among children admitted for asthma; state smoke-free laws were not associated with significant changes in discharge rates among children admitted for asthma, besides the effect o

nd county law: 0·51%§ (−1·04 to 2·06) NA County-level smoke-free laws were associated with a one percentage point decrease in discharge rates among children admitted for asthma; state smoke-free laws were not associated with significant changes in discharge rates among children admitted for asthma, besides the effect o f county laws Mackay (2010)51 Public places and workplaces (including restaurants and bars) NR 21 415 4·4% (3·3 to 5·5) NA −19·5% (−22·4 to −16·5) National smoke-free public places and workplaces were associated with a gradual 20% decrease per year in paediatric emergency asthma admissions Millett (2013)55 Public places and workplaces (including restaurants and bars) NR 217 381 2·2% (2 to 3) −8·9% (−11 to −7) −3·4% (−4 to −2) National smoke-free public places and workplaces were associated with an immediate 9% decrease in emergency admissions to hospital for paediatric asthma, and a subsequent gradual 3% decrease per year Rayens (2008)59 Most businesses open to the public (including restaurants and bars)§§ 395 116 5322 12·7% −18·0% (−29·0 to −4·0) NA The county-wide smoke-free law in most public places was associated with an 18% decrease in emergency department visits for paediatric asthma Shetty (2011)61 All workplaces except restaurants and bars: 100% smoke-free Any smoke-free workplace, restaurant, or bar law NR NR NR 100% smoke-free workplaces: 14·6% (3·7 to 25·5) Any smoke-free law: 9·0 (−1 to 19·1) NA State-wide or region-wide 100% smoke-free workplaces were associated with a 15% increase in hospital admissions for children with asthma; there was no evidence for an association between any state-wide or region-wide smoke-free legislation and asthma admissions RTI admissions (upper and lower) Been, Millett (2015)16 Public places and workplace (including restaurants and bars) NR 1 651 675 NR −3·5% (−4·7 to −2·3) −0·5% (−0·9 to −0·1) National smoke-free legislation was associated with an immediate 4% reduction and an additional 0·5% annual reduction in childhood acute RTI hospital admissions Vicedo-Cabrera (2017)64 Public places and workplaces (including restaurants and bars), with several exceptions in the hospitality sector†† NR 29 345 NR 2·7% (–9·7 to 16·7) NA Federal smoke-free legislation was not associated with a significant change in RTI hospital admissions Upper RTI admissions Been, Millett (2015)16 Public places and workplaces (including restaurants and bars) NR 979 370 NR 1·9% (0·5 to 3·2) −1·9 (−2·3 to −1·5) National smoke-free legislation was assoc

(–9·7 to 16·7) NA Federal smoke-free legislation was not associated with a significant change in RTI hospital admissions Upper RTI admissions Been, Millett (2015)16 Public places and workplaces (including restaurants and bars) NR 979 370 NR 1·9% (0·5 to 3·2) −1·9 (−2·3 to −1·5) National smoke-free legislation was assoc iated with an initial immediate 2% increase in childhood upper RTI hospital admissions, followed by a gradual decrease of 2% per year Hawkins (2016)18 State or local 100% smoke-free workplaces or restaurants, or both NR 410 686 NR State: −2% (−6 to 2) Local: 6% (−2 to 14) NA State or local smoke-free workplaces or restaurants were not associated with significant changes in emergency department visits for upper RTIs Lower RTI admissions Been, Millett (2015)16 Public places and workplaces (including restaurants and bars) NR 672 305 NR −13·8% (−15·6 to −12·0) 0·2% (−0·6 to 0·9) National smoke-free legislation was associated with an immediate 14% reduction in childhood lower RTI hospital admissions Hawkins (2016)18 State or local 100% smoke-free workplaces or restaurants, or both NR 139 239 NR State: −8% (−13 to −4) Local: 3% (−6 to 12) NA State-wide smoke-free workplaces or restaurants were associated with an 8% decrease in emergency department visits for lower RTIs Lee (2016)19 Public places and workplaces (including restaurants) 691 480 75 870 NR −33·5% (−36·4 to −30·5) −13·9% (−16·0 to −11·7) Comprehensive smoke-free legislation was associated with an immediate 34% reduction in hospital admissions for childhood lower RTIs, and a subsequent gradual decrease of 14% per year NA=not applicable.

Public places and workplaces (including restaurants) 691 480 75 870 NR −33·5% (−36·4 to −30·5) −13·9% (−16·0 to −11·7) Comprehensive smoke-free legislation was associated with an immediate 34% reduction in hospital admissions for childhood lower RTIs, and a subsequent gradual decrease of 14% per year NA=not applicable. NR=not reported. RTI=respiratory tract infection. * Both smoke-free laws were accompanied by a tobacco tax increase and mass-media campaign. † Exceptions to this smoke-free law were: hotels, bars and restaurants, sports, arts and culture venues, amusement arcades, tobacconist shops, international passenger transport systems, private spaces, open air, and designated areas for smoking within each facility. ‡ The smoke-free law now included hospitality venues: hotels, bars and restaurants, sports, art and culture venues, amusement arcades, tobacconist shops, and international passenger transport systems. Designated smoking areas within each facility were still allowed. § Percentage point change. ¶ The single smoke-free law model includes either the step or slope change of a single smoke-free law into the model. ‖ The final was obtained by including all three step changes and all three slope changes in one model and removing the least significant factors one at a time. ** No association measures were reported. †† Authorised smoking in establishments smaller than 80 m2 and designated smoking areas in larger establishments.

¶ The single smoke-free law model includes either the step or slope change of a single smoke-free law into the model. ‖ The final was obtained by including all three step changes and all three slope changes in one model and removing the least significant factors one at a time. ** No association measures were reported. †† Authorised smoking in establishments smaller than 80 m2 and designated smoking areas in larger establishments. ‡‡ Different states passed different 100% smoke-free laws: workplaces, restaurants, and bars (eight states); restaurants and bars (two states); workplaces and restaurants (one state); and workplaces (one state). §§ Including, but not limited to restaurants, bars, bowling alleys, bingo halls, convenience stores, laundromats, and other business open to the public. Table 2 Association between implementation of smoking cessation services and primary outcomes

‡‡ Different states passed different 100% smoke-free laws: workplaces, restaurants, and bars (eight states); restaurants and bars (two states); workplaces and restaurants (one state); and workplaces (one state). §§ Including, but not limited to restaurants, bars, bowling alleys, bingo halls, convenience stores, laundromats, and other business open to the public. Table 2 Association between implementation of smoking cessation services and primary outcomes Details of intervention Population at risk (n) Events (n) Slope before intervention (% change in events per year) Direct change in events (step change, %; 95% CI) Sustained change in events per year (slope change, %; 95% CI) Summary of findings Preterm birth Jarlenski (2014)47 State adoption of one of two optional Medicaid enrolment policies, allowing more low-income pregnant women to receive prenatal care, including smoking cessation services (presumptive eligibility and the unborn child option)* 24 544 NR NR Overall: −1·4%§ (−4·7 to 2·0)Comprehensive: −2·2%§ (−5·9 to 1·5)Non-comprehensive: 1·3%§ (−2·4 to 5·1) NA Neither optional Medicaid enrolment policy was associated with significant changes in preterm birth Asthma exacerbations requiring hospital attendance Hawkins (2016)18 Health reform legislation that provided counselling for smoking cessation and tobacco cessation treatment to Medicaid recipients NR 112 808 NR 2% (−4 to 8) NA The state-wide health reform legislation in MA, USA, was not associated with significant changes in emergency department visits for asthma Upper RTI admissions Hawkins (2016)18 Health reform legislation that provided counselling for smoking cessation and tobacco cessation treatment to Medicaid recipients NR 337 628 NR −6% (−10 to −1) NA The state-wide health reform legislation in MA, USA, was associated with a 6% decrease in emergency department visits for upper RTIs Lower RTI admissions Hawkins (2016)18 Health reform legislation that provided counselling for smoking cessation and tobacco cessation treatment to Medicaid recipients NR 113 137 NR 0% (−6 to 6) NA The state-wide health reform legislation in MA, USA, was not associated with significant changes in emergency department visits for lower RTIs NR=not reported. NA=not applicable. RTI=respiratory tract infection.

* Presumptive eligibility: low-income pregnant women are presumed to be eligible for Medicaid, so they can receive care (including smoking cessation services) while their Medicaid applications are still pending. The unborn-child option: the state can consider a fetus a “targeted low-income child”, allowing coverage of prenatal care (including smoking cessation services) and delivery to low-income pregnant women, even if they cannot provide documentation of citizenship or residency. Table 3 Association between implementation of tobacco taxation and primary outcomes Details of intervention Population at risk (n) Events (n) Slope before intervention (% change in events per year) Direct change in events (step change, %; 95% CI) Sustained change in events per year (slope change, %; 95% CI) Summary of findings Preterm birth Hawkins (2014)46 Effect of cigarette excise tax increase (in USD$; December 2010 rates) on mothers, by years of maternal education 9 981 855 NR NR White mothers: 0–11 years, −0·07%§ (−0·11 to −0·02); 12 years, −0·02%§ (−0·05 to 0·01); 13–15 years, −0·01%§ (−0·03 to 0·00); ≥16 years, −0·00%§ (−0·01 to 0·01) per USD$ increase in tax NA Cigarette taxes were associated with a decrease in preterm birth among white mothers with the least amount of education Hawkins (2014)46 Effect of cigarette excise tax increase (in USD$; December 2010 rates) on mothers, by years of maternal education 2 722 846 NR NR Black mothers: 0–11 years, −0·08%§ (−0·14 to −0·03); 12 years, −0·04%§ (−0·07 to −0·01); 13–15 years, −0·03%§ (−0·05 to −0·01); ≥16 years, −0·01%§ (−0·01 to −0·00) per USD$ increase in tax NA Cigarette taxes were associated with a decrease in preterm births among black mothers with any level of education; among black mothers, there was a gradient across maternal education levels, with the largest decreases among mothers with the least amount of education Hawkins (2014)46 Effect of cigarette excise tax increase (in USD$; December 2010 rates) on mothers, by years of maternal education 2 444 673 NR NR Hispanic mothers: 0–11 years, 0·01%§ (−0·00 to 0·02); 12 years, −0·00%§ (−0·01 to 0·00); 13–15 years, −0·01%§ (−0·02 to 0·00); ≥16 years, −0·00%§ (−0·00 to 0·00) per USD$ increase in tax NA Cigarette taxes were not associated with significant changes in preterm births among Hispanic mothers with any level of education Hawkins (2014)46 Effect of cigarette excise tax increase (in USD$; December 2010 rates) on mothers, by years of maternal education 804 447 NR NR Asian/Pacific Islander mothers: 0–11 years, 0·01%§ (−0·01 to 0·04); 12 years, −0·01%§ (−0·01 to 0·00); 13–15 years, −0·00%§ (−0·01 to 0·01); ≥16 years, 0·00%§ (−0·00 to 0·00) per USD$ increase in tax NA Cigarette taxes were not associated with significant changes in

others, by years of maternal education 804 447 NR NR Asian/Pacific Islander mothers: 0–11 years, 0·01%§ (−0·01 to 0·04); 12 years, −0·01%§ (−0·01 to 0·00); 13–15 years, −0·00%§ (−0·01 to 0·01); ≥16 years, 0·00%§ (−0·00 to 0·00) per USD$ increase in tax NA Cigarette taxes were not associated with significant changes in preterm births among Asian/Pacific Islander mothers with any level of education Hawkins (2014)46 Effect of cigarette excise tax increase (in USD$; December, 2010, rates) on mothers, by years of maternal education 244 823 NR NR Native American/Alaska Native mothers: 0–11 years, −0·02%§ (−0·08 to 0·04); 12 years, 0·01%§ (−0·02 to 0·03);13–15 years, 0·00%§ (−0·03 to 0·03); ≥16 years, −0·01%§ (−0·02 to 0·01) per USD$ increase in tax NA Cigarette taxes were not associated with significant changes in preterm births among Native American/Alaska Native mothers with any level of education Markowitz (2013)53 Cigarette excise tax increase (in 2008 USD$)Cigarette price increase (in 2008 USD$) Maternal age <20 years: 54 132Maternal age 20–24 years: 101 723Maternal age 25–34 years: 183 763Maternal age >34 years: 53 109 Maternal age <20 years: 5413Maternal age 20–24 years: 7120Maternal age 25–34 years: 11 026Maternal age >34 years: 3718 NR Cigarette excise tax: maternal age <20 years, −2·0%§ (−4·0 to 0·0) per USD$ increase in tax;maternal age 20–24 years, −0·7%§ (−1·4 to −0·0) per USD$ increase in tax; maternal age 25–34 years, −0·2%§ (−1·0 to 0·6) per USD$ increase in tax; maternal age >34 years, −1·0%§ (−1·9 to −0·1) per USD$ increase in taxCigarette price: NR NA State-wide increases in cigarette excise tax were associated with a 0·7 percentage point decrease in preterm births among women aged 20–24 years, and a 1·0 percentage point decrease among women aged >34 years Asthma exacerbations requiring hospital attendance Hawkins (2016)18 Cigarette excise tax increase in USD$ NR 128 807 NR −5% (−11 to 1) per USD$ increase in tax NA State-wide increase in cigarette excise tax was not associated with significant changes in emergency department visits for paediatric asthma Landers (2014)49 Cigarette excise tax increase in USD$ NR NR Mean rate across all states and years: 9·02 per 10 000 (SD 9·66; range 0·00– 144·47) −0·53%§ (−0·99 to −0·06) per USD$ increase in tax NA State-wide increase in cigarette excise tax was associated with a 0·5 percentage point decrease in asthma discharge rates Ma (2013)50 USD$0·69 cigarette excise tax increase;USD$0·35 cigarette excise tax increase 28 49

rs: 9·02 per 10 000 (SD 9·66; range 0·00– 144·47) −0·53%§ (−0·99 to −0·06) per USD$ increase in tax NA State-wide increase in cigarette excise tax was associated with a 0·5 percentage point decrease in asthma discharge rates Ma (2013)50 USD$0·69 cigarette excise tax increase;USD$0·35 cigarette excise tax increase 28 49 8 070 702 771 0·04 USD$0·69 cigarette excise tax increase: −11·01% (−24·71 to 2·77);USD$0·35 cigarette excise tax increase: −22·02% (−33·46 to −9·95) USD$0·69 cigarette excise tax increase: 4·88% (1·29 to 8·59)USD$0·35 cigarette excise tax increase: −4·72% (−8·01 to −1·44) The first cigarette excise tax increase (USD$0·69) was not associated with significant immediate changes, but was associated with a significant, gradual increase in asthma-related hospital admissions of 0·5% per year; the second cigarette excise tax increase (USD$0·35) was associated with both a 22% immediate decrease as well as a gradual 5% decrease in asthma-related hospital admissions per year Upper RTI admissions Hawkins (2016)18 Cigarette excise tax increase in USD$ NR 410 686 NR −2% (−6% to 2%) per USD$ increase in tax NA State-wide increase in cigarette excise tax was not associated with significant changes in emergency department visits for upper RTIs Lower RTI admissions Hawkins (2016)18 Cigarette excise tax increase in USD$ NR 139 239 NR −9% (−16 to −2) per USD$ increase in tax NA State-wide increase in cigarette excise tax was associated with a 9% decrease in emergency department visits for lower RTIs NA=not applicable. NR=not reported. RTI=respiratory tract infection.

8 070 702 771 0·04 USD$0·69 cigarette excise tax increase: −11·01% (−24·71 to 2·77);USD$0·35 cigarette excise tax increase: −22·02% (−33·46 to −9·95) USD$0·69 cigarette excise tax increase: 4·88% (1·29 to 8·59)USD$0·35 cigarette excise tax increase: −4·72% (−8·01 to −1·44) The first cigarette excise tax increase (USD$0·69) was not associated with significant immediate changes, but was associated with a significant, gradual increase in asthma-related hospital admissions of 0·5% per year; the second cigarette excise tax increase (USD$0·35) was associated with both a 22% immediate decrease as well as a gradual 5% decrease in asthma-related hospital admissions per year Upper RTI admissions Hawkins (2016)18 Cigarette excise tax increase in USD$ NR 410 686 NR −2% (−6% to 2%) per USD$ increase in tax NA State-wide increase in cigarette excise tax was not associated with significant changes in emergency department visits for upper RTIs Lower RTI admissions Hawkins (2016)18 Cigarette excise tax increase in USD$ NR 139 239 NR −9% (−16 to −2) per USD$ increase in tax NA State-wide increase in cigarette excise tax was associated with a 9% decrease in emergency department visits for lower RTIs NA=not applicable. NR=not reported. RTI=respiratory tract infection. A national study from the Netherlands, comprising 1 980 727 births, found no change in perinatal mortality following a law to prohibit smoking in workplaces and on public transport, or following expansion of the law to include restaurants and bars.58 In a study from England, comprising 10 291 113 births, comprehensive smoke-free legislation in public places and workplaces (including restaurants and bars) was associated with a reduction in stillbirths (–7·8%; 95% CI −18·0 to −3·5) and neonatal deaths (–7·6%; −11·7 to −3·4).32 The overall effect on perinatal mortality (ie, stillbirths and early neonatal deaths combined) was not reported in this study. Therefore, no meta-analysis was possible with these two studies.

bars) was associated with a reduction in stillbirths (–7·8%; 95% CI −18·0 to −3·5) and neonatal deaths (–7·6%; −11·7 to −3·4).32 The overall effect on perinatal mortality (ie, stillbirths and early neonatal deaths combined) was not reported in this study. Therefore, no meta-analysis was possible with these two studies. 15 studies investigated the association between smoke-free legislation and preterm births.30, 31, 34, 37, 43, 44, 45, 46, 52, 53, 54, 56, 58, 62, 63 In the meta-analysis, smoke-free legislation was associated with a significant immediate reduction in preterm births (ten studies, 27 530 183 individuals; −3·77% [95% CI −6·37 to −1·16]; figure 2A). Two studies caused some funnel plot asymmetry suggestive of publication bias, but this asymmetry was unlikely to have affected our findings (appendix p 34). No additional gradual change in preterm births was evident (two studies, 1 316 633 individuals; −0·01% per year [95% CI −6·76 to 6·73]; figure 3A). One study47 examined the association between provision of smoking cessation services and preterm births. Medicaid enrolment policies permitting low-income pregnant women to receive smoking cessation services were not associated with a change in preterm births (table 2).47 Reductions in preterm birth were observed after tobacco tax increases among women in specific population subgroups in two studies.46, 53 One study reported tobacco taxation to be associated with reduced rates of preterm birth among white mothers with low levels of education and among black mothers irrespective of level of education (table 3).46 The other study reported a 0·7 percentage point decrease in preterm births per USD$ increase in tax among women aged 20–24 years, and a 1·0 percentage point decrease per USD$ increase in tax among women older than 34 years (table 3).53Figure 2 Meta-analysis of immediate changes in primary outcomes after implementation of smoke-free legislation

a 0·7 percentage point decrease in preterm births per USD$ increase in tax among women aged 20–24 years, and a 1·0 percentage point decrease per USD$ increase in tax among women older than 34 years (table 3).53Figure 2 Meta-analysis of immediate changes in primary outcomes after implementation of smoke-free legislation (A) Preterm birth. (B) Asthma exacerbations requiring hospital attendance. (C) Respiratory tract infections requiring hospital attendance. (D) Lower respiratory tract infections requiring hospital attendance. (E) Upper respiratory tract infections requiring hospital attendance. Figure 3 Meta-analysis of gradual changes in primary outcomes after implementation of smoke-free legislation (A) Preterm birth. (B) Asthma exacerbations requiring hospital attendance. (C) Lower respiratory tract infections requiring hospital attendance.

(A) Preterm birth. (B) Asthma exacerbations requiring hospital attendance. (C) Respiratory tract infections requiring hospital attendance. (D) Lower respiratory tract infections requiring hospital attendance. (E) Upper respiratory tract infections requiring hospital attendance. Figure 3 Meta-analysis of gradual changes in primary outcomes after implementation of smoke-free legislation (A) Preterm birth. (B) Asthma exacerbations requiring hospital attendance. (C) Lower respiratory tract infections requiring hospital attendance. Associations between smoke-free legislation and the incidence of hospital attendances for childhood asthma were reported in ten studies (table 1).18, 36, 38, 40, 42, 49, 51, 55, 59, 61 In the meta-analysis, both an immediate reduction in asthma exacerbations requiring hospital attendance (five studies, 684 826 events; −9·83% [95% CI −16·62 to −3·04]; figure 2B) and an additional gradual reduction were seen (four studies, 243 377 events; −5·94% per year [95% CI −11·48 to −0·41]; figure 3B). No change in asthma admissions was seen following a health reform legislation that provided smoking cessation services for Medicaid recipients in one study.18 Among three US studies18, 49, 50 with overlapping populations evaluating tobacco taxation and asthma exacerbations requiring hospital attendance, the study with the lowest risk of bias found no significant reductions following state-wide increases in cigarette excise tax (table 3).18

Medicaid recipients in one study.18 Among three US studies18, 49, 50 with overlapping populations evaluating tobacco taxation and asthma exacerbations requiring hospital attendance, the study with the lowest risk of bias found no significant reductions following state-wide increases in cigarette excise tax (table 3).18 The association between smoke-free legislation and the incidence of hospital admissions for acute respiratory tract infections was reported in four studies (table 1).16, 18, 19, 64 In the meta-analysis, an immediate reduction was seen in respiratory tract infections (upper and lower respiratory tract infections combined) requiring hospital attendance (two studies, 1 681 020 events; −3·45% [95% CI −4·64 to −2·25]; figure 2C). For the studies that reported specifically on lower respiratory tract infections, the meta-analysis showed an immediate reduction in admissions for lower respiratory tract infections following smoke-free legislation (three studies, 887 414 events; −18·48% [95% CI −32·79 to −4·17]; figure 2D). No additional gradual reduction in lower respiratory tract infections was observed (two studies; 748 175 events: −6·81% per year [95% CI −20·63 to 7·01]; figure 3C). No significant association between smoke-free legislation and admissions for upper respiratory tract infections was seen in the meta-analysis (two studies; 1 390 056 events; 0·42% [95% CI −3·28 to 4·13]; figure 2E). One study18 reported that a health reform legislation that provided smoking cessation services for Medicaid recipients was associated with an immediate −6% (95% CI −10 to −1) decrease in hospital admissions for childhood upper respiratory tract infection, but not in admissions for lower respiratory tract infection (table 2). The same study18 evaluated the effect of tobacco taxation, showing a −9% decrease (95% CI −16 to −2) in lower respiratory tract infections requiring admission to hospital per USD$ increase in cigarette excise tax at the state level (table 3).

ction, but not in admissions for lower respiratory tract infection (table 2). The same study18 evaluated the effect of tobacco taxation, showing a −9% decrease (95% CI −16 to −2) in lower respiratory tract infections requiring admission to hospital per USD$ increase in cigarette excise tax at the state level (table 3). We did not identify any studies assessing the effect of other MPOWER policies on child health. In sensitivity analyses, inclusion of non-EPOC studies in the meta-analyses or restriction of the primary analyses to studies with low to moderate risk of bias did not materially change the effect estimates for smoke-free legislation and our primary outcomes (appendix pp 35–39). Point estimates for the association between smoke-free legislation and our primary outcomes were generally much larger when subgroup analyses were restricted to studies assessing comprehensive smoke-free laws than when studies assessing partial smoke-free laws were analysed (preterm birth: seven studies, 9 355 359 individuals, −5·12% [95% CI −7·24 to −2·99]; hospital attendances for asthma: four studies, 556 019 events, −12·49% [–19·78 to −5·20]; appendix, pp 40–44).

ricted to studies assessing comprehensive smoke-free laws than when studies assessing partial smoke-free laws were analysed (preterm birth: seven studies, 9 355 359 individuals, −5·12% [95% CI −7·24 to −2·99]; hospital attendances for asthma: four studies, 556 019 events, −12·49% [–19·78 to −5·20]; appendix, pp 40–44). 11 studies assessed whether the association between implementation of tobacco control policies and child health varied according to indicators of socioeconomic status (appendix pp 45, 46).16, 29, 30, 33, 44, 46, 51, 54, 55, 62, 63 One study16 showed that the most deprived children experienced the largest gradual reduction in hospital admissions for respiratory tract infection following smoke-free legislation (–1·5% per year [95% CI −2·1 to −1·0]). In two studies,44, 46 improvements in perinatal outcomes were greater among babies born to parents with low levels of education following smoke-free legislation than among those born to parents with high levels of education,44 and among babies born to black mothers with any level of education and to white mothers with low levels of education following tobacco tax increases.46 Other studies did not identify a clear socioeconomic gradient in the association between tobacco control policies and child health.

nts with high levels of education,44 and among babies born to black mothers with any level of education and to white mothers with low levels of education following tobacco tax increases.46 Other studies did not identify a clear socioeconomic gradient in the association between tobacco control policies and child health. 27 studies assessed the association between tobacco control policies and secondary outcomes (appendix, pp 47–76). In the meta-analyses (appendix pp 77–85), smoke-free legislation was associated with immediate reductions in very preterm birth (five studies; 3 354 636 individuals; −9·99% [95% CI −15·74 to −4·24]), low birthweight (nine studies; 35 206 918 individuals, −2·77% [–4·36 to −1·19]), and small for gestational age births (eight studies; 27 649 380 individuals; −1·84% [–3·21 to −0·47]), a gradual reduction in very small for gestational age births (two studies; 1 298 276 individuals; −0·60% per year [–0·60 to −0·60]), and a small increase in birthweight (seven studies; 3 238 575 individuals; 12·45 g [95% CI 2·09–22·81]). No significant changes in other secondary outcomes were seen following smoke-free legislation. Legislation to promote prenatal care, including smoking cessation services for low-income pregnant women, was not associated with a change in small for gestational age births in one US study.47 In another US study,28 although such legislation was associated with increased duration of gestation, depending on time of enrolment (308 521 participants; 0·063 weeks [95% CI 0·008–0·118] among women who enrolled in the Medicaid insurance programme before or during pregnancy and 0·086 weeks [0·004–0·168] among women who enrolled during pregnancy), it was not associated with a change in birthweight (appendix pp 66, 67). One study showed reductions in extremely and very preterm births following tobacco tax increases,53 with two others also showing an increase in gestation.28, 35 Among five studies assessing the link between tobacco tax and birthweight, two showed a positive effect,39, 46 albeit of very small magnitude.

ndix pp 66, 67). One study showed reductions in extremely and very preterm births following tobacco tax increases,53 with two others also showing an increase in gestation.28, 35 Among five studies assessing the link between tobacco tax and birthweight, two showed a positive effect,39, 46 albeit of very small magnitude. Accordingly, only one of these five studies showed a reduction in low birthweight following tobacco tax increases.46 This study also found reductions in small for gestational age births; both associations were confined to low socioeconomic groups.46 In two studies assessing very low birthweight, no changes were seen following tobacco tax increases.39, 53 Tobacco taxes were associated with a decreased risk of infant mortality in two studies assessing this association.57, 60 In one of these studies, however, an increase in fetal deaths was also observed.60 One study showed significant reductions in paediatric asthma prevalence following tobacco tax increases.33

es.39, 53 Tobacco taxes were associated with a decreased risk of infant mortality in two studies assessing this association.57, 60 In one of these studies, however, an increase in fetal deaths was also observed.60 One study showed significant reductions in paediatric asthma prevalence following tobacco tax increases.33 Discussion This systematic review and meta-analysis provides considerable evidence indicating child health benefits associated with implementation of MPOWER policies. By pooling data of 27·5 million births, 685 000 hospital admissions for asthma, and 2·3 million hospital admissions for respiratory tract infections, we found a 3·7% reduction in preterm births, a 9·8% reduction in childhood hospital admissions for asthma, and an 18·5% reduction in hospital admissions for lower respiratory tract infections following implementation of smoke-free legislation. Subgroup analyses suggested that health benefits were increased when the most comprehensive laws were applied. We also identified several studies indicating that tobacco tax increases and governmental support for smoking cessation services could benefit child health. Taken together with substantial existing evidence on the effectiveness of tobacco control policies in improving adult health, these findings provide strong support for implementation of such policies comprehensively across the world.

es and governmental support for smoking cessation services could benefit child health. Taken together with substantial existing evidence on the effectiveness of tobacco control policies in improving adult health, these findings provide strong support for implementation of such policies comprehensively across the world. This study is, to our knowledge, the most comprehensive assessment done to date of the effect of tobacco control policies on perinatal and child health outcomes. On the basis of our previous work,15 and the challenges of evaluating governmental policies through randomised trials,69, 70 we anticipated that most eligible studies would be of quasi-experimental design. We therefore followed EPOC guidelines to restrict our primary analyses to study types that were considered to be at lowest risk of bias. We confirmed the robustness of our findings via a number of prespecified sensitivity analyses, which indicated that our findings were not sensitive to exclusion of studies with a high risk of bias or inclusion of purely observational studies. Our work builds on existing evidence since it focuses on all available evidence on the effect of tobacco control policies on perinatal and child health. The consistency of this evidence, in our view, supports the validity of our findings.

of studies with a high risk of bias or inclusion of purely observational studies. Our work builds on existing evidence since it focuses on all available evidence on the effect of tobacco control policies on perinatal and child health. The consistency of this evidence, in our view, supports the validity of our findings. However, our study has some limitations. The risks of residual confounding and bias in quasi-experimental studies—due to non-random allocation of the intervention and the absence of a control group—need to be considered when interpreting the results.71 Additional limitations include between-study heterogeneity in methodology, differences in follow-up duration and diagnosis ascertainment, the absence of assessment of the likely causal pathways between the policies and their health effects in several studies, and the low number of studies in each meta-analysis, which precluded assessment of publication bias for most outcomes and the use of meta-regression.

follow-up duration and diagnosis ascertainment, the absence of assessment of the likely causal pathways between the policies and their health effects in several studies, and the low number of studies in each meta-analysis, which precluded assessment of publication bias for most outcomes and the use of meta-regression. This study adds to our previous work.15 We identified an additional 24 studies on the effect of smoke-free legislation on child health, comprising additional data from more than 10 million births, 4·6 million GP diagnoses, and 2·2 million hospital admissions. These additional studies allowed us, for the first time, to identify the association between smoke-free legislation and reductions in severe respiratory tract infections, which is particularly relevant since respiratory tract infections account for the vast majority of the global burden of disease resulting from second-hand smoke exposure in children.1 We also broadened the scope of this study to include all MPOWER policies, identifying several studies on the effect of tobacco tax increases and smoking cessation services on child health. We also identified one study evaluating a tobacco control policy that could not be classified according to WHO's MPOWER Framework. Following an increase in the minimum legal age to purchase cigarettes from 18 years to 21 years in the US state of Pennsylvania, a 1·4 (95% CI −2·6 to −0·2) percentage point reduction in low birthweight was observed, which was largest among smoking mothers and associated with a significant reduction in prenatal cigarette consumption.72

ease in the minimum legal age to purchase cigarettes from 18 years to 21 years in the US state of Pennsylvania, a 1·4 (95% CI −2·6 to −0·2) percentage point reduction in low birthweight was observed, which was largest among smoking mothers and associated with a significant reduction in prenatal cigarette consumption.72 Socioeconomic disparities in smoking and related morbidity are widely documented and affect both adults and children. For example, such disparities were estimated to account for 38% of the inequality in stillbirths and 31% of the inequality in infant deaths in Scotland.73 Previous systematic reviews74, 75, 76 showed that, among MPOWER measures, tobacco taxation has the greatest potential to reduce socioeconomic disparities associated with smoking in both young people and adults. We identified some evidence suggesting a pro-equity effect of both tobacco taxation and smoke-free legislation on early-life health. Since smokers are over-represented among deprived communities, such relative benefits of tobacco control policies translate into larger absolute effects in children from low socioeconomic groups than in children from high socioeconomic groups.

ct of both tobacco taxation and smoke-free legislation on early-life health. Since smokers are over-represented among deprived communities, such relative benefits of tobacco control policies translate into larger absolute effects in children from low socioeconomic groups than in children from high socioeconomic groups. Given the inherent restrictions in attributing causality from quasi-experimental studies, it is important to interpret the findings in light of circumstantial evidence supporting the link between tobacco control policies and child health benefits. We have previously described the main likely causal pathways.23 Tobacco smoke exposure during fetal stages and childhood is associated with various adverse perinatal and child health outcomes.4, 5, 6, 8, 77, 78, 79, 80, 81, 82, 83 Several studies have shown substantial reductions in maternal smoking31, 34, 52, 56, 84, 85 and in second-hand smoke exposure among adults (including pregnant women) and children after implementation of tobacco control policies (appendix pp 86–89).12, 13, 86, 87, 88, 89, 90 Whereas smoke-free laws specifically target public spaces, various studies have shown subsequent increases in smoking cessation and reduced initiation,12, 91, 92 as well as changes in social norms leading to decreased smoking in the home environment,93, 94, 95, 96, 97 which is probably the primary source of second-hand smoke exposure among children. Our study provides further support for a causal association, since we found the largest decreases in our outcomes of interest when comprehensive smoke-free legislation was considered. This observation is suggestive of a dose–response association, which has previously also been identified for adult studies.98 Because of the low number of studies in individual meta-analyses we did not formally test for this interaction, and future efforts to do so might strengthen our findings as more evidence becomes available.

ation is suggestive of a dose–response association, which has previously also been identified for adult studies.98 Because of the low number of studies in individual meta-analyses we did not formally test for this interaction, and future efforts to do so might strengthen our findings as more evidence becomes available. The global health burden of tobacco use is tremendous and its total global economic cost is estimated to be around USD$1·4 trillion.99 Despite global progress in tobacco control, over a third of the world's population remains unprotected by any MPOWER policy at the recommended level.2, 11 This issue is important because 40–50% of children worldwide are regularly exposed to tobacco smoke, and tobacco control policies have substantial potential to reduce the associated burden of death and disease.1 This global burden is acknowledged by the prioritisation within SDG 3 of more effective FCTC implementation and its aim to reduce early-life mortality; our data now show that these initiatives can act synergistically. Because our effect estimates are expressed as relative changes, background prevalence of smoking and second-hand smoke exposure, and of the health outcomes evaluated, should be considered when extrapolating our findings to local contexts. We did not formally assess the comparative effectiveness or cost-effectiveness of different MPOWER policies. Tax increases are considered to be the most effective measure to reduce smoking prevalence,2 and although our review indicates that tobacco taxation is likely to be associated with child health benefits, the evidence was particularly strong for smoke-free legislation. Smoke-free laws are the tobacco control policy most strongly supported by the public and appear to be the most straightforward measure to protect child health, particularly when implemented comprehensively.12 The synergistic effect of various policies implemented at the highest recommended levels in reducing smoking prevalence should be considered when planning policy changes,14 which, when implemented as part of a strong tobacco control programme, can be highly cost-effective.100, 101 Ongoing monitoring is needed to continue to evaluate the effectiveness of policies aimed at reducing the impact of tobacco, in particular the effectiveness of novel endgame strategies targeted at ending rather than controlling the global tobacco epidemic.102

acco control programme, can be highly cost-effective.100, 101 Ongoing monitoring is needed to continue to evaluate the effectiveness of policies aimed at reducing the impact of tobacco, in particular the effectiveness of novel endgame strategies targeted at ending rather than controlling the global tobacco epidemic.102 Reports indicate that at least two of five people living in low-income and middle-income countries remain unprotected by any MPOWER policy measure,2 and that wide variations in implementation and compliance are present across these countries.103 This finding is of concern, since these countries have the largest burden of tobacco-related illness and death, and harbour nearly 80% of the world's smokers.2 We highlight an important gap in the literature as more research is required in low-income and middle-income countries to understand the effect of tobacco control policies in these regions. Modelling approaches are increasingly being used to estimate the effect of tobacco control policies in low-income and middle-income countries, and original studies are now becoming available.104, 105 Efforts are underway to address the current absence of a child health focus in this area, which will be essential to inform the global policy agenda. Furthermore, we found no studies specifically evaluating early-life health outcomes in relation to legislation to prohibit tobacco advertising and sponsorship, or warnings against the dangers of tobacco. Priority should be given to establishing a core set of outcomes related to perinatal and child health, alongside adult health, for all future studies examining the effect of tobacco control policies.

in relation to legislation to prohibit tobacco advertising and sponsorship, or warnings against the dangers of tobacco. Priority should be given to establishing a core set of outcomes related to perinatal and child health, alongside adult health, for all future studies examining the effect of tobacco control policies. In conclusion, given the positive findings of this systematic review it is crucial that the uptake of comprehensive tobacco control policies is accelerated worldwide to further protect children from the health hazards of tobacco smoke exposure,106 in parallel with efforts to evaluate the effectiveness of novel policy initiatives. Supplementary Material Supplementary appendix Acknowledgments This study was funded by the Scottish Government Chief Scientist Office (CSO), and by personal fellowships to JVB from the Netherlands Lung Foundation (4.2.14.063JO) and the Erasmus MC. AS is supported by the Farr Institute. CM is funded by an NIHR Research Professorship award. We thank Wichor Bramer for assistance in preparing the search strategy; Amanda Amos, Anna Gilmore, Stanton Glantz, Summer Hawkins, Zubair Kabir, David Levy, Daniel Mackay, and Sara Markowitz for providing advice as members of the expert panel; Ioannis Bakolis, Bianca Cox, Summer Hawkins, Julian Johnsen, Glenn Landers, Britt McKinnon, and Yelena Tarasenko for providing additional information or data upon request on behalf of all authors of their respective articles; and Daan Nieboer and Chris Weir for providing assistance in the statistical analyses.

t panel; Ioannis Bakolis, Bianca Cox, Summer Hawkins, Julian Johnsen, Glenn Landers, Britt McKinnon, and Yelena Tarasenko for providing additional information or data upon request on behalf of all authors of their respective articles; and Daan Nieboer and Chris Weir for providing assistance in the statistical analyses. Contributors JVB and AS secured funding for this work. JVB, JPM, CM, SB, and AS designed the study and wrote the protocol. TF and AK did the study search, study selection, data extraction, and risk of bias assessment. JVB supervised all the steps in the review process. TF did the data analysis and created the figures. All authors interpreted the findings. TF, AK, and JVB drafted the manuscript and appendix. AS supervised the writing, and JPM, CM, and SB provided feedback. Declaration of interests We declare no competing interests.

Introduction The epidemiology of cervical cancer in high-income countries is changing. In England, UK, women vaccinated against human papillomavirus (HPV) in 2008 at age 17 years have been invited to screening for the first time in 2016–17. Furthermore, the cervical screening programme is preparing for the introduction of HPV testing as the primary screening test.1 So, not only will cohorts of women entering the screening programme have a lower risk of cervical cancer, but the sensitivity of the screening test to precancer will be increased. Additionally, the European Commission has in April, 2016, granted marketing authorisation for two-dose Gardasil nine vaccination (ie, the nine-valent vaccine), which protects against HPV types that cause about 90% of cervical cancers compared with the current bivalent and quadrivalent vaccines that protect against about 70% of cervical cancers.2 Research in context Evidence before this study

Introduction The epidemiology of cervical cancer in high-income countries is changing. In England, UK, women vaccinated against human papillomavirus (HPV) in 2008 at age 17 years have been invited to screening for the first time in 2016–17. Furthermore, the cervical screening programme is preparing for the introduction of HPV testing as the primary screening test.1 So, not only will cohorts of women entering the screening programme have a lower risk of cervical cancer, but the sensitivity of the screening test to precancer will be increased. Additionally, the European Commission has in April, 2016, granted marketing authorisation for two-dose Gardasil nine vaccination (ie, the nine-valent vaccine), which protects against HPV types that cause about 90% of cervical cancers compared with the current bivalent and quadrivalent vaccines that protect against about 70% of cervical cancers.2 Research in context Evidence before this study We searched PubMed with the search terms “cancer screening” AND “projections” OR “cancer prevention” AND “projections” to identify studies with a similar study design, which would provide evidence for time trends that took into account cancer prevention activities. We did not use date or language restrictions. We noted that age period cohort models have been used to estimate the effect of screening by assuming that period effects reflect the impact of screening. However, such models cannot explore the effect of decreasing screening coverage. Dynamic transmission models consider the effects of human papillomavirus (HPV) prevalence and transmission, cervical screening, and HPV vaccination on cervical cancer incidence, and are almost always used to produce relative measures of effect. We are not aware of any dynamic transition models that allow for underlying changes in risk on the basis of cohort effect.

s of human papillomavirus (HPV) prevalence and transmission, cervical screening, and HPV vaccination on cervical cancer incidence, and are almost always used to produce relative measures of effect. We are not aware of any dynamic transition models that allow for underlying changes in risk on the basis of cohort effect. Added value of this study Our study combined an age period cohort model (underlying rates of disease) with a dynamic model to estimate the reduction in risk of cervical cancer following screening by attendance status and vaccination status. It not only allowed for the estimation of absolute changes to incidence of cervical cancer, but also allowed us to vary screening coverage and vaccine uptake to estimate the effect on absolute rates. In combining the three levels of modelling, this study is the first to explore the effect of introducing HPV testing, changes in screening coverage, and vaccine uptake on cervical cancer incidences over the next 25 years. The study provides the absolute impact these changes will have at a population level in England, UK. In particular, it highlights the need for continued innovation within the programme to ensure cervical cancer rates decrease. We project that given current coverage and vaccine uptake, and accounting for the introduction of HPV primary screening, rates of cervical cancer will see only a modest decrease in the next 25 years. Implications of all the available evidence

In combining the three levels of modelling, this study is the first to explore the effect of introducing HPV testing, changes in screening coverage, and vaccine uptake on cervical cancer incidences over the next 25 years. The study provides the absolute impact these changes will have at a population level in England, UK. In particular, it highlights the need for continued innovation within the programme to ensure cervical cancer rates decrease. We project that given current coverage and vaccine uptake, and accounting for the introduction of HPV primary screening, rates of cervical cancer will see only a modest decrease in the next 25 years. Implications of all the available evidence Evidence presented in this study suggests the cervical screening programme will need to adapt swiftly to the changing epidemiology of cervical cancer. Going forward, focus should be placed on scenarios that offer less intensive screening for vaccinated women and more on increasing coverage and incorporating new approaches to enhance current cervical screening in unvaccinated women. Organised cervical screening was introduced in England in 1988 to women aged 20–64 years. Since 2003, cytology-based screening is offered once every 3 years to women aged 25–49 years and once every 5 years to women aged 50–64 years. HPV bivalent vaccination was introduced in 2008 as a school-based programme to girls aged 12–13 years, although a catch-up cohort of women aged 14–18 years were also offered the vaccine.

3, cytology-based screening is offered once every 3 years to women aged 25–49 years and once every 5 years to women aged 50–64 years. HPV bivalent vaccination was introduced in 2008 as a school-based programme to girls aged 12–13 years, although a catch-up cohort of women aged 14–18 years were also offered the vaccine. In the next 25 years, women aged 50 years and older will not benefit from prophylactic vaccination, and with screening ceasing by age 64 years for women with negative results, one might expect the burden of cervical cancer to once again move towards older age groups. Furthermore, population projections for the UK estimate a substantial increase in the number of women older than 60 years: a 29% increase (from 9·4 million in 2012 to 12·1 million in 2037) in the number of women aged 60–74 years and a 90% increase in those older than 75 years (from 5 million to 9·5 million).3 Because the aim of the cervical screening programme is to prevent the development of cancer by identifying and treating precancerous disease, age-specific incidence estimates (rates and numbers of cases) of cervical cancer for the next 25 years will help policy makers and local commissioning groups to adequately determine where the demand on preventive services will be.

is to prevent the development of cancer by identifying and treating precancerous disease, age-specific incidence estimates (rates and numbers of cases) of cervical cancer for the next 25 years will help policy makers and local commissioning groups to adequately determine where the demand on preventive services will be. In this study, we aimed to estimate the age-specific incidence of cervical cancer in England over the next 25 years from 2015 to 2040 in four policy scenarios: no changes to current screening coverage or vaccine uptake and HPV primary testing from 2019 (ie, the status quo), changing the year in which HPV primary testing is introduced, introduction of the nine-valent vaccine, and changes to cervical screening coverage.

next 25 years from 2015 to 2040 in four policy scenarios: no changes to current screening coverage or vaccine uptake and HPV primary testing from 2019 (ie, the status quo), changing the year in which HPV primary testing is introduced, introduction of the nine-valent vaccine, and changes to cervical screening coverage. Methods Model design We did a data modelling study that combined results from three levels: population modelling of incidence trends (level 1), observable data from the individual level with use of a generalised linear model (level 2), and unobservable disease states modelled through a microsimulation (level 3). We chose this approach because the microsimulation model focuses on a single cohort of women progressing from age 12 years to 80 years. Hence, we needed age-specific cohort effects to model rates in the future. Additionally, where possible, we used individual level data because it is more accurate and requires fewer assumptions than microsimulation data. The appendix (p 1) summarises how the different components of the model were brought together. Our model software is available from the corresponding author, so that those interested can consider other scenarios.

individual level data because it is more accurate and requires fewer assumptions than microsimulation data. The appendix (p 1) summarises how the different components of the model were brought together. Our model software is available from the corresponding author, so that those interested can consider other scenarios. We started by estimating the risk of cervical cancer in the absence of screening on the basis of women's individual screening histories. We used data from cases with cervical cancer diagnosed between April, 2007, and March, 2012 (and women with no history of cervical cancer matched on age and area of residence as the control), from the audit of invasive cervical cancers,4 a case-control study nested in a cohort that included more than 90% of cervical cancers diagnosed in England along with their full screening histories. This study design allowed us to obtain absolute risks by fitting a generalised linear model using the binomial family and log link function for each age group and staging of cervical cancer (stages 1A, 1B, and 2+), weighted so the number of cases in the audit diagnosed in each year and age group matched the number of cases recorded in the Office for National Statistics (ONS) cancer registration statistics (2011–15).5 These estimates were multiplied by five to estimate the 5-year risk. This first step provided absolute risks per 100 000 women by age for 2011–15 in the absence of screening and vaccination (appendix p 1).

he number of cases recorded in the Office for National Statistics (ONS) cancer registration statistics (2011–15).5 These estimates were multiplied by five to estimate the 5-year risk. This first step provided absolute risks per 100 000 women by age for 2011–15 in the absence of screening and vaccination (appendix p 1). To allow the underlying cancer incidence to vary over the next 25 years in the absence of vaccination and screening, we used a modified age period cohort model to fit cancer incidence data from 1971 to 2013 (cervical cancer incidence data were provided by the National Cancer Registration and Analysis Service [request ODR_2014_335]), and extrapolated it to 2040. The model includes a period effect, to capture the effect of screening, which was set to zero for years before the introduction of organised screening in England in 1988. To estimate future incidences in the absence of screening (ie, a counterfactual baseline), we used the combined age and cohort effects only. We used these baseline rates to calculate age-specific cohort effects relative to the 2011–15 cohort. Further details can be found in the appendix (pp 2, 3). In the second step, this model was used to obtain the risk for each period given age relative to 2011–15 in a population not offered screening or vaccination (appendix p 1). Although the age cohort model could be used to provide estimates of the absolute risks of cervical cancer among unscreened women, we chose to use risks estimated from individual level data (from the audit), because stronger assumptions are required in the age cohort model. The absolute risks from the first step were multiplied by the cohort effects of the second step to provide absolute risks of cervical cancer for unscreened women for each 5-year period from 2016–20 to 2036–40.

rom individual level data (from the audit), because stronger assumptions are required in the age cohort model. The absolute risks from the first step were multiplied by the cohort effects of the second step to provide absolute risks of cervical cancer for unscreened women for each 5-year period from 2016–20 to 2036–40. To allow for the effect of screening in unvaccinated women, we used relative risks of being diagnosed with cervical cancer by screening history calculated from the audit of invasive cervical cancers4 (appendix p 1). Hence, the relative effect of screening is dependent on age and screening history, but does not depend on the underlying risk of the birth cohort. Screening histories were categorised as never screened (either no tests or tests only more than 15 years earlier), regularly screened (screened at least once every 3·5-year [5·5 years after age 50 years] period in the previous 15 years), and lapsed attender (screened within a 15-year interval but not often enough to be defined as regularly screened). We used data from the cervical screening programme statistics (2014–15)6 to establish the proportion of women by age group in each of the screening history categories. We consider this to be the current screening coverage by age group in England and consider current incidence to be the observed average rate reported between 2011 and 2015 (appendix p 10). Women are not offered cervical screening beyond age 65 years in England. We estimated the risk of cervical cancer as a function of age (65–69 years, 70–74 years, and 75–79 years) in women regularly and irregularly screened at age 50–64 years relative to that in those not screened in the same age group. The risk of cancer at age 65–79 years is thereby determined by the screening history between ages 50 and 64 years.

cervical cancer as a function of age (65–69 years, 70–74 years, and 75–79 years) in women regularly and irregularly screened at age 50–64 years relative to that in those not screened in the same age group. The risk of cancer at age 65–79 years is thereby determined by the screening history between ages 50 and 64 years. Unvaccinated women are screened with use of cytology until HPV primary screening is implemented. We have taken the relative effectiveness of primary HPV screening compared with cytology screening from previously published research (appendix p 1), which concluded that in addition to the cervical cancer already prevented by cytology-based screening a further 24% of currently observed cases can be prevented each year once HPV screening has been fully introduced.7 More details are presented in the appendix (p 4).

g from previously published research (appendix p 1), which concluded that in addition to the cervical cancer already prevented by cytology-based screening a further 24% of currently observed cases can be prevented each year once HPV screening has been fully introduced.7 More details are presented in the appendix (p 4). The introduction of HPV primary screening into the programme was made under four assumptions. Firstly, in 2019, the test will be introduced to all laboratories in England at the same time. Secondly, HPV screening will not prevent any additional cancers within 12 months of the test, and the first women to benefit are those who test positive for HPV but negative with use of cytology who will be recalled for repeat testing 12 months later. Thirdly, the full effect of switching to HPV primary screening will be observed (in terms of cancer prevention) at the next screening round (ie, once every 3 years for those aged 25–49 years and once every 5 years for those aged 50–64 years). In any given year, 33% of women younger than 50 years will be invited for screening (20% of those aged 50–64 years); therefore, the full effect of the primary HPV screening roll-out will not be seen for 6 years in women aged 25–29 years and for 10 years in those aged 50–64 years (appendix p 8). Lastly, the introduction of primary HPV screening will affect women 65 years and older in the cohort who received HPV testing at age 60–64 years. No woman aged 65–69 years will be tested with HPV before 2021, whereas from 2026 onwards all women aged 65–69 years who were screened at age 60–64 years would have benefited from an HPV test.

duction of primary HPV screening will affect women 65 years and older in the cohort who received HPV testing at age 60–64 years. No woman aged 65–69 years will be tested with HPV before 2021, whereas from 2026 onwards all women aged 65–69 years who were screened at age 60–64 years would have benefited from an HPV test. The effect of screening in vaccinated women was estimated from a microsimulation model8 (appendix pp 4–7). Vaccinated women in the model are offered screening with use of HPV testing with a 6-year interval. The effect of vaccination was estimated using HPV 16/18 vaccination (ie, the bivalent or quadrivalent vaccine) and nine-valent vaccination. Vaccination with the 16/18 vaccine was assumed to prevent all HPV 16 and 18 infections as well as 15% of non-HPV 16 and 18 infections,9 corresponding to 74·5% of cervical cancers, whereas the nine-valent vaccine was assumed to prevent 90% of cervical cancers.10 The 16/18 vaccine was introduced in England and in the model in 2008–09 to girls aged 12–13 years (born from Sept 1, 1995, to Aug 31, 1996) and to a catch-up cohort aged 14–18 years (born from Sept 1, 1990, to Aug 31, 1995). Coverage among girls aged 12–13 years was 86%.11 The assumed effect of vaccination on the catch-up cohort is summarised in the appendix (p 8).

in the model in 2008–09 to girls aged 12–13 years (born from Sept 1, 1995, to Aug 31, 1996) and to a catch-up cohort aged 14–18 years (born from Sept 1, 1990, to Aug 31, 1995). Coverage among girls aged 12–13 years was 86%.11 The assumed effect of vaccination on the catch-up cohort is summarised in the appendix (p 8). Changes to vaccine uptake at age 12 years cannot affect women aged 25–29 years until 2031. Before 2031, women were assumed to have a 16/18 vaccine uptake of 86% throughout the model. For women aged 30–34 years, vaccine uptake was allowed to vary from 2036. No changes in vaccine uptake were allowed for women aged 35–49 years, and no women older than 49 years were protected by vaccination by 2040. To calculate the incidence in the population as a whole (appendix p 1), we used the proportion of women in each age category and interval who have a specific combination of screening and vaccination history, and apply these as weights to the corresponding rates of cervical cancer to obtain a weighted average. To calculate absolute numbers of cancers from the incidence, we use the ONS population projections for England up to 2039.12 Average 5-yearly population estimates can be found in the appendix (p 9). Current cervical cancer rates were taken as an average of age-specific rates between 2011 and 2015,5 using the estimated mid-2014 population. Rates across ages were standardised with use of the European standard population.13

ngland up to 2039.12 Average 5-yearly population estimates can be found in the appendix (p 9). Current cervical cancer rates were taken as an average of age-specific rates between 2011 and 2015,5 using the estimated mid-2014 population. Rates across ages were standardised with use of the European standard population.13 To validate the model, we applied weights so the model exactly replicated observed age-specific incidence of cervical cancer in England between 2011 and 2015. We coded the microsimulation model in C++ (version 11), and we did the age period cohort model and data analyses using Stata (version 13.1).

ngland up to 2039.12 Average 5-yearly population estimates can be found in the appendix (p 9). Current cervical cancer rates were taken as an average of age-specific rates between 2011 and 2015,5 using the estimated mid-2014 population. Rates across ages were standardised with use of the European standard population.13 To validate the model, we applied weights so the model exactly replicated observed age-specific incidence of cervical cancer in England between 2011 and 2015. We coded the microsimulation model in C++ (version 11), and we did the age period cohort model and data analyses using Stata (version 13.1). Scenarios We explored four different policy scenarios using our model. The first scenario comprises no changes to current screening or vaccination coverage—ie, the so-called status quo scenario. In this scenario, age-specific screening coverage remains as currently observed up to 2040. Cytology screening is offered to women until 2019 when HPV screening is implemented. Implementation of HPV screening from 2019 means that a woman will first be offered this test between 2019 and 2023, depending on when she is next due to be screened. The second scenario considers the introduction of primary HPV screening. In this scenario, we compared cervical cancer incidences under the following scenarios: no HPV primary testing to assess effect of cytology only, the status quo (ie, primary HPV from 2019), bringing forward the introduction of HPV primary screening to 2017, and delaying the introduction until 2023. In the third scenario, unless otherwise stated, all assumptions remain same as the status quo scenario with the exception of the introduction of the nine-valent vaccine. Given that the government's contract for the supply of the four-valent vaccine ends in June, 2018, we have assumed that the earliest the nine-valent vaccine could be introduced to girls aged 12–13 years will be 2019.14 The last scenario involves changes to cervical screening coverage. Between 2011 and 2016, age-specific coverage decreased from 75·7% to 72·7% for those aged 25–64 years.6 Interventions to increase coverage have at best had a moderate effect (with the exception of offering vaginal self-sampling).15 Hence in addition to a scenario where screening coverage increases, we consider a decrease to show the importance of high population coverage. We hope that this decrease in coverage is not a realistic scenario, but one that it is important to understand the implications of.

ception of offering vaginal self-sampling).15 Hence in addition to a scenario where screening coverage increases, we consider a decrease to show the importance of high population coverage. We hope that this decrease in coverage is not a realistic scenario, but one that it is important to understand the implications of. When considering changes to screening coverage, it was increased or decreased from the current coverage at a steady rate to achieve the nominal coverage by 2031. The women who entered or exited the regularly screened group were equally split between the lapsed group and the never screened group. Details of how the coverage decreased in the scenario where screening is phased out completely can be found in the appendix (p 10). When screening coverage is increased, women who have never been screened become irregularly screened, and irregularly screened women become regularly screened. Similarly, when coverage is decreased, women who were regularly screened become irregularly screened, and irregularly screened women become never screened (ie, no screening in the past 15 years).

d, women who have never been screened become irregularly screened, and irregularly screened women become regularly screened. Similarly, when coverage is decreased, women who were regularly screened become irregularly screened, and irregularly screened women become never screened (ie, no screening in the past 15 years). Role of the funding source This study was funded by Jo's Cervical Cancer Trust and the Cancer Research UK. The views expressed are those of the authors and not those of the funders. The funders of the study had no role in study design, data collection, data analysis, data interpretation, writing of the report, or decision to submit for publication. All authors had full access to the data. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

role in study design, data collection, data analysis, data interpretation, writing of the report, or decision to submit for publication. All authors had full access to the data. 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 Figure 1 summarises the rates per 100 000 women per year and yearly average numbers of cervical cancer cases by year of diagnosis in the status quo scenario. Similar results by birth cohort are also illustrated in figure 1. The results from the status quo scenario show a shift in peak age of cancer diagnosis from 25–29 years in 2011–15 to 55–59 years (with a second peak at 75–79 years) in 2036–40. Unvaccinated women born between 1975 and 1990 were predicted to have a relatively high risk of cervical cancer throughout their lives. Consequently, cancers in these women will dominate cancer incidence over the next 30 years. By contrast, age-specific rates in women aged 30–34 years were reduced by 53% from 17·5 in 2011–15 to 8·3 per 100 000 women by 2036–40.Figure 1 Rates of and yearly averages of cervical cancer under the status quo scenario (A) Cervical cancer rates per 100 000 women per age and year of diagnosis. (B) Yearly average numbers of cervical cancer cases per age and year of diagnosis. (C) Cervical cancer rates per 100 000 women per birth cohort and age of diagnosis. (D) Yearly average numbers of cervical cancer cases per birth cohort and age of diagnosis.

l cancer rates per 100 000 women per age and year of diagnosis. (B) Yearly average numbers of cervical cancer cases per age and year of diagnosis. (C) Cervical cancer rates per 100 000 women per birth cohort and age of diagnosis. (D) Yearly average numbers of cervical cancer cases per birth cohort and age of diagnosis. Figure 2 shows the estimated reduction in European age-standardised rates per 100 000 women aged 25–64 years among the scenario for which HPV primary screening is introduced to the English cervical screening programme. Provided that HPV primary screening is introduced by 2023, we estimated that cervical cancer rates at ages 25–64 years could be reduced by 2·9 (19%) cases per 100 000 women by 2028–32 (from 15·1 in 2016 to 12·2) by introducing this new test. In absolute terms, delaying the introduction of HPV primary screening from 2017 to 2019 resulted in 1400 extra cervical cancers diagnosed in women aged 25–64 years between 2018 and 2028, because it takes 10 years for the full effect to be observed.Figure 2 Effect of the introduction of HPV primary testing on age-standardised cervical cancer rates in women aged 25–64 years HPV=human papillomavirus.

Figure 2 shows the estimated reduction in European age-standardised rates per 100 000 women aged 25–64 years among the scenario for which HPV primary screening is introduced to the English cervical screening programme. Provided that HPV primary screening is introduced by 2023, we estimated that cervical cancer rates at ages 25–64 years could be reduced by 2·9 (19%) cases per 100 000 women by 2028–32 (from 15·1 in 2016 to 12·2) by introducing this new test. In absolute terms, delaying the introduction of HPV primary screening from 2017 to 2019 resulted in 1400 extra cervical cancers diagnosed in women aged 25–64 years between 2018 and 2028, because it takes 10 years for the full effect to be observed.Figure 2 Effect of the introduction of HPV primary testing on age-standardised cervical cancer rates in women aged 25–64 years HPV=human papillomavirus. Scenarios discussed from here onwards assume primary HPV screening is introduced in 2019, with all women (aged 25–64 years) offered screening by HPV testing by 2023. By 2021, all women aged 25–29 years will have been offered vaccination against HPV at age 12–13 years; therefore, changes to screening coverage make only a slight effect on cervical cancer rates in this age group. Cervical cancer rates per 100 000 women decreased by 55% from 20·9 in 2011–15 to 9·5 in 2036–40 among women aged 25–29 years vaccinated against HPV types 16 and 18 and by 71% from 20·9 in 2011–15 to 6·1 in 2036–40 among those receiving the nine-valent vaccine (figure 3). The additional benefit of the introduction of the nine-valent vaccine compared with continuing vaccination against HPV types 16 and 18 by 2036–40 was a further reduction of 36% in cervical cancer rates (ie, from 9·5 to 6·1 per 100 000 women in 2036–40). In this age group, the biggest modifier of cervical cancer risk in the future was vaccine uptake. In the most extreme scenario in which uptake of the nine-valent vaccine decreased to 40% by 2036–40, rates of cervical cancer were 144% higher (14·9 per 100 000 women) than the scenario in which uptake remained at 86% (6·1 per 100 000 women; figure 3).Figure 3 Effect of vaccine type, vaccine uptake, and screening coverage on cervical cancer rates

in which uptake of the nine-valent vaccine decreased to 40% by 2036–40, rates of cervical cancer were 144% higher (14·9 per 100 000 women) than the scenario in which uptake remained at 86% (6·1 per 100 000 women; figure 3).Figure 3 Effect of vaccine type, vaccine uptake, and screening coverage on cervical cancer rates Vaccination was introduced in England, UK, to girls aged 12–13 years in 2008–09 with a catch-up cohort aged 17–18 years (uptake in this cohort was poor). Not all women aged 30–34 years will have been vaccinated with the nine-valent vaccine by 2040, 24% will have received the four-valent vaccine instead. Maintaining or increasing cervical screening coverage until 2025 in women aged 30–34 years will be important because few women will have been vaccinated before becoming sexually active. In fact, the effect of changes in coverage among this age group can be seen as early as 2016–20 (figure 3). The additional benefit of introducing the nine-valent vaccine compared with continuing HPV 16/18 vaccination by 2036–40 was a 28% reduction in cervical cancer rates (from 8·3 to 6·0 per 100 000 women). Among those vaccinated with the nine-valent vaccine cervical cancer rates decreased by 66% (from 17·5 to 6·0 per 100 000 women) and by 53% in those vaccinated against HPV types 16 and 18 (from 17·5 to 8·3 per 100 000 women) by 2036–40. If nine-valent vaccine uptake were to decrease to 40% in girls aged 12–13 years we would see its effect on cervical cancer incidences from 2036 onwards with a 127% increase in rates (from 6·0 when uptake is 86% to 13·6 per 100 000 women).

ated against HPV types 16 and 18 (from 17·5 to 8·3 per 100 000 women) by 2036–40. If nine-valent vaccine uptake were to decrease to 40% in girls aged 12–13 years we would see its effect on cervical cancer incidences from 2036 onwards with a 127% increase in rates (from 6·0 when uptake is 86% to 13·6 per 100 000 women). The current screening coverage and changes to regular screening of cervical cancer by age groups in England is summarised in table 1. Table 2 summarises the cumulative number of cancers and European age-standardised incidences in women aged 25–79 years. Because cervical screening is already preventing the majority of cervical cancers, the effect of screening was most apparent in the scenarios where the coverage decreased. Despite vaccination and primary HPV screening with current screening coverage, the European age-standardised rates of cervical cancer at ages 25–79 years showed only a moderate 10% decrease by 2036–40 (from 12·8 in 2011–15 to 11·6 per 100 000 women in 2036–40).Table 1 Cervical cancer screening coverage by age group under various scenarios

and primary HPV screening with current screening coverage, the European age-standardised rates of cervical cancer at ages 25–79 years showed only a moderate 10% decrease by 2036–40 (from 12·8 in 2011–15 to 11·6 per 100 000 women in 2036–40).Table 1 Cervical cancer screening coverage by age group under various scenarios Current screening coverage* Absolute percentage coverage change (every 5 years) to achieve 85% regularly screened by 2031 Absolute percentage coverage change (every 5 years) to achieve 50% regularly screened by 2031 Absolute percentage coverage change (every 5 years) to achieve 20% regularly screened by 2031 Regularly screened Lapsed attender Never screened Total population Regularly screened Lapsed attender or never screened Regularly screened Lapsed attender or never screened Regularly screened Lapsed attender or never screened 25–29 years† 63·5% ·· 36·5% 2 076 444 5·4% −5·4% −3·4% 3·4% −10·9% 10·9% 30–34 years 70·4% 14·8% 14·8% 2 043 366 3·7% −1·8% −5·1% 2·6% −12·7% 6·3% 35–39 years 73·1% 17·4% 9·5% 1 831 058 3·0% −1·5% −5·8% 2·9% −13·3% 6·6% 40–44 years 75·1% 17·6% 7·3% 1 877 290 2·5% −1·2% −6·3% 3·1% −13·8% 6·9% 45–49 years 75·2% 17·9% 6·9% 1 923 869 2·5% −1·2% −6·3% 3·2% −13·8% 6·9% 50–54 years 80·8% 12·1% 7·1% 1 781 782 1·1% −0·5% −7·7% 3·9% −15·2% 7·6% 55–59 years 74·6% 17·1% 8·3% 1 451 314 2·6% −1·3% −6·2% 3·1% −13·7% 6·8% 60–64 years 72·4% 17·9% 9·7% 1 225 440 3·2% −1·6% −5·6% 2·8% −13·1% 6·6% 65–69 years 72·4% 17·9% 9·7% 1 207 164 3·2% −1·6% −5·6% 2·8% −13·1% 6·6% 70–74 years 72·4% 17·9% 9·7% 909 787 3·2% −1·6% −5·6% 2·8% −13·1% 6·6% 75–79 years 72·4% 17·9% 9·7% 776 335 3·2% −1·6% −5·6% 2·8% −13·1% 6·6% * Observed in the 2014–15 cervical screening programme statistics;6 regularly screened defined as test within 3·5 years in those aged 25–49 years and within 5 years in those aged 50–64 years.

2·4% 17·9% 9·7% 909 787 3·2% −1·6% −5·6% 2·8% −13·1% 6·6% 75–79 years 72·4% 17·9% 9·7% 776 335 3·2% −1·6% −5·6% 2·8% −13·1% 6·6% * Observed in the 2014–15 cervical screening programme statistics;6 regularly screened defined as test within 3·5 years in those aged 25–49 years and within 5 years in those aged 50–64 years. † Because women are first invited for screening at 25 years of age, women in this age group cannot be lapsed attenders. Table 2 European age-standardised cumulative number of cancers and incidences per 100 000 women aged 25–79 years in various screening coverage scenarios and calendar years at diagnosis 2016–20 2021–25 2026–30 2031–35 2036–40 N Rate per 100 000 N Rate per 100 000 N Rate per 100 000 N Rate per 100 000 N Rate per 100 000 Currently observed (status quo) 947 14·0 863 12·8 785 11·6 794 11·8 782 11·6 85% coverage 906 13·4 786 11·7 685 10·1 663 9·8 653 9·7 50% coverage 1043 15·5 1049 15·5 1034 15·3 1131 16·8 1112 16·5 20% coverage 1161 17·2 1273 18·9 1333 19·8 1532 22·7 1506 22·3 Phasing out* 2463 36·5 2498 37·0 2363 35·0 2527 37·4 2558 37·9 * In this scenario, screening is no longer offered from 2016 onwards; it takes several years for women who had been regularly screened to eventually become never screened because they must become lapsed attenders first. Hence, some degree of protection against cervical cancer remains in this population until 2031–35 (appendix p 10).

ario, screening is no longer offered from 2016 onwards; it takes several years for women who had been regularly screened to eventually become never screened because they must become lapsed attenders first. Hence, some degree of protection against cervical cancer remains in this population until 2031–35 (appendix p 10). Discussion As a public health policy, HPV immunisation will deliver the biggest reduction in cervical cancer diagnosed. Provided vaccine uptake is maintained, and even without the introduction of the nine-valent vaccination, cervical cancer rates in women aged 25–34 years will decrease by more than 50%. Introduction of the nine-valent vaccine from 2019 would decrease cancer rates by a further 36% in women aged 25–29 years and 28% in those aged 30–34 years by 2036–40. However, in the next 25 years, the vaccination strategy will have no direct effect on women born before 1991 who were not vaccinated before HPV exposure. In the short term, the timeliness of the introduction of HPV primary screening into the screening programme will be the most important determinant of the potential reduction in the number of cervical cancers diagnosed among unvaccinated women. Unfortunately, the risk of acquiring an HPV infection that will progress to cancer has increased in unvaccinated individuals born since 1960 (data sourced from the Genitourinary Medicine Clinic Activity dataset and Public Health England [KC60]), suggesting that current screening coverage is not sufficient to maintain—much less reduce—cervical cancer incidence in the next 20 years.

progress to cancer has increased in unvaccinated individuals born since 1960 (data sourced from the Genitourinary Medicine Clinic Activity dataset and Public Health England [KC60]), suggesting that current screening coverage is not sufficient to maintain—much less reduce—cervical cancer incidence in the next 20 years. Age period cohort models have been used to estimate the impact of screening by assuming that observed period effects reflect the outcome of screening.16, 17 Dynamic transmission models18 consider the effects of HPV prevalence and transmission, cervical screening, and HPV vaccination on cervical cancer incidence. For absolute efficacy, they need to be carefully calibrated to each population but relative efficacy is largely invariant to calibration. The current study combined an age cohort model with a dynamic model to estimate the effect of screening, taking into account screening attendance and vaccination coverage. We used population-based data to estimate the risk of being diagnosed with cervical cancer by screening attendance. This method enabled the estimation of absolute changes to cervical cancer incidence and allowed us to vary screening coverage and vaccine uptake to estimate absolute effect on rates. We know of no similar studies with cancer as an outcome, although we did identify a similar study with stroke as the main outcome.19 Other studies of cervical cancer in the UK use just one type of modelling; these include an economic evaluation of HPV vaccination,20 a dynamic model exploring the effect of HPV vaccination,21 and a microsimulation study considering screening in England following the introduction of nine-valent vaccination.22 None of them have estimated age-specific absolute risks by calendar time.

modelling; these include an economic evaluation of HPV vaccination,20 a dynamic model exploring the effect of HPV vaccination,21 and a microsimulation study considering screening in England following the introduction of nine-valent vaccination.22 None of them have estimated age-specific absolute risks by calendar time. The age cohort model used to underpin this research was adapted to ensure that cervical cancer rates in the absence of screening were not predicted to increase substantially beyond the greatest rates observed historically. This constraint was particularly important for more recent cohorts, which have seen rates in young women more than double in the past 15 years. We have assumed, on the basis of published literature,23, 24 that opportunistic screening had no effect on cervical cancer incidence before 1988. However, if screening before 1988 did have an effect, our age period cohort model will have underestimated cervical cancer rates in the absence of screening up to 2040. This effect would mostly affect birth cohorts born before 1955.

24 that opportunistic screening had no effect on cervical cancer incidence before 1988. However, if screening before 1988 did have an effect, our age period cohort model will have underestimated cervical cancer rates in the absence of screening up to 2040. This effect would mostly affect birth cohorts born before 1955. It was recently estimated, using individual level screening history data, that in the absence of screening, cervical cancer incidences in England would be 2·5 times higher than current rates.25 The projected increases in incidence from the age cohort model used in our study suggest a similar (3·1-times) increase in rates (appendix pp 2, 3). Although the benefit of screening in the model could be overestimated because of a higher underlying risk of cervical cancer in women who do not attend screening, given the similarity between the age cohort estimates and those from individual level screening data, the magnitude of the overestimation must be small.

3). Although the benefit of screening in the model could be overestimated because of a higher underlying risk of cervical cancer in women who do not attend screening, given the similarity between the age cohort estimates and those from individual level screening data, the magnitude of the overestimation must be small. The model does not take into account the effect of herd immunity. Because of sexual mixing patterns, the majority of the impact of herd immunity would occur in women born after 1985, so its effect on overall rates will be small. It will be more substantial when looking at age-specific rates, particularly where vaccine uptake decreases. We also do not account for type replacement (eg, from HPV types 16 and 18 to HPV 59), which would diminish the effect of HPV 16/18 vaccination.26 Furthermore, the model does not take into account changes to the effectiveness of cervical screening by cytology. On the one hand, with the prevalence of disease decreasing among vaccinated women, and less cytology taking place following the introduction of HPV screening, it is possible that the cytology workforce might become less skilled, leading to a decreased effect of screening. On the other hand, fewer cytoscreeners looking only at HPV-positive samples (in which high-grade precancerous lesions will be more common) might increase the sensitivity of cytology.

ction of HPV screening, it is possible that the cytology workforce might become less skilled, leading to a decreased effect of screening. On the other hand, fewer cytoscreeners looking only at HPV-positive samples (in which high-grade precancerous lesions will be more common) might increase the sensitivity of cytology. The multidisciplinary nature of the screening programme means that a great number of organisations and individuals are involved in delivering the programme, which has resulted in extensive piloting and planning so that there has been wide variation in when different programmes have introduced HPV testing in primary cervical screening. For example, Kaiser Premanente Northern California introduced primary HPV screening in combination with primary cytology in 2003.27 The Netherlands, Sweden, and Australia are expected to roll out HPV primary screening by the end of 2017.28, 29 In England, six pilot sites covering about 5% of the population have been offering primary HPV screening since 2013. In 2015, baseline results from the pilot sites confirmed that screening by HPV primary screening achieved a higher detection rate of CIN 2 (cervical squamous intraepithelial neoplasia 2) or worse than does cytology screening.30 National roll-out is planned for 2019.31 The complexity of implementing any new intervention into a successful screening programme and the potential for harm has in the past led to a cautious reaction to change. For example, the roll-out of liquid-based cytology took almost 5 years.32 In our model, we estimated that failure to introduce HPV testing nationally in 2017 will lead to the missed opportunity of preventing about 1400 cervical cancer cases between 2018 and 2028.

rm has in the past led to a cautious reaction to change. For example, the roll-out of liquid-based cytology took almost 5 years.32 In our model, we estimated that failure to introduce HPV testing nationally in 2017 will lead to the missed opportunity of preventing about 1400 cervical cancer cases between 2018 and 2028. Increasing screening coverage in unvaccinated women will remain a considerable challenge for the programme. We have assumed that increasing coverage brings in never-screened women and lapsed attenders in equal proportions. Engagement of women who have never attended screening is substantially harder than convincing women who have previously engaged with the programme.33, 34 There is concern that cervical cancer will become a disease of the underprivileged. Marginalised women are less likely to be vaccinated and have lower awareness of screening than those who are not marginalised.35 Engagement with these women is essential.

ncing women who have previously engaged with the programme.33, 34 There is concern that cervical cancer will become a disease of the underprivileged. Marginalised women are less likely to be vaccinated and have lower awareness of screening than those who are not marginalised.35 Engagement with these women is essential. Once HPV primary screening is fully introduced, screening intervals are expected to increase to at least 5-yearly at all ages.36 The cost-effectiveness of delivering HPV screening once every 5 years to vaccinated women needs to be reassessed because there is evidence to suggest that one test every 10 years might be sufficient.37 Furthermore, it seems possible that the programme will use new technologies, such as genotyping or DNA methylation testing, which will allow for individual risk-profiling with variable time to next screening invitation depending on risk. Vaccination of women older than 25 years has been proposed; however, the protection offered to these women appears to be considerably lower than the protection observed in women aged 12–13 years.38 Clear screening campaign messages will become essential if we are to continue engaging women with screening. To meet these challenges, policy makers need to ensure that there is a well thought-out mechanism to introduce change into the screening programme swiftly and effectively without compromising quality. Supplementary Material Supplementary appendix

Once HPV primary screening is fully introduced, screening intervals are expected to increase to at least 5-yearly at all ages.36 The cost-effectiveness of delivering HPV screening once every 5 years to vaccinated women needs to be reassessed because there is evidence to suggest that one test every 10 years might be sufficient.37 Furthermore, it seems possible that the programme will use new technologies, such as genotyping or DNA methylation testing, which will allow for individual risk-profiling with variable time to next screening invitation depending on risk. Vaccination of women older than 25 years has been proposed; however, the protection offered to these women appears to be considerably lower than the protection observed in women aged 12–13 years.38 Clear screening campaign messages will become essential if we are to continue engaging women with screening. To meet these challenges, policy makers need to ensure that there is a well thought-out mechanism to introduce change into the screening programme swiftly and effectively without compromising quality. Supplementary Material Supplementary appendix Acknowledgments This study was funded by Jo's Cervical Cancer Trust and Cancer Research UK (C8162/A16872), and presents independent research commissioned by Jo's Cervical Cancer Trust. The views expressed are those of the authors and not those of the funders.

To meet these challenges, policy makers need to ensure that there is a well thought-out mechanism to introduce change into the screening programme swiftly and effectively without compromising quality. Supplementary Material Supplementary appendix Acknowledgments This study was funded by Jo's Cervical Cancer Trust and Cancer Research UK (C8162/A16872), and presents independent research commissioned by Jo's Cervical Cancer Trust. The views expressed are those of the authors and not those of the funders. Contributors PS conceived and designed the study. AC interpreted the data and wrote the first draft. FP fitted the age period cohort model. PW coded the microsimulation model. RL designed and analysed the microsimulation model and combined the various components of the analysis. All authors edited the report and approved the final version. Declaration of interests All authors have received grants from Jo's Cervical Cancer Trust and Cancer Research UK, during the conduct of the study. PS has also received personal fees from Hologic and non-financial support from PreventX, outside the submitted work.

Introduction Prevention of unintended pregnancy is a primary strategy to reduce adverse pregnancy and birth outcomes related to Zika virus infection.1, 2 Puerto Rico has the highest number of symptomatic Zika virus infections in the USA and US territories, including infections in women.3 Additionally, 65% of pregnancies in Puerto Rico are unintended, and about 138 000 of the 715 000 women aged 15–44 years in Puerto Rico are at risk for unintended pregnancy.4 5–10% of the pregnancies with laboratory-confirmed Zika virus infection that were reported to the US Zika Pregnancy Registry resulted in a fetus or infant with Zika-virus-associated birth defects, and the full range of adverse development outcomes is not yet known.5 The threat of severe birth defects associated with Zika virus infection during pregnancy underscores the importance of contraception to prevent unintended pregnancies. However, a review of existing data and in-depth interviews with key informants early in the Zika virus outbreak in March, 2016, demonstrated that contraceptive access in Puerto Rico was limited by reduced availability of the full range of reversible methods, high out-of-pocket costs, insufficient provider reimbursement, logistical barriers that limit same-day provision, lack of patient education, and shortage of providers trained in insertion, removal, and management of long-acting reversible contraception (LARC), which includes intrauterine devices and contraceptive implants.4 LARC is a highly effective, safe, cost-effective, and user-friendly method of contraception that reduces unintended pregnancy and abortion.6, 7, 8, 9 In 2002–14, LARC use in the USA increased from 2·4% to 14·3% of women using contraception.10 However, LARC use in Puerto Rico was low before the Zika virus outbreak, with estimates indicating that less than 1% of women using contraception used a LARC method.4

on that reduces unintended pregnancy and abortion.6, 7, 8, 9 In 2002–14, LARC use in the USA increased from 2·4% to 14·3% of women using contraception.10 However, LARC use in Puerto Rico was low before the Zika virus outbreak, with estimates indicating that less than 1% of women using contraception used a LARC method.4 Research in context Evidence before this study We searched PubMed for articles published on or before April 1, 2016, using the terms “Contraceptive Choice Project”, “Zika and family planning”, and “Zika and contraception”. The Contraceptive CHOICE Project was a prospective cohort study of 10 000 women of reproductive age in St Louis, MO, USA, who wanted to prevent pregnancy and initiate a new method of contraception. The study was designed to introduce and promote the use of long-acting reversible contraception (LARC) methods, and the results showed that 65% of participating women chose LARC methods when cost, provider, and facility barriers were removed. In a report from April 1, 2016, early in Puerto Rico's 2016–17 Zika virus outbreak, women in the country were shown to have a high unmet need for contraception, high incidence of unintended pregnancy, poor access to contraception, and the highest number of Zika infections in the USA and US territories. We did not identify any studies that described a contraception-focused programme as part of the response to the Zika virus outbreak. Added value of this study

We searched PubMed for articles published on or before April 1, 2016, using the terms “Contraceptive Choice Project”, “Zika and family planning”, and “Zika and contraception”. The Contraceptive CHOICE Project was a prospective cohort study of 10 000 women of reproductive age in St Louis, MO, USA, who wanted to prevent pregnancy and initiate a new method of contraception. The study was designed to introduce and promote the use of long-acting reversible contraception (LARC) methods, and the results showed that 65% of participating women chose LARC methods when cost, provider, and facility barriers were removed. In a report from April 1, 2016, early in Puerto Rico's 2016–17 Zika virus outbreak, women in the country were shown to have a high unmet need for contraception, high incidence of unintended pregnancy, poor access to contraception, and the highest number of Zika infections in the USA and US territories. We did not identify any studies that described a contraception-focused programme as part of the response to the Zika virus outbreak. Added value of this study The Zika Contraception Access Network (Z-CAN) is the first to describe the large-scale implementation of a comprehensive programme to rapidly expand access to contraceptives during a major public health emergency response. The programme was implemented quickly and was able to serve more women than previous projects based on expansion of contraceptive access. Z-CAN included introduction to and education about LARC methods for both providers and patients with no previous exposure to or experience with these newer contraceptive methods.

se. The programme was implemented quickly and was able to serve more women than previous projects based on expansion of contraceptive access. Z-CAN included introduction to and education about LARC methods for both providers and patients with no previous exposure to or experience with these newer contraceptive methods. Implications of all the available evidence This large and rapidly established contraception programme could be replicated in other areas with serious and complex public health emergencies to ensure that unintended births are averted. Although this programme was developed to prevent unintended pregnancies and birth defects associated with Zika virus infection, avoiding unintended pregnancy is an important strategy for a wide variety of public health responses, particularly in view of frequent disruptions in care and services in emergency settings.

. Although this programme was developed to prevent unintended pregnancies and birth defects associated with Zika virus infection, avoiding unintended pregnancy is an important strategy for a wide variety of public health responses, particularly in view of frequent disruptions in care and services in emergency settings. Recognising the importance of contraceptive access during the Zika virus outbreak, the National Foundation for the Centers for Disease Control and Prevention (CDCF), with technical assistance from the Centers for Disease Control and Prevention (CDC) and in collaboration with a diverse group of stakeholders and private donors, established the Zika Contraception Access Network (Z-CAN) in Puerto Rico. Z-CAN was a short-term response (from May, 2016, to September, 2017) for rapid implementation of reversible contraceptive services in a complex emergency setting. Z-CAN aimed to build a network of health-care providers trained in client-centred contraceptive counselling and same-day provision of the full range of reversible contraceptive methods (including LARC) at no cost to women who choose to delay or avoid pregnancy, and to raise awareness in women and families of contraception as a primary prevention measure to reduce adverse pregnancy and birth outcomes related to Zika virus infection. In addition to access barriers, a history of coerced sterilisation and concern for unethical testing of oral contraceptives in Puerto Rico were important considerations in programme design.11, 12

ontraception as a primary prevention measure to reduce adverse pregnancy and birth outcomes related to Zika virus infection. In addition to access barriers, a history of coerced sterilisation and concern for unethical testing of oral contraceptives in Puerto Rico were important considerations in programme design.11, 12 Here we describe the Z-CAN programme design and implementation activities and the baseline characteristics of the first 21 124 women served through Z-CAN. Methods Programme design and implementation Z-CAN was designed to address gaps in contraceptive access and service provision in Puerto Rico as a preventive measure to reduce the effect of Zika virus on infants. The development of Z-CAN included several strategies to rapidly reduce access barriers to contraception in Puerto Rico's health system, strengthen infrastructure to support the Z-CAN programme, and work towards the sustainability of reversible contraceptive services after the Z-CAN programme ends (figure 1).Figure 1 Zika Contraception Access Network (Z-CAN) major milestones, 2016–17

educe access barriers to contraception in Puerto Rico's health system, strengthen infrastructure to support the Z-CAN programme, and work towards the sustainability of reversible contraceptive services after the Z-CAN programme ends (figure 1).Figure 1 Zika Contraception Access Network (Z-CAN) major milestones, 2016–17 CDC=Centers for Disease Control and Prevention. CMS=Centers for Medicare and Medicaid Services. HRSA=Health Resources and Service Administration. OPA=Office of Population Affairs. FLASOG=Federacion Latinoamericana de Sociedades de Obstetricia y Ginecologia. ACOG=American College of Obstetricians and Gynecologists. SOGC=The Society of Obstetricians and Gynecologists. PRDOH=Puerto Rico Department of Health. AO=Administrative Order. HHS OIG=Health and Human Services Office of the Inspector General. FDA=US Food and Drug Administration. VA=Veterans Administration.

Ginecologia. ACOG=American College of Obstetricians and Gynecologists. SOGC=The Society of Obstetricians and Gynecologists. PRDOH=Puerto Rico Department of Health. AO=Administrative Order. HHS OIG=Health and Human Services Office of the Inspector General. FDA=US Food and Drug Administration. VA=Veterans Administration. The development of strong partnerships was crucial in the design and implementation of Z-CAN. The programme was built with a network of partners including federal agencies, territorial health agencies, private corporations, and domestic philanthropic and non-profit organisations in the continental USA and Puerto Rico. Private donors provided product commitments to CDCF for the full range of reversible contraceptive methods (including LARC methods). CDCF established a plan for contraception procurement and distribution adherent to US Food and Drug Administration (FDA) and territorial guidelines and for private donations through CDCF-supported provider reimbursement and infrastructure costs to ensure contraception was available to women at no cost.

g LARC methods). CDCF established a plan for contraception procurement and distribution adherent to US Food and Drug Administration (FDA) and territorial guidelines and for private donations through CDCF-supported provider reimbursement and infrastructure costs to ensure contraception was available to women at no cost. The gaps in contraceptive access and service provision4 meant that it was necessary to build provider and staff capacity in contraception knowledge, counselling, and initiation and management, including the insertion and removal of LARC. Z-CAN recruited doctors and clinic staff (nurses and clinic administrators) from all public health regions and nearly all municipalities on the island who practised in private and publicly funded clinics and who were interested in receiving training in the provision of contraception.13 Doctors and clinic staff were not recruited from municipalities with no community health centres, government facilities, or private practices providing women's health care. Doctors and staff were recruited through the Puerto Rico section of the American College of Obstetricians and Gynecologists, Puerto Rico Obstetrics and Gynecology, the Puerto Rico Department of Health, the Puerto Rico Primary Care Association, the Puerto Rico Health Insurance Administration, and Medicaid-managed care organisations. Before Z-CAN, none of the participating clinics routinely provided levonorgestrel-releasing intrauterine devices or contra-ceptive implants, and access to copper intrauterine devices was very limited. A 1-day comprehensive training course offered participants an overview of Zika virus (including the risk of sexual transmission and the importance of condom use for disease prevention), a tested curriculum on client-centred contraceptive counselling, didactic information about the full range of reversible contraceptives, a review of evidence-based contraceptive guidelines,14, 15 practical training in insertion and removal of intrauterine devices (providers were observed on three to five simulations),13 an FDA-approved etonogestrel implant training, and a overview of Z-CAN policies and procedures.

ull range of reversible contraceptives, a review of evidence-based contraceptive guidelines,14, 15 practical training in insertion and removal of intrauterine devices (providers were observed on three to five simulations),13 an FDA-approved etonogestrel implant training, and a overview of Z-CAN policies and procedures. Provider reimbursement for these services was previously identified as barriers to contraception access.4 Through Z-CAN, private donations were used to provide a level of provider reimbursement that was commensurate with Medicaid reimbursement rates in the continental USA. This reimbursement covered client-centred contraceptive counselling for women and their partners, if desired, and method provision. If a LARC was provided, the reimbursement fee was bundled to include both insertion and removal at the time of the insertion visit to ensure that women could have their LARC devices removed when desired at no cost. After initial training, a Z-CAN staff member and a family planning specialist proctored providers and clinic staff to ensure delivery of high-quality care. Proctoring visits consisted of: direct observation of contraceptive counselling, at least one insertion of an intrauterine device, and staff interaction with patients; review of data collection, inventory tracking, and billing procedures; and a clinic audit to ensure that supplies, space, equipment, and security were sufficient to participate in Z-CAN. If provider, staff, and clinic met all readiness criteria, they were authorised to receive contraceptive products and to begin offering Z-CAN services.

lection, inventory tracking, and billing procedures; and a clinic audit to ensure that supplies, space, equipment, and security were sufficient to participate in Z-CAN. If provider, staff, and clinic met all readiness criteria, they were authorised to receive contraceptive products and to begin offering Z-CAN services. Data collection and analysis Women learned of Z-CAN through providers, word of mouth, and a health education campaign involving community engagement activities, Z-CAN materials, posters in health centres, a campaign website, and a Facebook page. Non-sterilised women of reproductive age were eligible to receive Z-CAN services, irrespective of age or insurance status. All Z-CAN services were provided free of charge. At the initial Z-CAN visit, women were assigned a unique identification number. Providers and clinic staff recorded women's demographic information, reproductive and contraception histories, and their chosen contraceptive method. Data were submitted without personal identifying information to the Z-CAN programme and entered into a REDCap database hosted on a secure server.16 The data presented here are descriptive characteristics of programme providers and women receiving Z-CAN services. To examine factors associated with choosing and receiving a LARC method, we estimated unadjusted and adjusted prevalence ratios with 95% CI. Data were analysed using SAS-callable SUDAAN version 11.0.0 to account for clustering of patients within clinic-provider dyads.

ramme providers and women receiving Z-CAN services. To examine factors associated with choosing and receiving a LARC method, we estimated unadjusted and adjusted prevalence ratios with 95% CI. Data were analysed using SAS-callable SUDAAN version 11.0.0 to account for clustering of patients within clinic-provider dyads. The CDC's Public Health Ethics Committee (PHEC) provided internal consultation during the programme and project design to ensure no conflicts of interest existed and to address any ethical concerns.17 The Public Health Ethics Conflict of Interest Work Group, part of the CDC Zika Response Emergency Operations Center and comprised of individuals from the PHEC, reviewed the Z-CAN programme proposal during its design phase and recommended that the programme offer the full range of reversible contraceptive methods and have measures in place to prevent coercion of women.

Group, part of the CDC Zika Response Emergency Operations Center and comprised of individuals from the PHEC, reviewed the Z-CAN programme proposal during its design phase and recommended that the programme offer the full range of reversible contraceptive methods and have measures in place to prevent coercion of women. As part of the Z-CAN programme monitoring plan, women were invited to participate in a 10 min self-administered online survey within 2 weeks of their initial visit. Z-CAN-trained clinic staff collected contact information from women who did not opt out of being contacted for future surveys. Women were invited to participate in the survey via email or text message; those without online access could complete the survey on the telephone with programme staff. The survey measured whether participants received free same-day access to the contraceptive method of their choice after receiving comprehensive counselling, patient perception of the received quality of care, and satisfaction with their chosen method and services. Perception of quality of care was measured using the validated interpersonal quality of family planning care scale,18 comprised of 11 items measured using a five-point Likert scale (a score of 1 means poor; a score of 5 means excellent; appendix). No personal identifiers were collected, and unique identification numbers were used to merge survey responses with initial visit data. Women were considered non-respondents if they did not complete the survey within 3 weeks after confirmed receipt of email or text message invitation and after up to three outreach attempts. Responses were collected through Survey Monkey online software, and respondents received a US$10 electronic gift card. We used SAS version 9.3 to compare baseline characteristics of survey respondents and non-respondents.

ter confirmed receipt of email or text message invitation and after up to three outreach attempts. Responses were collected through Survey Monkey online software, and respondents received a US$10 electronic gift card. We used SAS version 9.3 to compare baseline characteristics of survey respondents and non-respondents. The Z-CAN programme and patient satisfaction survey were determined by CDC to be non-research public health practice activities and thus exempt from Institutional Review Board review. The programme did not obtain consent from women served by Z-CAN providers. The women received a letter at their initial visit that described the follow-up contact planned for programme monitoring purposes and were given the opportunity to opt out. Women who did not opt out were invited to participate in the patient satisfaction survey. If a woman chose to participate in the survey, she did so by consenting to the survey within the online environment. Role of the funding source The philanthropic donors to CDCF had no role in programme design, data collection, data analysis, data interpretation, or writing of the report. CDC provided technical assistance in collaboration with CDCF for programme design and implementation. The cor-responding author had full access to all of the data and the final responsibility to submit for publication.

e in programme design, data collection, data analysis, data interpretation, or writing of the report. CDC provided technical assistance in collaboration with CDCF for programme design and implementation. The cor-responding author had full access to all of the data and the final responsibility to submit for publication. Results Training for providers took place between April 30, 2016, and Dec 6, 2016. 177 doctors, including nine resident doctors training in obstetrics and gynaecology, each participated in one of the eight Z-CAN training sessions. Of those who completed training, 153 practising doctors (141 obstetrician gynaecologists and 12 family doctors or paediatricians) agreed to participate in Z-CAN, completed proctoring visits, and received contraceptive supplies to provide Z-CAN services. The characteristics of providers are listed in table 1. 139 clinics across the island participated in the Z-CAN project (figure 2). The Z-CAN programme design, scale-up, and implementation occurred rapidly across the island, and the first Z-CAN contraception services were offered on May 4, 2016.Figure 2 Puerto Rico Zika Contraception Access Network clinics *Includes 17 community health centres and 23 satellite clinics. Source: Zika Contraception Access Network as of Sept 23, 2017. Table 1 Characteristics of Zika Contraception Access Network (Z-CAN) providers and the first 21 124 women enrolled in the Z-CAN programme, as of Aug 15, 2017

Results Training for providers took place between April 30, 2016, and Dec 6, 2016. 177 doctors, including nine resident doctors training in obstetrics and gynaecology, each participated in one of the eight Z-CAN training sessions. Of those who completed training, 153 practising doctors (141 obstetrician gynaecologists and 12 family doctors or paediatricians) agreed to participate in Z-CAN, completed proctoring visits, and received contraceptive supplies to provide Z-CAN services. The characteristics of providers are listed in table 1. 139 clinics across the island participated in the Z-CAN project (figure 2). The Z-CAN programme design, scale-up, and implementation occurred rapidly across the island, and the first Z-CAN contraception services were offered on May 4, 2016.Figure 2 Puerto Rico Zika Contraception Access Network clinics *Includes 17 community health centres and 23 satellite clinics. Source: Zika Contraception Access Network as of Sept 23, 2017. Table 1 Characteristics of Zika Contraception Access Network (Z-CAN) providers and the first 21 124 women enrolled in the Z-CAN programme, as of Aug 15, 2017 n/N (%) Provider characteristics Provider type Obstetrician-gynaecologist 141/153 (92%) Family doctor 10/153 (7%) Paediatrician 2/153 (1%) Practice type Private practice 102/153 (67%) Community health centre* 38/153 (25%) Public health clinic† 3/153 (2%) Academic clinic‡ 10/153 (7%) Participant characteristics Age, years ≤20 4539/21 124 (22%) 21–24 6057/21 124 (29%) 25–34 7759/21 124 (37%) ≥35 2558/21 124 (12%) Relationship status Single 8887/21 124 (42%) Married or partnered 11 979/21 124 (57%) Education ≤12 years 7895/21 124 (37%) College degree 11 024/21 124 (52%) Graduate degree 1941/21 124 (9%) Insurance status Private or other 8813/21 124 (42%) Public 10 786/21 124 (51%) None 1111/21 124 (5%) Previous livebirth 0 7762/21 124 (37%) ≥1 12 491/21 124 (59%) Breastfeeding at time of initial visit No 17 213/21 124 (82%) Yes 3350/21 124 (16%) Did not want to conceive in the next year 20 829/21 124 (95%) Received same-day services 20 110/21 124 (95%) Did not receive a contraceptive method at initial visit 959/21 124 (5%) Undecided or not ready 410/959 (43%) Might be pregnant 217/959 (23%) Desired method out of stock 97/959 (10%) Medical reason 83/959 (9%) Reason not specified 78/959 (8%) Did not want a contraceptive method 37/959 (4%) Continuing current method 26/959 (3%) Pregnant 11/959 (1%) Proportions might not add up to 100% because of missing data.

eady 410/959 (43%) Might be pregnant 217/959 (23%) Desired method out of stock 97/959 (10%) Medical reason 83/959 (9%) Reason not specified 78/959 (8%) Did not want a contraceptive method 37/959 (4%) Continuing current method 26/959 (3%) Pregnant 11/959 (1%) Proportions might not add up to 100% because of missing data. * Funded by the Health Resources and Services Administration. † Funded by the Puerto Rico Department of Public Health. ‡ Affiliated with the University of Puerto Rico. As of Aug 15, 2017, data were available for 21 124 women who had attended an initial visit in the Z-CAN programme (table 1). The mean age of participants was 26 years (SD 6·66).

* Funded by the Health Resources and Services Administration. † Funded by the Puerto Rico Department of Public Health. ‡ Affiliated with the University of Puerto Rico. As of Aug 15, 2017, data were available for 21 124 women who had attended an initial visit in the Z-CAN programme (table 1). The mean age of participants was 26 years (SD 6·66). The distribution of contraception methods used by women before and after joining the Z-CAN programme is shown in figure 3. Before their initial Z-CAN visit, most women used either no method or one of the least effective contraceptive methods (condoms, sponge, withdrawal, spermicide, or fertility awareness methods), and only a small proportion of women used one of the most effective methods (male sterilisation, intrauterine device, or implant; figure 3). At their visit, more than 14 259 (68%) women chose and received a LARC method and 5250 (25%) women chose oral contraceptive pills or other moderately effective hormonal contraception (eg, depot medroxyprogesterone acetate injection). Of the 959 (5%) women who did not receive a contraceptive method, the most common reasons were being undecided on method preference or not ready to receive the method that day, pregnancy could not be ruled out, or the desired method was not in stock (table 1). Of the 14 259 women who chose and received a LARC method, 7167 (50%) women received a levonorgestrel-releasing intrauterine device, 5031 (35%) women received an etonogestrel implant, and 2061 (14%) women received a copper intrauterine device. Women were more likely to choose and receive a LARC method if they had a college degree, had no insurance, had at least one livebirth, used a most effective contraceptive method before Z-CAN, and saw a Z-CAN provider in private practice or a public health or academic clinic, after adjustment for all other characteristics (table 2). Women aged 25 years or more and women using a moderately effective contraceptive method before Z-CAN were less likely to choose and receive a LARC method. Results were similar when the analysis was restricted to women who received a contraceptive method at their initial visit.Figure 3 Contraceptive method use by women before and after their initial visit to a Zika Contraception Access Network (Z-CAN) provider in Puerto Rico, as of Aug 15, 2017 (N=21 124)

e a LARC method. Results were similar when the analysis was restricted to women who received a contraceptive method at their initial visit.Figure 3 Contraceptive method use by women before and after their initial visit to a Zika Contraception Access Network (Z-CAN) provider in Puerto Rico, as of Aug 15, 2017 (N=21 124) Proportions might not add up to 100% because of missing data. Most effective contraceptive methods include intrauterine devices, implants, and partner sterilisation. Less than 1% of women using these methods will get pregnant during the first year of typical use. Moderately effective contraceptive methods include injectables, pills, patch, ring, and diaphragm. 6–12% of women using these methods will get pregnant during the first year of typical use. Least effective birth control methods include male and female condoms, withdrawal, sponge, fertility awareness methods, and spermicides. Least effective birth control methods have a failure rate of 18 or more pregnancies per 100 women who use these methods each year. The Centres for Disease Control and Prevention have produced an overview of the effectiveness of family planning methods. Methods provided by Z-CAN included intrauterine devices, implants, injectables, pills, patch, ring, and male condoms. Table 2 Factors associated with choosing and receiving a LARC method among the first 21 124 women enrolled in the Zika Contraception Access Network (Z-CAN) programme, as of Aug 15, 2017

Proportions might not add up to 100% because of missing data. Most effective contraceptive methods include intrauterine devices, implants, and partner sterilisation. Less than 1% of women using these methods will get pregnant during the first year of typical use. Moderately effective contraceptive methods include injectables, pills, patch, ring, and diaphragm. 6–12% of women using these methods will get pregnant during the first year of typical use. Least effective birth control methods include male and female condoms, withdrawal, sponge, fertility awareness methods, and spermicides. Least effective birth control methods have a failure rate of 18 or more pregnancies per 100 women who use these methods each year. The Centres for Disease Control and Prevention have produced an overview of the effectiveness of family planning methods. Methods provided by Z-CAN included intrauterine devices, implants, injectables, pills, patch, ring, and male condoms. Table 2 Factors associated with choosing and receiving a LARC method among the first 21 124 women enrolled in the Zika Contraception Access Network (Z-CAN) programme, as of Aug 15, 2017 LARC (n=14 259) Other contraceptive method (n=6810) Unadjusted prevalence ratio, 95% CI Adjusted prevalence ratio, 95% CI* Age, years ≤20 2930/14 125 (21%) 1594/6734 (24%) Referent Referent 21–24 4176/14 125 (30%) 1868/6734 (28%) 1·07, 1·03–1·10† 1·00, 0·97–1·03 25–34 5305/14 125 (38%) 2435/6734 (36%) 1·06, 1·02–1·10† 0·93, 0·90–0·97† ≥35 1714/14 125 (12%) 837/6734 (12%) 1·04, 0·98–1·10 0·85, 0·80–0·92† Relationship status Single 5717/14 106 (41%) 3148/6709 (47%) Referent Referent Married or partnered 8389/14 106 (60%) 3561/6709 (53%) 1·09, 1·04–1·14† 0·99, 0·95–1·04 Education ≤12 years 5258/14 094 (37%) 2617/6712 (39%) Referent Referent College degree 7585/14 094 (54%) 3411/6712 (51%) 1·03, 1·00–1·07 1·04, 1·01–1·08† Graduate degree 1251/14 094 (9%) 684/6712 (10%) 0·97, 0·91–1·03 1·02, 0·96–1·08 Insurance status Private or other 5827/13 970 (42%) 2968/6689 (44%) Referent Referent Public 7326/13 970 (52%) 3429/6689 (51%) 1·03, 0·97–1·09 0·97, 0·91–1·02 None 817/13 970 (6%) 292/6689 (4%) 1·11, 1·05–1·18† 1·11, 1·05–1·17† Previous livebirth 0 4301/13 688 (31%) 3431/6511 (53%) Referent Referent 1 or more 9387/13 688 (69%) 3080/6511 (47%) 1·35, 1·27–1·44† 1·40, 1·31–1·48† Currently breastfeeding No 11 271/13 884 (81%) 5892/6626 (89%) Referent Referent Yes 2613/13 884 (19%) 734/6626 (11%) 1·19, 1·14–1·24† 1·03, 0·99–1·08 Effectiveness of contraceptive method used before Z-CAN‡ None 6357/14 097 (45%) 2909/6683 (44%) Referent Referent Least 4451/14 097 (32%) 1757/6683 (26%) 1·05, 0·98–1·11 1·05, 0·99–1·11 Moderately 2666/14 097 (19%) 1874/6683 (28%) 0·86, 0·82–0·89† 0·90, 0·86–0·94† Most 623/14 097 (4%) 143/6683 (2%) 1·19, 1·12–1·25† 1·13, 1·06–1·21† Clinic type Community health clinic 2154/14 259 (15%) 1521/6810 (22%) Referent Referent Private practice or other 12 105/14 259 (85%) 5289/6810 (78%) 1·19, 1·06–1·33† 1·19, 1·07–1·33† Data are n/N (%) unless indicated otherwise. LARC=long-acting reversible contraceptive.

143/6683 (2%) 1·19, 1·12–1·25† 1·13, 1·06–1·21† Clinic type Community health clinic 2154/14 259 (15%) 1521/6810 (22%) Referent Referent Private practice or other 12 105/14 259 (85%) 5289/6810 (78%) 1·19, 1·06–1·33† 1·19, 1·07–1·33† Data are n/N (%) unless indicated otherwise. LARC=long-acting reversible contraceptive. * Each characteristic in the table was adjusted for all other characteristics. † 95% CI does not include 1. ‡ Least effective contraceptive methods include condoms for men and women, withdrawal, sponge, fertility awareness methods, and spermicides. Moderately effective contraceptive methods include injectables, pills, patch, ring, and diaphragm. Most effective contraceptive methods include intrauterine devices, implants, and partner sterilisation. Sterilised women were not eligible for Z-CAN services.

hdrawal, sponge, fertility awareness methods, and spermicides. Moderately effective contraceptive methods include injectables, pills, patch, ring, and diaphragm. Most effective contraceptive methods include intrauterine devices, implants, and partner sterilisation. Sterilised women were not eligible for Z-CAN services. The satisfaction survey began on Oct 28, 2016. By July 21, 2017, 9829 women had received invitations to complete the patient satisfaction survey, and 3489 (36%) women had responded (2482 women responded by email invitation, 1006 women responded by text message invitation, and one woman responded by phone administration). We were able to link initial visit data to survey data for 3439 (99%) respondents. Respondents differed from non-respondents with respect to age, insurance status, and type of method received; compared with non-respondents, respondents overall were slightly older, had private insurance, and chose a more effective method during their visit. 3489 women participated in the patient satisfaction survey, but not all women completed every question of the survey. 3068 (93%) of the 3294 women who answered the question about their satisfaction with services were very satisfied, 203 (6%) women were somewhat satisfied, and 23 (1%) women were not satisfied. 3216 (93%) of the 3478 women who answered the question about receiving the method they were most interested in after receiving counselling did receive the method they were most interested in. Of the 3040 women who completed every item on the 11-item interpersonal quality of family planning care scale, 2382 (78%) respondents rated their care as excellent or very good on all 11 items. Results from individual items measuring quality of care are summarised in the appendix.

e method they were most interested in. Of the 3040 women who completed every item on the 11-item interpersonal quality of family planning care scale, 2382 (78%) respondents rated their care as excellent or very good on all 11 items. Results from individual items measuring quality of care are summarised in the appendix. Discussion In Puerto Rico, the combination of a high incidence of Zika virus infection, a high incidence of unintended pregnancy, and low use of highly effective contraception necessitated programmatic efforts to improve con-traceptive access as a primary prevention strategy to reduce adverse pregnancy and birth outcomes related to Zika virus infection. The Z-CAN programme shows the feasibility of implementing a programme to increase access to the full range of reversible contraception, including LARC methods, within a complex public health response. Z-CAN also shows that it is possible to build capacity quickly with standardised and targeted training sessions and limited mentoring of committed providers and to provide high-quality, comprehensive contraceptive services in an emergency response.

ption, including LARC methods, within a complex public health response. Z-CAN also shows that it is possible to build capacity quickly with standardised and targeted training sessions and limited mentoring of committed providers and to provide high-quality, comprehensive contraceptive services in an emergency response. Contraception has an important role in the Zika response because Zika virus infection during pregnancy increases the risk for microcephaly and other severe birth defects.2 Contraception could be a key response strategy in other public health emergencies in which prenatal exposures pose a severe risk to pregnant women and their infants.19 Guidance for rapid reproductive health assessment and programme implementation in emergency settings is available, but existing tools position contraception services as post-emergency activities rather than services to be implemented in the emergency phase.20 Z-CAN shows that with concerted effort, commitment, dedicated resources, and recognition of the benefits of giving women the option to prevent pregnancy during a time of crisis, it is possible to prioritise and implement effective contraceptive provision early in an emergency response.

emented in the emergency phase.20 Z-CAN shows that with concerted effort, commitment, dedicated resources, and recognition of the benefits of giving women the option to prevent pregnancy during a time of crisis, it is possible to prioritise and implement effective contraceptive provision early in an emergency response. Contraceptive use and provision in Puerto Rico before the Z-CAN programme was limited by policy, financial, and logistical barriers.4, 21 Most of the 21 124 women seen by the Z-CAN programmme chose and received a LARC method, and most of these women were not using an effective method of contraception before Z-CAN; these findings suggest that when barriers to access are removed (eg, cost, limited service points, and lack of providers), most women who wanted to prevent pregnancy during the Zika virus outbreak chose a highly effective method of contraception. The choice of a LARC method was more likely in women who had previously given birth than in nulliparous women. Intrauterine devices are generally safe for all women, including nulliparous women.14 Providers might have misconceptions about the safety of intrauterine devices in nulliparous women, which have been shown to be associated with infrequent provision,22 emphasising the opportunity for providers to include LARC methods in counselling and eligibility determinations for all women seeking contraception. Although use of LARC methods by women using contraception in the USA is low (14%),10 our findings are consistent with those from other demonstration projects9, 23 that removed barriers to LARC access such as cost, provider availability, geographic access, and comprehensive contraception counselling. Women who chose a short-acting method were given up to 6 months advanced supply. Women who perceived a return visit to receive additional contraceptive supplies as a barrier might have inadvertently been incentivised to choose a LARC method. However, results from the patient satisfaction survey suggested that most women left their initial Z-CAN visit with the method they were most interested in receiving. In the context of the Zika virus outbreak, improved access to contraception has the potential to decrease unintended pregnancies and the number of adverse pregnancy and birth outcomes related to Zika virus infection.1, 4, 24

st women left their initial Z-CAN visit with the method they were most interested in receiving. In the context of the Zika virus outbreak, improved access to contraception has the potential to decrease unintended pregnancies and the number of adverse pregnancy and birth outcomes related to Zika virus infection.1, 4, 24 On the basis of results from multiple large-scale programmes and research studies to reduce barriers to contraceptive access, we anticipated that Z-CAN services would lead to an increase in LARC use. Because of their many advantages, including high effectiveness, safety, reversibility, user ease, high user satisfaction, and cost-effectiveness, LARC methods are crucial in public health efforts to decrease unintended pregnancies. However, issues of perceived or actual provider coercion of women to choose LARC methods (or refuse LARC removals), particularly based on age, race, and class, have been reported.25, 26 The historical context of unethical contraceptive practices and research in Puerto Rico and concerns for reproductive coercion with LARC provision were important considerations in programme design. An important element of the Z-CAN training and proctoring for all providers and clinic staff was to develop competency in delivering high-quality, patient-centred contraceptive counselling that facilitated autonomous decision making.13 Respondents to the satisfaction survey indicated high satisfaction with Z-CAN services, and nearly all women received the method they were most interested in after counselling, suggesting that participants received high-quality and patient-centred services through Z-CAN. The Z-CAN programme evaluation will include additional follow-up surveys of women participating in the programme to further assess quality of and satisfaction with Z-CAN services.

ey were most interested in after counselling, suggesting that participants received high-quality and patient-centred services through Z-CAN. The Z-CAN programme evaluation will include additional follow-up surveys of women participating in the programme to further assess quality of and satisfaction with Z-CAN services. Through partnership and collaboration with a diverse group of stakeholders, Z-CAN reduced barriers to contraception as part of the public health response to the Zika virus outbreak and expanded the capacity of Puerto Rico's health-care system to integrate same-day access to contraceptive services into normal clinic practice. Z-CAN efforts to build sustainability with key stakeholders include building the capacity of a broad network of providers who can provide access to contraception, raising awareness in women of reproductive age in Puerto Rico about the availability of contraceptive methods, expanding the number of contraceptive service access sites, eliminating prior authorisation requirements and cost-sharing in health insurance plans, and discussing continued availability of LARC methods in Puerto Rico through pricing negotiations and development of a sustainable supply chain with manufacturers. Although the total cost to implement, sustain, or replicate the Z-CAN programme is difficult to calculate, the most expensive aspects of the programme were provision of the contraceptive methods (almost all of which were donated in the case of Z-CAN) and provider reimbursement for services. Different contexts will have different cost challenges, but the financing requirements of these crucial aspects might be substantial and should be considered in programme design and sustainability planning. Successful sustainability will be achieved if the elimination of the most pressing barriers addressed by Z-CAN is maintained.

ntexts will have different cost challenges, but the financing requirements of these crucial aspects might be substantial and should be considered in programme design and sustainability planning. Successful sustainability will be achieved if the elimination of the most pressing barriers addressed by Z-CAN is maintained. This programme has several strengths. To our knowledge, Z-CAN is the first contraception access programme developed as a primary prevention strategy to mitigate the effect of a Zika virus outbreak, and it is the first contraception access programme as a primary intervention to prevent adverse pregnancy and birth outcomes in the context of a public health emergency response. The Z-CAN programme contains important elements of both rapid programme design and implementation and sustainability planning and can be adapted to other settings in which improving contraceptive access could enhance the response to an emergency. The strong partnerships between programme teams and stakeholders in Puerto Rico and the high demand for contraceptive services also strengthened the programme.

lementation and sustainability planning and can be adapted to other settings in which improving contraceptive access could enhance the response to an emergency. The strong partnerships between programme teams and stakeholders in Puerto Rico and the high demand for contraceptive services also strengthened the programme. The Z-CAN programme and this study also have several limitations. Although Z-CAN had broad coverage across the island, the programme was not able to provide services in municipalities without health-care infrastructure, so some women had to travel outside their municipality to access care. Because of the rapid design and implementation of Z-CAN and the specific threat of Zika virus to maternal and child health, our results are not readily generalisable to non-emergency situations. The response rate to the patient satisfaction survey was low, and the results of the survey might not be generalisable to all women who received Z-CAN services. The programme was implemented to serve women throughout the risk period for Zika virus transmission, while working towards sustainability of high-quality and accessible contraceptive services. Although the design and implementation phases were relatively fast, rate-limiting steps (eg, design of a procurement and distribution system for donated contraceptive methods) slowed the delivery of services in the early phases of the programme. In view of the challenges of procurement and payment of LARC methods, reaching a level of sustainability of contraceptive services that closely mimics Z-CAN will probably be difficult.21

ment and distribution system for donated contraceptive methods) slowed the delivery of services in the early phases of the programme. In view of the challenges of procurement and payment of LARC methods, reaching a level of sustainability of contraceptive services that closely mimics Z-CAN will probably be difficult.21 Z-CAN was designed as a short-term response for rapid implementation of contraceptive services in a complex emergency setting. Z-CAN has established an extensive network of providers in Puerto Rico and has served more than 21 000 women seeking to prevent pregnancy during the risk period for Zika virus infection. The programme might have prevented unintended pregnancies and birth defects related to Zika virus infections during the outbreak. Mosquito-borne transmission of Zika virus has reached 95 countries worldwide and all but two countries in the Latin America Caribbean Region.27 On the basis of these preliminary results, Z-CAN is a model programme that could be replicated or adapted in these settings as part of emergency preparedness and response efforts. Additionally, Z-CAN's design and implementation could be refined and adapted in other non-emergent settings, in which increased access to contraception could improve health outcomes. For the effectiveness of family planning methods see https://www.cdc.gov/reproductivehealth/unintendedpregnancy/pdf/contraceptive_methods_508.pdf Supplementary Material Supplementary appendix

Z-CAN was designed as a short-term response for rapid implementation of contraceptive services in a complex emergency setting. Z-CAN has established an extensive network of providers in Puerto Rico and has served more than 21 000 women seeking to prevent pregnancy during the risk period for Zika virus infection. The programme might have prevented unintended pregnancies and birth defects related to Zika virus infections during the outbreak. Mosquito-borne transmission of Zika virus has reached 95 countries worldwide and all but two countries in the Latin America Caribbean Region.27 On the basis of these preliminary results, Z-CAN is a model programme that could be replicated or adapted in these settings as part of emergency preparedness and response efforts. Additionally, Z-CAN's design and implementation could be refined and adapted in other non-emergent settings, in which increased access to contraception could improve health outcomes. For the effectiveness of family planning methods see https://www.cdc.gov/reproductivehealth/unintendedpregnancy/pdf/contraceptive_methods_508.pdf Supplementary Material Supplementary appendix Acknowledgments This study was funded by the National Foundation for the Centers for Disease Control and Prevention. Funding via the CDC Foundation was made possible by the Bill & Melinda Gates Foundation, Bloomberg Philanthropies, the William and Flora Hewlett Foundation, Pfizer Foundation, and the American College of Obstetricians and Gynecologists. The CDC Foundation also secured large-scale donations, offers of contraceptive products or services from Allergan, Medicines360, Americares and Janssen, Bayer, Merck & Co, Mylan, Pfizer, Teva Pharmaceuticals, Church & Dwight, RB, The National Campaign to Prevent Teen and Unplanned Pregnancy, Upstream USA, and MarketVision, Culture Inspired Marketing.

large-scale donations, offers of contraceptive products or services from Allergan, Medicines360, Americares and Janssen, Bayer, Merck & Co, Mylan, Pfizer, Teva Pharmaceuticals, Church & Dwight, RB, The National Campaign to Prevent Teen and Unplanned Pregnancy, Upstream USA, and MarketVision, Culture Inspired Marketing. The Zika Contraception Access Network would not have been possible without the support of the CDC Foundation and technical assistance from the CDC. We acknowledge the work of (in alphabetical order) Kate Agin, Dayna Alexander, Tanya Alvarez, Subhashini Babu, Jason Baker, Wanda Barfield, Sarah David, Kennis Dees, Romeo Galang, Luis Garcia, Kim Holt, Tochukwu Igbo, Jenna Klockenbrink, Kinzie Lee, Rui Li, Bradford Lord, Karla Moreno, Elizabeth Pantino McClune, Pierce Nelson, Verla Neslund, Kara Polen, Nicki Roth, Wendy Ruben, Samantha Sater, Joseph Segovia, Carrie Shapiro-Mendoza, Claire Stinson, Jasmin Taylor, Ruben Torrez, Maria del Carmen Vidal, Lee Warner, and John Zimmerman. We acknowledge the collaborative contributions of (in alphabetical order) the American College of Obstetricians and Gynecologists, the Beyond the Pill Program at the Bixby Center for Global Reproductive Health (University of California, San Francisco School of Medicine), Health Resources and Services Administration, Health Resources and Services Administration Office of Regional Operations, the Puerto Rico Department of Health, Puerto Rico Obstetrics and Gynecology, the Puerto Rico Primary Care Association, The National Campaign to Prevent Teen and Unplanned Pregnancy, and Upstream USA.

ealth Resources and Services Administration, Health Resources and Services Administration Office of Regional Operations, the Puerto Rico Department of Health, Puerto Rico Obstetrics and Gynecology, the Puerto Rico Primary Care Association, The National Campaign to Prevent Teen and Unplanned Pregnancy, and Upstream USA. Contributors EL, LR, SH, LBZ, MTF, MIR, and ENB-B contributed to literature search, study design, data collection, analysis, and interpretation, and manuscript preparation. NB contributed to literature search, study design, and data collection. MAH, JM, and DJJ contributed to literature search, study design, data interpretation, and manuscript preparation. The Z-CAN Working Group members collectively contributed to the literature search, study design, intervention implementation, and data collection. Z-CAN Working Group Lisa Koonin, Pierina Cordero, Ricardo Torres, Brenda Rivera, Claritsa Malave, Alicia Suarez, Yari Vale, Linette Sanchez, Brandon Talley, Laura Angel, Reema Bhakta, Turquoise Sidibe, Zipatly V Mendoza, Rachel Powell, Melissa Bennett, Katherine B. Simmons, Naomi Tepper, Jamie Krashin, Anna Brittain, Euna M August, Kathryn M Curtis, Maura Whiteman, Jackie Rosenthal, Caitlin Green, Charity Ntansah, Anna Fulton, Heather Clayton, Esteban Galarza, Carla Agosto, Luz Marilyn Colón López, Madelyn Rodriguez, Brian D Montalvo Martínez, Jeamy Rodriguez Coss, Martha Cañellas Garcia, Elvin Class Rodriguez, Juan L Cantres, Nilda Moreno Ruiz, Stephanie Rivera, Elizabeth Sotomayor, Ricardo Melendez, Susanna N Visser, Melody Stevens, and Von Nguyen.

layton, Esteban Galarza, Carla Agosto, Luz Marilyn Colón López, Madelyn Rodriguez, Brian D Montalvo Martínez, Jeamy Rodriguez Coss, Martha Cañellas Garcia, Elvin Class Rodriguez, Juan L Cantres, Nilda Moreno Ruiz, Stephanie Rivera, Elizabeth Sotomayor, Ricardo Melendez, Susanna N Visser, Melody Stevens, and Von Nguyen. Declaration of interests We declare no competing interests. The findings and conclusions in this report are those of the authors and do not necessarily represent the official position of the CDC.

Introduction Reducing socioeconomic inequalities in childhood and adolescent obesity is an important public policy goal because of its multiple long-term adverse health consequences.1, 2 A priority is to understand how health inequalities have changed over time and understand whether policy goals of health inequality reduction are being met.3 Although socioeconomic inequalities in childhood overweight have been documented in high-income countries,4 it remains unclear how these inequalities have changed across generations;5 interpretation of existent data is limited by the short timespan of previous investigations (eg, 5–10 years), gaps in timespans investigated, and methodological differences across studies. Research in context Evidence before this study

Introduction Reducing socioeconomic inequalities in childhood and adolescent obesity is an important public policy goal because of its multiple long-term adverse health consequences.1, 2 A priority is to understand how health inequalities have changed over time and understand whether policy goals of health inequality reduction are being met.3 Although socioeconomic inequalities in childhood overweight have been documented in high-income countries,4 it remains unclear how these inequalities have changed across generations;5 interpretation of existent data is limited by the short timespan of previous investigations (eg, 5–10 years), gaps in timespans investigated, and methodological differences across studies. Research in context Evidence before this study We searched PubMed for articles and reviews published between Jan 1, 1960, and Oct 9, 2017, using the search terms “body mass index” OR “obesity” AND “socioeconomic” OR “inequality” OR “disparity”. We screened published articles by title and abstract to identify relevant studies of how socioeconomic inequalities in body-mass index (BMI) or obesity risk had changed across time. The studies cited in this report are not an exhaustive list of existing research. Published systematic reviews have found many studies that document the existence of socioeconomic inequalities in childhood and adolescence BMI or obesity risk in high-income countries. However, evidence for how these inequalities have changed over time is typically short term and cross-sectional in nature, does not examine the composite parts of BMI (ie, weight and height), and does not examine whether inequalities differ in magnitude across the outcome distribution. To inform public policy—and specific concerns regarding the adverse long-term consequences of childhood obesity and its socioeconomic inequality—robust and nationally representative evidence is required to examine how inequalities have changed in response to shifting policy and societal factors.

e outcome distribution. To inform public policy—and specific concerns regarding the adverse long-term consequences of childhood obesity and its socioeconomic inequality—robust and nationally representative evidence is required to examine how inequalities have changed in response to shifting policy and societal factors. Added value of this study We used four British historic longitudinal studies to examine trends in socioeconomic inequalities in BMI from 1953 to 2015. This study provides added value by enabling a long-run investigation of socioeconomic inequalities in BMI, and more recent data than previously available. Most existing evidence is cross-sectional in nature; however, we used longitudinal data and found that absolute socioeconomic inequalities in BMI widened from childhood to adolescence. We also examined the different components of BMI that yielded new policy-relevant evidence; absolute height inequalities have narrowed in subsequent generations whereas weight inequalities have reversed (ie, changed direction). Finally, we examine how inequalities in these outcomes differ across the outcome distribution using quantile regression, in which we observed that socioeconomic inequalities in BMI were found at the median but were systematically larger at higher BMI quantiles than at lower quantiles. Implications of all available evidence

We used four British historic longitudinal studies to examine trends in socioeconomic inequalities in BMI from 1953 to 2015. This study provides added value by enabling a long-run investigation of socioeconomic inequalities in BMI, and more recent data than previously available. Most existing evidence is cross-sectional in nature; however, we used longitudinal data and found that absolute socioeconomic inequalities in BMI widened from childhood to adolescence. We also examined the different components of BMI that yielded new policy-relevant evidence; absolute height inequalities have narrowed in subsequent generations whereas weight inequalities have reversed (ie, changed direction). Finally, we examine how inequalities in these outcomes differ across the outcome distribution using quantile regression, in which we observed that socioeconomic inequalities in BMI were found at the median but were systematically larger at higher BMI quantiles than at lower quantiles. Implications of all available evidence The emergence and widening of socioeconomic inequalities in BMI in children and adolescents up to 2015 suggests a renewed need for effective policies to reduce obesity and its socioeconomic inequality in current and future generations; previous policies have not been adequate, and existing policies are unlikely to be either. Without effective intervention, socioeconomic inequalities in BMI are anticipated to widen further throughout adulthood and disproportionately affect those who have higher BMI, leading to decades of adverse health and economic consequences.

licies have not been adequate, and existing policies are unlikely to be either. Without effective intervention, socioeconomic inequalities in BMI are anticipated to widen further throughout adulthood and disproportionately affect those who have higher BMI, leading to decades of adverse health and economic consequences. Additionally, several important aspects of the nature of socioeconomic inequalities in BMI remain poorly understood; in particular, the extent to which socioeconomic inequalities have changed across the composite parts of BMI (ie, weight and height). Because lower socioeconomic position has been associated with shorter childhood height,6 changes in BMI might be attributable to changes in weight or height, or both. Understanding both components separately is important because of the association between shorter height in childhood and increased premature mortality risk,7 and the association between shorter height in adulthood and increased adult cardiovascular disease risk.8 Socioeconomic inequalities in these constituent parts might have changed in different ways over time. For example, improvements in population micronutrient intake and reductions in early life infections might have occurred, as suggested by secular trends towards taller childhood height from 1957 to 2012,9 and suggestive evidence for reduced prevalence of stunting in Britain.7 These changes might have also led to narrower height inequalities in recent decades, yet increases in total calorie consumption associated with the recent obesity burden might have led to the emergence and widening of weight and thus BMI inequalities from the 1980s onwards. BMI inequalities might also lead to narrower height inequalities, because obesity can increase the pace of pubertal development.10 Additionally, existing evidence for how BMI inequalities have changed over time is typically from repeated cross-sectional studies, limiting the understanding of the ages at which inequalities emerge or widen. Finally, the effect of socioeconomic inequalities on the population distribution of BMI, height, and weight is not well understood. A trend towards an increasing BMI across time has been observed,9, 11 and this increase could be disproportionately attributable to people in disadvantaged socioeconomic groups.12

r widen. Finally, the effect of socioeconomic inequalities on the population distribution of BMI, height, and weight is not well understood. A trend towards an increasing BMI across time has been observed,9, 11 and this increase could be disproportionately attributable to people in disadvantaged socioeconomic groups.12 Therefore in this study, we aimed to examine trends in socioeconomic inequalities in BMI, weight, and height across childhood to adolescence using data from four British birth cohort studies, enabling a long-run comparison from 1953 to 2015. We hypothesised that socioeconomically disadvantaged groups had lower bodyweight and shorter height than socioeconomically advantaged groups born in the mid-to-late 20th century; among those born in the early 21st century, we hypothesised that differences in bodyweight would have reversed and that height differences would be narrower. Methods Study design and samples We used data from four longitudinal, observational, British birth cohort studies. These cohorts have been previously described in detail elsewhere,9, 13, 14 and they were designed to be nationally representative when initiated in 1946 (MRC National Survey of Health and Development [1946 NSHD]), 1958 (National Child Development Study [1958 NCDS]), 1970 (British Cohort Study [1970 BCS]), and 2001 (Millennium Cohort Study [2001 MCS]). We categorised individuals born in the mid-to-late 20th century (ie, the 1946 NSHD, 1958 NCDS, and 1970 BCS) as the earlier-born cohorts, and those born in the early 21st century (ie, the 2001 MCS) as the later-born cohort.

8 NCDS]), 1970 (British Cohort Study [1970 BCS]), and 2001 (Millennium Cohort Study [2001 MCS]). We categorised individuals born in the mid-to-late 20th century (ie, the 1946 NSHD, 1958 NCDS, and 1970 BCS) as the earlier-born cohorts, and those born in the early 21st century (ie, the 2001 MCS) as the later-born cohort. To aid comparability, we restricted the analyses to singleton births in England, Scotland, and Wales in 1946, 1958, 1970, and 2001; however, only singletons were sampled in the 1946 NSHD. Weight, height, and BMI measurements As described elsewhere,9 BMI (kg/m2) was derived and harmonised in each study from measured weight and height. These three outcomes were obtained at the following ages: 7 years, 11 years, and 15 years in the 1946 NSHD; 7 years, 11 years, and 16 years in the 1958 NCDS; 10 years and 16 years in the 1970 BCS; and 7 years, 11 years, and 14 years in the 2001 MCS.

derived and harmonised in each study from measured weight and height. These three outcomes were obtained at the following ages: 7 years, 11 years, and 15 years in the 1946 NSHD; 7 years, 11 years, and 16 years in the 1958 NCDS; 10 years and 16 years in the 1970 BCS; and 7 years, 11 years, and 14 years in the 2001 MCS. Ascertainment of socioeconomic position Childhood socioeconomic position was indicated by the father's (occupational) social class, reported at the ages of 10–11 years. To assist cross-cohort comparability, the Registrar-General's Social Classes was used to classify social class by occupational group: I (professional), II (managerial and technical), IIIN (skilled non-manual), IIIM (skilled manual), IV (partly skilled), and V (unskilled).15 The 1990 classification schema was used for all cohorts except for the 1946 NSHD, for which the 1970 version was used because of the absence of a conversion schema. Additionally, the 1970 classification schema was used if historic source data were not retrievable. Those in the armed forces or who were unemployed were not assigned a social class. We used mother-figure occupational class when no father-figure was present in the household or for which no valid father-figure occupational class data were available in the 2001 MCS, because of recent increases in this type of family composition.

he armed forces or who were unemployed were not assigned a social class. We used mother-figure occupational class when no father-figure was present in the household or for which no valid father-figure occupational class data were available in the 2001 MCS, because of recent increases in this type of family composition. Statistical analysis To account for differences in the exact age of measurement across cohorts, we calculated age-centred BMI, weight, and height values at ages 7 years, 11 years, and 15 years using predictions from cohort-specific linear regression models of age regressed on these outcomes (ie, BMI, weight, and height). We found little evidence for gender differences in associations between social class and these outcomes (gender × socioeconomic position interactions); as such, we conducted gender-pooled and gender-adjusted models, which were consistent with previous analyses that used the 2001 MCS.16, 17, 18 To provide single quantifications of inequalities, we converted social class to ridit scores (ranging from 0 to 1) calculated separately in each cohort. The socioeconomic position coefficient in linear regression—the slope index of inequality—is interpreted as the estimated absolute (mean) difference (absolute inequality) in outcome between the lowest and highest socioeconomic position. This method enables comparisons even when the proportion of participants differs in each socioeconomic position category across cohorts. We also used multilevel models to examine whether absolute inequalities systematically changed by age; age × ridit score interaction terms were included in models with outcome measurements (level 1) nested within individuals (level 2). Additionally, we specified a random intercept and random slope, as well as modelled age as a linear term. In the 1970 BCS, we did not specify random effects because the maximal number of observations was two (participants were aged 10 years and 16 years).

come measurements (level 1) nested within individuals (level 2). Additionally, we specified a random intercept and random slope, as well as modelled age as a linear term. In the 1970 BCS, we did not specify random effects because the maximal number of observations was two (participants were aged 10 years and 16 years). We used conditional quantile regression to examine associations between social class ridit scores and outcomes at specific quantiles of distributions of BMI, weight, and height. We obtained estimates and plotted them at the following quantiles: fifth, tenth, 25th, 50th (median), 75th, 90th, and 95th. Additionally, we used multinomial regression to examine associations between social class ridit scores and International Obesity Task Force BMI thresholds; these thresholds are age-specific cutpoints designed to correspond to adult BMI cutpoints of thinness (<18·5 kg/m2), normal weight (18·5 to <25 kg/m2), overweight (25 to <30 kg/m2), and obesity (≥30 kg/m2).19 We tested the associations between socioeconomic position and this categorical outcome on both the absolute scale (difference in predicted probability of each outcome—ie, risk difference) and relative scale (relative risk ratio, with normal BMI as the reference).

overweight (25 to <30 kg/m2), and obesity (≥30 kg/m2).19 We tested the associations between socioeconomic position and this categorical outcome on both the absolute scale (difference in predicted probability of each outcome—ie, risk difference) and relative scale (relative risk ratio, with normal BMI as the reference). We repeated all analyses using maternal education attainment (ascertained at age 6 years in the 1946 NSHD, at birth in the 1958 NCDS and 1970 BCS, and at 9 months of age in the 2001 MCS) instead of the father's social class, which has previously been shown to be related to anthropometric outcomes in the included cohorts (eg, the 2001 MCS16, 17). These analyses might also provide a means of triangulation for causal inference, since consistency of findings based on both maternal education and paternal social class indicators suggest that findings are not solely explained by confounding factors acting on one parent figure. We used two measures: a binary indicator of whether the mother had left education at the mandatory leaving age (14 years from 1918, 15 years from 1944, and 16 years from 1972), and the age the mother left full-time education (in 10-year age groups from <13 years to ≥23 years, measured at age 16 years in the 1958 NCDS). We modelled both of these measures as ridit scores to aid comparability. To examine whether differences in ethnic composition affected results of cross-cohort comparisons, we repeated linear regression and quantile analyses restricted to white participants only.

to ≥23 years, measured at age 16 years in the 1958 NCDS). We modelled both of these measures as ridit scores to aid comparability. To examine whether differences in ethnic composition affected results of cross-cohort comparisons, we repeated linear regression and quantile analyses restricted to white participants only. We weighted the analyses where appropriate to account for the survey design of the 1946 NSHD and 2001 MCS, whereas analyses using the 1970 BCS and 1958 NCDS were not weighted because no subgroups were over or under sampled. We did all analyses using Stata (version 15.0). Data sharing The harmonised BMI dataset is freely available to download at the UK Data Archive. Additionally, all original datasets from the 1958 NCDS, 1970 BCS, and 2001 MCS are freely available to download at the UK Data Archive. Additional data from the 1946 NSHD are made freely available to researchers who submit data requests to the NSHD Data Archive. Role of the funding source The funders 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 harmonised BMI dataset is freely available to download at the UK Data Archive. Additionally, all original datasets from the 1958 NCDS, 1970 BCS, and 2001 MCS are freely available to download at the UK Data Archive. Additional data from the 1946 NSHD are made freely available to researchers who submit data requests to the NSHD Data Archive. Role of the funding source The funders 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 England, Scotland, and Wales, 5362 singleton births were enrolled in 1946, 17 202 in 1958, 17 290 in 1970, and 16 404 in 2001. All participants in the 1946 NSHD were white, as were 14 407 (98·7%) of 14 603 in the 1958 NCDS (2599 had missing ethnicity data), and 13 671 (95·2%) of 14 354 in the 1970 BCS (2936 had missing ethnicity data). In the 2001 MCS, only 13 208 (80·5%) of 16 404 were white. The table summarises the sample sizes for analyses by age group in each cohort. In 2318 participants in the 1958 NCDS, the 1970 classification schema was used because of irretrievable historic source data (Spearman's correlation coefficient between social class derived by 1970 and 1990 schema r=0·77; p<0·0001). The proportion of participants with no father figure at the age of 10–11 years was low in the early cohorts: unmarried women were not sampled in the 1946 NSHD and less than 3·6% of participants had no father figure at the age of 10–11 years in both the 1958 NCDS and 1970 BCS. No valid father-figure occupational class data were available for 2430 participants in the 2001 MCS; low mother-figure class was associated in expected directions with low maternal education attainment (p<0·0001) and low father's social class (p<0·0001).Table Averages and socioeconomic differences in BMI, weight, and height during childhood to adolescence

lass data were available for 2430 participants in the 2001 MCS; low mother-figure class was associated in expected directions with low maternal education attainment (p<0·0001) and low father's social class (p<0·0001).Table Averages and socioeconomic differences in BMI, weight, and height during childhood to adolescence Year n BMI (kg/m2) Weight (kg) Height (cm) Mean (SD) Median (IQR) SEP difference, SII (95% CI) Mean (SD) Median (IQR) SEP difference, SII (95% CI) Mean (SD) Median (IQR) SEP difference, SII (95% CI) Children aged 7 years 1946 NSHD 1953 3510 15·8 (1·5) 15·6 (14·9–16·6) 0·0 (−0·2 to 0·2) 22·5 (3·1) 22·2 (20·4–24·2) –1·4 (−1·9 to −0·9) 119·3 (5·6) 119·4 (116·4–123·2) –3·9 (−4·6 to −3·1) 1958 NCDS 1965 10 650 15·7 (1·8) 15·5 (14·7–16·5) 0·1 (−0·1 to 0·2) 22·7 (3·7) 22·2 (20·3–24·4) –1·1 (−1·3 to −0·8) 120·8 (5·8) 120·6 (117·3–124·1) –3·0 (−3·4 to −2·6) 2001 MCS 2008 8340 16·4 (2·2) 16·0 (15·0–17·2) 0·5 (0·3 to 0·7) 24·5 (4·6) 23·7 (21·4–26·6) 0·3 (−0·1 to 0·8) 122·6 (5·3) 122·4 (119·0–126·1) –1·2 (−1·7 to −0·8) Children and adolescents aged 11 years 1946 NSHD 1957 3629 17·4 (2·4) 17·0 (15·9–18·4) –0·1 (−0·4 to 0·3) 34·9 (6·5) 33·7 (30·5–37·8) –2·0 (−3·0 to −1·1) 141·0 (6·9) 140·6 (135·9–145·7) –4·1 (−5·1 to −3·2) 1958 NCDS 1969 11 193 17·3 (2·6) 16·7 (15·6–18·3) 0·0 (−0·2 to 0·1) 35·1 (7·3) 33·6 (30·1–38·6) –1·8 (−2·3 to −1·3) 142·3 (7·1) 142·2 (137·6–146·9) –3·5 (−3·9 to −3·0) 1970 BCS 1980 11 231 17·4 (2·1) 17·1 (16·0–18·4) 0·1 (0·0 to 0·3) 35·8 (5·3) 35·0 (32·0–38·7) –1·0 (−1·3 to −0·6) 142·2 (6·4) 142·0 (137·9–146·3) –2·7 (−3·1 to −2·3) 2001 MCS 2012 8820 18·9 (3·4) 18·2 (16·5–20·7) 1·3 (0·9 to 1·6) 40·5 (9·4) 38·9 (33·8–45·4) 2·1 (1·2 to 2·9) 145·7 (6·9) 145·5 (141·2–150·3) –1·2 (−1·7 to −0·6) Children and adolescents aged 15 years 1946 NSHD 1961 3262 20·4 (2·8) 20·0 (18·5–21·7) 0·2 (−0·2 to 0·7) 53·3 (9·2) 52·5 (47·0–58·2) –1·9 (−3·3 to −0·5) 162·2 (8·0) 162·2 (157·1–167·3) –4·0 (−5·2 to −2·9) 1958 NCDS 1973 8824 20·2 (2·9) 19·7 (18·3 −21·5) 0·4 (0·1 to 0·6) 53·7 (9·7) 52·5 (47·3–58·8) –1·3 (−2·1 to −0·6) 161·7 (8·5) 161·3 (155·6–167·2) –3·3 (−3·9 to −2·8) 1970 BCS 1986 6649 20·2 (3·1) 19·7 (18·2–21·6) 0·6 (0·3 to 0·9) 53·5 (10·1) 52·3 (46·8–58·9) –0·5 (−1·3 to 0·4) 161·4 (9·4) 161·3 (154·7–167·7) –3·0 (−3·6 to −2·3) 2001 MCS 2015 7393 21·7 (3·9) 20·9 (19·0–23·5) 1·4 (1·0 to 1·8) 60·9 (12·1) 59·2 (52·8–70·0) 2·4 (1·2 to 3·6) 168·9 (7·9) 168·5 (163·3–174·0) –1·7 (−2·4 to −1·0) Means and SII are gender-adjusted, and outcomes were age-centred at 7 years, 11 years

) –0·5 (−1·3 to 0·4) 161·4 (9·4) 161·3 (154·7–167·7) –3·0 (−3·6 to −2·3) 2001 MCS 2015 7393 21·7 (3·9) 20·9 (19·0–23·5) 1·4 (1·0 to 1·8) 60·9 (12·1) 59·2 (52·8–70·0) 2·4 (1·2 to 3·6) 168·9 (7·9) 168·5 (163·3–174·0) –1·7 (−2·4 to −1·0) Means and SII are gender-adjusted, and outcomes were age-centred at 7 years, 11 years , and 15 years. BMI=body-mass index. NSHD=Medical Research Council National Survey of Health and Development. NCDS=National Child Development Study. BCS=British Cohort Study. MCS=Millennium Cohort Study. SEP=socioeconomic position (or social class characterised by father's occupation). SII=slope index of inequality.

) –0·5 (−1·3 to 0·4) 161·4 (9·4) 161·3 (154·7–167·7) –3·0 (−3·6 to −2·3) 2001 MCS 2015 7393 21·7 (3·9) 20·9 (19·0–23·5) 1·4 (1·0 to 1·8) 60·9 (12·1) 59·2 (52·8–70·0) 2·4 (1·2 to 3·6) 168·9 (7·9) 168·5 (163·3–174·0) –1·7 (−2·4 to −1·0) Means and SII are gender-adjusted, and outcomes were age-centred at 7 years, 11 years , and 15 years. BMI=body-mass index. NSHD=Medical Research Council National Survey of Health and Development. NCDS=National Child Development Study. BCS=British Cohort Study. MCS=Millennium Cohort Study. SEP=socioeconomic position (or social class characterised by father's occupation). SII=slope index of inequality. Some descriptive trends were observed when comparing the 2001 MCS with the 1946 NSHD, 1958 NCDS, and 1970 BCS: BMI, weight, and height values were higher in the 2001 MCS than in the earlier-born cohorts (table; appendix pp 1, 2). The composition of socioeconomic position also differed across the cohorts: proportions in the managerial to technical occupational groups were highest in the 2001 MCS, and skilled manual occupations were lowest in this cohort (appendix p 3). Post compulsory education attendance was higher in the 2001 MCS cohort than in the earlier-born cohorts. Missing socioeconomic position and missing BMI data were more frequent in the 2001 MCS than in the earlier-born cohorts, whereas missing maternal education was more frequent in the 1946 NSHD than in the 2001 MCS (appendix p 3). For example, 1425 (26·6%) of the 5362 sampled at birth in the 1946 NSHD did not have their BMI measured at age 11 years compared with 5408 (33·0%) of the 16 404 in the 2001 MCS. Reasons for these missing measurements included death, emigration, and loss to follow-up. Missing data at 14–16 years were in most instances more frequent in participants of low socioeconomic position at 10–11 years and with high preceding BMI (appendix p 4).

rs compared with 5408 (33·0%) of the 16 404 in the 2001 MCS. Reasons for these missing measurements included death, emigration, and loss to follow-up. Missing data at 14–16 years were in most instances more frequent in participants of low socioeconomic position at 10–11 years and with high preceding BMI (appendix p 4). In the earlier-born cohorts, lower socioeconomic position at all ages was associated with lower weight, whereas only in the 2001 MCS was lower social class associated with higher weight (table). These differences did not systematically differ by age in the 1946 NSHD, 1958 NCDS, or 1970 BCS cohorts. However, in the 2001 MCS cohort, weight disparities became larger from childhood to adolescence (pinteraction=0·001 for social class × age; appendix p 5). In the 2001 MCS, inequalities in weight were present at the median, and became increasingly larger at higher quantiles (figure 1B). For example when comparing lowest with highest social class at age 11 years, there was a difference of 1·40 kg (95% CI 0·44–2·35) at the 50th weight percentile whereas a difference of 4·88 kg (2·66–7·10) was observed at the 90th weight percentile. Inequalities in weight were comparatively similar across quantiles in the earlier-born cohorts. These findings were similar at age 15 years (figure 2B).Figure 1 Estimated differences in BMI (A), weight (B), and height (C) in children aged 11 years in the lowest social class compared with the highest social class* (slope index of inequality)

were comparatively similar across quantiles in the earlier-born cohorts. These findings were similar at age 15 years (figure 2B).Figure 1 Estimated differences in BMI (A), weight (B), and height (C) in children aged 11 years in the lowest social class compared with the highest social class* (slope index of inequality) Data are quantile regression estimates at different quantiles of the outcome distribution. Error bars are 95% CI. Coefficients are interpreted analogously to linear regression—eg, Q50 shows the median difference in BMI comparing the lowest with highest social class. BMI=body-mass index. NSHD=MRC National Survey of Health and Development. NCDS=National Child Development Study. BCS=British Cohort Study. MCS=Millennium Cohort Study. *Social class characterised by father's occupation. Figure 2 Estimated differences in BMI (A), weight (B), and height (C) in children aged 15 years in the lowest social class compared with the highest social class* (slope index of inequality) Data are quantile regression estimates at different quantiles of the outcome distribution. Error bars are 95% CI. Coefficients are interpreted analogously to linear regression—eg, Q50 shows the median difference in BMI comparing the lowest with highest social class. BMI=body-mass index. NSHD=MRC National Survey of Health and Development. NCDS=National Child Development Study. BCS=British Cohort Study. MCS=Millennium Cohort Study. *Social class characterised by father's occupation.

regression—eg, Q50 shows the median difference in BMI comparing the lowest with highest social class. BMI=body-mass index. NSHD=MRC National Survey of Health and Development. NCDS=National Child Development Study. BCS=British Cohort Study. MCS=Millennium Cohort Study. *Social class characterised by father's occupation. In all cohorts, lower socioeconomic position was associated with shorter height, yet the absolute magnitude of this difference narrowed in each subsequent cohort (table). These associations became more negative (ie, height disparities widened in absolute terms) with age in the 2001 MCS (pinteraction=0·002) but not in the earlier-born cohorts (pinteraction=1·00 in the 1946 NSHD, pinteraction=0·29 in the 1958 NCDS, and pinteraction=0·51 in the 1970 BCS for social class × age; appendix p 5). In all cohorts, inequalities in height did not appear to systematically differ across the quantiles for those aged 11 years or 15 years (Figure 1, Figure 2).

In all cohorts, lower socioeconomic position was associated with shorter height, yet the absolute magnitude of this difference narrowed in each subsequent cohort (table). These associations became more negative (ie, height disparities widened in absolute terms) with age in the 2001 MCS (pinteraction=0·002) but not in the earlier-born cohorts (pinteraction=1·00 in the 1946 NSHD, pinteraction=0·29 in the 1958 NCDS, and pinteraction=0·51 in the 1970 BCS for social class × age; appendix p 5). In all cohorts, inequalities in height did not appear to systematically differ across the quantiles for those aged 11 years or 15 years (Figure 1, Figure 2). There was little evidence for socioeconomic inequality in mean BMI at age 7 years or 11 years in the 1946 NSHD, 1958 NCDS, or 1970 BCS; however, inequalities were present in the 2001 MCS at ages 7 years and 11 years (table). Mean BMI differences by socioeconomic position were present in all cohorts except the 1946 NSHD at age 15 years, and this difference remained substantially larger in the 2001 MCS. Inequalities generally widened with age from 7 years or 10 years to 15 years (social class × age interaction terms were positive in all cohorts: pinteraction=0·159 in the 1946 NSHD, pinteraction=0·004 in the 1958 NCDS, pinteraction=0·001 in the 1970 BCS, and pinteraction<0·001 in the 2001 MCS; figure 3; appendix p 5).Figure 3 BMI across childhood to adolescence by social class* in four British birth cohort studies

interaction terms were positive in all cohorts: pinteraction=0·159 in the 1946 NSHD, pinteraction=0·004 in the 1958 NCDS, pinteraction=0·001 in the 1970 BCS, and pinteraction<0·001 in the 2001 MCS; figure 3; appendix p 5).Figure 3 BMI across childhood to adolescence by social class* in four British birth cohort studies Lines are estimated BMI and widths of the shaded area are 95% CIs at each age among women, estimated with multilevel general linear regression models (the appendix shows the full model estimates). BMI=body-mass index. NSHD=MRC National Survey of Health and Development. NCDS=National Child Development Study. BCS=British Cohort Study. MCS=Millennium Cohort Study. *Social class characterised by father's occupation.

mated with multilevel general linear regression models (the appendix shows the full model estimates). BMI=body-mass index. NSHD=MRC National Survey of Health and Development. NCDS=National Child Development Study. BCS=British Cohort Study. MCS=Millennium Cohort Study. *Social class characterised by father's occupation. Quantile regression analyses suggested that in the 2001 MCS at age 11 years, inequalities in BMI were present at the median and increasingly became larger at higher quantiles (eg, when comparing lowest with highest social class, a difference of 0·98 kg/m2 [95% CI 0·63–1·33] at the 50th BMI percentile and a difference of 2·54 kg/m2 [1·85–3·22] at the 90th percentile; figure 1A). Such patterns were also found in all cohorts at age 15 years (figure 2A), but the magnitude of inequalities was larger in the 2001 MCS than in the earlier-born cohorts. These findings were consistent with visual inspection of BMI distributions in different socio-economic groups, suggesting more skewness in BMI distributions among lower socioeconomic groups, and results from multinomial regression analyses, which suggested that lower social class was associated with increased absolute risk of overweight or obesity at age 15 years in cohorts born in 1958–2001, but no increased absolute risk of thinness (appendix p 6). When examined in the absolute scale (differences in predicted probabilities of each BMI category), inequalities in overweight or obesity were higher in the 2001 MCS than in the earlier-born cohorts, which was consistent with the main findings (appendix p 6). These patterns of results were similar when the UK 1990 growth reference thresholds were used (data not shown).

n predicted probabilities of each BMI category), inequalities in overweight or obesity were higher in the 2001 MCS than in the earlier-born cohorts, which was consistent with the main findings (appendix p 6). These patterns of results were similar when the UK 1990 growth reference thresholds were used (data not shown). Results were similar when maternal education attainment was used as an alternative indicator of socioeconomic position (appendix pp 1, 2, 7–9). Findings were similar when restricted to only white participants (eg, quantile regression estimates for those in the 2001 MCS aged 11 years shown in the appendix [p 10]). Discussion In four national British birth cohorts with data spanning from 1953 to 2015, socioeconomic inequalities in weight reversed: lower socioeconomic position was associated with lower weights in the 1946, 1958, and 1970 cohorts but in the 2001 cohort it was associated with higher weight; whereas lower socioeconomic position was associated with shorter height in all cohorts but the absolute magnitude of this difference narrowed in each subsequent cohort. The magnitude of absolute inequalities in BMI differed in each cohort as a result, and was larger and apparent earlier in childhood in the 2001 MCS than in the earlier-born cohorts (ie, the 1946 NSHD, 1958 NCDS, and 1970 BCS). BMI differences widened from childhood to adolescence in all cohorts except the 1946 NSHD. These findings were consistent when using both father's social class and maternal education as indicators of socioeconomic position.

d in the 2001 MCS than in the earlier-born cohorts (ie, the 1946 NSHD, 1958 NCDS, and 1970 BCS). BMI differences widened from childhood to adolescence in all cohorts except the 1946 NSHD. These findings were consistent when using both father's social class and maternal education as indicators of socioeconomic position. Our findings are consistent with existing cross-sectional evidence from the UK, suggesting that relative inequalities in obesity or BMI have increased in recent decades. This trend was observed in the analysis of obesity inequalities from 1997 to 2005 among children aged 5–10 years in the Health Survey for England,20 in BMI inequalities among those aged 4–5 years and 10–11 years in 2007–08 and 2011–12 in the National Child Measurement Programme,12 and in obesity inequalities among those aged 10–11 years in the 1970 BCS and 2001 MCS (1980 compared with 2011).21 Here, we show that BMI inequalities have persisted to 2015, and that in multiple generations absolute and relative BMI inequalities widened with age from childhood to adolescence. We also show, as suggested in recent cross-sectional data,12 that BMI inequalities were larger at the higher end of the BMI distribution. These findings are consistent with, and might partially explain, the observed positive skew of the population BMI distribution in later-born cohorts.9, 11

m childhood to adolescence. We also show, as suggested in recent cross-sectional data,12 that BMI inequalities were larger at the higher end of the BMI distribution. These findings are consistent with, and might partially explain, the observed positive skew of the population BMI distribution in later-born cohorts.9, 11 Our finding that height inequalities have narrowed is consistent with suggestive evidence from a study comparing height inequalities among the 1958 NCDS and their children.22 Narrowing of height differences has also been reported in other countries that have had substantial economic and nutritional changes.23

m childhood to adolescence. We also show, as suggested in recent cross-sectional data,12 that BMI inequalities were larger at the higher end of the BMI distribution. These findings are consistent with, and might partially explain, the observed positive skew of the population BMI distribution in later-born cohorts.9, 11 Our finding that height inequalities have narrowed is consistent with suggestive evidence from a study comparing height inequalities among the 1958 NCDS and their children.22 Narrowing of height differences has also been reported in other countries that have had substantial economic and nutritional changes.23 Considerable changes took place in the period investigated (1953–2015) in Britain, including changes to several factors that might have ultimately influenced diet and physical activity, which are the plausible yet equivocal mediators of BMI and height inequalities.24, 25, 26 Diets in both the prenatal and postnatal periods are likely to contribute to BMI and height inequalities,27 and British diets have changed considerably. World War 2-related food rationing continued up to 1954 in the UK; compared with population diet in the 1990s, rationing-based diets were characterised by higher consumption of vegetables, lower consumption of sugar and soft drinks, and higher consumption of fat as a proportion of energy intake.28 Despite rationing, socioeconomic inequalities in diet were documented at age 4 years in the 1946 NSHD, in which children of lower socioeconomic groups consumed fewer total calories as well as fruit and vegetables, and thus fewer micronutrients such as zinc and potassium than those of higher socioeconomic position.29 These differences in diet might underlie the association between low socioeconomic position and both lower weight and shorter height in the 1946 NSHD.30 From 1953 to 2015, inequalities in micronutrient intake might have reduced leading to narrower height inequalities;31 inequalities in total calorie consumptions are likely to have reversed over time leading to those of lower socioeconomic position having higher weight, BMI, and obesity risk.

height in the 1946 NSHD.30 From 1953 to 2015, inequalities in micronutrient intake might have reduced leading to narrower height inequalities;31 inequalities in total calorie consumptions are likely to have reversed over time leading to those of lower socioeconomic position having higher weight, BMI, and obesity risk. Income and wealth inequality have increased since the 1970s,32 and some evidence suggests that the price of healthy food items has increased in recent decades.33 These changes might also have contributed to the emergence and widening of BMI inequalities. Among later-born cohorts, a study found inequalities in childhood exercise participation and sedentary behaviour,18 but such inequalities might be weaker or not present in earlier-born cohorts that predate the routine collection of such data. Increases in BMI among adults from the 1980s onwards,9 combined with the persistence of adult BMI inequalities,14 might have indirectly contributed to increases in BMI and weight inequalities among children, since in all cohorts higher parental BMI was associated with higher offspring BMI.34

ne collection of such data. Increases in BMI among adults from the 1980s onwards,9 combined with the persistence of adult BMI inequalities,14 might have indirectly contributed to increases in BMI and weight inequalities among children, since in all cohorts higher parental BMI was associated with higher offspring BMI.34 Additionally, we observed that inequalities in BMI were larger at the higher end of the distribution than at the midpoint (median) or lower end of the distribution. These findings could be explained by unmeasured modifiers that acted to increase the magnitude of BMI inequality. For example, individuals who were more susceptible to higher BMI (for environmental or genetic reasons, or both) might have been more susceptible to the adverse effects of socioeconomic disadvantage,35 leading to inequalities in BMI being larger at the higher end of the BMI distribution, as observed in our quantile regression results.

ividuals who were more susceptible to higher BMI (for environmental or genetic reasons, or both) might have been more susceptible to the adverse effects of socioeconomic disadvantage,35 leading to inequalities in BMI being larger at the higher end of the BMI distribution, as observed in our quantile regression results. Our study had several strengths that included the use of four national birth cohort studies, enabling investigation of long-term trends in BMI, weight, and height. Inferences regarding cross-cohort comparisons were strengthened by the use of harmonised socioeconomic and anthropometric data, comparable sample restrictions, and analyses that accounted for the differing sampling design in each cohort. However, although the study samples used were generally large, they were underpowered to evaluate socioeconomic inequalities in thinness,16 or differences across subgroups of race or ethnicity, in whom both past and future trends in inequalities might differ. Additionally, the 30-year gap from 1970 to 2001, in which no national birth cohorts were conducted, prevents the investigation of such cohorts born in this period.

economic inequalities in thinness,16 or differences across subgroups of race or ethnicity, in whom both past and future trends in inequalities might differ. Additionally, the 30-year gap from 1970 to 2001, in which no national birth cohorts were conducted, prevents the investigation of such cohorts born in this period. Despite the use of national data, our data potentially provide inexact approximations of existing health inequalities in Britain. Inequalities in fat might have been underestimated because BMI consists of, but does not distinguish, fat and lean mass; and because lower socioeconomic position has been associated with higher fat yet not associated with lean mass in children.36 Our analyses were designed to maximise cross-cohort comparability, including the use of harmonised socioeconomic position data for the father's social class and maternal education. Our findings were consistent across both of these indicators. This consistent finding was encouraging because each indicator had differing strengths and weaknesses. Compared with maternal education, social class had more missing data but was measured at a similar age and arguably contained more information across the socioeconomic distribution in each cohort. Consistency of findings also suggests that the results are not solely explained by confounding due to factors affecting one particular parent (eg, maternal health); however, the potential of confounding due to other shared factors cannot be ruled out. Although the same social class categories were used in all cohorts, the use of the 1970 schema to derive this categorisation in the 1946 NSHD but not the other cohorts could theoretically affect comparisons. However, we expect that this difference was unlikely to have a major effect on findings since the 1970 and 1990 versions were strongly correlated. The use of these indicators is potentially at the expense of obtaining the most informative estimates of inequality available in individual cohorts (eg, detailed parental education and household income data in the 2001 MCS). The use of slope indices of inequality aided comparisons of inequality by accounting for differences in the proportions of participants in the socioeconomic position categories in each cohort. However, as with all studies investigating trends in socioeconomic inequalities, changes in the selection into different socioeconomic groups might differ over time.

aided comparisons of inequality by accounting for differences in the proportions of participants in the socioeconomic position categories in each cohort. However, as with all studies investigating trends in socioeconomic inequalities, changes in the selection into different socioeconomic groups might differ over time. In this scenario, even when the statistical estimates of inequality are comparable (as in the slope index of inequality), interpretation might not be.

aided comparisons of inequality by accounting for differences in the proportions of participants in the socioeconomic position categories in each cohort. However, as with all studies investigating trends in socioeconomic inequalities, changes in the selection into different socioeconomic groups might differ over time. In this scenario, even when the statistical estimates of inequality are comparable (as in the slope index of inequality), interpretation might not be. Missing data, which might be due to death, emigration from Britain, dropout, and refusal to participate, might have also introduced bias into the inequality estimates. Attrition in longitudinal studies is generally greatest in those of lower socioeconomic position and higher BMI, and greater attrition of this type has been shown to lead to a reduction in the magnitude of observed health inequalities.37 Since this pattern of missing data was also found in the 2001 MCS but not the 1946 NSHD, we might have underestimated the increase in BMI inequalities over time, although accounting for missing data in the 2001 MCS has been reported to not substantially alter findings.17 Because of a teachers strike in 1986, missing anthropometric data were particularly substantial at age 16 years in the 1970 BCS, which might have primarily affected statistical power rather than biasing estimates since the cause of missing data was possibly unrelated to participant characteristics. Although our findings were similar when using multilevel models that enable those with incomplete information to be included in analyses (under the assumption of missing at random), as in all observational studies, we cannot rule out the possibility that missing data are non-ignorable and therefore might upwardly or downwardly bias inequality estimates.

when using multilevel models that enable those with incomplete information to be included in analyses (under the assumption of missing at random), as in all observational studies, we cannot rule out the possibility that missing data are non-ignorable and therefore might upwardly or downwardly bias inequality estimates. Our results suggest that the total effect of previous policies has been insufficient in preventing the emergence and widening of BMI inequalities in childhood and adolescence from 1953 to 2015. In Britain, numerous policy initiatives have been created to tackle obesity since 1991, which have differed in their ambition, funding, implementation, and suitability for evaluation.38 Our results show that powerful influence of the obesogenic environment has disproportionately affected socioeconomically disadvantaged children from 1953 to 2015. Our results reinforce the need for new approaches, particularly given absolute increases in BMI inequality with age. Without effective intervention, these inequalities are anticipated to widen further throughout adulthood in the 2001 MCS and future cohorts,14 with considerable public health and economic implications.39, 40

results reinforce the need for new approaches, particularly given absolute increases in BMI inequality with age. Without effective intervention, these inequalities are anticipated to widen further throughout adulthood in the 2001 MCS and future cohorts,14 with considerable public health and economic implications.39, 40 Globally, the need to reduce childhood obesity prevalence and its socioeconomic inequality has been repeatedly noted, yet policy responses are often assessed as ineffectual or inappropriately focused on individual or family agency rather than upstream societal factors.41, 42 Current policies in the UK, include the so-called Change4Life, a social marketing campaign aimed at families and individuals; and a forthcoming tax on soft drinks (the Soft Drinks Industry Levy) that notably excludes other sugary drinks and food items. Although the empirical evidence for what could reduce population-level obesity and its inequality is scarce,39 committed cross-government action is required on legislative changes rather than voluntary suggestions, which might help reverse the obesogenic environment—eg, further legislative incentives to food manufacturers to reduce sugar and fat content in food and drinks, as well as the advertising of such foods to children and parents, while incentivising the sale of healthier alternatives.

es rather than voluntary suggestions, which might help reverse the obesogenic environment—eg, further legislative incentives to food manufacturers to reduce sugar and fat content in food and drinks, as well as the advertising of such foods to children and parents, while incentivising the sale of healthier alternatives. Finally, our results of quantile regression analyses have potential policy implications. Because socioeconomic inequalities appear to disproportionately affect those of higher BMI, an additional effective means of reducing socioeconomic inequalities in BMI might be to target those of particularly high BMI. Alternatively, assuming a causal link, reducing socioeconomic inequalities in society might benefit population health by dis-proportionately lowering BMI in those with particularly high BMI values. By contrast, socioeconomic inequalities in height were similar across the distribution of height. The fact that socioeconomic inequalities in childhood and adolescent height have persisted up to 2015 suggests that new policies are required to reduce them. The narrowing of absolute height inequalities that we observed might have been favourable to public health. However, increases in BMI inequality are likely to have a greater adverse effect, because absolute risk of cardiovascular disease attributable to height is small compared with that to BMI, and taller stature might not always benefit health (eg, it is associated with increased risk of some types of cancers43).

c health. However, increases in BMI inequality are likely to have a greater adverse effect, because absolute risk of cardiovascular disease attributable to height is small compared with that to BMI, and taller stature might not always benefit health (eg, it is associated with increased risk of some types of cancers43). In conclusion, between the late 20th and early 21st centuries, socioeconomic inequalities in weight reversed (ie, changed direction) and those in height narrowed, whereas inequalities in BMI and obesity emerged and widened. These substantial changes highlight the powerful impact of societal changes on child and adolescent growth and the insufficiency of previous policies in preventing obesity and its socioeconomic inequality. New and effective policies are required to reduce BMI inequalities in current and future children and adolescents. Without effective interventions, these inequalities are anticipated to widen further throughout adulthood. For the UK Data Archive see http://www.data-archive.ac.uk For the NSHD Data Archive see http://www.nshd.mrc.ac.uk/data.aspx Supplementary Material Supplementary appendix

In conclusion, between the late 20th and early 21st centuries, socioeconomic inequalities in weight reversed (ie, changed direction) and those in height narrowed, whereas inequalities in BMI and obesity emerged and widened. These substantial changes highlight the powerful impact of societal changes on child and adolescent growth and the insufficiency of previous policies in preventing obesity and its socioeconomic inequality. New and effective policies are required to reduce BMI inequalities in current and future children and adolescents. Without effective interventions, these inequalities are anticipated to widen further throughout adulthood. For the UK Data Archive see http://www.data-archive.ac.uk For the NSHD Data Archive see http://www.nshd.mrc.ac.uk/data.aspx Supplementary Material Supplementary appendix Acknowledgments This project is part of a collaborative research programme entitled Cohorts and Longitudinal Studies Enhancement Resources (CLOSER). This programme is funded by the UK Economic and Social Research Council (ES/K000357/1). The UK Medical Research Council (MRC) provides core funding for the MRC National Survey of Health and Development as well as funding to RH and DK (MC_UU_12019/1 and MC_UU_12019/2). RH and DK work at the MRC Unit for Lifelong Health and Ageing at UCL but have no involvement in funding decisions for the MRC. The UK Economic and Social Research Council provides core funding for the National Child Development Study and the British Cohort Study (ES/M001660/1). WJ acknowledges support from the National Institute for Health Research Leicester Biomedical Research Centre, which is a partnership between University Hospitals of Leicester NHS Trust, Loughborough University, and the University of Leicester. DB is supported by the UK Economic and Social Research Council (ES/M008584/1 and ES/M001660/1) and the Academy of Medical Sciences/the Wellcome Trust Springboard Health of the public in 2040 Award (HOP001\1025). WJ is supported by an MRC New Investigator Research Grant (MR/P023347/1). We thank Brian Dodgeon and Sam Parsons (UCL Institute of Education, London, UK) for preparing the harmonised social class and maternal education data, and Vanessa Moulton (UCL Institute of Education, London, UK) for commenting on an earlier draft of this manuscript.

n MRC New Investigator Research Grant (MR/P023347/1). We thank Brian Dodgeon and Sam Parsons (UCL Institute of Education, London, UK) for preparing the harmonised social class and maternal education data, and Vanessa Moulton (UCL Institute of Education, London, UK) for commenting on an earlier draft of this manuscript. Contributors DB wrote the first draft, analysed the data, and is guarantor of the analysis. DB, WJ, LL, and RH designed the analyses. All authors edited and revised the paper, contributed to the interpretation of data, approved the final version, and agreed to be accountable for all aspects of this work. Declaration of Interests We declare no competing interests.

Introduction Differences between ethnic groups have been reported internationally for a range of health outcomes.1 Although genetic differences might make a minor contribution, much of this variation is associated with social determinants creating inequalities.2 Contributors to such differences include socioeconomic status, health-related behaviours, culture, racism, and health care. When such differences are deemed avoidable and unfair (eg, if they arise through health promotion activities), they can be considered health inequities, which should be addressed. Universal health coverage aims to provide health care to all, of sufficient quality to be effective, without imposing financial hardship.3 However, universal health coverage cannot fully address inequities in health care.4 Practices that could result in inequitable health care, though not quantified, include service provision in the dominant language only and dietary advice that is insufficiently adapted to diverse cultures. Culturally competent health care seeks to address such practices. Barriers to accessing effective health care can exacerbate ethnic inequalities in health from other causes, with the delivery of culturally appropriate and accessible health care potentially narrowing inequalities.5 Similarly, health gains from public health interventions might differ by ethnicity.6 Research in context Evidence before this study

Universal health coverage aims to provide health care to all, of sufficient quality to be effective, without imposing financial hardship.3 However, universal health coverage cannot fully address inequities in health care.4 Practices that could result in inequitable health care, though not quantified, include service provision in the dominant language only and dietary advice that is insufficiently adapted to diverse cultures. Culturally competent health care seeks to address such practices. Barriers to accessing effective health care can exacerbate ethnic inequalities in health from other causes, with the delivery of culturally appropriate and accessible health care potentially narrowing inequalities.5 Similarly, health gains from public health interventions might differ by ethnicity.6 Research in context Evidence before this study We searched MEDLINE and Embase for studies published in English from database inception to Feb 16, 2015, using keywords including “ethnicity”, “race”, “migrants”, “amenable mortality”, “avoidable deaths”, “ambulatory care sensitive conditions”, “avoidable hospitalisations”, and “unplanned readmissions”. We updated this search on Oct 12, 2017. Previous studies were mostly in the USA and New Zealand, with little investigation in European settings. Most studies focused on mortality-based outcomes, finding increased risks of amenable mortality among ethnic minorities. Similarly, studies of avoidable hospital admissions found elevated risks among ethnic or racial minorities. We found no studies that included both primary and secondary care.

on in European settings. Most studies focused on mortality-based outcomes, finding increased risks of amenable mortality among ethnic minorities. Similarly, studies of avoidable hospital admissions found elevated risks among ethnic or racial minorities. We found no studies that included both primary and secondary care. Added value of this study Our study investigated ethnic inequalities across a range of dimensions of health-system performance. South Asian groups had relatively high incidence of avoidable hospital admissions compared with the white Scottish ethnic group, suggesting barriers to high quality primary care. Other indicators reflecting quality of hospital care showed less variation, with only slightly increased unplanned readmissions among Pakistani men and women, and no differences in length of stay. By contrast with the existing scientific literature on amenable and preventable mortality, we found little evidence of adverse outcomes in ethnic minorities in Scotland. However, the white Scottish ethnic group had higher preventable and amenable mortality than several ethnic minority groups, with their higher burden partially accounted for by socioeconomic status. Implications of all the available evidence

Our study investigated ethnic inequalities across a range of dimensions of health-system performance. South Asian groups had relatively high incidence of avoidable hospital admissions compared with the white Scottish ethnic group, suggesting barriers to high quality primary care. Other indicators reflecting quality of hospital care showed less variation, with only slightly increased unplanned readmissions among Pakistani men and women, and no differences in length of stay. By contrast with the existing scientific literature on amenable and preventable mortality, we found little evidence of adverse outcomes in ethnic minorities in Scotland. However, the white Scottish ethnic group had higher preventable and amenable mortality than several ethnic minority groups, with their higher burden partially accounted for by socioeconomic status. Implications of all the available evidence As reflected in amenable and preventable mortality, there is little evidence of inequitable care for ethnic minorities in NHS Scotland, unlike the few high-income countries that have investigated the topic. For some ethnic groups, especially south Asians, there was evidence that avoidable hospital admissions were relatively high. Our study shows that assessing equity of health system performance across several dimensions simultaneously is feasible using national linked administrative data and is necessary to gain a rounded understanding. We identified some priority areas for improvement—in particular, a greater focus on reducing the high burden of preventable deaths among the white Scottish ethnic group and improved quality of primary care for the Pakistani ethnic group. Future international research on addressing health inequalities by ethnicity, and other aspects of social stratification, should consider studying a range of indicators, rather than relying on mortality outcomes alone. Future research should seek explanations for such patterns, including comorbidities and disease risk factors.

nternational research on addressing health inequalities by ethnicity, and other aspects of social stratification, should consider studying a range of indicators, rather than relying on mortality outcomes alone. Future research should seek explanations for such patterns, including comorbidities and disease risk factors. Measuring the performance of health systems is challenging, with a range of indicators used (panel; appendix).7 Amenable mortality assesses the quality of a health system by identifying conditions that theoretically should not result in death if timely and effective health care is provided (eg, appendicitis, bacterial meningitis, and ischaemic heart disease).8, 9 Preventable mortality defines causes of death that can be prevented by effective public health policies (eg, lung cancer and ischaemic heart disease). Deaths can be both amenable and preventable because these categories are not mutually exclusive (appendix). Avoidable mortality combines amenable and preventable deaths.Panel Indicators of health-system performance used Mortality indicators Mortality-based outcomes classify deaths on the basis of their causes into: • Amenable deaths, which should not occur if effective health care is received in a timely manner • Preventable deaths, which should not occur if effective public health measures are implemented

Measuring the performance of health systems is challenging, with a range of indicators used (panel; appendix).7 Amenable mortality assesses the quality of a health system by identifying conditions that theoretically should not result in death if timely and effective health care is provided (eg, appendicitis, bacterial meningitis, and ischaemic heart disease).8, 9 Preventable mortality defines causes of death that can be prevented by effective public health policies (eg, lung cancer and ischaemic heart disease). Deaths can be both amenable and preventable because these categories are not mutually exclusive (appendix). Avoidable mortality combines amenable and preventable deaths.Panel Indicators of health-system performance used Mortality indicators Mortality-based outcomes classify deaths on the basis of their causes into: • Amenable deaths, which should not occur if effective health care is received in a timely manner • Preventable deaths, which should not occur if effective public health measures are implemented • Avoidable deaths, which could be avoided through full implementation of effective health care and effective public health measures; because some causes of death can be coded under both amenable deaths and preventable deaths, the total number of avoidable deaths is less than the sum of the first two categories Primary care indicators

• Avoidable deaths, which could be avoided through full implementation of effective health care and effective public health measures; because some causes of death can be coded under both amenable deaths and preventable deaths, the total number of avoidable deaths is less than the sum of the first two categories Primary care indicators Hospital admissions that could have been avoided if effective primary care were delivered are referred to as avoidable hospital admissions or, alternatively, ambulatory care-sensitive conditions. These can be subdivided into acute avoidable hospital admissions (related to acute illnesses) and chronic avoidable hospital admissions (related to chronic conditions). Secondary care indicators The quality of secondary care was assessed primarily by investigating unplanned readmissions, with length of stay also assessed to assess reasons for potential variation in unplanned readmissions. • Unplanned readmissions: a readmission within 30 days of discharge suggests poor quality hospital care because it could reflect poorly planned discharge processes • Length of stay: to investigate whether the reason for differences in unplanned readmissions was early discharges among specific ethnic groups, length of stay for selected health conditions was checked to investigate potential early discharge from hospital

• Unplanned readmissions: a readmission within 30 days of discharge suggests poor quality hospital care because it could reflect poorly planned discharge processes • Length of stay: to investigate whether the reason for differences in unplanned readmissions was early discharges among specific ethnic groups, length of stay for selected health conditions was checked to investigate potential early discharge from hospital The adequacy of primary care has been investigated by identifying causes of hospital admissions that should be prevented by effective and timely primary care (eg, asthma and compulsory psychiatric admission, both conditions with high rates of hospital admission without clear evidence of extra morbidity in the community).10 Secondary care quality has been assessed by studying unplanned readmissions and the length of hospital stay.11 By convention, readmission within 30 days is the preferred indicator. No standard lengths of stay exist but they can be compared by subgroup. Importantly, comparative assessments across different dimensions of health-system performance remain rare. To minimise deaths, hospital admissions, and readmissions, actions might be required within the health-care system by public health services or policy, or by broader services and policies outside the health system.

mportantly, comparative assessments across different dimensions of health-system performance remain rare. To minimise deaths, hospital admissions, and readmissions, actions might be required within the health-care system by public health services or policy, or by broader services and policies outside the health system. Health policy is a devolved responsibility in the four nations comprising the UK.12 The UK's National Health Service (NHS) places considerable importance on providing equitable care to all groups, and NHS Scotland has a responsibility to provide an equitable service to all ethnic groups under the Equality Act 2010. However, the contribution of variation in health-care performance on ethnic differences in health outcomes has not previously been investigated in Scotland. We investigated whether delivery of health care and health policy was equitable across ethnic groups by studying quality outcomes based on mortality and hospital admission in the Scottish Health and Ethnicity Linkage Study (SHELS). The health status of the Scottish population has been deteriorating during the past 50 years relative to other nations with similar economies.13, 14 Scotland is a high-income country providing comprehensive health care free at the point of use for people with residency status, irrespective of ethnic group, country of birth, or nationality. Scotland is characterised by strong political support for ethnic equity from its Government and health agencies.

Health policy is a devolved responsibility in the four nations comprising the UK.12 The UK's National Health Service (NHS) places considerable importance on providing equitable care to all groups, and NHS Scotland has a responsibility to provide an equitable service to all ethnic groups under the Equality Act 2010. However, the contribution of variation in health-care performance on ethnic differences in health outcomes has not previously been investigated in Scotland. We investigated whether delivery of health care and health policy was equitable across ethnic groups by studying quality outcomes based on mortality and hospital admission in the Scottish Health and Ethnicity Linkage Study (SHELS). The health status of the Scottish population has been deteriorating during the past 50 years relative to other nations with similar economies.13, 14 Scotland is a high-income country providing comprehensive health care free at the point of use for people with residency status, irrespective of ethnic group, country of birth, or nationality. Scotland is characterised by strong political support for ethnic equity from its Government and health agencies. In general (ie, internationally), ethnic minority groups whether immigrants or offspring fare worse (or are thought to fare worse) in health-care quality and outcomes. In Scotland, the limited quantitative evidence for such a conclusion, mostly from the SHELS, has been conflicting, with the picture varying by outcome and ethnic group. With regard to total mortality, life expectancy, and common cancers, non-white ethnic groups in Scotland are advantaged,15, 16, 17 although Indian and Pakistani populations have more cardiovascular disease than the white Scottish population.18 For a wide range of other outcomes, the picture is mixed. National-level, reliable, current data on health-related risk factors are not available in Scotland, placing reliance on the Health Surveys for England 1999 and 2004. The datasets on health-related behaviours that are available, including from Scotland, paint a complex picture (eg, a low prevalence of smoking but a high prevalence of physical inactivity in most non-white ethnic minority groups).19

ailable in Scotland, placing reliance on the Health Surveys for England 1999 and 2004. The datasets on health-related behaviours that are available, including from Scotland, paint a complex picture (eg, a low prevalence of smoking but a high prevalence of physical inactivity in most non-white ethnic minority groups).19 We expected important ethnic group differences in the outcomes reported in this study but did not specify the direction in our analysis plan; however, in view of perceptions about barriers to health care, cultural competence, and the possibility of racism, the performance of the health service would be expected to be suboptimum for non-white ethnic minority populations. Methods Study design and population In this population-based cohort study, we assessed three dimensions of health-system performance (mortality-based health outcomes, primary care, and secondary care [panel]) by following a prespecified analysis plan that has also been published online.16 The SHELS is a population-based retrospective cohort.15, 20 In brief, using names, addresses, sex, and dates of birth, Scotland's Census 2001 was securely linked to the Community Health Index, a register of patients using the Scottish NHS. Personal details were replaced by encrypted Community Health Index numbers and encrypted Census numbers that were used to link to Census data, including ethnicity and health outcomes such as death records.

nd's Census 2001 was securely linked to the Community Health Index, a register of patients using the Scottish NHS. Personal details were replaced by encrypted Community Health Index numbers and encrypted Census numbers that were used to link to Census data, including ethnicity and health outcomes such as death records. A total of 4·86 million people completed Scotland's Census 2001 (on April 29, 2001), with 4·62 million (95%) successfully linked. The quality of linkage was high, with a false-positive linkage proportion of 0·08%.21 The cohort was censored at the time of death, departure from NHS Scotland to other parts of the UK, or the end of the follow-up period (April 30, 2013). The Multi-centre Research Ethics Committee for Scotland and the Privacy Advisory Committee of NHS National Services Scotland gave approval for this study. Assessment of ethnicity Ethnicity was reported in the Census with 14 predefined categories.20 One person completes the form on behalf of each person in the household, although optional individual forms for people older than 16 years were available. A high proportion of people in the cohort completed the form (95·7%); we received 100% completeness from the National Records of Scotland who applied their specifically developed imputation methods, using the records of similar people to predict answers.22

individual forms for people older than 16 years were available. A high proportion of people in the cohort completed the form (95·7%); we received 100% completeness from the National Records of Scotland who applied their specifically developed imputation methods, using the records of similar people to predict answers.22 To minimise the risk of disclosure (based on the rule that a minimum of six events were needed for each category of ethnic group), we combined several categories. We reduced the 14 census categories to ten for the mortality-based outcomes: white Scottish, other white British (primarily English and Welsh), white Irish, other white, any mixed background, Indian, Pakistani, other south Asian (combining Bangladeshi with other south Asian), African origin (combining black African, Caribbean, black Scottish, and other black), and Chinese. The “all other ethnic group” category was removed due to the difficulty of interpreting results for this heterogeneous group.

ackground, Indian, Pakistani, other south Asian (combining Bangladeshi with other south Asian), African origin (combining black African, Caribbean, black Scottish, and other black), and Chinese. The “all other ethnic group” category was removed due to the difficulty of interpreting results for this heterogeneous group. Outcomes We defined amenable, preventable, and avoidable mortality outcomes using the guidance of the Office for National Statistics,23 and defined diagnostic codes in death certificates using the tenth revision of the International Statistical Classification of Diseases and Related Health Problems (ICD 10; appendix). Some causes of death are both amenable to medical treatment and preventable through health policy, and these are not mutually exclusive categories; therefore, the number of avoidable deaths is less than the total of preventable and amenable deaths. In view of the large contribution of deaths from ischaemic heart disease, we repeated analyses for amenable mortality with these deaths excluded. Using the NHS Outcomes Framework,24 we coded avoidable hospital admissions into any avoidable hospital admission, avoidable hospital admission for an acute condition, and avoidable hospital admission for a chronic condition (appendix). We defined unplanned readmissions as being admitted to hospital for non-elective reasons within 30 days of a previous admission.

Using the NHS Outcomes Framework,24 we coded avoidable hospital admissions into any avoidable hospital admission, avoidable hospital admission for an acute condition, and avoidable hospital admission for a chronic condition (appendix). We defined unplanned readmissions as being admitted to hospital for non-elective reasons within 30 days of a previous admission. Covariates The National Records Scotland extracted age, sex, country of birth, and indicators of socioeconomic status from the Census. We created a dichotomous variable for country of birth (born in the UK or Ireland vs born elsewhere). We used three variables for dimensions of socioeconomic status, which were informed by our method for selecting such variables.25 The Scottish Index for Multiple Deprivation is a rank-based measure of deprivation, which is calculated for small areas (known as data zones and comprising a median of 750 individuals). Highest educational attainment typically reflects early adulthood socioeconomic status and was defined on the basis of the individual's highest educational attainment in people aged 16–74 years, and on the basis of the highest educational attainment of the household for the younger and elderly groups (for whom educational status was not collected). The socioeconomic circumstances of a child or elderly person living with a person with higher education were assumed to be better than of those of a child or elderly person who is not living with a person with higher education. Highest educational attainment was categorised as none, low, and high. Household tenure (homeowner vs non-homeowner) was used to reflect property-related household wealth.

partly related to poorer socioeconomic circumstances. The findings on avoidable hospital admissions suggest that some south Asian ethnic groups need improved access to primary care or improved quality of primary care, or both. Further investigation of the variations should provide insights to support a policy response. A barrier to investigating how well health systems meet the needs of ethnic minorities has been the paucity of data for quantitative analysis, with ethnicity recording being poor in many countries.26 Most such research has been in the USA and New Zealand.27 In the USA, where ethnicity and race are more strongly related to socioeconomic circumstances than in the UK, studies have found higher amenable mortality among black than white populations.28, 29 Similarly, ethnic and racial minorities in the USA have been found to have an increased risk of ambulatory care-sensitive conditions30, 31 and lower levels of receipt of high-quality treatment white ethnic groups.32 In New Zealand, amenable mortality makes an important contribution to the greater mortality risk among the Pacific peoples when compared with the European/other reference population (accounting for 26·2% of the disparity in men and 33·8% in women in 2001–04, unlike the opposite pattern of findings reported in our study).33, 34 Important variations in amenable mortality have also been reported in Singapore, a country with three major ethnic groups.35 The evidence is especially scarce within Europe. A 2017 systematic review27 found no studies of avoidable hospital admissions within Europe by ethnicity, race, or migration status, and only three longitudinal studies worldwide. A study by de Bruijne and colleagues,11 which found ethnic variations in unplanned readmissions and excess length of stay in the Netherlands, seems to be the only study to investigate the quality of health-care delivery by ethnicity in Europe. By contrast with our study, the authors suggest that shortcomings in the quality of the hospital care exist among people who were not born in western countries who have moved to the Netherlands.

on with higher education were assumed to be better than of those of a child or elderly person who is not living with a person with higher education. Highest educational attainment was categorised as none, low, and high. Household tenure (homeowner vs non-homeowner) was used to reflect property-related household wealth. Statistical analysis We followed the prespecified analysis plan with some minor deviation outlined in this paper. Avoidable mortality applied to causes of death at either all ages (eg, HIV/AIDS, complications of the perinatal period, and injuries) or between 0 and 74 years, with the exception of death from diabetes, which applies for those aged 0–49 years. We calculated rate ratios (RRs) with 95% CIs using Poisson regression, with robust variance for the mortality outcomes and with person-years at risk as the denominator. We multiplied RRs by 100 to be interpretable as percentages, as per our analysis plan. The white Scottish ethnic group was the reference category (100). Age-adjusted RRs were calculated by sex. Adjustments for country of birth and socioeconomic status were then added separately and in combination. Socioeconomic status variables were categorical and no specific functional forms were assumed; for example, across the quintile categories of deprivation. We derived age-adjusted mortality by ethnic group from age-adjusted RRs multiplied by the crude mortality rate of the white Scottish population to calculate the percentage contribution of amenable, preventable, and avoidable deaths to the overall mortality rates.

Statistical analysis We followed the prespecified analysis plan with some minor deviation outlined in this paper. Avoidable mortality applied to causes of death at either all ages (eg, HIV/AIDS, complications of the perinatal period, and injuries) or between 0 and 74 years, with the exception of death from diabetes, which applies for those aged 0–49 years. We calculated rate ratios (RRs) with 95% CIs using Poisson regression, with robust variance for the mortality outcomes and with person-years at risk as the denominator. We multiplied RRs by 100 to be interpretable as percentages, as per our analysis plan. The white Scottish ethnic group was the reference category (100). Age-adjusted RRs were calculated by sex. Adjustments for country of birth and socioeconomic status were then added separately and in combination. Socioeconomic status variables were categorical and no specific functional forms were assumed; for example, across the quintile categories of deprivation. We derived age-adjusted mortality by ethnic group from age-adjusted RRs multiplied by the crude mortality rate of the white Scottish population to calculate the percentage contribution of amenable, preventable, and avoidable deaths to the overall mortality rates. We used similar analytical approaches for hospital admission-based outcomes. We analysed people aged 19 years and older on the basis of the NHS outcome framework definition for any avoidable hospital admissions, calculating total avoidable hospital admissions during follow-up. We identified each outcome (all, chronic, or acute) from the main hospital discharge diagnosis. The denominator at risk was all hospital admissions during the follow-up (excluding unplanned readmissions), with the numerator being all unplanned readmissions. We used logistic regression because the model was a better fit than Poisson regression.

e (all, chronic, or acute) from the main hospital discharge diagnosis. The denominator at risk was all hospital admissions during the follow-up (excluding unplanned readmissions), with the numerator being all unplanned readmissions. We used logistic regression because the model was a better fit than Poisson regression. We analysed length of stay for chronic heart disease, lung cancer, and pneumonia in adults aged 20 years and older, following the SHELS standard age groups. We calculated geometric means for the length of stay in days, comparing ethnic groups using linear regression models of log-transformed lengths of stay. These data were not adjusted for comorbidity because of restrictions on data access. We analysed data using SAS version 9.4 (SAS Institute Inc, Cary, NC, USA). As required for disclosure control purposes, denominators and numbers of events are presented rounded to the nearest five, but regression models were estimated using exact numbers. Data sharing: the full dataset may be available to researchers via an application to NHS Scotland's Public Benefit and Privacy Panel and to National Records Scotland. Role of the funding source The funders had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all intermediate outputs, with the study statisticians (GC, LW, and MS) having access to the full study datasets. All authors had final responsibility for the decision to submit for publication.

n, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all intermediate outputs, with the study statisticians (GC, LW, and MS) having access to the full study datasets. All authors had final responsibility for the decision to submit for publication. Results 4 607 393 people were included in this analytical sample, of whom 2 409 344 (52·3%) were women (appendix). During 50·5 million person-years of follow-up, 84 705 amenable deaths, 138 065 preventable deaths, 166 245 avoidable deaths, 1·17 million avoidable hospital admissions, and 587 740 unplanned readmissions occurred. Characteristics of the SHELS cohort, including population size by ethnic group, are given in the appendix. In Scotland, in 2001, the other white British, white Irish, and other white minority populations were relatively large, but each non-white group had fewer than 10 000 men or 10 000 women, except for the Pakistani group (which accounted for about 2% of total population). The mean ages of non-white groups were all lower than that of the white Scottish group (39·6 years; 38 years in men and 41 years in women), especially that of the any mixed background group (22·6 years; 21 years in men and 24 years in women). Many non-white people were born in the UK or Ireland (for example, 60% of Pakistani women). Socioeconomic status was highest for the other white British group and varied for non-white ethnic groups, partly dependent on the indicator and sex.

tay in the Netherlands, seems to be the only study to investigate the quality of health-care delivery by ethnicity in Europe. By contrast with our study, the authors suggest that shortcomings in the quality of the hospital care exist among people who were not born in western countries who have moved to the Netherlands. Our study has several strengths, including simultaneous assessment of six dimensions of health-system performance in primary and secondary care. Comparison of patterns across these has indicated how complex issues are and the challenges for articulating health policy to respond to ethnic inequalities. We studied virtually the whole population, thereby minimising the non-response that occurs when using survey-based data. The private hospital sector is small in Scotland and it is required to supply data to the Information Services Division. Our study has allowed heterogeneity among usually broadly defined ethnic groups, such as within the south Asian group, to be investigated.

ny mixed background group (22·6 years; 21 years in men and 24 years in women). Many non-white people were born in the UK or Ireland (for example, 60% of Pakistani women). Socioeconomic status was highest for the other white British group and varied for non-white ethnic groups, partly dependent on the indicator and sex. Large ethnic differences were seen for amenable mortality, preventable mortality, and avoidable mortality (table). The white Scottish group had higher age-adjusted amenable mortality than most ethnic minority groups, although rates were imprecisely estimated for some ethnic groups (appendix). The high contribution of preventable deaths to overall mortality in the white Scottish group was substantial (table). By contrast, the lowest contribution of preventable deaths to all-cause mortality, after accounting for age differences, was among Chinese men and women. Although Indian and south Asian men and African-origin and Indian women had a relatively high proportion of preventable deaths, these represented lower absolute burdens of death due to lower all-cause mortality compared with white Scottish people.Table Amenable, preventable, and avoidable age-adjusted mortality rates per 100 000 person-years and their contribution to all-cause mortality, by sex and ethnic group

proportion of preventable deaths, these represented lower absolute burdens of death due to lower all-cause mortality compared with white Scottish people.Table Amenable, preventable, and avoidable age-adjusted mortality rates per 100 000 person-years and their contribution to all-cause mortality, by sex and ethnic group Total mortality Amenable mortality Preventable mortality Avoidable mortality Deaths Person-years Age-adjusted rate Age-adjusted rate Percentage of all deaths Age-adjusted rate Percentage of all deaths Age-adjusted rate Percentage of all deaths Men White Scottish 200 725 21 179 755 947·7 209·6 22·1 359·6 37·9 427·3 45·1 Other white British 12 975 1 571 080 684·8 133·3 19·5 224·1 32·7 265·6 38·8 White Irish 2710 202 190 946·5 205·1 21·7 345·3 36·5 410·4 43·4 Other white 1930 278 515 765·9 146·6 19·1 241·3 31·5 290·0 37·9 Any mixed background 195 56 265 1055·2 211·7 20·1 380·6 36·1 438·4 41·5 Indian 265 65 945 593·3 159·0 26·8 228·6 38·5 259·5 43·7 Pakistani 435 146 430 626·5 195·1 31·1 196·0 31·3 252·3 40·3 Other south Asian 135 35 500 753·7 183·8 24·4 298·9 39·7 348·9 46·3 African origin 130 32 160 846·5 170·2 20·1 259·8 30·7 340·7 40·2 Chinese 195 68 685 495 82·4 16·6 152·1 30·7 188·9 38·2 Women White Scottish 179 955 22 581 190 796·9 148·9 18·7 228·1 28·6 279·9 35·1 Other white British 11 155 1 644 435 599·5 103·3 17·2 147·3 24·6 180·8 30·2 White Irish 2470 216 905 697·5 123·5 17·7 188·0 27·0 234·1 33·6 Other white 1735 319 915 607·4 108·2 17·8 147·0 24·2 183·7 30·2 Any mixed background 145 59 970 708 129·4 18·3 182·9 25·8 247·7 35·0 Indian 155 59 925 483·5 124·6 25·8 132·1 27·3 161·6 33·4 Pakistani 295 143 940 588·1 130·1 22·1 147·1 25·0 184·6 31·4 Other south Asian 90 28 610 737·8 117·1 15·9 171·4 23·2 204·2 27·7 African origin 90 28 590 648·3 155·7 24·0 186·4 28·8 246·6 38·0 Chinese 175 68 010 524 90·9 17·3 119·3 22·8 164·8 31·5

25·8 247·7 35·0 Indian 155 59 925 483·5 124·6 25·8 132·1 27·3 161·6 33·4 Pakistani 295 143 940 588·1 130·1 22·1 147·1 25·0 184·6 31·4 Other south Asian 90 28 610 737·8 117·1 15·9 171·4 23·2 204·2 27·7 African origin 90 28 590 648·3 155·7 24·0 186·4 28·8 246·6 38·0 Chinese 175 68 010 524 90·9 17·3 119·3 22·8 164·8 31·5 In men, the any mixed background (RR 101·0, 95% CI 71·5–142·7) and white Irish (RR 97·9, 79·1–121·1) ethnic groups had similar age-adjusted amenable mortality to the white Scottish (figure 1; appendix); the Chinese (RR 39·3, 26·8–57·7) and other white British (RR 69·9, 58·5–83·6) ethnic groups had the lowest mortality. In women, the Chinese (RR 61·1, 44·2–84·4), other white British (RR 69·4, 59·9–80·3), and other white (RR 72·7, 61·5–85·9) ethnic groups also had lower age-adjusted amenable mortality.Figure 1 Amenable mortality by ethnic group Shown are RRs multiplied by 100 (so that they can be interpreted as percentage differences between ethnic groups) and adjusted for age; age and socioeconomic status; and age, socioeconomic status, and country of birth. The white Scottish group is the reference category. Bars represent 95% CIs. RR=rate ratio.

In men, the any mixed background (RR 101·0, 95% CI 71·5–142·7) and white Irish (RR 97·9, 79·1–121·1) ethnic groups had similar age-adjusted amenable mortality to the white Scottish (figure 1; appendix); the Chinese (RR 39·3, 26·8–57·7) and other white British (RR 69·9, 58·5–83·6) ethnic groups had the lowest mortality. In women, the Chinese (RR 61·1, 44·2–84·4), other white British (RR 69·4, 59·9–80·3), and other white (RR 72·7, 61·5–85·9) ethnic groups also had lower age-adjusted amenable mortality.Figure 1 Amenable mortality by ethnic group Shown are RRs multiplied by 100 (so that they can be interpreted as percentage differences between ethnic groups) and adjusted for age; age and socioeconomic status; and age, socioeconomic status, and country of birth. The white Scottish group is the reference category. Bars represent 95% CIs. RR=rate ratio. Adjustment for socioeconomic status attenuated, but did not abolish, the difference between the white Scottish and some other ethnic groups (particularly other white British, other white, and Indian) for both men and women. For example, the RR for amenable mortality in other white British men relative to the white Scottish population changed from 63·6 (95% CI 53·9–75·1) to 80·3 (76·4–84·4) after adjustment for socioeconomic status, whereas ad-justment for country of birth resulted in less attenuation (64·0, 54·2–75·6). Adjustment for country of birth moderately attenuated ethnic differences for the other white, Indian, other south Asian, African, Pakistani, and Chinese groups in both sexes.

3 (76·4–84·4) after adjustment for socioeconomic status, whereas ad-justment for country of birth resulted in less attenuation (64·0, 54·2–75·6). Adjustment for country of birth moderately attenuated ethnic differences for the other white, Indian, other south Asian, African, Pakistani, and Chinese groups in both sexes. The findings for amenable mortality, excluding ischaemic heart disease, were broadly similar (appendix). However, the any mixed background ethnic group no longer had high mortality, but this finding was based on few events and was imprecisely estimated. The patterns of preventable mortality by ethnic group were similar to those of amenable mortality, although differences tended to be larger (figure 2; appendix). The white Scottish population again had high proportions of age-adjusted preventable mortality in both men (359·6) and women (228·1) per 100 000 person-years, accounting for 37·9% and 28·6% of deaths, respectively, similar to the any mixed background group. The Chinese, Indian, other white, and other white British populations had lower age-adjusted risks than the white Scottish in both sexes. By contrast with amenable mortality, the Pakistani group had much lower preventable mortality than the white Scottish group (RR 54·5 [95% CI 45·2–65·8] in men; RR 64·5 [52·2–79·8] in women). Again, adjustment for socioeconomic status reduced the magnitude of ethnic differences.Figure 2 Preventable mortality by ethnic group

ontrast with amenable mortality, the Pakistani group had much lower preventable mortality than the white Scottish group (RR 54·5 [95% CI 45·2–65·8] in men; RR 64·5 [52·2–79·8] in women). Again, adjustment for socioeconomic status reduced the magnitude of ethnic differences.Figure 2 Preventable mortality by ethnic group Shown are RRs multiplied by 100 (so that they can be interpreted as percentage differences between ethnic groups) and adjusted for age; age and socioeconomic status; and age, socioeconomic status, and country of birth. The white Scottish group is the reference category. Bars represent 95% CIs. RR=rate ratio. Results for avoidable mortality were similar to those for preventable mortality (figure 3; appendix). In men, overall levels of avoidable mortality were highest among white Scottish, any mixed background, and white Irish groups, with the elevated mortality in white Scottish men partly accounted for by differences in socioeconomic status.Figure 3 Avoidable mortality by ethnic group Shown are RRs multiplied by 100 (so that they can be interpreted as percentage differences between ethnic groups) and adjusted for age; age and socioeconomic status; and age, socioeconomic status, and country of birth. The white Scottish group is the reference category. Bars represent 95% CIs. RR=rate ratio.

Results for avoidable mortality were similar to those for preventable mortality (figure 3; appendix). In men, overall levels of avoidable mortality were highest among white Scottish, any mixed background, and white Irish groups, with the elevated mortality in white Scottish men partly accounted for by differences in socioeconomic status.Figure 3 Avoidable mortality by ethnic group Shown are RRs multiplied by 100 (so that they can be interpreted as percentage differences between ethnic groups) and adjusted for age; age and socioeconomic status; and age, socioeconomic status, and country of birth. The white Scottish group is the reference category. Bars represent 95% CIs. RR=rate ratio. Patterns for avoidable hospital admissions differed from those seen for mortality-based outcomes (figure 4; appendix), with avoidable hospital admissions among some ethnic minority groups much higher than in the white Scottish group. For all-cause avoidable hospital admissions, the Pakistani population had the highest RRs (140·6 [95% CI 131·9–150·0] in men; 141·0 [129·0–154·1] in women), with Bangladeshi and Indian groups also having high RRs in men, but not in women. Adjustment for socioeconomic status tended to increase differences between the white Scottish group and these ethnic minority groups, as did adjustment for country of birth. The other white British and other white ethnic groups had lower avoidable hospital admissions, with adjustment for socioeconomic status attenuating the difference compared with the white Scottish group in both men and women. Caribbean men had lower risk of avoidable hospital admission (RR 65·1, 95% CI 49·9–84·9) than the white Scottish group, whereas Caribbean women did not (RR 96·3, 72·2–128·4). The RRs for the Chinese ethnic group were consistently low, too. Similar ethnic patterns were seen for both acute and chronic avoidable hospital admissions, although ethnic differences were larger for chronic avoidable hospital admissions (appendix).Figure 4 Avoidable hospital admissions by ethnic group

3, 72·2–128·4). The RRs for the Chinese ethnic group were consistently low, too. Similar ethnic patterns were seen for both acute and chronic avoidable hospital admissions, although ethnic differences were larger for chronic avoidable hospital admissions (appendix).Figure 4 Avoidable hospital admissions by ethnic group Shown are RRs multiplied by 100 (so that they can be interpreted as percentage differences between ethnic groups) and adjusted for age; age and socioeconomic status; and age, socioeconomic status, and country of birth. The white Scottish group is the reference category. Bars represent 95% CIs. RR=rate ratio. Differences in unplanned readmissions were slight (appendix). In men, the Pakistani ethnic group had higher odds ratios (ORs) for unplanned readmissions (OR 113·8, 95% CI 104·0–124·6), as did the Indian ethnic group after adjustment for covariates. Pakistani women also had slightly higher ORs after adjustment. The other white British group had lower ORs for unplanned readmission in both men and women, but adjustment for socioeconomic status substantially narrowed the difference with the white Scottish comparison group (in men, age-adjusted OR 86·0 [95% CI 83·4–88·7] vs age and socioeconomic status-adjusted OR 95·3 [92·5–98·1]; in women, age-adjusted OR 89·7 [87·2–92·2] vs age and socioeconomic status-adjusted OR 97·3 [94·6–100·0]). Length of stay for hospital admissions due to three diagnoses showed no consistent differences between ethnic groups (appendix).

Differences in unplanned readmissions were slight (appendix). In men, the Pakistani ethnic group had higher odds ratios (ORs) for unplanned readmissions (OR 113·8, 95% CI 104·0–124·6), as did the Indian ethnic group after adjustment for covariates. Pakistani women also had slightly higher ORs after adjustment. The other white British group had lower ORs for unplanned readmission in both men and women, but adjustment for socioeconomic status substantially narrowed the difference with the white Scottish comparison group (in men, age-adjusted OR 86·0 [95% CI 83·4–88·7] vs age and socioeconomic status-adjusted OR 95·3 [92·5–98·1]; in women, age-adjusted OR 89·7 [87·2–92·2] vs age and socioeconomic status-adjusted OR 97·3 [94·6–100·0]). Length of stay for hospital admissions due to three diagnoses showed no consistent differences between ethnic groups (appendix). Discussion Our study is a nationwide assessment using a set of measures of health-system performance by ethnicity. We found no evidence of systematically poorer health outcomes across ethnic minority groups in Scotland. Instead, we found that the white Scottish ethnic group had a relatively high burden of amenable and preventable deaths compared with other ethnic groups, possibly partly related to poorer socioeconomic circumstances. The findings on avoidable hospital admissions suggest that some south Asian ethnic groups need improved access to primary care or improved quality of primary care, or both. Further investigation of the variations should provide insights to support a policy response.

g the non-response that occurs when using survey-based data. The private hospital sector is small in Scotland and it is required to supply data to the Information Services Division. Our study has allowed heterogeneity among usually broadly defined ethnic groups, such as within the south Asian group, to be investigated. This study has some limitations. Although the indicators in this study are widely used, their validity remains unclear and they might be better viewed as an indicator for further investigation being needed.7, 36 The Office for National Statistics classification we used is not the only one, and has the drawback that the categories of amenable and preventable deaths are not mutually exclusive.37 Other potential indicators of health service performance exist but have not been included in this study (eg, quality of life, disability, and aspects of morbidity). Although our study is large, the number of events within some ethnic groups is small and some important ethnic inequalities might have been missed. Further assessment of the any mixed background ethnic group would be particularly important. Adjustment for further socioeconomic covariates (such as occupation38) could result in slightly differing findings, but given our structured approach to identifying appropriate socioeconomic covariates,25 our main results are unlikely to change. We did not adjust for disease risk factors (such as hypertension, obesity, and smoking) because of a scarcity of such data on a national scale in Scotland.18 Ideally, such adjustments will be done in the future. We did not examine the results for those born in and outside the UK separately because of insufficient statistical power, but would recommend such research for the future once the number of outcomes has increased over time and greater statistical power is available. Some people living in Scotland in 2001 were not linked to our database and some would have outcomes outside Scotland (eg, death abroad following travel or re-emigration). We cannot quantify the effects of incomplete data. We did not examine readmission by diagnosis. The generalisability of our results to the rest of the world is unclear. In view of changes in Scotland's ethnic group composition since 2001, this research warrants replication using linkage to the 2011 Census.

emigration). We cannot quantify the effects of incomplete data. We did not examine readmission by diagnosis. The generalisability of our results to the rest of the world is unclear. In view of changes in Scotland's ethnic group composition since 2001, this research warrants replication using linkage to the 2011 Census. The UK has performed poorly when compared with similar high-income countries in improving premature mortality.39 We add an important equity dimension to that analysis. On the basis of the limited available scientific literature, NHS Scotland seems to provide more equitable care for ethnic minorities than other high-income countries, which have attempted to assess the equity of health-system performance by ethnicity. Scotland's focus on ethnic equity within health policy might hold clues to realising improvements elsewhere. Our study suggests areas for improvement. A focus on improving the quality of primary care for ethnic minority groups (particularly south Asians) is warranted, especially because previous studies suggest suboptimum primary care might contribute to poor asthma and mental health outcomes.40, 41 In the context of long-term conditions such as asthma, there is a need to ensure appropriate self-management education, highlight the importance of regular use of preventive treatments, and ensure that individuals can detect early signs of a deterioration in their clinical condition and then know how to respond effectively.42 Improvements in diabetes care might be similarly important.43

d to ensure appropriate self-management education, highlight the importance of regular use of preventive treatments, and ensure that individuals can detect early signs of a deterioration in their clinical condition and then know how to respond effectively.42 Improvements in diabetes care might be similarly important.43 Our study emphasises the importance of studying several dimensions of health-system and health-policy performance in view of differences in the patterns shown by ethnic group (eg, between avoidable mortality and hospital admission). The complexity inherent in these patterns has also been found in other outcomes studied in SHELS. For example, excepting the Chinese population in whom health is generally good, some health outcomes are better and some worse in ethnic minority groups than in the reference white Scottish group (eg, cancers tend to be low compared with the reference group in most ethnic minority groups but cardiovascular diseases are commoner in south Asian groups).15, 21 White Scottish populations fare worse in several outcomes but have the highest level of breast cancer screening, an indicator of health care.44 Generalisations are difficult and evaluations need to be made for each outcome and for each ethnic group. This principle will likely apply outside of Scotland. The high preventable mortality among the white Scottish group requires a policy response that extends beyond the NHS to improve the social determinants of health, including health-related risk factors such as alcohol, tobacco, and poor diet. Such analysis benefits from examination through the lens of ethnic variations.

The high preventable mortality among the white Scottish group requires a policy response that extends beyond the NHS to improve the social determinants of health, including health-related risk factors such as alcohol, tobacco, and poor diet. Such analysis benefits from examination through the lens of ethnic variations. Supplementary Material Supplementary appendix Acknowledgments This study was supported by the Chief Scientist's Office (grant number CZH/4/878) and supplementary funding from the UK National Health Service (NHS) Health Scotland. The Information Services Division of NHS National Services Scotland and the National Records of Scotland both made in-house contributions to the study. Additionally, SVK acknowledges funding from an NHS Research Scotland Senior Clinical Fellowship (SCAF/15/02), the Medical Research Council (MC_UU_12017/13 and MC_UU_12017/15), and the Chief Scientist's Office (SPHSU13 and SPHSU15). We would also like to acknowledge Colin Fischbacher as a co-applicant who helped to set the study up; Chris Povey as a co-investigator who had the idea of linking the census data to health data and did most of the linkage of census to Community Health Index; Alex Stannard who advised on census data; David Clark who advised and assisted with linkage to the Information Services Division databases; and Jamie Pearce who advised on data analysis, particularly in relation to social and economic variables. We thank Theresa Kirkpatrick for secretarial and administrative assistance.

lex Stannard who advised on census data; David Clark who advised and assisted with linkage to the Information Services Division databases; and Jamie Pearce who advised on data analysis, particularly in relation to social and economic variables. We thank Theresa Kirkpatrick for secretarial and administrative assistance. Contributors SVK conceived the idea for the study and drafted the manuscript. GC, LW, and MS conducted the statistical analyses. RSB was principal investigator for the study. LG was chair of the mortality and hospitalisation subgroup of the Scottish Health and Ethnicity Linkage Study. All authors contributed to the study design and interpretation of results, critically revised the manuscript, and approved the final version of the paper. Declaration of interests We declare no competing interests.

Introduction Sexually transmitted infections such as chlamydia are thought to be an important cause of infertility worldwide, and WHO estimates that there are 131 million new chlamydia infections each year in total (including undetected cases).1 Their Global Health Sector Strategy on Sexually Transmitted Infections recommends chlamydia prevalence monitoring in high-risk groups, including adolescents.1 Other studies have recommended definition of acceptable local targets for chlamydia prevalence; reduction of chlamydia infections to reduce the incidence of pelvic inflammatory disease (PID);2 and establishment of surveillance systems to investigate the effects of control policies on PID and its complications.3 However, prevalence monitoring is challenging because most chlamydia infections are asymptomatic4 and therefore not detected by syndromic case reporting systems. WHO and the European Centre for Disease Prevention and Control (ECDC) both note that the best strategies to control and monitor chlamydia infections have yet to be established, and they encourage further research into these areas.1, 5

ons are asymptomatic4 and therefore not detected by syndromic case reporting systems. WHO and the European Centre for Disease Prevention and Control (ECDC) both note that the best strategies to control and monitor chlamydia infections have yet to be established, and they encourage further research into these areas.1, 5 Programmes of screening or widespread testing for chlamydia form part of national sexual health provision in several countries around the world.5, 6 Early modelling studies predicted that screening would be highly effective, but more recent studies have been more conservative and have increasingly considered the role of partner notification as well as widespread testing.7 A 2016 Cochrane review8 highlighted a controlled trial in the Netherlands that found low screening uptake and no change in the proportion of tested individuals who were positive for chlamydia (ie, positivity) after three annual screening invitations; whereas by contrast, a trial in female sex workers in Peru found a reduction in prevalence after 4 years. In the USA, the 2-yearly cycles of surveys for the NHANES studies indicated a decrease in chlamydia prevalence from 1999 to 2008 in people aged 14–39 years overall, but no change in women aged 15–25 years, which is the population targeted for routine screening.9 Overall prevalence remained similar over the three cycles from 2007 to 2012.10 Research in context Evidence before this study

Programmes of screening or widespread testing for chlamydia form part of national sexual health provision in several countries around the world.5, 6 Early modelling studies predicted that screening would be highly effective, but more recent studies have been more conservative and have increasingly considered the role of partner notification as well as widespread testing.7 A 2016 Cochrane review8 highlighted a controlled trial in the Netherlands that found low screening uptake and no change in the proportion of tested individuals who were positive for chlamydia (ie, positivity) after three annual screening invitations; whereas by contrast, a trial in female sex workers in Peru found a reduction in prevalence after 4 years. In the USA, the 2-yearly cycles of surveys for the NHANES studies indicated a decrease in chlamydia prevalence from 1999 to 2008 in people aged 14–39 years overall, but no change in women aged 15–25 years, which is the population targeted for routine screening.9 Overall prevalence remained similar over the three cycles from 2007 to 2012.10 Research in context Evidence before this study Widespread chlamydia screening has been implemented in several countries, but the extent of its effects is unclear. A 2016 Cochrane review investigating the effects of chlamydia screening searched the Cochrane Sexually Transmitted Infections Group Specialised Register and other registries with the terms (variations on and synonyms of) “genital chlamydia infection” and “screening” to Feb 14, 2016. The authors searched for randomised controlled trials (RCTs) done in adults (people older than 13 years), which compared a chlamydia screening intervention with usual care and reported one of the following as a primary outcome: chlamydia prevalence; pelvic inflammatory disease in women; epididymitis in men; or incidence of preterm delivery. The review found two trials investigating the effect of chlamydia screening on population chlamydia prevalence. A controlled trial in the Netherlands found low screening uptake and no change in chlamydia positivity after three annual screening invitations, whereas a trial in female sex workers in Peru found a reduction in prevalence after 4 years. Both trials were assessed as providing low-quality evidence. England's extensive National Chlamydia Screening Programme (NCSP) was rolled out between 2003 and 2008, but population-based surveys in 1999–2001 and 2010–12 found little change in chlamydia prevalence in young people, although confidence intervals were wide. Comprehensive estimates of annual numbers of chlamydia tests and diagnoses in young people in England from 2000 to 2015, by sex and age group, have recently become available. However, no study has shown how these data correspond to changes in chlamydia prevalence.

young people, although confidence intervals were wide. Comprehensive estimates of annual numbers of chlamydia tests and diagnoses in young people in England from 2000 to 2015, by sex and age group, have recently become available. However, no study has shown how these data correspond to changes in chlamydia prevalence. Added value of this study In our study, we provide a more-detailed picture of the year-to-year changes in chlamydia prevalence before, during, and after the period over which the NCSP was rolled out and test coverage increased (2003–08). We estimated prevalence by age group and sex each year from 2000 to 2015 using a recently developed evidence synthesis method and newly published data for test coverage and numbers of diagnoses, combined with information on natural history and care-seeking behaviour. Additionally, we estimated the average duration of infection, which clearly declined year-on-year in both sexes as screening activity increased, and particularly after full-scale implementation of NCSP. Since 2010, rates of testing and diagnosis have reduced, and modest increases in inferred prevalence and the average duration of infection have occurred in all sex and age groups assessed. Implications of all the available evidence

In our study, we provide a more-detailed picture of the year-to-year changes in chlamydia prevalence before, during, and after the period over which the NCSP was rolled out and test coverage increased (2003–08). We estimated prevalence by age group and sex each year from 2000 to 2015 using a recently developed evidence synthesis method and newly published data for test coverage and numbers of diagnoses, combined with information on natural history and care-seeking behaviour. Additionally, we estimated the average duration of infection, which clearly declined year-on-year in both sexes as screening activity increased, and particularly after full-scale implementation of NCSP. Since 2010, rates of testing and diagnosis have reduced, and modest increases in inferred prevalence and the average duration of infection have occurred in all sex and age groups assessed. Implications of all the available evidence Our analysis provides evidence for a reduction in chlamydia prevalence and average duration of infection in England associated with large-scale population screening, which occurred in both sexes. Reduction in average duration of infection is the better indicator of screening programme success because prevalence is also affected by changes in population sexual risk behaviour. Our evidence synthesis using data from a national testing programme complements existing data from clinical trials to improve understanding of the effects of chlamydia screening. Declines in testing in England since 2010, alongside partial reversals of health gains, are concerning.

in population sexual risk behaviour. Our evidence synthesis using data from a national testing programme complements existing data from clinical trials to improve understanding of the effects of chlamydia screening. Declines in testing in England since 2010, alongside partial reversals of health gains, are concerning. In England, the National Chlamydia Screening Programme (NCSP) has offered chlamydia testing to men and women aged 15–24 years since 2003, with full nationwide implementation completed in 2008. Screening is commissioned locally, and offered in settings including by general practitioners, sexual health services, pharmacies, and online. In 2016, 20·7% of the eligible population was tested and 9·1% of tests were positive, corresponding to a national detection rate of 1882 per 100 000 people aged 15–24 years. However, chlamydia prevalence in young people was similar between population-based prevalence surveys done in 1999–2001 (National Surveys of Sexual Attitudes and Lifestyles [Natsal]-2) and in 2010–12 (Natsal-3).11 Prevalence in women aged 18–24 years was 3·1% (95% CI 1·8–5·2) in 1999–2001 and 3·2% (2·2–4·6) in 2010–12. The corresponding estimates for men were 2·9% (1·3–6·3) in 1999–2001 and 2·6% (1·7–4·0) in 2010–12.11 Surveys done once per decade provide little information on prevalence trends, and over time, marked changes have occurred in uptake of chlamydia testing in England.

1·8–5·2) in 1999–2001 and 3·2% (2·2–4·6) in 2010–12. The corresponding estimates for men were 2·9% (1·3–6·3) in 1999–2001 and 2·6% (1·7–4·0) in 2010–12.11 Surveys done once per decade provide little information on prevalence trends, and over time, marked changes have occurred in uptake of chlamydia testing in England. Surveillance data provide more detailed information on trends, but no system in England had recorded complete annual numbers of chlamydia tests and diagnoses by sex and age until the Chlamydia Testing Activity Dataset became available in 2012. However, a study published in 2017, which combined data from several surveillance systems, has provided estimates for these quantities before 2012.12 In this analysis we use a model-based framework to estimate changes in chlamydia prevalence in England each year from 2000 to 2015 as diagnostic test coverage changed before, during, and after the rollout of the NCSP. Methods Study design and data sources This study was a modelling analysis that used previously published estimates and Public Health England data. Numbers of chlamydia tests and diagnoses in England for the period 2000–12 were estimated by Chandra and colleagues,12 who reported minimum and maximum estimates. For 2013 onwards, comprehensive data have been published by Public Health England.13 Our definition of testing coverage is that used by the NCSP: the annual number of chlamydia tests divided by the number of sexually active individuals in the relevant age and sex group. Mid-year population estimates came from the Office for National Statistics.14

ive data have been published by Public Health England.13 Our definition of testing coverage is that used by the NCSP: the annual number of chlamydia tests divided by the number of sexually active individuals in the relevant age and sex group. Mid-year population estimates came from the Office for National Statistics.14 We estimated chlamydia prevalence and incidence in England in men and women aged 15–24 years from 2000 to 2015 using a method that synthesises data on chlamydia testing and diagnosis with information on care-seeking behaviour and infection natural history. Details of the method are published elsewhere.15 We did analyses separately by sex and age group (men and women, ages 15–19 years and 20–24 years). For 2000–12, we did separate calculations using the minimum and maximum testing and diagnosis estimates.12

ormation on care-seeking behaviour and infection natural history. Details of the method are published elsewhere.15 We did analyses separately by sex and age group (men and women, ages 15–19 years and 20–24 years). For 2000–12, we did separate calculations using the minimum and maximum testing and diagnosis estimates.12 Statistical analysis The method is based on a model of a closed population at steady state, in which uninfected individuals who become infected move into either a symptomatic-infected or an asymptomatic-infected state. Symptomatic individuals seek treatment, whereas asymptomatic infections are detected through asymptomatic testing programmes. By allowing for different rates of testing in individuals who are symptomatic and non-symptomatic (either uninfected or with an asymptomatic infection), the model accounts for the changing proportion of tests done in each of these groups. We used data from previous studies to parameterise natural history and care-seeking behaviour (appendix), which remain the same from year to year. The observed test coverage and annual diagnoses per capita, which typically change from year to year, can then be used to estimate the proportion of the population in each state, and hence the chlamydia prevalence. The strength of this method is that it provides an estimate of prevalence, rather than relying on proxies such as diagnoses (which are affected by the amount of testing) or positivity (which is affected by the risk profile of those tested).

proportion of the population in each state, and hence the chlamydia prevalence. The strength of this method is that it provides an estimate of prevalence, rather than relying on proxies such as diagnoses (which are affected by the amount of testing) or positivity (which is affected by the risk profile of those tested). Prevalence estimates were then used to calculate estimated year-on-year prevalence changes in each sex. For each sampled set of natural history parameters, for each year, the prevalence sample for the previous year was subtracted from the sample for the current year to give a sample for the change in prevalence. Together, the samples provide a posterior distribution for the year-to-year change. Because our method is based on a mechanistic model, it can also be used to establish other quantities, including the average duration of infections. The average duration of infections was calculated by dividing prevalence by incidence. All computation and visualisation of results was done in the Python language (version 2.7.10), and results and figures can be reproduced using the code available online. Role of the funding source The funders of this study had no role in study design, data collection, analysis, or interpretation, or writing of the report. The corresponding author had full access to all the data in the study and the final responsibility to submit for publication.

All computation and visualisation of results was done in the Python language (version 2.7.10), and results and figures can be reproduced using the code available online. Role of the funding source The funders of this study had no role in study design, data collection, analysis, or interpretation, or writing of the report. The corresponding author had full access to all the data in the study and the final responsibility to submit for publication. Results Chlamydia test coverage increased in both sexes and age groups every year from 2000 to 2010, with both the maximum and the minimum test numbers (figure 1). From 2010 onwards, coverage decreased in all age and sex groups, but the decrease was greater in people aged 15–19 years than those aged 20–24 years. In men, diagnoses per capita increased until 2010, after which point they stayed at a similar level or began to decrease slightly (figure 1A, B). The trend of diagnoses in women was less certain than in men, but increased until around 2008 (figure 1C, D). In women aged 15–19 years, the diagnoses per capita began to decrease again after 2010; whereas in those aged 20–24 years, diagnoses remained fairly constant from 2009 onwards.Figure 1 Chlamydia tests and diagnoses in young people in England, 2000–15 (A) Men aged 15–19 years; (B) men aged 20–24 years; (C) women aged 15–19 years; and (D) women aged 20–24 years. Dotted lines show tests; solid lines show diagnoses. Pairs of lines show minimum and maximum numbers of tests and diagnoses. Data up to 2012 are from Chandra and colleagues;12 data from 2013 are from Public Health England.13

19 years; (B) men aged 20–24 years; (C) women aged 15–19 years; and (D) women aged 20–24 years. Dotted lines show tests; solid lines show diagnoses. Pairs of lines show minimum and maximum numbers of tests and diagnoses. Data up to 2012 are from Chandra and colleagues;12 data from 2013 are from Public Health England.13 In men, the model estimated that chlamydia prevalence was increasing in the years immediately preceding the introduction of the NCSP (figure 2A, B). The rollout of the NCSP coincided with a reversal of this trend and by 2009, when the programme was fully implemented nationally, the model estimated that prevalence was decreasing. In men aged 15–24 years, the prevalence decreases over the 2 years after full completion of NCSP (2008–10; appendix), estimated using minimum and maximum testing and diagnosis estimates, were 0·68 percentage points (95% credible interval 0·26–1·40; minimum) and 0·66 percentage points (0·25–1·37; maximum). However, prevalence increased in 2010–12, followed by only minor estimated changes in the years after (figure 2A, B).Figure 2 Inferred annual changes in chlamydia prevalence in young people in England, 2000–15

percentage points (95% credible interval 0·26–1·40; minimum) and 0·66 percentage points (0·25–1·37; maximum). However, prevalence increased in 2010–12, followed by only minor estimated changes in the years after (figure 2A, B).Figure 2 Inferred annual changes in chlamydia prevalence in young people in England, 2000–15 (A) Men aged 15–19 years; (B) men aged 20–24 years; (C) women aged 15–19 years; and (D) women aged 20–24 years. Each datapoint shows the change between estimates for one year and those for the subsequent year. For the later years the minimum and maximum estimates were identical so boxes show the same results. Coloured boxes show the 50% credible interval, the small horizontal lines show the median estimate, and whiskers show the 95% credible interval. The pattern of annual estimated prevalence changes in women was less clear than in men (figure 2C, D). Incomplete data collection before 200812 led to major uncertainty in the numbers of tests and diagnoses, making it difficult to say whether prevalence was increasing or decreasing year-on-year before and during the NCSP rollout. Following full implementation in 2008, however, the pattern was almost identical to that in men: an initial estimated reduction in prevalence, which was reversed in 2011–12, then followed by only minor estimated change. In women aged 15–24 years, the estimated decrease in prevalence over the period 2008–10 was 0·77 percentage points (0·45–1·27)—maximum and minimum test and diagnosis numbers were the same for women aged 15–24 years after 2008.

n in prevalence, which was reversed in 2011–12, then followed by only minor estimated change. In women aged 15–24 years, the estimated decrease in prevalence over the period 2008–10 was 0·77 percentage points (0·45–1·27)—maximum and minimum test and diagnosis numbers were the same for women aged 15–24 years after 2008. We compared prevalence estimated from surveillance data in 2000 and 2011 to the survey-based estimates from Natsal, and the estimates were in agreement (appendix).

n in prevalence, which was reversed in 2011–12, then followed by only minor estimated change. In women aged 15–24 years, the estimated decrease in prevalence over the period 2008–10 was 0·77 percentage points (0·45–1·27)—maximum and minimum test and diagnosis numbers were the same for women aged 15–24 years after 2008. We compared prevalence estimated from surveillance data in 2000 and 2011 to the survey-based estimates from Natsal, and the estimates were in agreement (appendix). Our results showed a decrease in the estimated median durations of infections in both sexes, with reductions every year, 2000–10 (figure 3). These estimated reductions began before the introduction of the NCSP and continued throughout and after its rollout until 2010, as rates of testing increased. In men aged 15–24 years, the posterior median decreases in the duration of infection from 2008 to 2010, estimated using the minimum or maximum testing and diagnosis figures, were 75 days (95% credible interval 17–255; minimum) and 74 days (17–247; maximum). In women aged 15–24 years, the decrease was 30 days (22–40). However, testing declined after 2010 and the estimated average duration of infections increased in turn. The estimated average duration of infections was in general longer in men than in women, and had greater uncertainty, because of a longer and less-certain duration of untreated infections in men.16 However, changes in average duration (data not shown) had less uncertainty than the absolute duration estimates (figure 3), and clear evidence supported the direction of change, showing annual reductions in 2000–10, followed by increases in 2010–15.Figure 3 Median absolute estimated duration of chlamydia infection in young people in England, 2000–15

ation (data not shown) had less uncertainty than the absolute duration estimates (figure 3), and clear evidence supported the direction of change, showing annual reductions in 2000–10, followed by increases in 2010–15.Figure 3 Median absolute estimated duration of chlamydia infection in young people in England, 2000–15 (A) Men aged 15–19 years; (B) men aged 20–24 years; (C) women aged 15–19 years; and (D) women aged 20–24 years. Shading shows 95% credible interval for estimated median duration of infection. The two lines and shading show estimates from the minimum and maximum numbers of tests and diagnoses; dark shading shows overlap of minimum and maximum.12 Discussion Our analysis suggests that in both men and women, chlamydia prevalence and the average duration of infections fell during the period immediately following the full-scale implementation of the NCSP in 2008. However, after 2010 our analysis estimated that prevalence increased and then stabilised in both sexes, whereas duration of infection increased—slightly in women, and substantially in men.

the average duration of infections fell during the period immediately following the full-scale implementation of the NCSP in 2008. However, after 2010 our analysis estimated that prevalence increased and then stabilised in both sexes, whereas duration of infection increased—slightly in women, and substantially in men. The best population-based estimates for chlamydia prevalence in England come from the Natsal-2 survey of 1999–2001 and the Natsal-3 survey of 2010–12.11, 17 Comparison of these surveys found little evidence of any prevalence change, although small numbers of infections meant the CIs were wide. Our results in people aged 15–19 years and 20–24 years are consistent with these estimates but add to the picture by using annual figures from surveillance data to examine estimated changes in prevalence in the years between the surveys, and since Natsal-3.

although small numbers of infections meant the CIs were wide. Our results in people aged 15–19 years and 20–24 years are consistent with these estimates but add to the picture by using annual figures from surveillance data to examine estimated changes in prevalence in the years between the surveys, and since Natsal-3. Testing and diagnoses increased in men in the years between the two Natsal surveys, and we estimated annual increases in prevalence followed by declines. Increases before 2004 were driven in our analysis by increasing incidence, probably due to behavioural changes. Rising prevalence could have contributed to the increased testing, with more infected men leading to more symptomatic men seeking tests. The later declines in prevalence coincided with the full-scale implementation of the NCSP. In women before 2008, the large differences between minimum and maximum estimates of numbers of tests and diagnoses make the trend in estimated changes in prevalence unclear, but clear estimated reductions in prevalence followed full-scale implementation of the NCSP in 2008. In women, the prevalence changes estimated using the maximum estimates of tests and diagnoses have a more plausible trajectory than those from the minimum estimates, showing a less abrupt change in testing and diagnoses from 2007 to 2008 and more consistent changes in prevalence. These estimates suggest that prevalence fell as testing increased. Absolute changes were small (<1%), but nonetheless important as a proportion of prevalence, which was also low (around 2–3% in men and women aged 18–24 years in both Natsal-217 and Natsal-311).

rom 2007 to 2008 and more consistent changes in prevalence. These estimates suggest that prevalence fell as testing increased. Absolute changes were small (<1%), but nonetheless important as a proportion of prevalence, which was also low (around 2–3% in men and women aged 18–24 years in both Natsal-217 and Natsal-311). In both men and women, the greatest estimated reductions in prevalence were between 2008 and 2010. This period corresponded to the greatest annual increases in testing. Using the maximum estimates for chlamydia testing and diagnosis rates, the most marked estimated increases in prevalence occurred in 2011–12. This corresponded to a decrease in testing in all four age and sex groups.

lence were between 2008 and 2010. This period corresponded to the greatest annual increases in testing. Using the maximum estimates for chlamydia testing and diagnosis rates, the most marked estimated increases in prevalence occurred in 2011–12. This corresponded to a decrease in testing in all four age and sex groups. Attempting to assess the performance of a chlamydia screening programme through changes in prevalence is difficult because of confounding effects of changes in sexual behaviour. Increases in risky behaviour would attenuate reductions in prevalence, while reductions in risky behaviour would exaggerate the prevalence reductions caused by screening. Regardless of changes in sexual behaviour, a successful screening programme would reduce the average duration of infections—thereby reducing prevalence, incidence of onward transmission, and incidence of sequelae to lower levels than would have occurred without screening. We estimated a continuous decline in the average duration of infection in both sexes from 2001 to 2010 because of increased asymptomatic testing, despite estimated increases and then decreases in prevalence during that time. Based on estimated average duration of infection, the NCSP appears to have been successful, although we note some reversal of progress in the years since 2010, coinciding with declining rates of testing. The decline in infection duration before the NCSP is probably due to increasing awareness of chlamydia by clinicians leading to the increase in testing and diagnosis ahead of the official implementation of NCSP in 2003–08, with national screening having been formally considered from 1998 and pilot studies done since 1999.18

ecline in infection duration before the NCSP is probably due to increasing awareness of chlamydia by clinicians leading to the increase in testing and diagnosis ahead of the official implementation of NCSP in 2003–08, with national screening having been formally considered from 1998 and pilot studies done since 1999.18 We estimated a longer duration of infection in men than women, partly due to slower natural clearance of untreated, asymptomatic infections in men than women,16 and partly due to less frequent asymptomatic screening in men than women. These two factors outweigh the larger proportion of incident infections in men than in women that are symptomatic and prompt testing.

en than women, partly due to slower natural clearance of untreated, asymptomatic infections in men than women,16 and partly due to less frequent asymptomatic screening in men than women. These two factors outweigh the larger proportion of incident infections in men than in women that are symptomatic and prompt testing. Additional work on the natural history and pathophysiology of chlamydia infection would complement improved surveillance to provide a better understanding of the short-term and long-term effects of chlamydia screening on population sexual and reproductive health.2, 3, 15, 16, 19 Good evidence supports the idea that chlamydia infection and pelvic inflammatory disease increase a woman's risk of future reproductive morbidity.2 Reproductive complications might take many years to occur, depending on when a woman becomes, or attempts to become, pregnant. As time goes on, however, data on reproductive morbidity could complement analysis of testing and diagnosis data. Some concerns exist that chlamydia screening and treatment might lead to more repeat infections, potentially increasing risk of reproductive morbidity compared with a single, untreated infection.3 However, whether the higher risk associated with repeat infections2 is due to increased exposure time or an increased per-episode risk is unknown. No data on repeat infections is available to investigate this hypothesis further. Chlamydia might induce some degree of immunity to future infections,3 but this would not affect any of our estimates because our methods estimate the average force of infection across all uninfected people that reproduces the recorded annual numbers of tests and diagnoses.

ilable to investigate this hypothesis further. Chlamydia might induce some degree of immunity to future infections,3 but this would not affect any of our estimates because our methods estimate the average force of infection across all uninfected people that reproduces the recorded annual numbers of tests and diagnoses. The precision of our estimates of infection duration is limited by uncertainty in chlamydia's natural history,15 particularly the clearance rate of untreated infections in men,16 and the proportion of infections that become symptomatic in both sexes. However, these uncertainties have less of an effect on estimated changes in prevalence and duration of infection than on absolute estimates. Similarly, posterior distributions for both prevalence and the average duration of infection are insensitive to changes in symptomatic treatment seeking (appendix).

in both sexes. However, these uncertainties have less of an effect on estimated changes in prevalence and duration of infection than on absolute estimates. Similarly, posterior distributions for both prevalence and the average duration of infection are insensitive to changes in symptomatic treatment seeking (appendix). Uncertainty also exists in the numbers of chlamydia tests and diagnoses before 2012, particularly in women.12 The true values for tests and diagnoses in women are probably closer to the maximum estimates, since the minimum estimates have an abrupt transition in 2007–08, and the maximum estimates provide a smoother pattern of prevalence changes in the years 2000–10, which has fewer turning points and is closer to the pattern in men. Correlations have been reported between infection risk factors and testing behaviour,20 which are not formally accounted for in our estimation approach because they are not recorded in the surveillance data, but sensitivity analysis has indicated that these have a very small effect on prevalence estimates.15

rn in men. Correlations have been reported between infection risk factors and testing behaviour,20 which are not formally accounted for in our estimation approach because they are not recorded in the surveillance data, but sensitivity analysis has indicated that these have a very small effect on prevalence estimates.15 With constrained public health budgets worldwide, more-robust assessments of chlamydia screening programme performance are needed urgently. Population-based surveys such as Natsal11, 17 are valuable but can only be done infrequently, which limits the information they can provide to assess the performance of chlamydia screening programmes. The partial reversal of screening benefits in England since 2010, associated with a reduction in testing activity, points towards a need to maintain screening to maintain control of infection. A combination of effective prevention (eg, condom use promotion), case-finding, and treatment, including both screening and partner notification and management, are needed for effective chlamydia control. With improved surveillance data, the efficiency of screening might be increased by better targeting of those at greatest risk of infection, to reduce the volume of testing required to achieve a similar diagnosis rate and to address health inequalities. Ultimately, planning a broad and effective strategy for chlamydia control requires health economic analysis, considering the cost of the different options as well as their effectiveness. Before the introduction of NCSP, detailed mathematical and economic modelling was done;21 now, with more than a decade of data since the implementation of NCSP, we recommend revisiting these models to plan the most cost-effective chlamydia control strategies, incorporating prevention, screening, and partner notification.

fore the introduction of NCSP, detailed mathematical and economic modelling was done;21 now, with more than a decade of data since the implementation of NCSP, we recommend revisiting these models to plan the most cost-effective chlamydia control strategies, incorporating prevention, screening, and partner notification. We advocate monitoring of test coverage and diagnoses, which can be used to estimate chlamydia prevalence, incidence, duration of infection and changes in these quantities with the method we have used in this study. The method15 can readily incorporate improved surveillance data (eg, recording whether the infection was symptomatic, the duration of any symptoms, why the patient was tested, and information on risk behaviour) to increase the precision of estimates and compare performance in different social groups to assess inequalities. Additionally, we endorse the recommendation for the monitoring of chlamydia-related complications, including PID.3 Other causes of PID, such as gonorrhoea infection, could also be included in the analysis subject to the required data being collected. In summary, this study provides evidence that increased chlamydia testing in England has reduced the prevalence of chlamydia infection and the average duration of infections in both men and women, but that these benefits have partly reversed since 2010. This evidence supports chlamydia screening as a control strategy, in conjunction with other measures, including effective partner notification.

England has reduced the prevalence of chlamydia infection and the average duration of infections in both men and women, but that these benefits have partly reversed since 2010. This evidence supports chlamydia screening as a control strategy, in conjunction with other measures, including effective partner notification. For more information on the Chlamydia Testing Activity Dataset see https://www.gov.uk/guidance/chlamydia-testingactivity-dataset-ctad For the code to reproduce the results and figures from this study see https://github.com/joanna-lewis/ct_trends Supplementary Material Supplementary appendix Acknowledgments JL and PJW thank the UK National Institute for Health Research (NIHR) Health Protection Research Unit in Modelling Methodology at Imperial College London, in partnership with Public Health England (HPRU-2012–10080) for funding. PJW also thanks the Medical Research Council (MR/K010174/1) for funding. The views expressed are those of the authors and not necessarily those of the UK Department of Health, MRC, National Health Service, NIHR, or Public Health England. Contributors Both authors conceived the study. JL did the analyses and wrote the first draft of the report. Both authors contributed substantially to study conception and design, acquisition and interpretation of data, and revision and editing of the manuscript. Declaration of interests Both authors declare no competing interests.

For 2006–10, Perviz Asaria and colleagues1 assessed what proportion of deaths in England with acute myocardial infarction as an underlying or contributing cause were in people admitted to a hospital in the 28 days before death, and whether acute myocardial infarction was one of the recorded diagnoses in such hospital admissions. The authors also wanted to set acute myocardial infarction deaths with a 28 day antecedent of hospital-admission in broader context: that of people admitted to hospital for whom an acute myocardial infarction event was diagnosed. Asaria and colleagues estimated that as many deaths with acute myocardial infarction as underlying cause had occurred without having had a 28-day antecedent hospital admission as had occurred within 28 days of hospitalised acute myocardial infarction-events. In such a study, dates and definitions matter. In the methods section of their paper, Asaria and colleagues note: “We use the term hospital admission to refer to a continuous spell of care”. Therein lies statistical danger because admission occurs on a specific date, whereas hospitalisation has a variable duration. Moreover, although the date of transition from first to second finished consultant episode (FCE) within a single hospitalisation is recorded by NHS Digital, the date of diagnosis of acute myocardial infarction within the first or subsequent FCE is not.

n a specific date, whereas hospitalisation has a variable duration. Moreover, although the date of transition from first to second finished consultant episode (FCE) within a single hospitalisation is recorded by NHS Digital, the date of diagnosis of acute myocardial infarction within the first or subsequent FCE is not. By contrast, for evaluating Scotland’s National Naloxone Policy,2 28 day look-backs from opioid-related deaths (ORDs) in 2006–2010 were precisely defined: to prison release, with day of prison-release as day 1 so that prison-release-day ORDs were counted;3,4 and to hospital-discharge with day after hospital-discharge as day 1 so that deaths on or at admission were excluded.3,5,6 Editors should allow authors of record-linkage studies properly to convey in their methods the careful, logical definitions of at-risk periods that must have been explicit in their linkage-protocol; and to do so without short-circuiting. Definitions need to be sufficiently precise that others, internationally, who seek to do validation studies, can replicate the authors’ approach and test deviations in potentially influential respects. How dates and definitions were specified in the linkage-protocol matter as people other than the research team typically prepare the linkage database: as specified in the protocol.

nationally, who seek to do validation studies, can replicate the authors’ approach and test deviations in potentially influential respects. How dates and definitions were specified in the linkage-protocol matter as people other than the research team typically prepare the linkage database: as specified in the protocol. By focusing on deaths during 2006–10 that were registered with Office for National Statistics (ONS) by March 31, 2012, Asaria and colleagues circumvented the problem that late registration of deaths in England and Wales poses for record-linkage study-teams.7 However, the confidential inquiries that the investigators propose could be delayed because deaths with a specific underlying cause cannot be sampled until registered with the ONS, which, for inquest-deaths, might be months or years after the date of death.8 In Scotland, fact of death must be registered within 8 days of death having been ascertained.7 I lead for the Royal Statistical Scoiety on the need for legislation to end the late registration of deaths in England and Wales.

Introduction Greece entered an austerity programme in 2010 as part of a financial bailout imposed by the European Commission and the European Central Bank in coalition with the International Monetary Fund.1 Its recession has been compared with the USA's great depression of 1929–39, with most of the budget allocated to debt payoff, and contraction of national gross domestic product (GDP).2 Compounded by a shrinking economy, health expenditure in Greece also proportionally shrank from 9·8% of GDP in 2008 to 8·1% in 2014.2, 3, 4 Reduced health spending, required as part of the austerity programme, has been criticised for not containing specific provisions to safeguard the National Health System,4, 5 a system instituted in the 1980s as part of the national programme of compulsory social insurance and through which most residents of Greece receive care. By contrast, health expenditure within the EU rose from 9·4% of GDP in 2008 to 10% in 2014.6

taining specific provisions to safeguard the National Health System,4, 5 a system instituted in the 1980s as part of the national programme of compulsory social insurance and through which most residents of Greece receive care. By contrast, health expenditure within the EU rose from 9·4% of GDP in 2008 to 10% in 2014.6 Following the onset of austerity measures, multiple reports noted adverse health trends, increasing out-of-pocket health expenditure, and unmet health-care needs.7, 8, 9 According to the Hellenic Centre for Disease Control and Prevention, tuberculosis rates have risen among native-born Greeks.8 HIV incidence nearly doubled from 2010 to 2012, reaching 10·4 per 100 000 population and prompting reinstatement of syringe distribution programmes; subsequently, HIV incidence decreased from 2012 to 2016, back to 5·7 per 100 000.10 Increasing rates of major depression and suicidality have been documented,4, 11 along with stagnation in maternal, infant, and child mortality.3, 12, 13 Research in context Evidence before this study

Following the onset of austerity measures, multiple reports noted adverse health trends, increasing out-of-pocket health expenditure, and unmet health-care needs.7, 8, 9 According to the Hellenic Centre for Disease Control and Prevention, tuberculosis rates have risen among native-born Greeks.8 HIV incidence nearly doubled from 2010 to 2012, reaching 10·4 per 100 000 population and prompting reinstatement of syringe distribution programmes; subsequently, HIV incidence decreased from 2012 to 2016, back to 5·7 per 100 000.10 Increasing rates of major depression and suicidality have been documented,4, 11 along with stagnation in maternal, infant, and child mortality.3, 12, 13 Research in context Evidence before this study Following the global financial crisis, Greece entered into an austerity programme in 2010 that has substantially affected health expenditures. Reports from the Hellenic Statistical Authority, the European Centers for Disease Control, and published studies have documented untoward effects on public health following the onset of austerity measures. However, these studies have generally been limited in scope, including being limited to examining only a few causes of illness or mortality, examining only official statistics, and not continuing beyond the first few years of austerity. To better understand the current state of public health in Greece, its trajectory pre-2010 and post-2010, and the broader impacts of financial crisis and austerity on health requires comprehensive evaluation of multiple health-related domains, including all-cause and cause-specific mortality, non-fatal health loss, risk factors, health spending, and demographics.

lic health in Greece, its trajectory pre-2010 and post-2010, and the broader impacts of financial crisis and austerity on health requires comprehensive evaluation of multiple health-related domains, including all-cause and cause-specific mortality, non-fatal health loss, risk factors, health spending, and demographics. Added value of this study

lic health in Greece, its trajectory pre-2010 and post-2010, and the broader impacts of financial crisis and austerity on health requires comprehensive evaluation of multiple health-related domains, including all-cause and cause-specific mortality, non-fatal health loss, risk factors, health spending, and demographics. Added value of this study This study expands on what is known about the effects of the austerity-driven reductions in national health spending on population health in Greece by utilising data from the most comprehensive effort to date to estimate summary measures of global population health—the Global Burden of Diseases, Injuries, and Risk Factors (GBD) initiative. Previous research has focused on the reporting of short-term mortality and morbidity trends that were observed within the first years of the implementation of the austerity programme that limits their sensitivity to detect adverse health outcomes with longer incubation times. In addition to assessing risk factors for the aforementioned health changes, this study investigates the temporal linkage between health financing and health loss until 2015. We did a comprehensive analysis of annual changes in population health in Greece from 2000 to 2016 using regional comparators that allowed us to capture region-specific divergent temporal trends, and we assessed these trends in conjunction with population structure data and health financing measures. The combination of these data sources enabled us to identify evidence of a disproportionate decrement in the health of Greeks as compared with regional populations from 2000 to 2010 (pre-austerity era) to those from 2010 to 2016 (post-austerity era), which was concordant with decreases in national health spending. Further comparison of crude and age-standardised mortality rates indicated accelerated population ageing in Greece compared with Cyprus, with major population shrinkage in the 15–34 years that might be associated to indirect effects of the crisis, such as longstanding unemployment, reduced wages, and stringent taxation schedules.

parison of crude and age-standardised mortality rates indicated accelerated population ageing in Greece compared with Cyprus, with major population shrinkage in the 15–34 years that might be associated to indirect effects of the crisis, such as longstanding unemployment, reduced wages, and stringent taxation schedules. Implications of all the available evidence These findings point to multifaceted effects of the financial crisis in Greece: direct effects of reduced health spending on mortality and to a lesser extent morbidity age and temporal trends, and indirect effects related to a major shift in demographics with population ageing emerging as a public health concern. The increase in total deaths in children younger than 5 years and older adults with increase in causes sensitive to resource availability (eg, access to screening and urgent care) suggest that the health system requires substantial restructuring to cope with the effects that the financial crisis has had on resource availability, resource allocation, and population structure.

ears and older adults with increase in causes sensitive to resource availability (eg, access to screening and urgent care) suggest that the health system requires substantial restructuring to cope with the effects that the financial crisis has had on resource availability, resource allocation, and population structure. Despite the temporal relationship with austerity measures, these reports were criticised for not being supported by national estimate data,14 or being only partly associated with the economic downturn.15 Furthermore, some published communications reported less dramatic health changes in Greece, including actual improvements in cardiovascular mortality.16, 17 The discordance of estimates derived from different data sources, such as between the Hellenic Statistical Authority (ELSTAT) and the EU's statistical office (EUROSTAT), has been further suggested as an obstacle to obtaining pragmatic estimates for temporal changes in health indicators in Greece.18

lar mortality.16, 17 The discordance of estimates derived from different data sources, such as between the Hellenic Statistical Authority (ELSTAT) and the EU's statistical office (EUROSTAT), has been further suggested as an obstacle to obtaining pragmatic estimates for temporal changes in health indicators in Greece.18 As a sovereign nation neighbouring Greece that shares the Greek ethnicity (with a smaller fraction of Turkish Cypriots, Armenians, and Maronites), language, and socioeconomic structure at large, Cyprus represents as close a direct comparator to Greece as exists in western Europe. Although in former years large proportions of the population adhered to a Mediterranean diet, dietary habits are changing rapidly, with increasing obesity, persistent smoking, and alcohol use despite public health efforts to reduce consumption. Similar to Greece, a steep rise in HIV was documented in Cyprus post-2010,19 but the prevalence and trends of other conditions in Cyprus have been much less explored. Cyprus also felt the profound impacts from the recession and entered a financial bailout programme in the 2012–13 period in the aftermath of the Greek financial crisis, although recovery was faster than in Greece.19

s post-2010,19 but the prevalence and trends of other conditions in Cyprus have been much less explored. Cyprus also felt the profound impacts from the recession and entered a financial bailout programme in the 2012–13 period in the aftermath of the Greek financial crisis, although recovery was faster than in Greece.19 Documenting of the effect of diseases on population health is very demanding from the standpoint of both surveillance and analytical methods. The only comprehensive effort to quantify global population health is the Global Burden of Diseases, Injuries, and Risk Factors Study (GBD).20 GBD is therefore the natural platform for comparison between different locations. In this report, we use the GBD 2016 results to explore temporal trends in health loss, risk factors, and health financing in Greece from 2000 to 2016, comparing with those of Cyprus and western Europe overall. Methods Overview and metrics GBD 2016 quantified multiple measures of health loss for 333 causes and 84 risk factors for each of 195 countries and territories, 23 age groups, and both sexes from 1990 to 2016.21 For the present analysis, we compared Greece, Cyprus (estimates refer to the Republic of Cyprus only), and the GBD region of western Europe (Andorra, Austria, Belgium, Cyprus, Denmark, Finland, France, Germany, Greece, Iceland, Ireland, Israel, Italy, Luxembourg, Malta, the Netherlands, Norway, Portugal, and Spain). Our comparisons are primarily for both sexes combined and for wider-ranging age groups than those used in the analysis.

gion of western Europe (Andorra, Austria, Belgium, Cyprus, Denmark, Finland, France, Germany, Greece, Iceland, Ireland, Israel, Italy, Luxembourg, Malta, the Netherlands, Norway, Portugal, and Spain). Our comparisons are primarily for both sexes combined and for wider-ranging age groups than those used in the analysis. We compared rates across locations to control for population size, age standardisation to control for age structure, and comparative rankings and disease rates from more granular age groups to examine specific causes and risk factors more closely. We compared trends using the annualised rate of change (ARC), computed as the final estimates, divided by initial estimates, then divided by number of years, using the natural log transformation of this figure. Detailed methods for the overall study and for each specific cause and risk are available in the GBD 2016 summary publications,21, 22, 23, 24, 25 each of which is GATHER compliant. The next sections are brief descriptions of methods for deriving each of the GBD metrics included in the present analysis; these metrics include deaths, health expenditure, years lived with disability (YLDs), and summary exposure values (SEV) for risk factors. The GBD study's protocol has been approved by the research ethics board at the University of Washington (UW). The GBD study shall be conducted in full compliance with UW policies and procedures, as well as applicable federal, state, and local laws.

We compared rates across locations to control for population size, age standardisation to control for age structure, and comparative rankings and disease rates from more granular age groups to examine specific causes and risk factors more closely. We compared trends using the annualised rate of change (ARC), computed as the final estimates, divided by initial estimates, then divided by number of years, using the natural log transformation of this figure. Detailed methods for the overall study and for each specific cause and risk are available in the GBD 2016 summary publications,21, 22, 23, 24, 25 each of which is GATHER compliant. The next sections are brief descriptions of methods for deriving each of the GBD metrics included in the present analysis; these metrics include deaths, health expenditure, years lived with disability (YLDs), and summary exposure values (SEV) for risk factors. The GBD study's protocol has been approved by the research ethics board at the University of Washington (UW). The GBD study shall be conducted in full compliance with UW policies and procedures, as well as applicable federal, state, and local laws. All-cause mortality and causes of death Mortality estimation methods are described in detail elsewhere.21, 22 Briefly, all-cause mortality was estimated primarily from vital registration data, adjusted for completeness in Greece, Cyprus, and each of the other countries of western Europe, all of which had uninterrupted vital registration time series from pre-1990 to either 2014 or 2015. Spatiotemporal effects and covariates were used to extend estimates to 2016. Under-5 mortality and mortality for ages 15–59 years were estimated separately and linked using empirical life tables to derive age-specific and sex-specific all-cause mortality. The lowest observed risk of death for each age group in total populations of greater than 5 million was summed to construct a global standard life expectancy, which was 86·6 years at birth for GBD 2016.

imated separately and linked using empirical life tables to derive age-specific and sex-specific all-cause mortality. The lowest observed risk of death for each age group in total populations of greater than 5 million was summed to construct a global standard life expectancy, which was 86·6 years at birth for GBD 2016. Cause-specific deaths and years of life lost (YLLs) were estimated for underlying causes of mortality, ensuring that the sum of all specific causes is equal to all-cause mortality. Each death was assigned to a single cause. Vital registration and census death data were adjusted for incompleteness and misclassification (eg, prostate cancer in a female); non-specific and intermediate codes (eg, sepsis, heart failure, unknown causes) were redistributed using age-specific, sex-specific, and geography-specific statistical redistribution methods before modelling. The most commonly used estimation method was with the Cause of Death Ensemble model (CODEm). CODEm uses a train-test-test approach, first testing all combinations of selected country-level covariates and their relation to in-sample data, then ranking those component models based on out-of-sample predictive validity to construct weighted ensembles. All ensemble models are again ranked on a second round of out-of-sample predictive validity to select a final model. Cause-specific fractions were multiplied by all-cause mortality estimates to calculate cause-specific deaths, then scaled with all other causes to match all-cause mortality. YLLs were calculated by multiplying age-specific deaths by global standard life expectancy at age of death.

tive validity to select a final model. Cause-specific fractions were multiplied by all-cause mortality estimates to calculate cause-specific deaths, then scaled with all other causes to match all-cause mortality. YLLs were calculated by multiplying age-specific deaths by global standard life expectancy at age of death. Health expenditure Methods for estimating total health expenditure, government health expenditure, out-of-pocket health expenditure, and prepaid public health expenditure have been described previously.26 Briefly, health spending data were extracted for 1995 to 2014 in national currency units and were divided by GDP reported by WHO, then multiplied by GDP per capita in 2015 purchasing-power-parity-adjusted currency.26 We added a segmented correlation analysis (Spearman's rank correlation coefficient) per financial period of interest (2000–09; 2010–14) to assess the relation between trends in health expenditure and all-cause mortality.27

ed by WHO, then multiplied by GDP per capita in 2015 purchasing-power-parity-adjusted currency.26 We added a segmented correlation analysis (Spearman's rank correlation coefficient) per financial period of interest (2000–09; 2010–14) to assess the relation between trends in health expenditure and all-cause mortality.27 Non-fatal health loss as expressed by YLDs Epidemiological data from systematic literature reviews, health surveys, surveillance systems, disease registries, and hospital and claims databases were used to generate cause-specific and sequela-specific prevalence and incidence estimates using a variety of modelling approaches, of which Bayesian meta-regression compartmental modelling in DisMod-MR 2.1 was the most common.28 Disability weights for each unique health state were derived from population surveys of more than 60 000 respondents completed for GBD 2010 and GBD 2013.29, 30 A microsimulation framework was then used to adjust for comorbidity, and YLDs for each cause were calculated by multiplying prevalence and corresponding disability weights for each sequela of each cause.23

te were derived from population surveys of more than 60 000 respondents completed for GBD 2010 and GBD 2013.29, 30 A microsimulation framework was then used to adjust for comorbidity, and YLDs for each cause were calculated by multiplying prevalence and corresponding disability weights for each sequela of each cause.23 Risk factor estimation and SEVs The GBD 2016 comparative risk assessment (CRA) framework classified each of 84 risk factors and clusters of risk factors into one of three categories: behavioural, environmental and occupational, or metabolic. Data on risk factor exposure levels were identified, evaluated, and modelled using similar approaches to non-fatal models, with added emphasis on accurately fitting distributions of exposure for continuous and polytomous risk factors. Quantitative relative risk was estimated for each risk-outcome pair, and population attributable fraction statistics were calculated using standard GBD CRA methods.25 SEVs represent risk-weighted sums of total exposure in each population, scaled from 0 to 100.

ons of exposure for continuous and polytomous risk factors. Quantitative relative risk was estimated for each risk-outcome pair, and population attributable fraction statistics were calculated using standard GBD CRA methods.25 SEVs represent risk-weighted sums of total exposure in each population, scaled from 0 to 100. Uncertainty estimation Uncertainty for each metric was derived from 1000 draws from the distribution of each estimation step by age, sex, and location for each year included in the GBD 2016 analysis; lower and upper uncertainty intervals (UIs) represent the ordinal 25th and 975th draws of each quantity. UIs allow final estimates to reflect the combined uncertainty of multiple modelling steps. UIs for mortality and YLLs reflect uncertainty in regression coefficients, uncertainty due to sampling and non-sampling error in the cause-of-death data, uncertainty due to various model specifications, and uncertainty in the levels of all-cause mortality. UIs for YLDs reflect uncertainty in prevalence estimates, distribution of severity within each cause, and disability weight valuations. Aggregation of uncertainty across age, sex, and location was done on each draw assuming no correlation.31 Role of the funding source The funders of the study had no role in the study design, data collection, data analysis, data interpretation, writing of the report, or the decision to submit the article for publication. The authors had access to the data in the study and had final responsibility for the decision to submit for publication.

ing source The funders of the study had no role in the study design, data collection, data analysis, data interpretation, writing of the report, or the decision to submit the article for publication. The authors had access to the data in the study and had final responsibility for the decision to submit for publication. Results Mortality and causes of death pre-austerity and post-austerity The all-age, all-cause mortality rate in Greece was 1174·9 (95% UI 1107·4–1243·2) deaths per 100 000 in 2016 compared with 997·8 (975·4–1018) in 2010 and 944·5 (923·1–964·5) in 2000. This finding corresponded to a 0·55% (95% UI·24 to 0·85) ARC from 2000 to 2010, followed by a 2·72% (1·65 to 3·74) annualised increase from 2010 to 2016—a five-times greater ARC post-austerity, with evidence of continuing acceleration (table). The rise in the ARC for all-age mortality was also threefold higher in Greece post-austerity than the 0·86% (95% UI 0·54 to 1·17) rise seen across western Europe for the same period, which was in the opposite direction of the global estimate of a 0·7% (−0·92 to −0·48) fall in all-age mortality from 2010 to 2016. The ARC for all-age mortality in Cyprus increased by 0·65% annually for 2010 to 2016, which was more favourable than regional and global estimates, but was still a retrogression compared with 2000–10 estimates (ARC −1·57% [–1·91 to −1·18]). In terms of age-standardised mortality, Greece's ARC was among the worst-performing countries in western Europe from 2010 to 2016 (appendix), with a reduction from −1·61 (−1·91 to −1·30) from 2000 to 2010 to a marginally negative ARC of −0·87 (−2·03 to 0·20).Table Mortality estimates for Greece, Cyprus, western Europe, and globally in 2000, 2010, and 2016

y, Greece's ARC was among the worst-performing countries in western Europe from 2010 to 2016 (appendix), with a reduction from −1·61 (−1·91 to −1·30) from 2000 to 2010 to a marginally negative ARC of −0·87 (−2·03 to 0·20).Table Mortality estimates for Greece, Cyprus, western Europe, and globally in 2000, 2010, and 2016 Number of deaths (all ages) Rate of death (all ages, per 100 000) Rate of death (age-standardised, per 100 000) Greece 2000 103 470 (101 126–105 663) 944·5 (923·1 to 964·5) 630·6 (616·4 to 643·8) 2010 111 720 (109 215–113 992) 997·8 (975·4 to 1018) 537·1 (525·1 to 548·1) 2016 127 694 (120 356–135 108) 1174·9 (1107·4 to 1243·2) 509·7 (478·9 to 541·4) ARC 2000–10 .. 0·55% (0·24 to 0·85) −1·61 (−1·91 to −1·30) ARC 2010–16 .. 2·72% (1·65 to 3·74) −0·87 (−2·03 to 0·20) Cyprus 2000 6236 (6059–6427) 893·2 (867·8–920·5) 740·3 (718·8 to 762·8) 2010 6431 (6227–6658) 763·6 (739·5 to 790·7) 566·6 (548·5 to 586·6) 2016 7229 (6919–7566) 793·9 (759·8 to 830·9) 538·1 (514·6 to 563·8) ARC 2000–10 .. −1·57% (−1·91 to −1·18) −2·67 (−3·03 to −2·29) ARC 2010–16 .. 0·65% (0·13 to 1·14) −0·86 (−1·4 to −0·36) Western Europe 2000 3 840 157 (3 803 360–3 875 763) 971 (961·7 to 980·0) 624·6 (618·5 to 630·5) 2010 3 876 317 (3 837 211–3 913 200) 929 (919·6 to 937·8) 508·7 (503·4 to 513·8) 2016 4 189 406 (4 104 536–4 272 881) 978·3 (958·5 to 997·8) 475·1 (465·3 to 484·9) ARC 2000–10 .. −0·44% (−0·57 to −0·3) −2·05 (−2·18 to −1·91) ARC 2010–16 .. 0·86% (0·54 to 1·17) −1·14 (−1·48 to −0·81) Global 2000 51 464 936 (51 060 928–51 865 787) 842·2 (835·6 to 848·8) 1111·3 (1103·3 to 1119·5) 2010 53 304 029 (52 825 670–53 789 494) 771·6 (764·6 to 778·6) 928 (920·3 to 935·7) 2016 54 698 580 (54 028 682–55 514 892) 739·9 (730·9 to 751·0) 832·7 (822·7 to 845·0) ARC 2000–10 .. −0·88% (−0·98 to −0·78) −1·80% (−1·89 to −1·71) ARC 2010–16 .. −0·70% (−0·92 to −0·48) −1·81% (−2·02 to −1·59) Ranges are smallest to highest. ARC=annualised rate of change.

) 771·6 (764·6 to 778·6) 928 (920·3 to 935·7) 2016 54 698 580 (54 028 682–55 514 892) 739·9 (730·9 to 751·0) 832·7 (822·7 to 845·0) ARC 2000–10 .. −0·88% (−0·98 to −0·78) −1·80% (−1·89 to −1·71) ARC 2010–16 .. −0·70% (−0·92 to −0·48) −1·81% (−2·02 to −1·59) Ranges are smallest to highest. ARC=annualised rate of change. Mortality trends in Greece were especially unfavourable in adults 15 years or older, with the largest increases seen in those aged 70 years or older (figure 1). Disaggregation of trends in age-specific mortality revealed that, although Greece had slower improvement than Cyprus and western Europe in most age groups from 2000 to 2010, the ARC of age-specific all-cause mortality was similar across all three locations in ages 15 and older from 2010 to 2016, albeit worse from birth to age 14 in Greece (appendix). In parallel, there was evidence of a rapid change in population structure from 2010 to 2016, with a reduction of population in age groups of 15–34 years and increases in population age groups of 55–59 years and 75 years and older (appendix). In comparison with Greece's ageing population, all population age groups over 20 years of age in Cyprus recorded slight increases during the same period (appendix).Figure 1 All-cause mortality by age group from 1990–2016 in Greece, Cyprus, and western Europe

age groups of 55–59 years and 75 years and older (appendix). In comparison with Greece's ageing population, all population age groups over 20 years of age in Cyprus recorded slight increases during the same period (appendix).Figure 1 All-cause mortality by age group from 1990–2016 in Greece, Cyprus, and western Europe As for specific causes of death, adverse effects of medical treatment, self-harm, and several types of cancer stood out as consistently increasing in Greece across all ages (figure 2). Within specific age groups, other causes are apparent, with rapid increases in deaths due to neonatal haemolytic disease and neonatal sepsis in children younger than 5 years, and prominent increases in self-harm among adolescents and young adults. Greek adults aged 15–49 years had increased mortality due to HIV, several treatable neoplasms, all types of cirrhosis, neurological disorders (eg, multiple sclerosis, motor neuron disease), chronic kidney disease, and most types of cardiovascular disease except for ischaemic heart disease and stroke (appendix). This result contrasts with Cyprus, where drug use was the only top ten cause of death that increased in the 15–49 years age groups, and with western Europe, where no causes increased from 2010 to 2016. In adults aged 70 years or older, only a subset of causes of death increased in Cyprus and western Europe, but nearly all increased—and increased more rapidly—in Greece between 2010 and 2016 (appendix).Figure 2 Proportional distribution of causes of death for age 15–49 years (A) and ≥70 years (B) in Greece

010 to 2016. In adults aged 70 years or older, only a subset of causes of death increased in Cyprus and western Europe, but nearly all increased—and increased more rapidly—in Greece between 2010 and 2016 (appendix).Figure 2 Proportional distribution of causes of death for age 15–49 years (A) and ≥70 years (B) in Greece Figure shows all level 3 Global Burden of Diseases, Injuries, and Risk Factors causes of death with colouring of causes that have increased from 2010 to 2016, as deaths per 100 000. IHD=ischaemic heart disease. LRI=lower respiratory infection. TB=tuberculosis. Mech=exposure to mechanical forces. Fire=fire injuries. Hodgkin=Hodgkin's lymphoma. HTN HD=hypertensive heart disease. Other cardio=other cardiovascular disease. Endocar=endocarditis. Nasoph=nasopharyngeal disease. Hemog=haemoglobinopathies and haemolytic anemias. COPD=chronic obstructive pulmonary disease. CKD=chronic kidney disease. Cirr=cirrhosis. Other neuro=other neurological disease. MS=multiple sclerosis. ALS=amyotrophic lateral sclerosis. PUD=peptic ulcer disease. Other unint=other unintentional injuries. Gall bile= gallbladder and biliary diseases.

d haemolytic anemias. COPD=chronic obstructive pulmonary disease. CKD=chronic kidney disease. Cirr=cirrhosis. Other neuro=other neurological disease. MS=multiple sclerosis. ALS=amyotrophic lateral sclerosis. PUD=peptic ulcer disease. Other unint=other unintentional injuries. Gall bile= gallbladder and biliary diseases. Health expenditure Total health expenditure in Greece increased consistently from 2000 to 2008 (the year Greece's austerity programme commenced), at which point total health expenditure was cut by US$130·32 per capita and has since remained in decline (figure 3). When assessing the associations between mortality rates and total health expenditure in Greece, two distinct patterns were observed: during the 2000–08 period, an interrupted increase in total health expenditure was associated with a slow incremental increase in total mortality (Spearman's rs=0·88), whereas from 2009 to 2014, this relationship was reversed (Spearman's rs=–0·99) as the rate of increase in all-cause mortality hastened in the setting of compound decreases in total health expenditure. The same trend was noted for total health expenditure in Cyprus, but the cuts were not as substantial and there was a 1-year lag compared with Greece. Associations between total health expenditure and mortality were also different in Cyprus (appendix), because all-cause death rates decreased continuously until 2012. Examining more granular components of total health expenditure, including government, out-of-pocket, and prepaid public health expenditure revealed broadly similar trends (appendix). Prepaid public health expenditure dropped by 14% from 2008 to 2009, consistent with the onset of austerity measures, but increased to pre-austerity levels thereafter, which could be consistent with the pursuit of private health care due to inefficiencies of the public health system (appendix).Figure 3 Mortality according to total health expenditure in Greece

d by 14% from 2008 to 2009, consistent with the onset of austerity measures, but increased to pre-austerity levels thereafter, which could be consistent with the pursuit of private health care due to inefficiencies of the public health system (appendix).Figure 3 Mortality according to total health expenditure in Greece YLDs pre-austerity and post-austerity Trends in non-fatal health loss were less dramatic than those of mortality. A consistent pattern for causes of YLDs was noted across Greece, Cyprus, and all of western Europe. Low-back and neck pain, migraine, depressive disorders, anxiety disorders, and skin diseases were the top five causes of YLDs in all areas for both 2000 and 2016, whereas oral disorders (caries of permanent teeth and chronic periodontal disease) and tension-type headache were the most prevalent conditions (appendix). HIV increased in both Cyprus and Greece from 2000 to 2016 in adults.

orders, and skin diseases were the top five causes of YLDs in all areas for both 2000 and 2016, whereas oral disorders (caries of permanent teeth and chronic periodontal disease) and tension-type headache were the most prevalent conditions (appendix). HIV increased in both Cyprus and Greece from 2000 to 2016 in adults. Risk factors Age-standardised SEVs for each of the most detailed risk factors, which represent risk-weighted sums of total exposure in 2000, 2010, and 2016 show that Greece has higher exposure to several risks than either Cyprus or western Europe, including smoking, ambient ozone pollution, high body-mass index (BMI), and diet low in omega-3 and polyunsaturated fatty acids. Behavioural and metabolic risk factors accounted for the majority of attributable YLDs in Greece and Cyprus in 2016. The appendix shows the trends in age-standardised YLDs due to these factors from 2000 to 2016. In western Europe, 936·2 (663·2–1283·9) YLDs per 100 000 were attributable to metabolic risks in 2016, which was a slight decrease compared with 2010. The corresponding 2016 estimates for Greece and Cyprus were 892·1 (632·3–1229·2) and 1055·2 (749·7–1436·4) per 100 000 YLDs attributable to metabolic risks, respectively. The burden attributable to behavioural risks recorded a steady fall for western Europe and Cyprus from 2000 to 2016, whereas an opposite trend was observed in Greece, with an overall increase of 2% in YLDs attributable to behavioural risks from 2000 to 2016. Disaggregation of metabolic risks revealed an increase in YLDs attributable mainly to high BMI, high fasting plasma glucose, and high blood pressure in Greece since 2000. The risk attributable to tobacco continuously increased in Greece since 2000, ranking first among behavioural risks with a 2·25% increase in YLD percentage attributable to tobacco from 2010 to 2016. Alcohol consumption and dietary risks ranked second and third, respectively. By contrast, during the same period, YLDs attributable to tobacco reduced by 8·1% in Cyprus and 13·2% in western Europe.