Browse the corpus

†Adjusted for age, sex, year of diagnosis of depression, severity of depression, deprivation, smoking status, alcohol intake, ethnic group (white/not recorded or non-white), coronary heart disease, diabetes, hypertension, cancer, epilepsy/seizures, hypothyroidism, osteoarthritis, asthma/chronic obstructive pulmonary disease, stroke/transient ischaemic attack, rheumatoid arthritis, osteoporosis, liver disease, renal disease, obsessive-compulsive disorder, statins, non-steroidal anti-inflammatory drugs, aspirin, antihypertensives, anticonvulsants, hypnotics/anxiolytics, oral contraceptives, hormone replacement therapy, antipsychotics, bisphosphonates, anticoagulants. ‡Daily doses could not be evaluated for some prescriptions. §Test for trend uses continuous values of dose. We found no significant trends with dose in the three drug classes (table 2). A significant increase in the rate of arrhythmia occurred in the first 28 days after starting treatment with tricyclic and related antidepressants (adjusted hazard ratio 1.99, 1.27 to 3.13; P=0.003), as well as a significant reduction from 84 days after starting selective serotonin reuptake inhibitors (0.78, 0.66 to 0.92; P=0.004).

increase in the rate of arrhythmia occurred in the first 28 days after starting treatment with tricyclic and related antidepressants (adjusted hazard ratio 1.99, 1.27 to 3.13; P=0.003), as well as a significant reduction from 84 days after starting selective serotonin reuptake inhibitors (0.78, 0.66 to 0.92; P=0.004). In the analysis of the 11 most commonly prescribed drugs, we found significant differences between the drugs overall (P=0.004) but no significant difference between the four tricyclic and related antidepressants (P=0.22) or the five selective serotonin reuptake inhibitors (P=0.39), although we saw a significantly decreased risk for fluoxetine (adjusted hazard ratio 0.74, 0.59 to 0.92; P=0.008) and some indication of an increased risk for lofepramine (1.67, 1.01 to 2.76; P=0.05) compared with periods of no antidepressant treatment (fig 1). Fig 1 Adjusted hazard ratios (compared with periods of non-use of antidepressants) for arrhythmia for individual antidepressant drugs over 5 years’ follow-up. SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant

Introduction Change in sexual behaviour is the cornerstone of HIV prevention, yet relatively little research and development has been invested in interventions aimed at behaviour change in any setting.1 School based HIV/AIDS programmes for young people in sub-Saharan Africa have generally not been rigorously evaluated but are often weakly designed, and evaluations suggest they have little impact on sexual behaviours.2 In this respect they are no different from programmes for adolescents in other countries that have rarely shown sustained behaviour change.3 Three randomised controlled trials have been conducted in Africa with both behavioural and other interventions (management of sexually transmitted infection, microfinance, community action, or health service strengthening) in community and school settings.4 5 6 Two studies showed no effectiveness in prevention of sexually transmitted infections,5 6 but one had a positive effect on self reported behaviour.5 The third showed that the intervention was associated with a reduced prevalence of curable sexually transmitted infections. The incidence of herpes simplex type 2 virus (HSV-2) was lower in the group that had only a behavioural intervention but not in the group that had both this and treatment for sexually transmitted infections, and so the authors do not attribute the effect in the behavioural arm to success of their intervention.4 The failure to show a biological impact is a particularly important weakness as there are known limitations to the validity of self reported change in sexual behaviour, with a potential for interventions to bias reporting towards socially desirable behaviours, and because an effect on sexually transmitted infections is the ultimate objective of these interventions.5 7 8

larly important weakness as there are known limitations to the validity of self reported change in sexual behaviour, with a potential for interventions to bias reporting towards socially desirable behaviours, and because an effect on sexually transmitted infections is the ultimate objective of these interventions.5 7 8 None of the intervention programmes previously evaluated was established and widely used before the research was conducted. This is potentially an important weakness as development of interventions is an iterative process, and interventions are generally strengthened by being more extensively tested and adapted.9 In this respect Stepping Stones is a quite different intervention as it has been widely used for many years.10 It was originally developed for use in Uganda in 1995 and has been used in over 40 countries, adapted for 17 settings (including South Africa in 199811), translated into 13 languages, and used with hundreds of thousands of individuals.12 It is almost certainly the most widely used intervention of its kind in the world. Stepping Stones is a participatory HIV prevention programme that aims to improve sexual health through building stronger, more gender equitable relationships. We conducted a trial to assess the impact of Stepping Stones on the incidence of HIV and HSV-2 and sexual practices among men and women in rural areas in the Eastern Cape province of South Africa.

HIV prevention programme that aims to improve sexual health through building stronger, more gender equitable relationships. We conducted a trial to assess the impact of Stepping Stones on the incidence of HIV and HSV-2 and sexual practices among men and women in rural areas in the Eastern Cape province of South Africa. Methods Recruitment and randomisation In this randomised trial we used a cluster design because the intervention is delivered to groups. The setting was historically a subsistence farming region within a radius of 1.5 hours’ drive from the town of Mthatha, where contemporary households are primarily supported by contributions from family working elsewhere, grants, and pensions. The area has two sizeable towns, seven small towns, and many villages. There are 12 hospitals, and most villages have a clinic that distributes free condoms. The unit of randomisation was a geographically defined area in which we recruited one pair of single sex groups. Details of the study design have been described previously.13

two sizeable towns, seven small towns, and many villages. There are 12 hospitals, and most villages have a clinic that distributes free condoms. The unit of randomisation was a geographically defined area in which we recruited one pair of single sex groups. Details of the study design have been described previously.13 The 70 study clusters comprised 64 villages and six townships. Eligible locations were about 10 km from the nearest cluster (to minimise contamination of study arms), had a senior or junior secondary school, and a community willing to participate (established through a process of community mobilisation13). Clusters were grouped into seven strata, with one stratum comprising the townships and six having the villages grouped according to proximity to particular roads. Within each stratum, equal numbers of clusters were allocated to each arm. The study statistician (JL) based in Pretoria, who had no knowledge of the study area, randomly generated the allocation sequence for each stratum. The project manager (MN) and field work coordinators in Mthatha identified and randomised the clusters and then enrolled participants. There was no blinding and for logistical reasons randomisation was done before village recruitment.

ledge of the study area, randomly generated the allocation sequence for each stratum. The project manager (MN) and field work coordinators in Mthatha identified and randomised the clusters and then enrolled participants. There was no blinding and for logistical reasons randomisation was done before village recruitment. In each cluster we recruited about 20 men and 20 women volunteers. Those eligible were aged 16-23, normally resident in the village where they were at school, and mature enough to understand the study and the consent process. There was a difference between the actual and intended age of participants, which is discussed in detail elsewhere.13 Most were recruited from schools. In each cluster recruitment started with general community mobilisation, and the study was explained to key local figures. In most villages the chief (or his representative) called a monthly community meeting. Typically the staff member attended and made a brief presentation and then took questions from the community, including from parents of potential participants. After the community meeting project staff went to the school to raise interest in the study and invited possible participants to a meeting. Here they explained the study to a group of about 60 young men and women in the targeted age group. Names were taken and the group was asked to decide on the 40 people who were most likely to be able to participate in the study. The presenter read aloud the study’s consent form to the 40 and gave an opportunity for questions. The form explained the procedures that would occur in some detail. After the group presentation we asked for confirmation that there was still general interest in participation and asked the young people to talk with their families before committing themselves. Each potential participant received a Xhosa language leaflet describing the study in terms understandable to a lay audience. Those who decided to participate were asked to report at an assigned time anywhere from two to seven days later. At that time, they provided and signed formal informed consent forms and study recruitment was finalised.

ant received a Xhosa language leaflet describing the study in terms understandable to a lay audience. Those who decided to participate were asked to report at an assigned time anywhere from two to seven days later. At that time, they provided and signed formal informed consent forms and study recruitment was finalised. We used the method of Hayes and Bennett14 to calculate the sample size—that is, the number of clusters required in each arm. The calculation assumed that the effect (as measured by incidence rate ratio) would be homogeneous for men and women and assumed a two year cumulative HIV incidence rate (averaged over men and women) in the control arm of 12% and that two year incidence results would be obtained for at least 14 women and 14 men per cluster (thus allowing up to 30% of women and 30% of men to be either lost to follow-up or to be HIV positive at baseline). To calculate the sample size we needed an estimate of k, the coefficient of variation between clusters for the outcome measure; we used k=0.35 on the basis of an analysis of the results for the Eastern Cape of the 1999 national antenatal HIV seroprevalence survey, in which the clusters were antenatal clinics. A sample size of 35 clusters per trial arm would then give more than 80% power to detect as significant at the 5% level a 50% reduction in HIV incidence.

n the basis of an analysis of the results for the Eastern Cape of the 1999 national antenatal HIV seroprevalence survey, in which the clusters were antenatal clinics. A sample size of 35 clusters per trial arm would then give more than 80% power to detect as significant at the 5% level a 50% reduction in HIV incidence. Intervention and implementation We compared the impact of the South African Stepping Stones (second edition)15 and the control intervention for groups of men and women on incident infections of HIV and HSV-2 and sexual behaviours. Our study was an effectiveness trial, rather than an efficacy trial, with the programme implemented as if in a broader community roll-out.13 The interventions were facilitated by project staff, who were employed by our partner non-governmental organisation the Planned Parenthood Association of South Africa (PPASA), and trained, supervised, and shown how to implement the programmein accordance with its practices. Facilitators were the same sex as the participants and either the same age or a little older. Most had further education or had undergone life skills training and were selected, in part, for their open mindedness and gendersensitivity. After three weeks of training and two practice groups, 11 facilitators delivered the Stepping Stones intervention. Another four, who were trained for four days, administered the control intervention. These two groups of facilitators were trained and supervised separately to reduce contamination.

and gendersensitivity. After three weeks of training and two practice groups, 11 facilitators delivered the Stepping Stones intervention. Another four, who were trained for four days, administered the control intervention. These two groups of facilitators were trained and supervised separately to reduce contamination. Stepping Stones uses participatory learning approaches, including critical reflection, roleplay, and drama and draws the everyday reality of participants’ lives into the sessions. It is delivered to single sex groups, which are run in parallel, and has 13 three hour long sessions that are complemented by three meetings of male and female peer groups and a final community meeting. The programme spanned about 50 hours and ran for six to eight weeks. The sessions covered how we act and what shapes our actions; sex and love; conception and contraception; taking risks and sexual problems; unwanted pregnancy; sexually transmitted diseases and HIV; safer sex and condoms; gender based violence; motivations for sexual behaviour; dealing with grief and loss; and communication skills. The sessions were mainly held on school premises after school hours.

tion and contraception; taking risks and sexual problems; unwanted pregnancy; sexually transmitted diseases and HIV; safer sex and condoms; gender based violence; motivations for sexual behaviour; dealing with grief and loss; and communication skills. The sessions were mainly held on school premises after school hours. Our South African adaptation has a slightly different content from the Welbourn original and was not used in a community development context. Welbourn recommended working with older men and women in each community as well as young people and suggested that peer groups be encouraged to continue to meet after the end of the workshops.10 We did not implement either of these components as it would have added greatly to the cost and we wanted to test a delivery model that we thought could be more easily funded for roll-out. The control intervention was a single three hour session on HIV, safer sex, and condoms. The content was taken from Stepping Stones.

We did not implement either of these components as it would have added greatly to the cost and we wanted to test a delivery model that we thought could be more easily funded for roll-out. The control intervention was a single three hour session on HIV, safer sex, and condoms. The content was taken from Stepping Stones. We administered questionnaires and collected blood samples before the intervention (baseline) and after about one and two years. The baseline interviews and intervention were staggered over a 12 month period (March 2003-March 2004), as was each round of follow-up. Participants were located for repeat interviews by using details collected at enrolment. If they had moved within the study area, they were interviewed in their new location or invited to come to the office. We also went to Cape Town, East London, and Gauteng province to conduct interviews with migrants. All participants were given 20 rand (about £1.30, €1.60, $2.50) after each interview.

ted at enrolment. If they had moved within the study area, they were interviewed in their new location or invited to come to the office. We also went to Cape Town, East London, and Gauteng province to conduct interviews with migrants. All participants were given 20 rand (about £1.30, €1.60, $2.50) after each interview. We had an active community advisory board and data safety and monitoring board. After the group discussions, participants were asked to sign an informed consent form to participate on the day of the interview. The consent form was in two parts with consent to participate in the trial separated from consent for the blood tests. A trained nurse counsellor provided counselling before HIV testing to groups of eight to 10 people after they had enrolled in the study, signed consent for the interview, and completed the baseline questionnaire. Counselling typically involved five minutes of information provided by the nurse followed by 20 minutes of questions. Afterwards participants signed consent for the HIV test (they could raise issues privately then if they wanted) and they were asked whether they wanted to be told their results. If so, a study nurse gave them test results with counselling some weeks later. Participants could change their mind and get their results at any stage. Those with positive results were told their CD4 counts and screened for medical problems. They were also referred to local health services and HIV support groups according to a referral algorithm that took into account their clinical condition and the available community services as well as locally accepted standards of care. The Medical Research Council paid for lunch, transport, and consultation fees for HIV positive participants accessing health services. The study nurses supported participants with social problems and HIV related problems throughout the course of the study, referring them to social workers or health facilities as appropriate. During the study anti-retroviral drugs became available in the public sector and at this point the consent form was changed to ask participants who had opted not to collect their result if they would like to be told if they tested positive.

he study, referring them to social workers or health facilities as appropriate. During the study anti-retroviral drugs became available in the public sector and at this point the consent form was changed to ask participants who had opted not to collect their result if they would like to be told if they tested positive. Laboratory methods The primary outcome measure was HIV incidence, determined through blood tests at baseline and at 12 and 24 months. All blood tests were conducted blind to the treatment arm. HIV status was assessed with two rapid tests by using the World Health Organization’s testing algorithm.16 We used the Determine (Abbott Diagnostics, Johannesburg) screening test and retested samples with positive results with Uni-gold (Trinity Biotech, Dublin, Ireland). We carried out an HIV-1 enzyme linked immunosorbent assay (ELISA) (Genscreen) followed by two confirmatory ELISAs (Vironostika and Murex 1.2.0) if the sample was positive for HIV to clarify any indeterminate results. Towards the end of the second round of interviews collection of dried blood spots was introduced (in 357 cases) as it was easier logistically and more acceptable for participants. In the third round of interviews most blood was collected as dried blood spots (n=1530). These were tested with a screen ELISA (Genscreen) and positive results were confirmed with a second ELISA (Vironostika). Participants were not given a choice of the method of blood collection, which was chosen purely on fieldwork logistics. There was no difference between arms in the use of dried blood spots at 24 months but there was a small difference (7%) at 12 months, with more in the control arm. The dried blood spot method has been used extensively in South African populations over the past few years but has not been specifically validated for South Africa. The HIV tests on the dried blood spots were optimised by use of paired serum samples and dried blood. The National Institute for Communicable Diseases participates in a programme supported by the Center for Communicable Diseases Control that has shown the methods used can optimally identify HIV-1 from dried blood spots.

he HIV tests on the dried blood spots were optimised by use of paired serum samples and dried blood. The National Institute for Communicable Diseases participates in a programme supported by the Center for Communicable Diseases Control that has shown the methods used can optimally identify HIV-1 from dried blood spots. We used two glycoprotein G based HSV-2 ELISAs to test for herpes infection, Kalon (Kalon Biological, Aldershot, UK) and HerpeSelect Immunoblot IgG (Focus Technologies, Cypress, Ca, USA). We used an additional test, CAPTIA herpes simplex virus (HSV) IgG type specific ELISAs to resolve discrepant results. The testing for HSV-2 on dried blood spots was optimised with paired serum samples and dried blood spots as described in Hofgrefe et al.17 The CAPTIA HSV-2 assay had not been validated for use as a confirmatory assay at the time of the study. We assessed the impact of the intervention on behaviour and attitudes with a questionnaire administered in Xhosa. Table 1 describes the outcome measures, indicators, and assessment and further details can be found elsewhere.13 Table 1  Outcome measures in Stepping Stones programme for HIV prevention

We used two glycoprotein G based HSV-2 ELISAs to test for herpes infection, Kalon (Kalon Biological, Aldershot, UK) and HerpeSelect Immunoblot IgG (Focus Technologies, Cypress, Ca, USA). We used an additional test, CAPTIA herpes simplex virus (HSV) IgG type specific ELISAs to resolve discrepant results. The testing for HSV-2 on dried blood spots was optimised with paired serum samples and dried blood spots as described in Hofgrefe et al.17 The CAPTIA HSV-2 assay had not been validated for use as a confirmatory assay at the time of the study. We assessed the impact of the intervention on behaviour and attitudes with a questionnaire administered in Xhosa. Table 1 describes the outcome measures, indicators, and assessment and further details can be found elsewhere.13 Table 1  Outcome measures in Stepping Stones programme for HIV prevention Indicator No of items* Expected direction of change because of intervention Primary outcome Incidence of HIV HIV seroconversion in individuals who were HIV negative at baseline NA Decrease Other outcomes Incidence of HSV-2 HSV-2 seroconversion in individuals who were HSV-2 negative at baseline NA Decrease No of partners No of main partners, one-off partners, and ongoing non-primary partners (makhwapheni) since last interview 3 Decrease Any transactional sex with casual partner Sex primarily motivated by material gain to female casual partner since last interview, defined as provision of food, cosmetics, clothes, transportation, items for children or family, school fees, somewhere to sleep, or cash. Giving for men and receiving for women 17 Decrease >1 incident of physical or sexual intimate partner violence More than one episode of physical or sexual intimate partner violence since last interview 9 Decrease Rape or attempted rape Rape or attempted rape of non-intimate partner since last interview 5 Decrease Correct condom use at last sex Use of condom at last sexual intercourse for each coital act without user error or breakage 6 Increase Any casual partner Any one-off partner or ongoing non-primary partner (makhwapheni) since last interview 2 Decrease Unwanted pregnancy Pregnancy since last interview with indication that at the time she became pregnant she wanted to become pregnant “later” or “not at all” 2 Decrease Depression Depression in past week measured with CES-D (cut-off point ≥21) 20 Decrease Problem drinking Problem drinking measured with AUDIT scale (cut-off point ≥9) 11 Decrease Ever misused drugs Misuse of cannabis, benzene, mandrax, injecting drugs, or other drug 5 Decrease *No of questions used to assess each indicator. NA=not applicable.

n past week measured with CES-D (cut-off point ≥21) 20 Decrease Problem drinking Problem drinking measured with AUDIT scale (cut-off point ≥9) 11 Decrease Ever misused drugs Misuse of cannabis, benzene, mandrax, injecting drugs, or other drug 5 Decrease *No of questions used to assess each indicator. NA=not applicable. Data analysis We followed an intention to treat approach, in which we included in the analysis all participants with evaluable data for the outcome measure under consideration. We stratified the analyses for incidence of HIV and HSV-2 by sex and carried out a test of homogeneity of treatment effect over the sexes. All other analyses were carried out separately for men and women. Participants were included in the analysis of the primary outcome only if they were HIV negative at baseline; those who had missing HIV results at baseline were excluded even if they had a subsequent negative result as they could not have been included if they had tested positive at the subsequent visit. For each participant we calculated the person years of exposure as the time from baseline to the last negative result if the person remained negative, or as the total time between any negative tests as well as half the time between the last negative and first positive tests. The primary analysis was carried out by fitting generalised linear mixed models (GLMMs) as advocated by Murray.18 The GLMM used a log link and assumed an underlying Poisson distribution and included terms for stratum, sex, age of respondent, baseline prevalence of HIV and HSV-2of the cluster for men and women, and treatment, with clusters being treated as a random effect. Homogeneity of the treatment effect over men and women was established by testing for a sex by treatment interaction. Generalised estimating equation (GEE) models were also fitted to test the robustness of the GLMMs. In addition we carried out cluster level analyses. Firstly, we calculated the cluster level incidence rate for each cluster, separately for men and women. These rates were compared between treatment arms by fitting a general linear model to the 140 cluster level rates (70 for men and 70 for women) with terms for the baseline cluster level prevalence, stratum, sex, and treatment arm. The results of fitting these models were used to estimate the number of HIV or HSV-2 infections prevented by the intervention over a two year period per 1000 participants.

to the 140 cluster level rates (70 for men and 70 for women) with terms for the baseline cluster level prevalence, stratum, sex, and treatment arm. The results of fitting these models were used to estimate the number of HIV or HSV-2 infections prevented by the intervention over a two year period per 1000 participants. In addition, we fitted a cluster level Poisson model using the number of events and the total person years of exposure for each cluster, with terms for cluster prevalence at baseline (separate for men and women), stratum, sex, and treatment arm. These results were used as a check on the results obtained from the GLMMs. In all cases, apart from reporting the number of infections prevented, the results presented are those from the GLMMs.

for each cluster, with terms for cluster prevalence at baseline (separate for men and women), stratum, sex, and treatment arm. These results were used as a check on the results obtained from the GLMMs. In all cases, apart from reporting the number of infections prevented, the results presented are those from the GLMMs. We analysed other outcomes separately for men and women and for the 12 and 24 month visits. “Correct condom use on last sex” was analysed as a binary outcome with a GLMM with a logit link and underlying binomial distribution (that is, a random effects logistic regression model). The model contained terms for stratum, age of the respondent, and treatment arm. Any casual partner since the last visit, transactional sex with a casual partner since the last visit (giving for men and receiving for women), more than one incident of physical or sexual abuse since the last visit (perpetration for men and receipt for women), unwanted pregnancy since then, and any rape or attempted rape against a non-partner since then (men only) were also treated as binary outcomes. As the likelihood of these events happening might be higher with longer periods between interviews, we included the time between interviews as a covariate. Thus for each outcome we fitted random effects logistic regression models with terms for stratum, age of the respondent, time since the last visit, and treatment arm. The number of sexual partners since the last interview was analysed by first applying a square root transformation to this outcome as this was found to be variance stabilising and to lead to approximately normally distributed residuals in an analysis ignoring the clustering. A mixed model was then fitted to the transformed outcome with terms for stratum, age of the respondent, time since the last visit, and treatment arm. The effects were back transformed for easier interpretation. We analysed depression (CES-D scale), problem drinking (AUDIT scale), and drug misuse using models similar to those for correct condom use. In all cases GEE models were also fitted to confirm the results of the GLMMs. In the case of drug misuse at month 24 for women (that is, having started drug misuse between the 12 month and 24 month interview) the results presented are those for the GEE model as the GLMM failed to converge because of the small number of women misusing drugs. In all other cases the results reported are those from the GLMM.

se of drug misuse at month 24 for women (that is, having started drug misuse between the 12 month and 24 month interview) the results presented are those for the GEE model as the GLMM failed to converge because of the small number of women misusing drugs. In all other cases the results reported are those from the GLMM. Results The figure shows the trial profile. No clusters were lost to follow-up. Twelve month follow-up rates for women with known HIV status at baseline were 75.8% and 75.3% in the intervention and control arms and 75.1% and 71.8% for men in the intervention and control arms, respectively. At 24 months, 73.1% (intervention) and 76.0% (control) of women with baseline HIV results were retested and 69.5% (intervention) and 69.2% (control) of men were tested again for HIV. Loss to follow-up was mainly because participants had moved and could not be located. At baseline, 9.8% of men and 6.3% of women had a main partner also in the study. Fig 1 Trial profile Eighteen participants died during the main study and one committed suicide in the pilot study (figure). Causes of death in the main study were interpersonal violence (six), suicide (three), injuries from traffic incidents (two), and a range of natural causes (seven), including AIDS (one). Four of the non-natural deaths were in the control arm and seven in the intervention arm. All deaths were investigated and none was linked to activities of the study. There were no other serious adverse events.

(three), injuries from traffic incidents (two), and a range of natural causes (seven), including AIDS (one). Four of the non-natural deaths were in the control arm and seven in the intervention arm. All deaths were investigated and none was linked to activities of the study. There were no other serious adverse events. From the available attendance registers (an incomplete set), 90 (16.8%) men and 63 (12.5%) women did not participate in any of the Stepping Stones sessions, and 189 (31.7%) men and 228 (35.7%) women did not attend the short intervention. Some 324 (60.7%) men and 298 (59.1%) women attended 75% or more of the Stepping Stones sessions, and 147 (27.5%) men and 128 (25.4%) women attended the complete programme. Table 2 shows the participants’ baseline characteristics. The two arms were similar for both sexes, although participants in the control arm were slightly more educated (P=0.09 for women, P=0.08 for men). Table 2  Social and demographic characteristics of two study arms. Figures are numbers (percentages)

From the available attendance registers (an incomplete set), 90 (16.8%) men and 63 (12.5%) women did not participate in any of the Stepping Stones sessions, and 189 (31.7%) men and 228 (35.7%) women did not attend the short intervention. Some 324 (60.7%) men and 298 (59.1%) women attended 75% or more of the Stepping Stones sessions, and 147 (27.5%) men and 128 (25.4%) women attended the complete programme. Table 2 shows the participants’ baseline characteristics. The two arms were similar for both sexes, although participants in the control arm were slightly more educated (P=0.09 for women, P=0.08 for men). Table 2  Social and demographic characteristics of two study arms. Figures are numbers (percentages) Women Men Intervention (n=715) Control (n=701) Intervention (n=694) Control (n=666) Age (years): 15-17 329 (46.0) 289 (41.2) 234 (33.7) 182 (27.3) 18-19 261 (36.5) 250 (35.7) 259 (37.3) 264 (39.8) 20-21 102 (14.3) 128 (18.3) 160 (23.1) 164 (24.6) 22-26 23 (3.2) 34 (4.8) 41 (5.9) 56 (8.4) Education grade: <9 67 (9.4) 40 (5.7) 112 (16.1) 79 (11.9) 9 343 (48.0) 248 (35.4) 335 (48.3) 243 (36.5) 10 243 (34.0) 279 (39.8) 188 (27.1) 229 (34.4) 11 58 (8.1) 121 (17.3) 55 (7.9) 105 (15.8) >11 4 (0.6) 13 (1.8) 4 (0.6) 10 (1.5) Current schooling 707 (98.9) 678 (96.7) 680 (98.0) 640 (96.2) HIV seroprevalence 70 (9.8) 90 (12.8) 12 (1.7) 14 (2.1) HSV-2 seroprevalence 194 (27.6) 213 (31.0) 70(10.3) 65 (10.0) Ever had sexual intercourse 655 (91.6) 633 (90.3) 654 (94.2) 624 (93.7) Correct use of condom on last sexual intercourse 266 (40.7) 288 (45.6) 292 (44.6) 303 (48.6) >2 sexual partners in past year 62 (9.5) 71 (11.2) 317 (48.5) 314 (50.3) Casual partner in past 12 months 138 (21.1) 146 (23.1) 379 (58.0) 378 (60.6) Ever had transactional sex 177 (27.1) 141 (22.3) 200 (30.6) 168 (26.9) More than one incident of physical/sexual intimate partner violence 177 (24.7) 157 (22.4) 100 (14.5) 96 (14.5) Rape or attempted rape — — 126 (18.2) 141 (21.3) Depression (CES-D) 117 (16.4) 103 (14.7) 45 (6.5) 54 (8.1) Problem drinking 28 (3.9) 19 (2.7) 170 (24.5) 171 (25.7) Ever misused drugs 37 (5.2) 53 (7.6) 256 (37.1) 262 (39.5) Table 3 shows the results for the comparison of incidence rates of HIV and HSV-2 between the two study arms. After adjustment for stratum, baseline HIV prevalence in the cluster, and age of the respondent, Stepping Stones had little effect on the incidence of HIV. The incidence of HSV-2 was significantly lower in the Stepping Stones arm than the control arm (incidence rate ratio 0.67, 95% confidence interval 0.46 to 0.97, P=0.036). This represents a 33% reduction in incidence and translates to 34.9 (1.6 to 68.2) infections being prevented over a two year period per 1000 people in the programme. There was no evidence of heterogeneity—that is, the effect of Stepping Stones on incidence of HSV-2 was similar for men and women.

0.46 to 0.97, P=0.036). This represents a 33% reduction in incidence and translates to 34.9 (1.6 to 68.2) infections being prevented over a two year period per 1000 people in the programme. There was no evidence of heterogeneity—that is, the effect of Stepping Stones on incidence of HSV-2 was similar for men and women. Table 3  Incidence of HIV and HSV-2 according to intervention Stepping Stones Control P value for homogeneity Adjusted* incidence rate ratio (95% CI) P value Coefficient of variation No of events Rate/100 person years No of events Rate/100 person years HIV Overall 72 3.46 81 4.07 0.56 0.95 (0.67 to 1.35) 0.78 1.02 Women 57 5.65 68 6.95 0.81 Men 15 1.40 13 1.29 1.60 HSV-2 Overall 57 3.24 75 4.62 0.91 0.67 (0.47 to 0.97) 0.036 1.13 Women 43 5.35 57 7.71 0.93 Men 14 1.46 18 2.04 1.58 *Adjusted for stratum, sex, participant’s age, and baseline cluster prevalence of HIV or HSV-2, respectively.

4.07 0.56 0.95 (0.67 to 1.35) 0.78 1.02 Women 57 5.65 68 6.95 0.81 Men 15 1.40 13 1.29 1.60 HSV-2 Overall 57 3.24 75 4.62 0.91 0.67 (0.47 to 0.97) 0.036 1.13 Women 43 5.35 57 7.71 0.93 Men 14 1.46 18 2.04 1.58 *Adjusted for stratum, sex, participant’s age, and baseline cluster prevalence of HIV or HSV-2, respectively. Table 4 shows the results of the analysis of the other outcomes for women. There was no evidence of difference in the expected direction between the two arms in any of these outcomes. At 12 months the proportion of women who had transactional sex with a casual partner since the first interview was higher in the Stepping Stones arm. It is worth noting, however, that there was little difference between the two arms in the proportions of women who had a casual partner and that the difference in the proportions having transactional sex with a casual partner had disappeared by month 24. There was slight evidence (P=0.11) that the incidence of pregnancy was higher in the Stepping Stones arm at 24 months. Table 4  Other outcomes at 12 and 24 months in women according to intervention

Table 4 shows the results of the analysis of the other outcomes for women. There was no evidence of difference in the expected direction between the two arms in any of these outcomes. At 12 months the proportion of women who had transactional sex with a casual partner since the first interview was higher in the Stepping Stones arm. It is worth noting, however, that there was little difference between the two arms in the proportions of women who had a casual partner and that the difference in the proportions having transactional sex with a casual partner had disappeared by month 24. There was slight evidence (P=0.11) that the incidence of pregnancy was higher in the Stepping Stones arm at 24 months. Table 4  Other outcomes at 12 and 24 months in women according to intervention Stepping Stones Control Effect* or adjusted odds ratio (95% CI) P value Coefficient of variation No of participants Mean* or proportion No of participants Mean* or proportion No of partners in past year: 12 months 558 1.32 546 1.32 0.0001 (−0.0024 to 0.0049) 0.74 0.18 24 months 536 1.19 547 1.19 0.0001 (−0.0012 to 0.0025) 0.73 0.12 Any transactional sex with a casual partner: 12 months 559 0.072 550 0.031 2.55 (1.19 to 5.46) 0.016 1.52 24 months 537 0.020 547 0.022 0.94 (0.41 to 2.18) 0.89 1.62 >1 incident of physical or sexual intimate partner violence: 12 months 559 0.184 550 0.207 0.87 (0.64 to 1.18) 0.36 0.55 24 months 537 0.147 547 0.135 1.14 (0.77 to 1.68) 0.51 0.72 Pregnancy: 12 months 537 0.145 518 0.129 1.14 (0.78 to 1.67) 0.49 0.70 24 months 521 0.144 534 0.116 1.45 (0.92 to 2.28) 0.11 0.82 Any casual partner: 12 months 553 0.228 544 0.204 1.14 (0.82 to 1.59) 0.42 0.61 24 months 534 0.183 547 0.163 1.17 (0.85 to 1.63) 0.34 0.45 Correct condom use at last sex: 12 months 533 0.552 516 0.558 0.96 (0.72 to 1.28) 0.79 0.29 24 months 520 0.575 534 0.596 0.90 (0.70 to 1.17) 0.45 0.24 Depression: 12 months 559 0.154 550 0.122 1.32 (0.92 to 1.89) 0.13 0.69 24 months 537 0.128 547 0.161 0.76 (0.51 to 1.15) 0.20 0.74 Problem drinking: 12 months 559 0.041 550 0.042 0.94 (0.45 to 1.95) 0.87 1.47 24 months 537 0.034 547 0.022 1.40 (0.61 to 3.17) 0.43 1.80 Ever misused drugs: 12 months 530 0.024 513 0.039 0.60 (0.29 to 1.28) 0.19 1.17 24 months 437 0.023 423 0.019 1.20 (0.51 to 2.83) 0.68 1.29 *For No of partners in past year only.

.74 Problem drinking: 12 months 559 0.041 550 0.042 0.94 (0.45 to 1.95) 0.87 1.47 24 months 537 0.034 547 0.022 1.40 (0.61 to 3.17) 0.43 1.80 Ever misused drugs: 12 months 530 0.024 513 0.039 0.60 (0.29 to 1.28) 0.19 1.17 24 months 437 0.023 423 0.019 1.20 (0.51 to 2.83) 0.68 1.29 *For No of partners in past year only. Some of the other outcomes for men did show differences in the hypothesised direction (table 5). A significantly lower proportion of men in the intervention arm reported having had transactional sex with a casual partner at 12 months, although this difference had disappeared by 24 months. The proportion of men who perpetrated physical or sexual intimate partner violence was significantly lower in the Stepping Stones arm at 24 months, and there was some evidence that it was also lower at 12 months. There was some evidence that a lower proportion of men in the Stepping Stones reported raping or attempting rape at 12 months and that a lower proportion had any casual partner at 12 months. A significantly lower proportion of men in the Stepping Stones arm reported problem drinking at 12 months, and there was some evidence that a lower proportion were depressed at 24 months and that a lower proportion initiated drug misuse between 12 and 24 months. Table 5  Results for men: other outcomes at 12 and 24 months

Some of the other outcomes for men did show differences in the hypothesised direction (table 5). A significantly lower proportion of men in the intervention arm reported having had transactional sex with a casual partner at 12 months, although this difference had disappeared by 24 months. The proportion of men who perpetrated physical or sexual intimate partner violence was significantly lower in the Stepping Stones arm at 24 months, and there was some evidence that it was also lower at 12 months. There was some evidence that a lower proportion of men in the Stepping Stones reported raping or attempting rape at 12 months and that a lower proportion had any casual partner at 12 months. A significantly lower proportion of men in the Stepping Stones arm reported problem drinking at 12 months, and there was some evidence that a lower proportion were depressed at 24 months and that a lower proportion initiated drug misuse between 12 and 24 months. Table 5  Results for men: other outcomes at 12 and 24 months Stepping Stones Control Effect* or adjusted odds ratio (95% CI) P value Coefficient of variation No of participants Mean* or proportion No of participants Mean* or proportion No of partners in past year: 12 months 531 2.28 500 2.51 −0.0078 (−0.033 to 0.0001) 0.063 0.24 24 months 501 2.15 474 2.39 −0.0045 (−0.023 to 0.0003) 0.12 0.25 Any transactional sex with a casual partner: 12 months 534 0.036 505 0.075 0.39 (0.17 to 0.92) 0.031 1.72 24 months 504 0.018 479 0.019 1.02 (0.39 to 2.65) 0.97 2.09 >1 incident of physical or sexual intimate partner violence: 12 months 534 0.114 505 0.149 0.73 (0.50 to 1.06 ) 0.099 0.63 24 months 504 0.062 479 0.096 0.62 (0.38 to 1.01) 0.054 0.67 Rape or attempted rape: 12 months 534 0.092 505 0.123 0.71 (0.47 to 1.06) 0.094 1.86 24 months 501 0.080 473 0.085 0.92 (0.53 to 1.58) 0.76 1.91 Impregnated any woman: 12 months 534 0.086 505 0.083 1.03 (0.66 to 1.62) 0.89 0.82 24 months 504 0.113 479 0.129 0.88 (0.60 to 1.31) 0.53 0.76 Correct condom use at last sex: 12 months 513 0.735 483 0.689 1.26 (0.92 to 1.74) 0.16 0.22 24 months 485 0.732 462 0.751 0.88 (0.64 to 1.21) 0.43 0.20 Any casual partner: 12 months 530 0.549 499 0.601 0.79 (0.59 to 1.05) 0.098 0.28 24 months 501 0.531 473 0.569 0.85 (0.62 to 1.15) 0.29 0.33 Depression: 12 months 534 0.022 505 0.046 0.45 (0.16 to 1.21) 0.11 2.05 24 months 504 0.028 479 0.050 0.52 (0.24 to 1.13) 0.097 1.49 Problem drinking: 12 months 534 0.198 505 0.265 0.68 (0.49 to 0.94) 0.021 0.57 24 months 504 0.266 479 0.257 1.10 (0.81 to 1.49) 0.56 0.52 Ever misused drugs: 12 months 337 0.16 310 0.16 1.07 (0.65 to 1.77) 0.78 0.49 24 months 232 0.065 215 0.12 0.50 (0.23 to 1.11) 0.088 0.47 *For No of partners in past year only.

1.49 Problem drinking: 12 months 534 0.198 505 0.265 0.68 (0.49 to 0.94) 0.021 0.57 24 months 504 0.266 479 0.257 1.10 (0.81 to 1.49) 0.56 0.52 Ever misused drugs: 12 months 337 0.16 310 0.16 1.07 (0.65 to 1.77) 0.78 0.49 24 months 232 0.065 215 0.12 0.50 (0.23 to 1.11) 0.088 0.47 *For No of partners in past year only. The aggregated cluster level analyses produced estimates and confidence intervals that were similar to those from the individual level analyses (results not shown). The per protocol analysis produced estimates that were similar to the intention to treat analysis, so the results are not shown. Discussion Participation in the Stepping Stones programme in South Africa did not reduce the incidence of HIV infection among young men and women aged 15/26 but was associated with a reduced incidence of herpes simplex type 2 (HSV-2). There was no evidence of any desired behaviour change in women. There was more transactional sex with a casual partner at 12 months (but not at 24 months) among women in the Stepping Stones arm, and there was a suggestion of more unwanted pregnancies at 24 months. Men in Stepping Stones reported less transactional sex at 12 months, less perpetration of intimate partner violence (significant at 24 months, suggested at 12 months), less problem drinking at 12 months, and less drug misuse at 24 months. There was a suggestion of change in several other outcomes in men, including fewer partners at 12 months, less likelihood of casual partners, less rape at 12 months, and less depression at 24 months.

ignificant at 24 months, suggested at 12 months), less problem drinking at 12 months, and less drug misuse at 24 months. There was a suggestion of change in several other outcomes in men, including fewer partners at 12 months, less likelihood of casual partners, less rape at 12 months, and less depression at 24 months. Strengths This was a randomised controlled trial had two biological outcomes. Few randomised controlled trials in Africa have evaluated behavioural interventions with biological outcomes and none has found clear evidence of effect. Our finding of an impact on HSV-2 infection (although not on HIV) is important for HIV prevention as Stepping Stones is a widely used intervention and HSV-2 is an important cofactor in heterosexual transmission of HIV. Meta-analysis indicates that people infected with HSV-2 have three times the risk of HIV infection.19

Our finding of an impact on HSV-2 infection (although not on HIV) is important for HIV prevention as Stepping Stones is a widely used intervention and HSV-2 is an important cofactor in heterosexual transmission of HIV. Meta-analysis indicates that people infected with HSV-2 have three times the risk of HIV infection.19 The impact of the intervention on incident infections of HSV-2 in women suggests that desirable behaviour change occurred in at least some women. A possible explanation for the lack of demonstrated impact on women’s behavioural outcomes is differential reporting bias—that is, under-reporting of sexual activity at baseline—with those who went through Stepping Stones becoming more forthright. This is a recognised problem with self reported behavioural outcomes.5 Alternatively Stepping Stones might have influenced unmeasured behaviour changes or choices of partners that protected against HSV-2 in previously unexposed women. Other authors have reflected that the ability of women to change their sexual behaviour in the context of unequal gender power relations is less than that of men.5 Young women are particularly at risk of being infected with HIV by older men (who have a higher age specific prevalence than younger men),20 and in these relationships the age differential further reduces women’s power. The prevalence of herpes is much higher in young men than that of HIV and so it is possible that some women were able to change their behaviour with younger male partners in a way that protected them from acquiring HSV-2 and this was somehow not reflected in the study’s behavioural outcomes. The findings from the qualitative research support this, as it was observed that women were sometimes able to change their behaviour with younger partners while not doing so with their older main partner. This raises the possibility of Stepping Stones having a positive longer term impact on women’s HIV risk beyond the period of observation of the study.

esearch support this, as it was observed that women were sometimes able to change their behaviour with younger partners while not doing so with their older main partner. This raises the possibility of Stepping Stones having a positive longer term impact on women’s HIV risk beyond the period of observation of the study. We observed changes in two other outcomes in women that were not in the intended direction. There was more transactional sex with a casual partner at 12 months among women in the Stepping Stones arm and there was a suggestion of more unwanted pregnancies at 24 months. Though the negative impact on transactional sex had resolved by 24 months, we suggest that particular care should be given to how transactional sex is discussed in groups of young women. Group discussions might have inadvertently encouraged transactional sex by reflecting it as at least common, if not standard, and an effective way of acquiring desired items. The attempts of facilitators to avoid being moralistic in discussions about transactional sex might have meant that the negative impacts were insufficiently emphasised. The changes in sexual and violent behaviour of men were supported by the findings of qualitative research. Stepping Stones is a behavioural intervention that, according to a recent classification of interventions by WHO, is “gender transformative” in that it seeks to transform gender roles and promote more gender equitable relationships between men and women. 21

of men were supported by the findings of qualitative research. Stepping Stones is a behavioural intervention that, according to a recent classification of interventions by WHO, is “gender transformative” in that it seeks to transform gender roles and promote more gender equitable relationships between men and women. 21 Our results suggest that it did lead to some change in violent and exploitative behaviour in men. Analyses performed on the baseline dataset showed that behaviours transformed by the intervention were those associated with perpetration of intimate partner violence,22 rape,23 and participation in transactional sex,24 and we hypothesise that these variables reflect particular ideas of masculinity. Evaluations of Stepping Stones in many other settings have documented an impact on men’s violence against their intimate partner,11 25 which further supports our study’s findings. Stepping Stones is one of few interventions with demonstrated effectiveness in reducing this. The clearer visibility of this reduction at 24 months when compared with 12 months is consistent with the findings of other interventions26 and suggests that this positive behaviour change is being strengthened over time. Exposure to intimate partner violence has been identified as an important risk factor for HIV in women27 and so the reduction in male violence might have a broader impact on HIV in their sexual partners well beyond the study setting. Many of the other changes in men’s behaviour were not sustained to 24 months, which points to the need for research to strengthen the intervention.

an important risk factor for HIV in women27 and so the reduction in male violence might have a broader impact on HIV in their sexual partners well beyond the study setting. Many of the other changes in men’s behaviour were not sustained to 24 months, which points to the need for research to strengthen the intervention. Weaknesses The trial has several weaknesses that might affect the interpretation of the results. Randomisation occurred before recruitment. No villages declined to participate in the study because of their allocation but some individuals did. In both control and intervention clusters we usually had more volunteers for the study than we were able to include so it was not possible to count how many people dropped out during selection because of the allocation as opposed to other reasons, including the need to restrict recruitment to a maximum of 40 per cluster. We noted, however, that some people (particularly women) who lived far from the schools where the sessions were held thought they could not attend the whole Stepping Stones programme. Some women were not allowed to take part because they had strict parents who expected them home quickly after school. It is possible that those in the Stepping Stones arm were in some ways more motivated. This might have differentially influenced the response to the interventions. It is difficult to know how this would have affected our results and generalisability thereof, but given the fairly modest results of this trial, especially for women, it seems unlikely that there was a substantial impact.

ways more motivated. This might have differentially influenced the response to the interventions. It is difficult to know how this would have affected our results and generalisability thereof, but given the fairly modest results of this trial, especially for women, it seems unlikely that there was a substantial impact. The generalisability of the study findings could be influenced by several aspects of the trial design. The scope of our intervention was deliberately constrained by affordability in the design of the evaluation. We thus did not evaluate the model of programme delivery originally intended by Welbourn,10 which includes groups of older adult participants and multiple groups within the same village. Having done so might have enhanced the overall impact of the programme. This model reflects the socioecological perspective, which has been advocated in HIV prevention.28 We designed the trial to measure the impact of Stepping Stones when delivered in a way that reflected the practices of local organisations that work with such programmes. In so doing, our intention was to give some indication of the likely impact of Stepping Stones outside a trial setting. Any weaknesses in delivery of the intervention were probably no greater than those normally found. Our findings are a measure of the difference in outcomes between the two arms. For ethical reasons we provided a reasonably substantial control intervention that focused on HIV prevention and was taken from the Stepping Stones intervention. We cannot exclude the possibility that it resulted in behaviour change, although given the difficulties researchers face in showing impact from behavioural interventions2 3 it would be surprising if the control intervention had a substantial impact.

on HIV prevention and was taken from the Stepping Stones intervention. We cannot exclude the possibility that it resulted in behaviour change, although given the difficulties researchers face in showing impact from behavioural interventions2 3 it would be surprising if the control intervention had a substantial impact. The assumptions we used in calculating the required sample size for the trial were too optimistic. The effect size used in the sample size calculation was large (50% reduction in HIV incidence) and the anticipated overall incidence of HIV was incorrect. In addition, although we used a larger value for the coefficient of variation between clusters than was used in the Mwanza trial, the value used (0.35) was in fact considerably smaller than the actual value of 1.02. Our stratification of the clusters did not help in reducing the variation in incidence rates between clusters, which shows the practical difficulties of stratifying on surrogate geographical variables. Sample size calculations in future evaluations of behavioural interventions should use a more modest estimate of expected effect size and realistic estimates of either the coefficient of variation (k) between clustersor the intracluster correlation. The choice of k was informed by an analysis of the results of the 1999 national antenatal seroprevalence survey for the Eastern Cape. We thought that this value of k might be optimistic as the clusters in the antenatal survey (being antenatal clinics) cover a larger geographical area than the clusters proposed for the Stepping Stones intervention. The final value of k, however, was a compromise. We though that if we used too large a value of k in our sample size calculations, this would discourage funders. One positive recommendation from this study, which has been supported by other evaluations of behavioural interventions, is that large sample sizes are required to assess the modest but important reductions in incidence of HIV that might result from behavioural interventions and that necessary funding should be provided.

e positive recommendation from this study, which has been supported by other evaluations of behavioural interventions, is that large sample sizes are required to assess the modest but important reductions in incidence of HIV that might result from behavioural interventions and that necessary funding should be provided. In the control arm slightly more blood specimens were collected by dried blood spot. Although this method for HIV testing was optimised, previous research has suggested that it might be slightly less sensitive than when serum is used.29 The impact of any such loss of sensitivity would have been to underestimate the true incidence of HIV in the control arm. There could have been contamination between arms, but serious contamination is unlikely as clusters were geographically separated and the total sample size was small compared with the overall population, so the likelihood of participants forming friendships with people from the other study arm was low. Despite considerable efforts to trace cohort members, about 15% failed to contribute any data to the biological outcomes and a quarter were untraceable at 24 months. Our follow-up rates compare favourably with those of similar trials—for example, Ross et al lost 27% to follow-up.5 As follow-up rates were similar in the intervention and control arms this is unlikely to have biased the results.

ontribute any data to the biological outcomes and a quarter were untraceable at 24 months. Our follow-up rates compare favourably with those of similar trials—for example, Ross et al lost 27% to follow-up.5 As follow-up rates were similar in the intervention and control arms this is unlikely to have biased the results. Implications The meaning of the study findings is determined by an assessment of whether this was a trial with negative or positive findings. Some would argue that Stepping Stones did not work because it failed to affect the incidence of HIV. Literature on evaluation of behavioural interventions, however, rarely disregards all other outcomes, and we have shown significant other effects. We analysed the other biological outcome, incidence of HSV-2, across both years and found a reduction in the intervention arm. Most of the changes suggested in other behaviours were not sustained to two years, as is commonly found with evaluations of behavioural interventions, and we endorse the view that for behaviour change to be meaningful it must be enduring.28 In contrast, the impact on perpetration of intimate partner violence seems to have been strengthened over the two years of follow-up. This is a pattern that is recognised in the behavioural science literature30; it results from people having had an opportunity over time to reflect on their behaviour or for the environment to reinforce behaviours. Both HSV-2 and intimate partner violence are established risk factors for HIV and so the observation that Stepping Stones had an effect is of some interest.

ural science literature30; it results from people having had an opportunity over time to reflect on their behaviour or for the environment to reinforce behaviours. Both HSV-2 and intimate partner violence are established risk factors for HIV and so the observation that Stepping Stones had an effect is of some interest. What is already known on the topic HIV prevention studies have had mixed results in terms of their impact on sexual behaviour Most studies have been conducted in Western countries; in sub-Saharan Africa no programme has been shown to reduce sexually transmitted infections What this study adds The Stepping Stones programme did not lower incidence of HIV but did reduce the incidence of herpes simplex type 2 virus and male perpetration of intimate partner violence Stepping Stones can affect some risk factors for HIV in young men and women

Most studies have been conducted in Western countries; in sub-Saharan Africa no programme has been shown to reduce sexually transmitted infections What this study adds The Stepping Stones programme did not lower incidence of HIV but did reduce the incidence of herpes simplex type 2 virus and male perpetration of intimate partner violence Stepping Stones can affect some risk factors for HIV in young men and women We thank the Planned Parenthood Association of South Africa Eastern Cape Branch, our partner in the study intervention; the National Institute for Communicable Diseases for quality control, testing, and storage of specimens; Nelisiwe Khuzwayo (community coordinator); Leslie Setheni, Veliswa Gobinduku, Busiswa Mketo, Yandisa Sikweyiya, Mthokozisi Madiya, Bongwekazi Rapiya, Sanele Mdlungu, Ayanda Mxekezo, Lungelo Mdekazi, Nocawe Mxinwa, Andiswa Njengele, Mvuyo Mayisela, Philiswa Bango, Nobapostile Malu, Lizo Tshona, Khanyisile Bakan, Linda Shute, Lindiwe Farlane, Siya Kave, Bantu Waka, Zoleka Mbange (field nurses and field workers); Bomkazi Mnombeli, Engela Gerber, Alta Hansen (data management, data entry, and secretarial support); Daniel Kayongo, UNITRA (advice on biological aspects of the study); and Chief Z S Mtirara and all the members of the community advisory board. Mary Koss advised on questionnaire design and aspects of the study implementation.

i Mnombeli, Engela Gerber, Alta Hansen (data management, data entry, and secretarial support); Daniel Kayongo, UNITRA (advice on biological aspects of the study); and Chief Z S Mtirara and all the members of the community advisory board. Mary Koss advised on questionnaire design and aspects of the study implementation. Contributors: RJ was the project leader throughout the trial; wrote the proposal, led on the study design, intervention adaptation, and questionnaires; managed the study; directly managed the project from September 2004-April 2006; and did much of the data management and led drafting of the paper. MN contributed to the design of the intervention and questionnaires, developed operating plans for the implementation of the trial, was the project manager from September 2002-August 2004, and contributed to interpretation of the data. JL was project statistician, responsible for statistical aspects of study design, data management, and data analysis. NJ contributed to the design of the intervention and questionnaires, implementation of the trial, and interpretation of the data. KD contributed to the design of the study, data management, interpretation of the findings, and drafting of the paper. AP designed the protocols related to HIV and HSV-2 testing and quality control of the biological side of the study, and supervised the laboratory tests and storage of specimens and interpretation of results. ND contributed to the management of the study and the interpretation of the data. All investigators contributed to writing this paper. JL is guarantor.

V and HSV-2 testing and quality control of the biological side of the study, and supervised the laboratory tests and storage of specimens and interpretation of results. ND contributed to the management of the study and the interpretation of the data. All investigators contributed to writing this paper. JL is guarantor. Funding: National Institute of Mental Health grant No MH 64882-01 and South African Medical Research Council. KD was funded from the Harry F Guggenheim Foundation and by the Emory Center for AIDS Research (P30 AI050409). Competing interests: None declared. Ethical approval: University of Pretoria and University of Witwatersrand Ethics Committees. Provenance and peer review: Not commissioned; externally peer reviewed. Cite this as: BMJ 2008;337:a506

Introduction Depression is a common and debilitating condition, which is often treated with antidepressants. Depression increases the risk of cardiovascular outcomes, but controversy exists as to whether use of antidepressants, particularly selective serotonin reuptake inhibitors, increases or reduces the risk.1 2 This is important because antidepressants are one of the most commonly prescribed types of drug worldwide, and their use is increasing.3 4 5 In the United States, antidepressants were the third most commonly used prescription drug in 2005-08, and their use had increased by almost 400% compared with 1988-946; in England, more than 53 million prescriptions for antidepressants were issued in 2013,7 nearly a twofold increase compared with a decade earlier.8 More than half (54%) of the prescriptions in England in 2013 were for selective serotonin reuptake inhibitors, including nearly 14 million prescriptions for the most commonly prescribed antidepressant citalopram.

ptions for antidepressants were issued in 2013,7 nearly a twofold increase compared with a decade earlier.8 More than half (54%) of the prescriptions in England in 2013 were for selective serotonin reuptake inhibitors, including nearly 14 million prescriptions for the most commonly prescribed antidepressant citalopram. Theoretically, antidepressants such as selective serotonin reuptake inhibitors may have effects on coagulation, and some studies have explored their cardioprotective effect. These studies have tended to be underpowered and explored outcomes in secondary care or other selected populations. Randomised controlled trials of antidepressants tend to be short term and underpowered to detect effects on cardiovascular outcomes, and observational studies of cardiovascular outcomes show conflicting results and many have not accounted for depression and so are susceptible to indication biases. The observational studies have either been restricted to or predominantly included older people, so uncertainty exists about associations in a younger age group, although antidepressants are often prescribed for depression in adults of working age. Antidepressants may have differential effects on cardiovascular outcomes according to age. A meta-analysis of 13 observational studies found that use of selective serotonin reuptake inhibitors was associated with a 40% increased risk of stroke, but this was significant only in studies restricted to older age groups and no significantly increased risk was seen in studies with no age restriction, although none of the studies specifically focused on a younger age group.9 Similarly, for myocardial infarction, uncertainty exists about an association with selective serotonin reuptake inhibitors. A large observational study in people aged 65 and over with depression found an increased risk of myocardial infarction with selective serotonin reuptake inhibitors,10 whereas other studies in broader age groups have found no association or reduced risks,11 12 13 which could be a result of differing age ranges or indication biases.

vational study in people aged 65 and over with depression found an increased risk of myocardial infarction with selective serotonin reuptake inhibitors,10 whereas other studies in broader age groups have found no association or reduced risks,11 12 13 which could be a result of differing age ranges or indication biases. The US Food and Drug Administration (FDA) issued a drug safety communication in 2011, stating that citalopram should not be prescribed at doses greater than 40 mg per day, based on findings of QT interval prolongation in a study of 119 participants who received different doses of citalopram.14 The European Medicines Agency issued a similar safety warning in 2011. Further studies have reported QT interval prolongation with citalopram and also with some other antidepressants such as escitalopram and amitriptyline.15 16 QT interval prolongation can lead to arrhythmias including potentially fatal torsades de pointes,17 but few studies have specifically assessed risk of arrhythmia for different antidepressant drugs. A cohort study in predominantly older men of two different selective serotonin reuptake inhibitor antidepressants found significantly lower risks of arrhythmia for doses of citalopram over 40 mg/day compared with doses of 1-20 mg/day, with similar findings for sertraline.18 A cohort study based on claims data in the United States found no significant differences in risk of ventricular arrhythmia/sudden death for 20 types of antidepressant drug compared with paroxetine, except for a higher risk in mirtazapine users.19

pared with doses of 1-20 mg/day, with similar findings for sertraline.18 A cohort study based on claims data in the United States found no significant differences in risk of ventricular arrhythmia/sudden death for 20 types of antidepressant drug compared with paroxetine, except for a higher risk in mirtazapine users.19 Few observational studies of cardiovascular effects have examined associations with individual drugs, so evidence for specific commonly prescribed antidepressants is lacking, especially in younger people, as is evidence in relation to duration and dose. We therefore carried out a cohort study in people aged 20 to 64 to investigate the associations between different antidepressant drugs and the risk of myocardial infarction, arrhythmia, and stroke/transient ischaemic attack and also examined both dose and duration of use. Methods The cohort study was designed to estimate associations between antidepressant treatment and several different adverse outcomes including arrhythmia, myocardial infarction, and stroke or transient ischaemic attack. Full details of the study design, outcomes, and methods can be found in the study protocol.20 Results relating to the epilepsy, suicide, and self harm outcomes have been published previously.21 22

d several different adverse outcomes including arrhythmia, myocardial infarction, and stroke or transient ischaemic attack. Full details of the study design, outcomes, and methods can be found in the study protocol.20 Results relating to the epilepsy, suicide, and self harm outcomes have been published previously.21 22 Study cohort The study cohort was selected from a large primary care database (QResearch, version 34). At the time of the study, the QResearch database contained the anonymised longitudinal health records of more than 12 million patients from more than 600 general practices across the United Kingdom, which record data using the Egton Medical Information Systems (EMIS) medical records computer system. Recorded information includes patients’ characteristics, clinical diagnoses, symptoms, and prescribed drugs.

alth records of more than 12 million patients from more than 600 general practices across the United Kingdom, which record data using the Egton Medical Information Systems (EMIS) medical records computer system. Recorded information includes patients’ characteristics, clinical diagnoses, symptoms, and prescribed drugs. The cohort included patients with a first computer recorded diagnosis of depression between the ages of 20 and 64 years at the time of diagnosis, from 1 January 2000 to 31 July 2011, as described previously.22 We identified patients with a diagnosis of depression by using diagnostic Read codes used in previous studies.10 23 24 Read codes are the clinical codes used in general practice in the United Kingdom. Patients were eligible for inclusion if their diagnosis of depression occurred at least 12 months after their registration with a study practice and the installation date of their practice’s EMIS computer system. We restricted our cohort to patients with a first recorded diagnosis of depression so that antidepressant prescribing during follow-up would not be influenced by any previous experiences and preferences that would be difficult to account for in the analyses. We used the 12 month inclusion criterion to ensure that the diagnosis of depression was not a retrospective recording of a previous diagnosis.

n so that antidepressant prescribing during follow-up would not be influenced by any previous experiences and preferences that would be difficult to account for in the analyses. We used the 12 month inclusion criterion to ensure that the diagnosis of depression was not a retrospective recording of a previous diagnosis. We excluded patients with a previous recorded diagnosis of depression; those with a diagnosis of schizophrenia, bipolar disorder, or another type of psychosis; and those who had received prescriptions for lithium or antimanic drugs. We also excluded patients if they had received prescriptions for an antidepressant before the study start date (1 January 2000), before their registration date, before they were aged 20, or more than 36 months before their first recorded diagnosis of depression. Temporary residents were also excluded. The patient’s study entry date was the earliest of the date of the first recorded diagnosis of depression or the date of the first prescription for an antidepressant. Participants in the cohort were followed up until the earliest of date of death, date of leaving the practice, or the end of the follow-up period (1 August 2012).

nt’s study entry date was the earliest of the date of the first recorded diagnosis of depression or the date of the first prescription for an antidepressant. Participants in the cohort were followed up until the earliest of date of death, date of leaving the practice, or the end of the follow-up period (1 August 2012). Outcomes The three outcomes for these analyses were arrhythmia, myocardial infarction, and stroke or transient ischaemic attack. We identified patients with these outcomes if they were recorded either on their general practice record using the relevant Read codes or on their linked Office of National Statistics cause of death record using ICD (international classification of diseases) diagnostic codes, based on codes used in previous studies,25 26 27 as listed in the web appendix. For the analysis of each separate outcome, we considered only the first event and excluded patients with a previous diagnosis of the outcome recorded at baseline. Exposures We extracted information on all prescriptions for antidepressants during follow-up. We calculated the duration of each prescription by dividing the number of tablets prescribed by the number to be taken each day.22 For the main analyses, we grouped antidepressant drugs according to the four main classes in the British National Formulary: tricyclic and related antidepressants, selective serotonin reuptake inhibitors, monoamine oxidase inhibitors, and other antidepressants. We classified prescriptions for different antidepressant drugs on the same date as combined prescriptions.

drugs according to the four main classes in the British National Formulary: tricyclic and related antidepressants, selective serotonin reuptake inhibitors, monoamine oxidase inhibitors, and other antidepressants. We classified prescriptions for different antidepressant drugs on the same date as combined prescriptions. We calculated the daily dose of each prescription by multiplying the number of tablets to be taken each day by the dose of each tablet, and we converted this to a defined daily dose to enable comparison of doses between antidepressant classes, using values assigned by the World Health Organization’s Collaborating Centre for Drug Statistics Methodology (www.whocc.no/atc_ddd_index). For some prescriptions, the dosing instructions were missing or not sufficiently detailed to allow calculation of a daily dose (<5% of total prescriptions). We also assessed the 11 most frequently prescribed individual antidepressant drugs.10 19 22

ng Centre for Drug Statistics Methodology (www.whocc.no/atc_ddd_index). For some prescriptions, the dosing instructions were missing or not sufficiently detailed to allow calculation of a daily dose (<5% of total prescriptions). We also assessed the 11 most frequently prescribed individual antidepressant drugs.10 19 22 Confounding variables We extracted data on variables considered to be potential risk factors for the cardiovascular outcomes or associated with the likelihood of receiving a particular antidepressant treatment, based on our previous study of antidepressants in people aged 65 or over.10 These were age at study entry (continuous); sex; year of diagnosis of depression (continuous); severity of index diagnosis of depression (categorised as mild, moderate, or severe, using the classification of Read codes for depression published by Martinez and colleagues23 and additional classification by a member of the study team (RM) of some Read codes for depression used in our study21 but not included in the study by Martinez); deprivation (Townsend deprivation score corresponding to the patient’s postcode, in fifths); smoking status (non-smoker, ex-smoker, light smoker (1-9 cigarettes/day), moderate smoker (10-19 cigarettes/day), heavy smoker (≥20 cigarettes/day), not recorded); alcohol intake (none, trivial (<1 unit/day), light (1-2 units/day), medium (3-6 units/day), heavy (7-9 units/day), very heavy (>9 units/day), not recorded); ethnic group (categorised into a binary variable of white/not recorded or non-white (comprising Indian, Pakistani, Bangladeshi, other Asian, black African, black Caribbean, Chinese, other including mixed)); comorbidities at baseline (individual binary variables for each of coronary heart disease, diabetes, hypertension, cancer, epilepsy/seizures, hypothyroidism, osteoarthritis, rheumatoid arthritis, asthma/chronic obstructive pulmonary disease, osteoporosis, liver disease, renal disease, obsessive-compulsive disorder); and use of other drugs at baseline (individual binary variables for each of antihypertensives, aspirin, statins, anticoagulants, non-steroidal anti-inflammatory drugs, anticonvulsants, hypnotics/anxiolytics, antipsychotics, bisphosphonates, oral contraceptives, hormone replacement therapy). In addition, for the arrhythmia and myocardial infarction outcomes, we adjusted for a diagnosis of stroke or transient ischaemic attack at baseline.

gulants, non-steroidal anti-inflammatory drugs, anticonvulsants, hypnotics/anxiolytics, antipsychotics, bisphosphonates, oral contraceptives, hormone replacement therapy). In addition, for the arrhythmia and myocardial infarction outcomes, we adjusted for a diagnosis of stroke or transient ischaemic attack at baseline. We included year of diagnosis of depression as a confounding variable to account for changes in prescribing patterns over time.

gulants, non-steroidal anti-inflammatory drugs, anticonvulsants, hypnotics/anxiolytics, antipsychotics, bisphosphonates, oral contraceptives, hormone replacement therapy). In addition, for the arrhythmia and myocardial infarction outcomes, we adjusted for a diagnosis of stroke or transient ischaemic attack at baseline. We included year of diagnosis of depression as a confounding variable to account for changes in prescribing patterns over time. Statistical analysis We used Cox’s proportional hazards models to estimate associations between the three outcomes and exposure to antidepressant drugs, treating antidepressant exposure as a time varying exposure to allow for patients starting and stopping and also changing between treatments during follow-up. We used robust standard errors to allow for clustering of patients within practices. We excluded patients from the analysis of each outcome if they had the outcome recorded at baseline. We classified patients as exposed to an antidepressant if no gaps of more than 90 days existed between the end of one prescription and the start of the next. If gaps of more than 90 days occurred, patients counted as exposed for the first 90 days and then unexposed for the remaining period. When patients stopped an antidepressant, we classified them as exposed for the first 90 days after the estimated date of stopping, so that outcomes occurring during withdrawal periods would be attributed to the antidepressant. The main analyses were based on the first five years of follow-up after study entry, and patients were censored at the earliest of five years after study entry, date of death, date of leaving the practice, or the end of the follow-up period in these analyses. We selected five years of follow-up for our main analyses as this would incorporate periods of long term treatment and also allow for more events to accrue than a shorter follow-up period would, so increasing the power of the study.

th, date of leaving the practice, or the end of the follow-up period in these analyses. We selected five years of follow-up for our main analyses as this would incorporate periods of long term treatment and also allow for more events to accrue than a shorter follow-up period would, so increasing the power of the study. The analyses calculated unadjusted and adjusted hazard ratios for each antidepressant class (tricyclic and related antidepressants, selective serotonin reuptake inhibitors, other antidepressants, combined treatment) compared with periods of no antidepressant treatment. The unexposed reference category included periods of unexposed time in patients treated at other periods of time during follow-up, as well as person years from patients who received no antidepressant treatment throughout follow-up, so the hazard ratios compare rates of the outcomes between exposed and unexposed periods of time throughout follow-up. Patients who received monoamine oxidase inhibitors at any time were excluded from these analyses, as the number in this category was small. We excluded patients with missing deprivation scores from the adjusted analyses. Analyses were carried out for time varying exposures of prescribed daily dose (categorised as ≤0.5, >0.5 and ≤1.0, and >1.0 defined daily doses), and we calculated tests for trend within each drug class by using dose as a continuous variable. Periods of exposure time for which daily dose was missing were excluded from the analysis of dose. We did additional analyses for time since starting treatment (categorised as no use or treatment duration of 1-28 days, 29-84 days, or ≥85 days) and time since stopping treatment (1-28 days, 29-84 days, and 85-182 days after the estimated date of stopping treatment) and for the 11 most commonly prescribed individual antidepressants, as in a previous study.10 Individual antidepressants were further categorised by dose (≤1 or >1 defined daily doses), and citalopram was also categorised as ≤20 mg/day, 20-39 mg/day, and ≥40 mg/day for an analysis of the arrhythmia outcome, in light of the FDA’s drug safety communication.28

individual antidepressants, as in a previous study.10 Individual antidepressants were further categorised by dose (≤1 or >1 defined daily doses), and citalopram was also categorised as ≤20 mg/day, 20-39 mg/day, and ≥40 mg/day for an analysis of the arrhythmia outcome, in light of the FDA’s drug safety communication.28 We used Wald’s significance tests to identify significant differences between antidepressant classes and between individual antidepressant drugs. We tested for interactions between class of antidepressant and age and sex. We assessed the proportional hazards assumption by using log minus log plots. As sensitivity analyses, we repeated the analyses including the entire follow-up period and did an analysis excluding patients who received no antidepressant prescriptions during follow-up.22 We repeated our main analyses using selective serotonin reuptake inhibitors as the comparison group for drug class, the middle dose category of selective serotonin reuptake inhibitors as the comparison group for drug dose, and citalopram (the most commonly prescribed antidepressant) as the comparison group for individual antidepressants.

n analyses using selective serotonin reuptake inhibitors as the comparison group for drug class, the middle dose category of selective serotonin reuptake inhibitors as the comparison group for drug dose, and citalopram (the most commonly prescribed antidepressant) as the comparison group for individual antidepressants. We also did an analysis restricted to the first year of follow-up; we did this because we had some evidence of non-proportional hazards over five years of follow-up, and also this time period more closely reflected the average duration of treatment. As a post hoc analysis, we also estimated adjusted hazard ratios separately using interaction terms for the 0-1 years, 1-3 years, and 3-5 years after the start of follow-up to further investigate changes in hazard ratios over time. We did these analyses for drug class and for only the five most frequently prescribed antidepressants owing to the smaller numbers of events in the later time periods. To examine the effect of adjusting for different confounding variables, we did additional analyses entering the variables in blocks. As a post hoc analysis, we used a stratified Cox model, with stratification by general practice to compare with our main models using robust standard errors to account for clustering by practice. We calculated absolute risks of the three outcomes over one year, accounting for the confounding variables by using the adjusted hazard ratios from the analyses based on one year of follow-up, according to the method described by Altman et al.29

We also did an analysis restricted to the first year of follow-up; we did this because we had some evidence of non-proportional hazards over five years of follow-up, and also this time period more closely reflected the average duration of treatment. As a post hoc analysis, we also estimated adjusted hazard ratios separately using interaction terms for the 0-1 years, 1-3 years, and 3-5 years after the start of follow-up to further investigate changes in hazard ratios over time. We did these analyses for drug class and for only the five most frequently prescribed antidepressants owing to the smaller numbers of events in the later time periods. To examine the effect of adjusting for different confounding variables, we did additional analyses entering the variables in blocks. As a post hoc analysis, we used a stratified Cox model, with stratification by general practice to compare with our main models using robust standard errors to account for clustering by practice. We calculated absolute risks of the three outcomes over one year, accounting for the confounding variables by using the adjusted hazard ratios from the analyses based on one year of follow-up, according to the method described by Altman et al.29 We included all eligible patients in the database in our analyses to maximise power. We used a P value of <0.01 (two tailed) to determine statistical significance. We used Stata (v12.1) for all analyses.

We calculated absolute risks of the three outcomes over one year, accounting for the confounding variables by using the adjusted hazard ratios from the analyses based on one year of follow-up, according to the method described by Altman et al.29 We included all eligible patients in the database in our analyses to maximise power. We used a P value of <0.01 (two tailed) to determine statistical significance. We used Stata (v12.1) for all analyses. Patient involvement No patients were involved in setting the research question or the outcome measures, nor were they involved in the design or implementation of the study. No patients were asked to advise on interpretation or writing up of results. Patient representatives from the QResearch Advisory Board have advised on dissemination of studies using QResearch data, including the use of lay summaries describing the research and its results. Results The initial cohort included 327 235 patients with a first diagnosis of depression made between the ages of 20 and 64, between 1 January 2000 and 31 July 2011. We excluded 88 272 (27.0%) patients because they had schizophrenia, bipolar disorder, or other psychoses; had been treated with lithium or antimanic drugs; or had received a prescription for an antidepressant before the study entry date, before age 20, or more than 36 months before their date of diagnosis of depression. This left 238 963 patients from 687 practices in the final study cohort.

r disorder, or other psychoses; had been treated with lithium or antimanic drugs; or had received a prescription for an antidepressant before the study entry date, before age 20, or more than 36 months before their date of diagnosis of depression. This left 238 963 patients from 687 practices in the final study cohort. The total length of follow-up was 1 307 326 person years. Among patients in the cohort 123 038 (51.5%) had at least five years of follow-up, with a median of 5.2 (interquartile range 2.5-8.2) years overall. The mean age of patients in the study cohort was 39.5 (SD 11.1) years, and 61% were women (table 1). Townsend deprivation score was missing for 8201 (3.4%) patients. Table 1 Characteristics of study cohort (n=238 963) at baseline. Values are numbers (percentages) unless stated otherwise

The total length of follow-up was 1 307 326 person years. Among patients in the cohort 123 038 (51.5%) had at least five years of follow-up, with a median of 5.2 (interquartile range 2.5-8.2) years overall. The mean age of patients in the study cohort was 39.5 (SD 11.1) years, and 61% were women (table 1). Townsend deprivation score was missing for 8201 (3.4%) patients. Table 1 Characteristics of study cohort (n=238 963) at baseline. Values are numbers (percentages) unless stated otherwise Characteristic Value Female sex 146 028 (61.1) Mean (SD) age, years 39.5 (11.1) Ethnic group: Recorded 136 624 (57.2) White/not recorded 227 451 (95.2) Non-white 11 512 (4.8) Depression severity (index diagnosis): Mild 171 208 (71.7) Moderate 59 140 (24.8) Severe 8615 (3.6) Smoking status*: Non-smoker 110 849 (47.5) Ex-smoker 35 132 (15.1) Current light smoker 24 104 (10.3) Current moderate smoker 40 546 (17.4) Current heavy smoker 22 659 (9.7) Not recorded 5673 Alcohol consumption*: Non-drinker 55 253 (27.2) Trivial (<1 unit/day) 77 579 (38.2) Light (1-2 units/day) 51 310 (25.3) Moderate (3-6 units/day) 14 482 (7.1) Heavy (7-9 units/day) 2174 (1.1) Very heavy (>9 units/day) 2391 (1.2) Not recorded 35 774 Townsend deprivation score in fifths*: 1 (least deprived) 45 021 (19.5) 2 46 207 (20.0) 3 48 293 (20.9) 4 47 063 (20.4) 5 (most deprived) 44 178 (19.1) Not recorded 8201 Comorbidities at baseline: Coronary heart disease 4109 (1.7) Diabetes 7371 (3.1) Hypertension 17 217 (7.2) Stroke/transient ischaemic attack 1741 (0.7) Arrhythmia 2373 (1.0) Any cancer 3810 (1.6) Asthma/chronic obstructive pulmonary disease 31 816 (13.3) Epilepsy/seizures 3325 (1.4) Hypothyroidism 5267 (2.2) Obsessive-compulsive disorder 494 (0.2) Osteoarthritis 7228 (3.0) Osteoporosis 867 (0.4) Liver disease 698 (0.3) Renal disease 549 (0.2) Rheumatoid arthritis 1301 (0.5) Drugs at baseline: Anticonvulsants 2672 (1.1) Antihypertensives 25 344 (10.6) Antipsychotics 836 (0.4) Anticoagulants 1073 (0.5) Aspirin 7159 (3.0) Bisphosphonates 854 (0.4) Hypnotics/anxiolytics 11 354 (4.8) Non-steroidal anti-inflammatory drugs 12 725 (5.3) Statins 10 823 (4.5) Oral contraceptives† 27 396 (18.8) Hormone replacement therapy† 7207 (4.9) *Percentages are out of total with recorded values.

otics 836 (0.4) Anticoagulants 1073 (0.5) Aspirin 7159 (3.0) Bisphosphonates 854 (0.4) Hypnotics/anxiolytics 11 354 (4.8) Non-steroidal anti-inflammatory drugs 12 725 (5.3) Statins 10 823 (4.5) Oral contraceptives† 27 396 (18.8) Hormone replacement therapy† 7207 (4.9) *Percentages are out of total with recorded values. †Percentage is for females only. Antidepressant treatment during follow-up During follow-up, 209 476 (87.7%) patients received a total of 3 337 336 antidepressant prescriptions. These comprised 2 379 668 (71.3%) prescriptions for selective serotonin reuptake inhibitors, 533 798 (16.0%) for tricyclic and related antidepressants, and 422 079 (12.7%) for the group of other antidepressants. In addition, 156 patients had received a total of 1791 (0.05%) prescriptions for monoamine oxidase inhibitors. There were 83 784 combined prescriptions for two or more different antidepressant drugs prescribed on the same day. The median duration of treatment during follow-up was 221 (interquartile range 79-590) days.

In addition, 156 patients had received a total of 1791 (0.05%) prescriptions for monoamine oxidase inhibitors. There were 83 784 combined prescriptions for two or more different antidepressant drugs prescribed on the same day. The median duration of treatment during follow-up was 221 (interquartile range 79-590) days. Among a total of 3 252 633 prescriptions (with combined prescriptions counting as single prescriptions), citalopram was the most commonly prescribed antidepressant (1 023 255 (31.5%) prescriptions) followed by fluoxetine (778 285; 23.9%), and then amitriptyline (236 416; 7.3%). Supplementary table A shows numbers of prescriptions for the 11 most commonly prescribed antidepressants, with information on prescribed daily doses. Distributions of baseline characteristics according to the first antidepressant prescribed for these 11 drugs have been presented in a previous paper.22 Associations with arrhythmia At baseline, 2373 patients had an existing diagnosis of arrhythmia. We excluded these patients from analysis of the arrhythmia outcome, along with the patients who received prescriptions for monoamine oxidase inhibitors, leaving 236 434 patients in the analysis cohort. During the first five years of follow-up, 1452 new diagnoses of arrhythmia were made, giving an incidence rate of 16.2 per 10 000 person years (20.1 per 10 000 in men and 13.8 per 10 000 in women).

the patients who received prescriptions for monoamine oxidase inhibitors, leaving 236 434 patients in the analysis cohort. During the first five years of follow-up, 1452 new diagnoses of arrhythmia were made, giving an incidence rate of 16.2 per 10 000 person years (20.1 per 10 000 in men and 13.8 per 10 000 in women). We found no significant associations with arrhythmia (at P<0.01) for any of the drug classes over five years compared with periods of no antidepressant treatment, as shown in table 2, although we saw some indication of a reduced hazard ratio for selective serotonin reuptake inhibitors (adjusted hazard ratio 0.84, 95% confidence interval 0.73 to 0.97; P=0.02) compared with no current use of antidepressants. In a direct comparison with selective serotonin reuptake inhibitors (supplementary table B), we found a significantly increased rate for the group of other antidepressants (adjusted hazard ratio 1.44, 1.12 to 1.85). Table 2 Unadjusted and adjusted hazard ratios for arrhythmia by antidepressant class, dose, and duration over 5 years’ follow-up

We found no significant associations with arrhythmia (at P<0.01) for any of the drug classes over five years compared with periods of no antidepressant treatment, as shown in table 2, although we saw some indication of a reduced hazard ratio for selective serotonin reuptake inhibitors (adjusted hazard ratio 0.84, 95% confidence interval 0.73 to 0.97; P=0.02) compared with no current use of antidepressants. In a direct comparison with selective serotonin reuptake inhibitors (supplementary table B), we found a significantly increased rate for the group of other antidepressants (adjusted hazard ratio 1.44, 1.12 to 1.85). Table 2 Unadjusted and adjusted hazard ratios for arrhythmia by antidepressant class, dose, and duration over 5 years’ follow-up No of events Person years Unadjusted hazard ratio (95% CI) Adjusted analysis† Hazard ratio (95% CI) P value Antidepressant class No current use 887 568 365 1.00 1.00 - TCAs 102 41 208 1.59 (1.29 to 1.96) 1.09 (0.88 to 1.35) 0.46 SSRIs 352 224 985 1.02 (0.89 to 1.18) 0.84 (0.73 to 0.97) 0.02 Other antidepressants 68 28 048 1.55 (1.23 to 1.95) 1.21 (0.96 to 1.54) 0.11 Combined antidepressants 10 4233 1.47 (0.75 to 2.89) 1.07 (0.54 to 2.09) 0.85 Antidepressant class and dose categories No current use 887 568 365 1.00 1.00 - TCAs: ≤0.5 DDD 51 23 506 1.37 (1.03 to 1.82) 0.89 (0.67 to 1.19) 0.44 >0.5 DDD/≤1.0 DDD 26 8400 2.03 (1.39 to 2.96) 1.35 (0.91 to 1.99) 0.14 >1.0 DDD 14 5306 1.66 (0.98 to 2.81) 1.32 (0.77 to 2.26) 0.31 Test for trend§  - - - - 0.15 SSRIs: ≤0.5 DDD 30 15 995 1.19 (0.82 to 1.71) 0.93 (0.64 to 1.35) 0.71 >0.5 DDD/≤1.0 DDD 236 157 668 0.97 (0.82 to 1.14) 0.79 (0.67 to 0.94) 0.007 >1.0 DDD 75 42 566 1.16 (0.91 to 1.49) 0.98 (0.76 to 1.26) 0.88 Test for trend§ - - - - 0.55 Others: ≤0.5 DDD 9 4026 1.40 (0.74 to 2.64) 0.98 (0.52 to 1.86) 0.95 >0.5 DDD/≤1.0 DDD 31 13 199 1.52 (1.08 to 2.15) 1.16 (0.81 to 1.65) 0.41 >1.0 DDD 20 8411 1.49 (0.97 to 2.29) 1.28 (0.84 to 1.97) 0.25 Test for trend§ - - - - 0.69 Antidepressant class by time since starting and stopping treatment No current or recent use 804 510 266 1.00 1.00 - TCAs: First 28 days 23 5482 2.56 (1.64 to 4.02) 1.99 (1.27 to 3.13) 0.003 29-84 days after starting 12 5400 1.36 (0.77 to 2.43) 1.04 (0.58 to 1.87) 0.89 ≥85 days after starting 44 18 941 1.52 (1.11 to 2.07) 0.91 (0.67 to 1.25) 0.57 1-28 days after stopping 11 3614 2.04 (1.15 to 3.62) 1.57 (0.86 to 2.86) 0.14 29-84 days after stopping 11 7030 1.02 (0.56 to 1.88) 0.85 (0.46 to 1.56) 0.60 85-182 days after stopping 15 10 711 1.00 (0.60 to 1.66) 0.79 (0.46 to 1.35) 0.39 SSRIs: First 28 days 44 20 639 1.31 (0.90 to 1.89) 1.23 (0.85 to 1.79) 0.28 29-84 days after starting 44 27 863 0.95 (0.66 to 1.37) 0.91 (0.63 to 1.32) 0.63 ≥85 days after starting 198 127 197 1.04 (0.88 to 1.23) 0.78 (0.66 to 0.92) 0.004 1-28 days after stopping 22 15 685 0.88 (0.58 to 1.36) 0.94 (0.61 to 1.44) 0.76 29-84 days after stopping 41 30 405 0.94 (0.70 to 1.26) 0.94

5 to 1.79) 0.28 29-84 days after starting 44 27 863 0.95 (0.66 to 1.37) 0.91 (0.63 to 1.32) 0.63 ≥85 days after starting 198 127 197 1.04 (0.88 to 1.23) 0.78 (0.66 to 0.92) 0.004 1-28 days after stopping 22 15 685 0.88 (0.58 to 1.36) 0.94 (0.61 to 1.44) 0.76 29-84 days after stopping 41 30 405 0.94 (0.70 to 1.26) 0.94 (0.69 to 1.27) 0.69 85-182 days after stopping 66 46 815 0.97 (0.75 to 1.27) 1.01 (0.77 to 1.33) 0.92 Others: First 28 days 7 2776 1.56 (0.75 to 3.23) 1.35 (0.65 to 2.80) 0.42 29-84 days after starting 7 3504 1.44 (0.71 to 2.91) 1.07 (0.50 to 2.30) 0.85 ≥85 days after starting 41 16 854 1.52 (1.13 to 2.04) 1.14 (0.85 to 1.54) 0.38 1-28 days after stopping 5 1573 2.00 (0.83 to 4.79) 1.86 (0.78 to 4.46) 0.16 29-84 days after stopping 6 3023 1.29 (0.58 to 2.88) 1.19 (0.54 to 2.65) 0.66 85-182 days after stopping 8 4537 1.16 (0.58 to 2.34) 1.09 (0.54 to 2.21) 0.80 DDD=defined daily dose; SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant. *Based on numbers in adjusted analysis.

(0.69 to 1.27) 0.69 85-182 days after stopping 66 46 815 0.97 (0.75 to 1.27) 1.01 (0.77 to 1.33) 0.92 Others: First 28 days 7 2776 1.56 (0.75 to 3.23) 1.35 (0.65 to 2.80) 0.42 29-84 days after starting 7 3504 1.44 (0.71 to 2.91) 1.07 (0.50 to 2.30) 0.85 ≥85 days after starting 41 16 854 1.52 (1.13 to 2.04) 1.14 (0.85 to 1.54) 0.38 1-28 days after stopping 5 1573 2.00 (0.83 to 4.79) 1.86 (0.78 to 4.46) 0.16 29-84 days after stopping 6 3023 1.29 (0.58 to 2.88) 1.19 (0.54 to 2.65) 0.66 85-182 days after stopping 8 4537 1.16 (0.58 to 2.34) 1.09 (0.54 to 2.21) 0.80 DDD=defined daily dose; SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant. *Based on numbers in adjusted analysis. †Adjusted for age, sex, year of diagnosis of depression, severity of depression, deprivation, smoking status, alcohol intake, ethnic group (white/not recorded or non-white), coronary heart disease, diabetes, hypertension, cancer, epilepsy/seizures, hypothyroidism, osteoarthritis, asthma/chronic obstructive pulmonary disease, stroke/transient ischaemic attack, rheumatoid arthritis, osteoporosis, liver disease, renal disease, obsessive-compulsive disorder, statins, non-steroidal anti-inflammatory drugs, aspirin, antihypertensives, anticonvulsants, hypnotics/anxiolytics, oral contraceptives, hormone replacement therapy, antipsychotics, bisphosphonates, anticoagulants. ‡Daily doses could not be evaluated for some prescriptions. §Test for trend uses continuous values of dose.

In the analysis of the 11 most commonly prescribed drugs, we found significant differences between the drugs overall (P=0.004) but no significant difference between the four tricyclic and related antidepressants (P=0.22) or the five selective serotonin reuptake inhibitors (P=0.39), although we saw a significantly decreased risk for fluoxetine (adjusted hazard ratio 0.74, 0.59 to 0.92; P=0.008) and some indication of an increased risk for lofepramine (1.67, 1.01 to 2.76; P=0.05) compared with periods of no antidepressant treatment (fig 1). Fig 1 Adjusted hazard ratios (compared with periods of non-use of antidepressants) for arrhythmia for individual antidepressant drugs over 5 years’ follow-up. SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant In an analysis of dose for individual antidepressants (table 3), rates of arrhythmia were not significantly increased for higher doses of citalopram (adjusted hazard ratio 1.08, 0.74 to 1.57, for doses >20 mg/day) or escitalopram (1.06, 0.52 to 2.16, for doses >10 mg/day), but we found a significant increase for lower doses of lofepramine (3.89, 1.92 to 7.90, for doses ≤105 mg/day) and a significantly reduced risk for lower doses of fluoxetine (0.72, 0.56 to 0.91, for doses ≤20 mg/day). Even for doses of citalopram of 40 mg/day or greater, we saw no significantly increased risk (adjusted hazard ratio 1.11, 0.72 to 1.71), although the number of events was small (n=28) (supplementary table C).

day) and a significantly reduced risk for lower doses of fluoxetine (0.72, 0.56 to 0.91, for doses ≤20 mg/day). Even for doses of citalopram of 40 mg/day or greater, we saw no significantly increased risk (adjusted hazard ratio 1.11, 0.72 to 1.71), although the number of events was small (n=28) (supplementary table C). Table 3 Unadjusted and adjusted hazard ratios for arrhythmia by individual drug categorised according to dose for 5 years’ follow-up*

day) and a significantly reduced risk for lower doses of fluoxetine (0.72, 0.56 to 0.91, for doses ≤20 mg/day). Even for doses of citalopram of 40 mg/day or greater, we saw no significantly increased risk (adjusted hazard ratio 1.11, 0.72 to 1.71), although the number of events was small (n=28) (supplementary table C). Table 3 Unadjusted and adjusted hazard ratios for arrhythmia by individual drug categorised according to dose for 5 years’ follow-up* Antidepressant drug No of events† Person years† Unadjusted hazard ratio (95% CI) Adjusted analysis‡ Hazard ratio (95% CI) P value No current use 887 568 365 1.00 1.00 - Tricyclic and related antidepressants Amitriptyline: ≤1 DDD 41 16 040 - - - Amitriptyline: >1 DDD 4 1442 - - - Dosulepin: ≤1 DDD 23 10 967 - - - Dosulepin: >1 DDD 1 205 - - - Lofepramine: ≤1 DDD 8 961 5.19 (2.55 to 10.54) 3.89 (1.92 to 7.90) <0.001 Lofepramine: >1 DDD 8 3394 1.49 (0.74 to 2.99) 1.17 (0.58 to 2.39) 0.66 Trazodone: ≤1 DDD 2 2139 - - - Trazodone: >1 DDD 1 19 - - - Selective serotonin reuptake inhibitors Citalopram: ≤1 DDD 115 72 340 1.04 (0.85 to 1.28) 0.82 (0.66 to 1.01) 0.06 Citalopram: >1 DDD 34 17 854 1.27 (0.88 to 1.83) 1.08 (0.74 to 1.57) 0.70 Escitalopram: ≤1 DDD 18 9068 1.31 (0.81 to 2.12) 1.04 (0.63 to 1.72) 0.88 Escitalopram: >1 DDD 7 3758 1.35 (0.69 to 2.64) 1.06 (0.52 to 2.16) 0.88 Fluoxetine: ≤1 DDD 91 68 345 0.84 (0.66 to 1.07) 0.72 (0.56 to 0.91) 0.007 Fluoxetine: >1 DDD 16 11 072 0.92 (0.56 to 1.53) 0.78 (0.48 to 1.27) 0.32 Paroxetine: ≤1 DDD 19 12 216 0.98 (0.62 to 1.57) 0.84 (0.53 to 1.34) 0.46 Paroxetine: >1 DDD 9 3398 1.72 (0.90 to 3.27) 1.47 (0.77 to 2.84) 0.25 Sertraline: ≤1 DDD 23 11 539 1.31 (0.86 to 2.01) 1.09 (0.70 to 1.68) 0.71 Sertraline: >1 DDD 9 6448 0.89 (0.47 to 1.70) 0.78 (0.41 to 1.49) 0.45 Others Mirtazapine: ≤1 DDD 20 7533 1.74 (1.13 to 2.70) 1.17 (0.75 to 1.84) 0.49 Mirtazapine: >1 DDD 6 1933 1.94 (0.89 to 4.23) 1.48 (0.67 to 3.26) 0.33 Venlafaxine: ≤1 DDD 18 8432 1.35 (0.86 to 2.12) 1.14 (0.72 to 1.81) 0.57 Venlafaxine: >1 DDD 14 6369 1.38 (0.82 to 2.32) 1.24 (0.74 to 2.08) 0.42 DDD=defined daily dose.

thers Mirtazapine: ≤1 DDD 20 7533 1.74 (1.13 to 2.70) 1.17 (0.75 to 1.84) 0.49 Mirtazapine: >1 DDD 6 1933 1.94 (0.89 to 4.23) 1.48 (0.67 to 3.26) 0.33 Venlafaxine: ≤1 DDD 18 8432 1.35 (0.86 to 2.12) 1.14 (0.72 to 1.81) 0.57 Venlafaxine: >1 DDD 14 6369 1.38 (0.82 to 2.32) 1.24 (0.74 to 2.08) 0.42 DDD=defined daily dose. DDD values are amitriptyline 75 mg/day; dosulepin 150 mg/day; lofepramine 105 mg/day; trazodone 300 mg/day; citalopram 20 mg/day; escitalopram 10 mg/day; fluoxetine 20 mg/day; paroxetine 20 mg/day; sertraline 50 mg/day; mirtazapine 30 mg/day; venlafaxine 100 mg/day. *Results only shown for drugs for which ≥5 events were recorded in both dose categories. †Based on numbers in adjusted analysis. ‡Adjusted for age, sex, year of diagnosis of depression, severity of depression, deprivation, smoking status, alcohol intake, ethnic group (white/not recorded or non-white), coronary heart disease, diabetes, hypertension, cancer, epilepsy/seizures, hypothyroidism, osteoarthritis, asthma/chronic obstructive pulmonary disease, stroke/transient ischaemic attack, rheumatoid arthritis, osteoporosis, liver disease, renal disease, obsessive-compulsive disorder, statins, non-steroidal anti-inflammatory drugs, aspirin, antihypertensives, anticonvulsants, hypnotics/anxiolytics, oral contraceptives, hormone replacement therapy, antipsychotics, bisphosphonates, anticoagulants.

attack, rheumatoid arthritis, osteoporosis, liver disease, renal disease, obsessive-compulsive disorder, statins, non-steroidal anti-inflammatory drugs, aspirin, antihypertensives, anticonvulsants, hypnotics/anxiolytics, oral contraceptives, hormone replacement therapy, antipsychotics, bisphosphonates, anticoagulants. Adjusted hazard ratios were similar when patients who had not received any prescriptions for antidepressants during follow-up were removed from the analysis (supplementary table D) and when the entire follow-up period was used (supplementary table E), although more associations were significant owing to larger numbers. When we used just the first year of follow-up (table 4), results were similar to the five year analysis, although the hazard ratio for combined antidepressant use was higher (adjusted hazard ratio 3.45, 1.24 to 9.57; P=0.017) and the association with fluoxetine was no longer statistically significant (0.79, 0.55 to 1.13; P=0.19). We found no indication of non-proportional hazards for the arrhythmia outcome; separate results for years 0-1, 1-3, and 3-5 of follow-up are shown in supplementary tables F and G. Table 4 Adjusted hazard ratios for arrhythmia, myocardial infarction, and stroke or transient ischaemic attack by antidepressant class, dose, and individual drug over first year of follow-up

Adjusted hazard ratios were similar when patients who had not received any prescriptions for antidepressants during follow-up were removed from the analysis (supplementary table D) and when the entire follow-up period was used (supplementary table E), although more associations were significant owing to larger numbers. When we used just the first year of follow-up (table 4), results were similar to the five year analysis, although the hazard ratio for combined antidepressant use was higher (adjusted hazard ratio 3.45, 1.24 to 9.57; P=0.017) and the association with fluoxetine was no longer statistically significant (0.79, 0.55 to 1.13; P=0.19). We found no indication of non-proportional hazards for the arrhythmia outcome; separate results for years 0-1, 1-3, and 3-5 of follow-up are shown in supplementary tables F and G. Table 4 Adjusted hazard ratios for arrhythmia, myocardial infarction, and stroke or transient ischaemic attack by antidepressant class, dose, and individual drug over first year of follow-up Arrhythmia Myocardial infarction Stroke/TIA No of events Adjusted hazard ratio† (95% CI) P value No of events Adjusted hazard ratio† (95% CI) P value No of events Adjusted hazard ratio† (95% CI) P value Antidepressant class No current use 127 1.00 - 90 1.00 - 113 1.00 - TCAs 39 1.16 (0.81 to 1.67) 0.42 25 1.09 (0.72 to 1.66) 0.68 33 1.01 (0.69 to 1.49) 0.94 SSRIs 141 0.86 (0.66 to 1.11) 0.24 63 0.58 (0.42 to 0.79) 0.001 118 0.83 (0.63 to 1.09) 0.18 Other antidepressants 20 1.33 (0.84 to 2.12) 0.23 9 0.81 (0.42 to 1.58) 0.54 16 1.15 (0.69 to 1.90) 0.59 Combined antidepressants 5 3.45 (1.24 to 9.57) 0.02 2 1.68 (0.43 to 6.65) 0.46 1 0.69 (0.10 to 4.96) 0.72 Antidepressant class and dose categories No current use 127 1.00 - 90 1.00 - 113 1.00 - TCAs: ≤0.5 DDD 21 0.98 (0.62 to 1.55) 0.92 12 0.86 (0.47 to 1.56) 0.62 18 0.87 (0.54 to 1.41) 0.58 >0.5 DDD/≤1.0 DDD 10 1.76 (0.92 to 3.35) 0.09 4 0.93 (0.35 to 2.50) 0.89 8 1.36 (0.66 to 2.78) 0.41 >1.0 DDD 4 1.22 (0.46 to 3.24) 0.69 3 1.29 (0.41 to 4.04) 0.66 4 1.26 (0.47 to 3.38) 0.65 Test for trend‡ - - 0.83 - - 0.47 - - 0.23 SSRIs: ≤0.5 DDD 11 0.95 (0.52 to 1.72) 0.85 5 0.76 (0.30 to 1.92) 0.56 7 0.73 (0.34 to 1.56) 0.42 >0.5 DDD/≤1.0 DDD 105 0.81 (0.62 to 1.08) 0.15 43 0.52 (0.37 to 0.73) <0.001 90 0.81 (0.61 to 1.09) 0.16 >1.0 DDD 21 1.07 (0.65 to 1.76) 0.79 11 0.75 (0.41 to 1.36) 0.34 17 0.99 (0.59 to 1.67) 0.98 Test for trend‡ - - 0.57 - - 0.42 - - 0.47 Others: ≤0.5 DDD 3 1.06 (0.34 to 3.32) 0.93 2 0.95 (0.23 to 3.96) 0.95 4 1.58 (0.57 to 4.35) 0.38 >0.5 DDD/≤1.0 DDD 13 1.65 (0.91 to 2.98) 0.10 3 0.53 (0.17 to 1.60) 0.26 7 0.95 (0.45 to 1.98) 0.88 >1.0 DDD 2 0.80 (0.20 to 3.20) 0.76 2 1.04 (0.26 to 4.17) 0.95 4 1.76 (0.66 to 4.73) 0.26 Test for trend‡ - - 0.51 - - 0.40 - - 0.72 Antidepressant drug No current use 130 1.00 90 1.00 113 1.00 TCAs: Amitriptyline 18 1.15 (0.69 to 1.94) 0.59 8 0.75 (0.37 to 1.55) 0.44 15 1.00 (0.59 to 1.70) 1.00 Dosulepin 8 0.73 (0.35 to 1.50) 0.39 8 1.07 (0.53 to 2.18) 0.85 12 1.12 (0.63 to 1.98) 0.70 Lofepramine 8 2.13 (1.05 to 4.33) 0.04 8 3.07 (1.50 to 6.26) 0.002 4 1.15 (0.43 to 3.11) 0.78 Trazodone 3 1.72 (0.53 to 5.56) 0.36 1 0.73 (0.10 to 5.19) 0.76 1 0.56 (0.08 to 3.72) 0.55 SSRIs

.55) 0.44 15 1.00 (0.59 to 1.70) 1.00 Dosulepin 8 0.73 (0.35 to 1.50) 0.39 8 1.07 (0.53 to 2.18) 0.85 12 1.12 (0.63 to 1.98) 0.70 Lofepramine 8 2.13 (1.05 to 4.33) 0.04 8 3.07 (1.50 to 6.26) 0.002 4 1.15 (0.43 to 3.11) 0.78 Trazodone 3 1.72 (0.53 to 5.56) 0.36 1 0.73 (0.10 to 5.19) 0.76 1 0.56 (0.08 to 3.72) 0.55 SSRIs : Citalopram 56 0.79 (0.57 to 1.10) 0.17 27 0.59 (0.39 to 0.91) 0.017 43 0.73 (0.51 to 1.05) 0.09 Escitalopram 9 1.01 (0.47 to 2.16) 0.99 4 0.67 (0.25 to 1.82) 0.43 5 0.63 (0.26 to 1.53) 0.31 Fluoxetine 48 0.79 (0.55 to 1.13) 0.19 18 0.44 (0.27 to 0.72) 0.001 56 1.06 (0.76 to 1.50) 0.72 Paroxetine 13 1.10 (0.61 to 1.99) 0.74 3 0.38 (0.12 to 1.22) 0.10 7 0.63 (0.28 to 1.38) 0.25 Sertraline 15 1.21 (0.71 to 2.07) 0.48 10 1.18 (0.64 to 2.20) 0.59 7 0.63 (0.30 to 1.35) 0.24 Others: Mirtazapine 8 1.20 (0.57 to 2.53) 0.62 5 0.91 (0.37 to 2.24) 0.84 12 1.85 (1.01 to 3.37) 0.04 Venlafaxine 11 1.64 (0.88 to 3.08) 0.12 4 0.89 (0.33 to 2.39) 0.81 3 0.51 (0.16 to 1.57) 0.24 All other antidepressants 3 0.90 (0.30 to 2.69) 0.85 1 0.46 (0.06 to 3.35) 0.44 2 0.64 (0.15 to 2.63) 0.53 Combined antidepressants 5 3.44 (1.24 to 9.55) 0.02 2 1.68 (0.43 to 6.64) 0.46 1 0.70 (0.10 to 4.97) 0.72 SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant; TIA=transient ischaemic attack. *Based on numbers in adjusted analysis.

: Citalopram 56 0.79 (0.57 to 1.10) 0.17 27 0.59 (0.39 to 0.91) 0.017 43 0.73 (0.51 to 1.05) 0.09 Escitalopram 9 1.01 (0.47 to 2.16) 0.99 4 0.67 (0.25 to 1.82) 0.43 5 0.63 (0.26 to 1.53) 0.31 Fluoxetine 48 0.79 (0.55 to 1.13) 0.19 18 0.44 (0.27 to 0.72) 0.001 56 1.06 (0.76 to 1.50) 0.72 Paroxetine 13 1.10 (0.61 to 1.99) 0.74 3 0.38 (0.12 to 1.22) 0.10 7 0.63 (0.28 to 1.38) 0.25 Sertraline 15 1.21 (0.71 to 2.07) 0.48 10 1.18 (0.64 to 2.20) 0.59 7 0.63 (0.30 to 1.35) 0.24 Others: Mirtazapine 8 1.20 (0.57 to 2.53) 0.62 5 0.91 (0.37 to 2.24) 0.84 12 1.85 (1.01 to 3.37) 0.04 Venlafaxine 11 1.64 (0.88 to 3.08) 0.12 4 0.89 (0.33 to 2.39) 0.81 3 0.51 (0.16 to 1.57) 0.24 All other antidepressants 3 0.90 (0.30 to 2.69) 0.85 1 0.46 (0.06 to 3.35) 0.44 2 0.64 (0.15 to 2.63) 0.53 Combined antidepressants 5 3.44 (1.24 to 9.55) 0.02 2 1.68 (0.43 to 6.64) 0.46 1 0.70 (0.10 to 4.97) 0.72 SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant; TIA=transient ischaemic attack. *Based on numbers in adjusted analysis. †Adjusted for age, sex, year of diagnosis of depression, severity of depression, deprivation, smoking status, alcohol intake, ethnic group (white/not recorded or non-white), coronary heart disease, diabetes, hypertension, cancer, epilepsy/seizures, hypothyroidism, osteoarthritis, asthma/chronic obstructive airways disease, stroke/transient ischaemic attack (except for the stroke/TIA outcome), rheumatoid arthritis, osteoporosis, liver disease, renal disease, obsessive-compulsive disorder, statins, non-steroidal anti-inflammatory drugs, aspirin, antihypertensives, anticonvulsants, hypnotics/anxiolytics, oral contraceptives, hormone replacement therapy, antipsychotics, bisphosphonates, anticoagulants.

utcome), rheumatoid arthritis, osteoporosis, liver disease, renal disease, obsessive-compulsive disorder, statins, non-steroidal anti-inflammatory drugs, aspirin, antihypertensives, anticonvulsants, hypnotics/anxiolytics, oral contraceptives, hormone replacement therapy, antipsychotics, bisphosphonates, anticoagulants. ‡Test for trend uses continuous values of dose. Associations with myocardial infarction At baseline, 1790 patients had a previous diagnosis of myocardial infarction recorded. We excluded these patients from analysis of the myocardial infarction outcome, along with the patients who received monoamine oxidase inhibitors, leaving 237 017 patients in the analysis cohort. During the first five years of follow-up, 772 new diagnoses of myocardial infarction were made, giving an incidence rate of 8.6 per 10 000 person years (16.2 per 10 000 in men and 3.9 per 10 000 in women). We found no significant association between antidepressant class and myocardial infarction over five years in the adjusted analysis (table 5) and no significant trends with dose. No clear pattern in risk according to different periods of time after starting or stopping antidepressant drugs was apparent, although rates were increased from 28 days after stopping tricyclic and related antidepressants. Table 5 Unadjusted and adjusted hazard ratios for myocardial infarction by antidepressant class, dose, and duration over 5 years’ follow-up

We found no significant association between antidepressant class and myocardial infarction over five years in the adjusted analysis (table 5) and no significant trends with dose. No clear pattern in risk according to different periods of time after starting or stopping antidepressant drugs was apparent, although rates were increased from 28 days after stopping tricyclic and related antidepressants. Table 5 Unadjusted and adjusted hazard ratios for myocardial infarction by antidepressant class, dose, and duration over 5 years’ follow-up No of events* Person years* Unadjusted hazard ratio (95% CI) Adjusted analysis† Hazard ratio (95% CI) P value Antidepressant class No current use 469 570 843 1.00 1.00 TCAs 63 41 295 1.83 (1.44 to 2.33) 1.20 (0.94 to 1.52) 0.14 SSRIs 182 225 863 1.02 (0.86 to 1.22) 0.85 (0.71 to 1.00) 0.06 Other antidepressants 33 28 144 1.39 (0.98 to 1.98) 1.00 (0.70 to 1.42) 0.98 Combined antidepressants 3 4224 0.84 (0.27 to 2.59) 0.57 (0.18 to 1.75) 0.32 Antidepressant class and dose categories No current use 469 570 843 1.00 1.00 TCAs: ≤0.5 DDD 31 23 555 1.59 (1.11 to 2.26) 1.02 (0.72 to 1.45) 0.89 >0.5 DDD/≤1.0 DDD 15 8412 2.15 (1.31 to 3.53) 1.29 (0.78 to 2.13) 0.32 >1.0 DDD 10 5318 2.24 (1.21 to 4.16) 1.59 (0.86 to 2.97) 0.14 Test for trend§ - - - - 0.35 SSRIs: ≤0.5 DDD 14 16 132 1.12 (0.68 to 1.86) 0.97 (0.57 to 1.63) 0.90 >0.5 DDD/≤1.0 DDD 110 158 252 0.89 (0.72 to 1.11) 0.73 (0.59 to 0.91) 0.005 >1.0 DDD 50 42 683 1.46 (1.11 to 1.92) 1.16 (0.88 to 1.54) 0.30 Test for trend§ - - - - 0.03 Others: ≤0.5 DDD 9 4041 2.65 (1.38 to 5.10) 1.80 (0.94 to 3.45) 0.08 >0.5 DDD/≤1.0 DDD 8 13 236 0.72 (0.36 to 1.43) 0.51 (0.26 to 1.02) 0.06 >1.0 DDD 11 8440 1.54 (0.86 to 2.78) 1.11 (0.61 to 2.00) 0.74 Test for trend§ - - - - 0.79 Antidepressant class by time since starting and stopping treatment No current or recent use 416 512 509 1.00 1.00 TCAs: First 28 days 6 5499 1.08 (0.48 to 2.44) 0.83 (0.37 to 1.86) 0.65 29-84 days after starting 5 5414 1.05 (0.44 to 2.51) 0.77 (0.32 to 1.83) 0.55 ≥85 days after starting 33 18 957 2.17 (1.56 to 3.00) 1.23 (0.89 to 1.71) 0.21 1-28 days after stopping 5 3627 1.60 (0.66 to 3.86) 1.30 (0.54 to 3.12) 0.56 29-84 days after stopping 13 7056 2.32 (1.32 to 4.06) 1.85 (1.05 to 3.23) 0.03 85-182 days after stopping 20 10 753 2.36 (1.47 to 3.78) 1.89 (1.18 to 3.02) 0.008 SSRIs: First 28 days 14 20 710 0.66 (0.35 to 1.25) 0.63 (0.32 to 1.22) 0.17 29-84 days after starting 14 27 967 0.59 (0.34 to 1.02) 0.56 (0.31 to 0.99) 0.05 ≥85 days after starting 109 127 711 1.12 (0.91 to 1.38) 0.84 (0.68 to 1.03) 0.10 1-28 days after stopping 20 15 744 1.64 (1.04 to 2.60) 1.66 (1.05 to 2.63) 0.03 29-84 days after stopping 22 30 521 0.96 (0.61 to 1.49) 1.00 (0.64 to 1.5

0.17 29-84 days after starting 14 27 967 0.59 (0.34 to 1.02) 0.56 (0.31 to 0.99) 0.05 ≥85 days after starting 109 127 711 1.12 (0.91 to 1.38) 0.84 (0.68 to 1.03) 0.10 1-28 days after stopping 20 15 744 1.64 (1.04 to 2.60) 1.66 (1.05 to 2.63) 0.03 29-84 days after stopping 22 30 521 0.96 (0.61 to 1.49) 1.00 (0.64 to 1.5 8) 0.98 85-182 days after stopping 33 47 004 0.95 (0.65 to 1.38) 0.99 (0.67 to 1.45) 0.95 Others: First 28 days 5 2788 1.91 (0.76 to 4.84) 1.52 (0.60 to 3.82) 0.37 29-84 days after starting 2 3514 0.67 (0.17 to 2.66) 0.53 (0.13 to 2.08) 0.36 ≥85 days after starting 20 16 908 1.44 (0.90 to 2.29) 0.96 (0.60 to 1.53) 0.87 1-28 days after stopping 1 1580 0.75 (0.11 to 5.35) 0.64 (0.09 to 4.54) 0.65 29-84 days after stopping 4 3036 1.64 (0.62 to 4.37) 1.38 (0.52 to 3.67) 0.52 85-182 days after stopping 5 4557 1.37 (0.56 to 3.33) 1.17 (0.48 to 2.85) 0.72 DDD=defined daily dose; SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant. *Based on numbers in adjusted analysis

8) 0.98 85-182 days after stopping 33 47 004 0.95 (0.65 to 1.38) 0.99 (0.67 to 1.45) 0.95 Others: First 28 days 5 2788 1.91 (0.76 to 4.84) 1.52 (0.60 to 3.82) 0.37 29-84 days after starting 2 3514 0.67 (0.17 to 2.66) 0.53 (0.13 to 2.08) 0.36 ≥85 days after starting 20 16 908 1.44 (0.90 to 2.29) 0.96 (0.60 to 1.53) 0.87 1-28 days after stopping 1 1580 0.75 (0.11 to 5.35) 0.64 (0.09 to 4.54) 0.65 29-84 days after stopping 4 3036 1.64 (0.62 to 4.37) 1.38 (0.52 to 3.67) 0.52 85-182 days after stopping 5 4557 1.37 (0.56 to 3.33) 1.17 (0.48 to 2.85) 0.72 DDD=defined daily dose; SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant. *Based on numbers in adjusted analysis †Adjusted for age, sex, year of diagnosis of depression, severity of depression, deprivation, smoking status, alcohol intake, ethnic group (white/not recorded or non-white), coronary heart disease, diabetes, hypertension, cancer, epilepsy/seizures, hypothyroidism, osteoarthritis, asthma/chronic obstructive pulmonary disease, stroke/transient ischaemic attack, rheumatoid arthritis, osteoporosis, liver disease, renal disease, obsessive-compulsive disorder, statins, non-steroidal anti-inflammatory drugs, aspirin, antihypertensives, anticonvulsants, hypnotics/anxiolytics, oral contraceptives, hormone replacement therapy, antipsychotics, bisphosphonates, anticoagulants. ‡Daily doses could not be evaluated for some prescriptions. §Test for trend uses continuous values of dose.

†Adjusted for age, sex, year of diagnosis of depression, severity of depression, deprivation, smoking status, alcohol intake, ethnic group (white/not recorded or non-white), coronary heart disease, diabetes, hypertension, cancer, epilepsy/seizures, hypothyroidism, osteoarthritis, asthma/chronic obstructive pulmonary disease, stroke/transient ischaemic attack, rheumatoid arthritis, osteoporosis, liver disease, renal disease, obsessive-compulsive disorder, statins, non-steroidal anti-inflammatory drugs, aspirin, antihypertensives, anticonvulsants, hypnotics/anxiolytics, oral contraceptives, hormone replacement therapy, antipsychotics, bisphosphonates, anticoagulants. ‡Daily doses could not be evaluated for some prescriptions. §Test for trend uses continuous values of dose. We found no significant associations (at P<0.01) for individual drugs in the adjusted analyses (fig 2) and no significant difference between the five selective serotonin reuptake inhibitors (P=0.27) or the four tricyclic and related antidepressants (P=0.26), although fluoxetine had an adjusted hazard ratio of 0.73 (0.54 to 0.98; P=0.04) and lofepramine had an adjusted hazard ratio of 2.02 (1.14 to 3.59; P=0.02), both compared with periods of no antidepressant treatment. Fig 2 Adjusted hazard ratios (compared with periods of non-use of antidepressants) for myocardial infarction for individual antidepressant drugs over 5 years’ follow-up. SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant

We found no significant associations (at P<0.01) for individual drugs in the adjusted analyses (fig 2) and no significant difference between the five selective serotonin reuptake inhibitors (P=0.27) or the four tricyclic and related antidepressants (P=0.26), although fluoxetine had an adjusted hazard ratio of 0.73 (0.54 to 0.98; P=0.04) and lofepramine had an adjusted hazard ratio of 2.02 (1.14 to 3.59; P=0.02), both compared with periods of no antidepressant treatment. Fig 2 Adjusted hazard ratios (compared with periods of non-use of antidepressants) for myocardial infarction for individual antidepressant drugs over 5 years’ follow-up. SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant Adjusted hazard ratios were similar when patients who had not received any antidepressant prescriptions during follow-up were removed from the analysis (supplementary table H) and when the entire follow-up period was used (supplementary table I). We saw some indication that hazard rates were not proportional over the five years of follow-up, and some differences in the hazard ratios were apparent when the analysis was restricted to the first year of follow-up compared with values over five years. In this one year analysis (table 4), we found a significantly reduced risk for selective serotonin reuptake inhibitors compared with no use of antidepressants (adjusted hazard ratio 0.58, 0.42 to 0.79; P=0.001); although overall no significant difference (at P<0.01) existed between the five selective serotonin reuptake inhibitors (P=0.11) or the four tricyclic and related antidepressants (P=0.03), a significant reduction occurred with fluoxetine (adjusted hazard ratio 0.44, 0.27 to 0.72; P=0.001) and a significant increase with lofepramine (3.07, 1.50 to 6.26; P=0.002). We found no significant associations with selective serotonin reuptake inhibitors in years 1-3 and 3-5 of follow-up (supplementary table F) or with fluoxetine (supplementary table G).

luoxetine (adjusted hazard ratio 0.44, 0.27 to 0.72; P=0.001) and a significant increase with lofepramine (3.07, 1.50 to 6.26; P=0.002). We found no significant associations with selective serotonin reuptake inhibitors in years 1-3 and 3-5 of follow-up (supplementary table F) or with fluoxetine (supplementary table G). Associations with stroke/transient ischaemic attack At baseline, 1741 patients had a diagnosis of stroke or transient ischaemic attack recorded. These patients were excluded from analysis of the stroke/transient ischaemic attack outcome, along with the patients who received monoamine oxidase inhibitors, leaving 237 067 patients in the analysis cohort. During the first five years of follow-up, 1106 new diagnoses of stroke or transient ischaemic attack were made, giving an incidence rate of 12.3 per 10 000 person years (17.3 per 10 000 in men and 9.3 per 10 000 in women). We found no significant associations between antidepressant class and stroke/transient ischaemic attack over five years and no significant trends (at P<0.01) with dose (table 6). A significant increase in risk occurred during the first 28 days after starting other antidepressants (adjusted hazard ratio 2.72, 1.45 to 5.08; P=0.002) and from 85 to 182 days after stopping tricyclic and related antidepressants (1.82, 1.21 to 2.74; P=0.004). Rates were also increased in the first 84 days after starting tricyclic and related antidepressants, although not significantly (at P<0.01).

her antidepressants (adjusted hazard ratio 2.72, 1.45 to 5.08; P=0.002) and from 85 to 182 days after stopping tricyclic and related antidepressants (1.82, 1.21 to 2.74; P=0.004). Rates were also increased in the first 84 days after starting tricyclic and related antidepressants, although not significantly (at P<0.01). Table 6 Unadjusted and adjusted hazard ratios for stroke or transient ischaemic attack by antidepressant class, dose, and duration over 5 years’ follow-up.

her antidepressants (adjusted hazard ratio 2.72, 1.45 to 5.08; P=0.002) and from 85 to 182 days after stopping tricyclic and related antidepressants (1.82, 1.21 to 2.74; P=0.004). Rates were also increased in the first 84 days after starting tricyclic and related antidepressants, although not significantly (at P<0.01). Table 6 Unadjusted and adjusted hazard ratios for stroke or transient ischaemic attack by antidepressant class, dose, and duration over 5 years’ follow-up. No of events* Person years* Unadjusted hazard ratio (95% CI) Adjusted analysis† Hazard ratio (95% CI) P value Antidepressant class No current use 610 570 879 1.00 1.00 TCAs 90 41 109 1.98 (1.56 to 2.52) 1.24 (0.98 to 1.58) 0.08 SSRIs 313 225 600 1.30 (1.12 to 1.51) 1.09 (0.93 to 1.27) 0.28 Other antidepressants 50 28 056 1.71 (1.30 to 2.25) 1.20 (0.91 to 1.60) 0.20 Combined antidepressants 11 4196 2.59 (1.47 to 4.55) 1.54 (0.86 to 2.78) 0.15 Antidepressant class and dose categories No current use 610 570 879 1.00 1.00 TCAs: ≤0.5 DDD 48 23 489 1.85 (1.36 to 2.50) 1.10 (0.81 to 1.49) 0.54 >0.5 DDD/≤1.0 DDD 24 8362 2.62 (1.76 to 3.88) 1.59 (1.06 to 2.37) 0.02 >1.0 DDD 12 5265 2.06 (1.13 to 3.76) 1.52 (0.84 to 2.76) 0.17 Test for trend§ - -  - - 0.27 SSRIs: ≤0.5 DDD 24 16 083 1.37 (0.88 to 2.11) 1.12 (0.72 to 1.73) 0.61 >0.5 DDD/≤1.0 DDD 216 158 042 1.28 (1.09 to 1.52) 1.06 (0.90 to 1.26) 0.47 >1.0 DDD 66 42 676 1.44 (1.12 to 1.87) 1.22 (0.94 to 1.59) 0.14 Test for trend§ - - - - 0.57 Others: ≤0.5 DDD 10 4017 2.25 (1.21 to 4.17) 1.54 (0.83 to 2.86) 0.17 >0.5 DDD/≤1.0 DDD 20 13 197 1.51 (0.99 to 2.29) 1.01 (0.65 to 1.57) 0.95 >1.0 DDD 13 8418 1.40 (0.82 to 2.38) 1.10 (0.65 to 1.87) 0.72 Test for trend§ - - - - 0.25 Antidepressant class by time since starting and stopping treatment No current or recent use 528 512 603 1.00 1.00 TCAs: First 28 days 14 5474 2.42 (1.35 to 4.37) 1.72 (0.95 to 3.10) 0.07 29-84 days after starting 16 5393 2.58 (1.56 to 4.26) 1.79 (1.08 to 2.97) 0.02 ≥85 days after starting 43 18 843 2.23 (1.64 to 3.02) 1.22 (0.90 to 1.67) 0.20 1-28 days after stopping 7 3619 1.78 (0.85 to 3.72) 1.37 (0.65 to 2.89) 0.40 29-84 days after stopping 10 7040 1.35 (0.72 to 2.53) 1.04 (0.56 to 1.95) 0.90 85-182 days after stopping 24 10 726 2.31 (1.54 to 3.47) 1.82 (1.21 to 2.74) 0.004 SSRIs: First 28 days 32 20 688 1.50 (0.96 to 2.36) 1.41 (0.89 to 2.23) 0.14 29-84 days after starting 34 27 938 1.04 (0.70 to 1.54) 1.00 (0.67 to 1.50) 0.99 ≥85 days after starting 183 127 522 1.46 (1.22 to 1.74) 1.10 (0.92 to 1.32) 0.30 1-28 days after stopping 22 15 737 1.36 (0.87 to 2.11) 1.43 (0.91 to 2.24) 0.12 29-84 days after stopping 38 30 508 1.21 (0.87 to 1.68) 1.32 (0.95 t

.23) 0.14 29-84 days after starting 34 27 938 1.04 (0.70 to 1.54) 1.00 (0.67 to 1.50) 0.99 ≥85 days after starting 183 127 522 1.46 (1.22 to 1.74) 1.10 (0.92 to 1.32) 0.30 1-28 days after stopping 22 15 737 1.36 (0.87 to 2.11) 1.43 (0.91 to 2.24) 0.12 29-84 days after stopping 38 30 508 1.21 (0.87 to 1.68) 1.32 (0.95 t o 1.85) 0.10 85-182 days after stopping 55 46 983 1.30 (0.98 to 1.74) 1.35 (1.01 to 1.81) 0.04 Others: First 28 days 10 2781 3.71 (2.04 to 6.75) 2.72 (1.45 to 5.08) 0.002 29-84 days after starting 7 3505 1.84 (0.88 to 3.84) 1.48 (0.70 to 3.10) 0.30 ≥85 days after starting 27 16 854 1.64 (1.13 to 2.39) 1.07 (0.72 to 1.58) 0.74 1-28 days after stopping 4 1574 2.40 (0.90 to 6.37) 2.15 (0.81 to 5.70) 0.13 29-84 days after stopping 2 3024 0.64 (0.16 to 2.53) 0.58 (0.15 to 2.28) 0.43 85-182 days after stopping 7 4542 1.76 (0.88 to 3.52) 1.43 (0.68 to 3.00) 0.35 DDD=defined daily dose; SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant. *Based on numbers in adjusted analysis.

o 1.85) 0.10 85-182 days after stopping 55 46 983 1.30 (0.98 to 1.74) 1.35 (1.01 to 1.81) 0.04 Others: First 28 days 10 2781 3.71 (2.04 to 6.75) 2.72 (1.45 to 5.08) 0.002 29-84 days after starting 7 3505 1.84 (0.88 to 3.84) 1.48 (0.70 to 3.10) 0.30 ≥85 days after starting 27 16 854 1.64 (1.13 to 2.39) 1.07 (0.72 to 1.58) 0.74 1-28 days after stopping 4 1574 2.40 (0.90 to 6.37) 2.15 (0.81 to 5.70) 0.13 29-84 days after stopping 2 3024 0.64 (0.16 to 2.53) 0.58 (0.15 to 2.28) 0.43 85-182 days after stopping 7 4542 1.76 (0.88 to 3.52) 1.43 (0.68 to 3.00) 0.35 DDD=defined daily dose; SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant. *Based on numbers in adjusted analysis. †Adjusted for age, sex, year of diagnosis of depression, severity of depression, deprivation, smoking status, alcohol intake, ethnic group (white/not recorded or non-white), coronary heart disease, diabetes, hypertension, cancer, epilepsy/seizures, hypothyroidism, osteoarthritis, asthma/chronic obstructive pulmonary disease, rheumatoid arthritis, osteoporosis, liver disease, renal disease, obsessive-compulsive disorder, statins, non-steroidal anti-inflammatory drugs, aspirin, antihypertensives, anticonvulsants, hypnotics/anxiolytics, oral contraceptives, hormone replacement therapy, antipsychotics, bisphosphonates, anticoagulants. ‡Daily doses could not be evaluated for some prescriptions. §Test for trend uses continuous values of dose. In the adjusted analysis of individual antidepressant drugs, we found no significant associations for any of the drugs (fig 3).

†Adjusted for age, sex, year of diagnosis of depression, severity of depression, deprivation, smoking status, alcohol intake, ethnic group (white/not recorded or non-white), coronary heart disease, diabetes, hypertension, cancer, epilepsy/seizures, hypothyroidism, osteoarthritis, asthma/chronic obstructive pulmonary disease, rheumatoid arthritis, osteoporosis, liver disease, renal disease, obsessive-compulsive disorder, statins, non-steroidal anti-inflammatory drugs, aspirin, antihypertensives, anticonvulsants, hypnotics/anxiolytics, oral contraceptives, hormone replacement therapy, antipsychotics, bisphosphonates, anticoagulants. ‡Daily doses could not be evaluated for some prescriptions. §Test for trend uses continuous values of dose. In the adjusted analysis of individual antidepressant drugs, we found no significant associations for any of the drugs (fig 3). Fig 3 Adjusted hazard ratios (compared with periods of non-use of antidepressants) for stroke or transient ischaemic attack for individual antidepressant drugs over 5 years’ follow-up. SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant

In the adjusted analysis of individual antidepressant drugs, we found no significant associations for any of the drugs (fig 3). Fig 3 Adjusted hazard ratios (compared with periods of non-use of antidepressants) for stroke or transient ischaemic attack for individual antidepressant drugs over 5 years’ follow-up. SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant Adjusted hazard ratios were similar when patients who had not received any prescriptions for antidepressants during follow-up were removed (supplementary table J) and when the entire follow-up period was used (supplementary table K), but they tended to be lower when just the first year of follow-up was used in the analysis (table 4). We saw some indication that hazard rates were not proportional over the five years of follow- up, with higher hazard ratios in the later periods of follow-up for tricyclic and related antidepressants and selective serotonin reuptake inhibitors (supplementary tables F and G). Additional analyses The results of analyses including confounding variables in blocks are shown in supplementary tables L to N, showing that adjustment for age, sex, deprivation, ethnic group, and year of diagnosis had a marked effect on hazard ratios, but additional adjustment for further blocks of variables had a relatively small effect. Results were similar to those of our main models which used robust standard errors when the Cox models were stratified by general practice.

sex, deprivation, ethnic group, and year of diagnosis had a marked effect on hazard ratios, but additional adjustment for further blocks of variables had a relatively small effect. Results were similar to those of our main models which used robust standard errors when the Cox models were stratified by general practice. Absolute risks Table 7 shows absolute risks of the three outcomes over one year by antidepressant class and for the individual drugs. Absolute risks of arrhythmia and myocardial infarction were highest for lofepramine (30 per 10 000 and 31 per 10 000, respectively), and for stroke/transient ischaemic attack they were highest for mirtazapine (24 per 10 000). However, the 95% confidence intervals for these values were wide and mainly overlapped with the other drugs. Table 7 Absolute risks of arrhythmia, myocardial infarction, and stroke or transient ischaemic attack over 1 year by antidepressant class and for individual drugs.

Absolute risks Table 7 shows absolute risks of the three outcomes over one year by antidepressant class and for the individual drugs. Absolute risks of arrhythmia and myocardial infarction were highest for lofepramine (30 per 10 000 and 31 per 10 000, respectively), and for stroke/transient ischaemic attack they were highest for mirtazapine (24 per 10 000). However, the 95% confidence intervals for these values were wide and mainly overlapped with the other drugs. Table 7 Absolute risks of arrhythmia, myocardial infarction, and stroke or transient ischaemic attack over 1 year by antidepressant class and for individual drugs. Treatment Absolute risk per 10,000 over 1 year (95% CI) Arrhythmia* Myocardial infarction† Stroke/TIA‡ No treatment 14 (11 to 17) 10 (8 to 12) 13 (11 to 16) Antidepressant class TCAs 16 (11 to 23) 11 (7 to 17) 13 (9 to 19) SSRIs 12 (9 to 16) 6 (4 to 8) 11 (8 to 14) Other antidepressants 19 (12 to 30) 8 (4 to 16) 15 (9 to 25) Combined antidepressants 48 (17 to 133) 17 (4 to 66) 9 (1 to 64) Antidepressant drug TCAs: Amitriptyline 16 (10 to 27) 8 (4 to 16) 13 (8 to 22) Dosulepin 10 (5 to 21) 11 (5 to 22) 15 (8 to 26) Lofepramine 30 (15 to 60) 31 (15 to 62) 15 (6 to 40) Trazodone 24 (7 to 78) 7 (1 to 52) 7 (1 to 48) SSRIs: Citalopram 11 (8 to 15) 6 (4 to 9) 10 (7 to 14) Escitalopram 14 (7 to 30) 7 (2 to 18) 8 (3 to 20) Fluoxetine 11 (8 to 16) 4 (3 to 7) 14 (10 to 19) Paroxetine 15 (9 to 28) 4 (1 to 12) 8 (4 to 18) Sertraline 17 (10 to 29) 12 (6 to 22) 8 (4 to 18) Others: Mirtazapine 17 (8 to 35) 9 (4 to 22) 24 (13 to 44) Venlafaxine 23 (12 to 43) 9 (3 to 24) 7 (2 to 20) All other antidepressants 13 (4 to 38) 5 (1 to 33) 8 (2 to 34) SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant; TIA=transient ischaemic attack.

29) 12 (6 to 22) 8 (4 to 18) Others: Mirtazapine 17 (8 to 35) 9 (4 to 22) 24 (13 to 44) Venlafaxine 23 (12 to 43) 9 (3 to 24) 7 (2 to 20) All other antidepressants 13 (4 to 38) 5 (1 to 33) 8 (2 to 34) SSRI=selective serotonin reuptake inhibitor; TCA=tricyclic and related antidepressant; TIA=transient ischaemic attack. *Absolute risks are adjusted for confounders listed in table 2. †Absolute risks are adjusted for confounders listed in table 5. ‡Absolute risks are adjusted for confounders listed in table 6. Discussion The main findings of this large population based cohort study were that selective serotonin reuptake inhibitors were not associated with an increased risk of arrhythmia, myocardial infarction, or stroke or transient ischaemic attack in a general population cohort of people with depression aged 20 to 64 and that risk of arrhythmia was not significantly increased in patients treated with citalopram even at high doses (40 mg/day and over), although numbers in this category were relatively small. We found some evidence that selective serotonin reuptake inhibitors were associated with a reduced risk of arrhythmia and myocardial infarction. Fluoxetine was associated with the lowest risks of these two outcomes, but overall no significant differences were seen between the selective serotonin reuptake inhibitors. The risk of arrhythmia was significantly increased in the first four weeks of starting tricyclic and related antidepressants, and the tricyclic drug lofepramine was associated with a significantly increased risk of myocardial infarction in the first year of follow-up.

etween the selective serotonin reuptake inhibitors. The risk of arrhythmia was significantly increased in the first four weeks of starting tricyclic and related antidepressants, and the tricyclic drug lofepramine was associated with a significantly increased risk of myocardial infarction in the first year of follow-up. Strengths and limitations of study This study included a large representative sample of people aged 20 to 64 diagnosed as having depression in the general UK population and had a long follow-up period. All eligible patients were included, so no bias due to non-response was present, and no recall bias occurred because data on prescriptions and confounding variables were recorded prospectively before the outcomes occurred. We reduced indication bias by restricting our cohort to include only patients with a diagnosis of depression, as depression itself is an established risk factor for cardiovascular outcomes,30 31 and separating the effects of antidepressant treatment from those of depression would otherwise be difficult. This means that our findings can be generalised only to people diagnosed as having depression.

ith a diagnosis of depression, as depression itself is an established risk factor for cardiovascular outcomes,30 31 and separating the effects of antidepressant treatment from those of depression would otherwise be difficult. This means that our findings can be generalised only to people diagnosed as having depression. Some bias may remain in comparisons between antidepressant drugs if the selection of a particular antidepressant was influenced by risk factors for the outcome, but we accounted for a large number of potential confounding variables in the analysis to reduce differences between comparison groups. The increased risk for lofepramine in some analyses may nevertheless reflect preferential selection of this drug in patients considered to be more prone to arrhythmias or heart disease, as this drug is viewed as being safer in overdose and less cardiotoxic than other tricyclic and related antidepressants.32 33 The increased risk of arrhythmia for low doses of lofepramine but not higher doses supports this, whereby patients at highest risk are treated with lower doses, although numbers of events were small in both dose categories. However, in a comparison of baseline characteristics of patients who received prescriptions for different antidepressants, we saw no indication that lofepramine was prescribed more frequently than other tricyclic antidepressants to patients with cardiovascular risk factors.22 For example, among patients whose first antidepressant prescription was for lofepramine, 1.1% had coronary heart disease compared with 2.1% for amitriptyline, and 0.8% had a previous stroke recorded compared with 1.0% for amitriptyline. Similarly, we observed no indication that fluoxetine was prescribed more frequently than other selective serotonin reuptake inhibitors to younger patients or patients with fewer cardiovascular risk factors. For example, the mean age of patients when first treated with fluoxetine was 38.8 years, compared with 39.8 for citalopram and 38.3 for paroxetine, and the proportion of patients with hypertension when first treated with fluoxetine was 6.7%, whereas for paroxetine it was 5.3%.22

with fewer cardiovascular risk factors. For example, the mean age of patients when first treated with fluoxetine was 38.8 years, compared with 39.8 for citalopram and 38.3 for paroxetine, and the proportion of patients with hypertension when first treated with fluoxetine was 6.7%, whereas for paroxetine it was 5.3%.22 Some residual confounding may still be present owing to variables that either were not recorded on the database, such as dietary factors and physical activity, or were not recorded in sufficient detail for their confounding effect to be completely removed by analysis. Although we adjusted for severity of depression, this was based on a basic classification of diagnostic Read codes for depression, as depression severity scores are not routinely recorded in general practice. Numbers of patients in the different non-white ethnic groups were small, so we combined these for inclusion in the analysis, which may contribute to residual confounding. Some misclassification of the antidepressant exposure variables will have occurred, as some patients may not have taken their prescribed antidepressant or may not have taken it at the prescribed dose. This misclassification could underestimate associations with drug use. Furthermore, although the cohort was large, the number of events was small for some of the antidepressant exposure categories. In particular, there were relatively few prescriptions for citalopram at doses of 40 mg/day or more (19% of citalopram prescriptions), and only 28 diagnoses of arrhythmia in this category, so the 95% confidence interval for risk of arrhythmia with high doses of citalopram is wide, and increases in risk of up to 71% cannot be excluded.

here were relatively few prescriptions for citalopram at doses of 40 mg/day or more (19% of citalopram prescriptions), and only 28 diagnoses of arrhythmia in this category, so the 95% confidence interval for risk of arrhythmia with high doses of citalopram is wide, and increases in risk of up to 71% cannot be excluded. The outcomes were not formally adjudicated in this study, but validation studies in other UK primary care databases have shown high levels of validity across a range of diseases, and we would expect levels of validity to be similar in QResearch.34 35 For example, Khan reported high positive predictive values in validation studies of acute myocardial infarction and cerebrovascular disease.35 A study validating diagnostic codes for ventricular arrhythmia and sudden cardiac death reported a positive predictive value of 93%.36 We included information from death certificates to identify additional patients with the outcomes, which will have increased ascertainment and reduced misclassification.

ular disease.35 A study validating diagnostic codes for ventricular arrhythmia and sudden cardiac death reported a positive predictive value of 93%.36 We included information from death certificates to identify additional patients with the outcomes, which will have increased ascertainment and reduced misclassification. Comparison with other studies Our results for arrhythmia are consistent with those of two other large cohort studies in finding no increased risk for citalopram,18 19 even at high doses, and our rates of arrhythmia are of the same order of magnitude. Our study adds new information on risks associated with other antidepressant drugs and on effects of duration of treatment. Our findings contrast to some extent with those of studies that have found QT interval prolongation in patients taking citalopram.14 15 16 One cross sectional study,15 which included 38 397 patients aged 18 and over with an electrocardiogram recorded after prescription of antidepressant or methadone, found that QT prolongation was associated with dose for citalopram, escitalopram, and amitriptyline but not for other antidepressants examined. A study of psychiatric inpatients aged 18 and over found that most people with QT prolongation had two or more risk factors for QT prolongation, such as hypokalaemia, HIV infection, abnormal T wave morphology, and alcohol or drug use disorders, and that citalopram (including escitalopram) was significantly associated with QT prolongation after adjustment for these factors.16 This lack of coherence may reflect the smaller numbers of arrhythmia outcomes in the cohort studies when split by antidepressant drug and dose. Thus, power to detect an increased risk among higher antidepressant dose categories is low in comparison with studies that measure QT interval in adults receiving different doses of antidepressants and treat it as a continuous outcome variable in the analyses.14 15 Torsades de pointes, which is the type of arrhythmia most closely related to QT interval prolongation, is extremely rare, so cohort studies (including ours) cannot rule out an association for this particular type of arrhythmia. Furthermore, a surrogate measure such as QT interval may not necessarily translate into an effect on a clinically important outcome such as arrhythmia.

losely related to QT interval prolongation, is extremely rare, so cohort studies (including ours) cannot rule out an association for this particular type of arrhythmia. Furthermore, a surrogate measure such as QT interval may not necessarily translate into an effect on a clinically important outcome such as arrhythmia. Our findings of an increased risk of arrhythmia in the first four weeks of starting a tricyclic antidepressant is consistent with several potential arrhythmias that can occur with tricyclic overdose in people with previously unsuspected cardiac abnormalities such as bundle branch block37 38; our findings are important, as few studies have examined this for prescribed doses of tricyclic antidepressants.

tricyclic antidepressant is consistent with several potential arrhythmias that can occur with tricyclic overdose in people with previously unsuspected cardiac abnormalities such as bundle branch block37 38; our findings are important, as few studies have examined this for prescribed doses of tricyclic antidepressants. In our previous study of antidepressants in people aged 65 and over with depression,10 25 we found a significantly increased risk of myocardial infarction with selective serotonin reuptake inhibitors but not with tricyclic or other antidepressants. Other observational studies have found similar results for selective serotonin reuptake inhibitors,39 40 whereas several have found no association11 12 41 42 or a reduced risk13 43 44; few studies have assessed risks for individual antidepressants. A meta-analysis of 16 observational studies concluded that use of neither selective serotonin reuptake inhibitors nor tricyclic antidepressants is associated with an increased risk of coronary heart disease,45 but only two studies were restricted to patients with depression. These contradictory findings are likely to be due to differences between studies, as they vary considerably in their sizes and inclusion criteria. Several studies either did not restrict their study sample to patients with depression or did not account for depression in the analysis and so are highly susceptible to indication bias because depression is a strong risk factor for cardiovascular disease11 12 13; some studies are only in older or postmenopausal populations10 39 42; and one was an interview based case-control study prone to recall bias.44 Why our results differ from those of our previous study in older people, which had a very similar study design, is unclear,10 but it could be due to the larger number of myocardial infarction events (n=2350) in the older cohort or increased susceptibility to side effects in older people resulting from age related pharmacokinetic changes,46 or the high prevalence of multimorbidity and use of concomitant drugs in older people may result in interactions giving different patterns of risk with antidepressant use.

vents (n=2350) in the older cohort or increased susceptibility to side effects in older people resulting from age related pharmacokinetic changes,46 or the high prevalence of multimorbidity and use of concomitant drugs in older people may result in interactions giving different patterns of risk with antidepressant use. Observational studies of antidepressants and stroke have shown a more consistent pattern; several studies have found an increased risk of stroke with selective serotonin reuptake inhibitor use.10 42 47 48 49 A systematic review and meta-analysis of 13 observational studies of selective serotonin reuptake inhibitors and stroke reported that selective serotonin reuptake inhibitors were associated with an increased risk of all types of stroke (overall adjusted odds ratio 1.40, 95% confidence interval 1.09 to 1.80) and that the risk was still significantly increased when the analysis was restricted to the studies in which potential confounding by depression was considered.9 In a subgroup analysis by age group, the combined odds ratio for all types of stroke associated with selective serotonin reuptake inhibitor use was significant only in the four studies restricted to people aged at least 50 years (overall adjusted odds ratio 1.58, 1.06 to 2.36),10 42 50 51 and no significantly increased risk was seen in studies with no age restriction (overall adjusted odds ratio 1.13, 0.91 to 1.39). This concurs with our findings in this study of no association between selective serotonin reuptake inhibitors and stroke in people aged 20 to 64 and of an increased risk in our previous study in people aged 65 and over.10

isk was seen in studies with no age restriction (overall adjusted odds ratio 1.13, 0.91 to 1.39). This concurs with our findings in this study of no association between selective serotonin reuptake inhibitors and stroke in people aged 20 to 64 and of an increased risk in our previous study in people aged 65 and over.10 Clinical implications and future research Prescription of antidepressants is a complex process, involving balancing of risks and benefits for different antidepressants and doses, accounting for severity of depression, and considering patients’ risk factors, comorbidities, and preferences. The results of this study in adults aged 20 to 64 are reassuring in light of recent concerns about citalopram and potential risk of arrhythmia; however, as only small numbers of patients were treated with high doses of citalopram, we cannot rule out the possibility of an increased risk. We suggest that high doses of citalopram should not be prescribed without a strong indication, particularly in patients with any risk factors for an increased QT interval. We also found no evidence that selective serotonin reuptake inhibitors are associated with an increased risk of myocardial infarction or stroke/transient ischaemic attack in this age group; they may even be associated with a reduced risk of myocardial infarction and arrhythmia, particularly for fluoxetine. The potential cardioprotective effects of selective serotonin reuptake inhibitors, particularly fluoxetine, warrant further investigation.

arction or stroke/transient ischaemic attack in this age group; they may even be associated with a reduced risk of myocardial infarction and arrhythmia, particularly for fluoxetine. The potential cardioprotective effects of selective serotonin reuptake inhibitors, particularly fluoxetine, warrant further investigation. The risk of arrhythmia was increased during the first 28 days of taking tricyclic and related antidepressants, and among the antidepressants studied lofepramine had the highest risks of arrhythmia, myocardial infarction, and stroke/transient ischaemic attack. This finding may reflect selective prescribing of lofepramine, as it is generally considered to be safer than other tricyclic and related antidepressants in overdose, but could also indicate increased risks when it is taken at doses typically prescribed in primary care. Further research using other designs such as the self controlled case series approach may help to elucidate this association.

rally considered to be safer than other tricyclic and related antidepressants in overdose, but could also indicate increased risks when it is taken at doses typically prescribed in primary care. Further research using other designs such as the self controlled case series approach may help to elucidate this association. Conclusions This large observational study has found no evidence that selective serotonin reuptake inhibitors are associated with an increased risk of arrhythmia, myocardial infarction, or stroke/transient ischaemic attack in people with depression aged 20 to 64, but some indication that they are associated with a reduced risk of myocardial infarction and arrhythmia, particularly for fluoxetine. Citalopram was not significantly associated with an increased risk of arrhythmia, even at higher doses, although the confidence interval was wide. These findings are reassuring in light of recent safety concerns about selective serotonin reuptake inhibitors. What is already known on this topic Depression is a common condition, and antidepressants—particularly selective serotonin reuptake inhibitors—are increasingly used in its treatment Rates of cardiovascular disease are higher in people with depression, but whether different antidepressant treatments increase or reduce these rates is unclear High doses of certain antidepressants, including citalopram, can cause QT prolongation, which may increase the risk of arrhythmia, but this is not established

What is already known on this topic Depression is a common condition, and antidepressants—particularly selective serotonin reuptake inhibitors—are increasingly used in its treatment Rates of cardiovascular disease are higher in people with depression, but whether different antidepressant treatments increase or reduce these rates is unclear High doses of certain antidepressants, including citalopram, can cause QT prolongation, which may increase the risk of arrhythmia, but this is not established What this study adds This study found no evidence that selective serotonin reuptake inhibitors as a class are associated with an increased risk of arrhythmia and stroke or transient ischaemic attack in people with depression aged 20 to 64 No evidence was found that citalopram is associated with a significantly increased risk of arrhythmia, even at high doses Some indication was seen of a reduced risk of myocardial infarction for selective serotonin reuptake inhibitors, particularly fluoxetine Web Extra Extra material supplied by the author Appendix Click here for additional data file. Supplementary tables Click here for additional data file. We acknowledge the contribution of practices that contribute to the QResearch, as well as Egton Medical Information Systems (EMIS) and the University of Nottingham for expertise in establishing, developing, and supporting the database. We acknowledge the Office of National Statistics for providing mortality data.

Click here for additional data file. We acknowledge the contribution of practices that contribute to the QResearch, as well as Egton Medical Information Systems (EMIS) and the University of Nottingham for expertise in establishing, developing, and supporting the database. We acknowledge the Office of National Statistics for providing mortality data. Contributors: CC, JH-C, RM, AA, and MM contributed to the overall conception and design of the study. CC wrote the first draft of this manuscript. JH-C did the data extraction. TH and CC did the statistical analyses. All authors contributed to interpretation of results and drafting of this manuscript. All authors read and approved the final manuscript. CC is the guarantor. Funding: The project was funded by the National Institute for Health Research (NIHR) School for Primary Care Research (project number 81). The funding body did not play a role in the study design, the writing of the manuscript, or the decision to submit the manuscript for publication. This paper presents independent research funded by the NIHR. The views expressed are those of the authors and not necessarily those of the NHS, the NIHR, or the Department of Health. RM’s contribution to the study has been funded through the NIHR Collaboration for Leadership in Applied Health Research and Care East Midlands (CLAHRC EM).

s independent research funded by the NIHR. The views expressed are those of the authors and not necessarily those of the NHS, the NIHR, or the Department of Health. RM’s contribution to the study has been funded through the NIHR Collaboration for Leadership in Applied Health Research and Care East Midlands (CLAHRC EM). Competing interests: All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf (available on request from the corresponding author) and declare: financial support from NIHR for the submitted work; JH-C is director of QResearch, which is a not for profit venture between the University of Nottingham and EMIS (commercial supplier of GP clinical systems); no financial relationships with any organisations that might have an interest in the submitted work in the previous three years; no other relationships or activities that could appear to have influenced the submitted work. Ethical approval: The project has been independently peer reviewed and accepted by the QResearch Scientific board and has been approved in accordance with the agreed procedure with the Trent Research Ethics Committee (reference number: MREC/03/4/021). Transparency declaration: The lead author (the manuscript’s guarantor) affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained.

ation: The lead author (the manuscript’s guarantor) affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained. Data sharing: The patient level data from the QResearch are specifically licensed according to its governance framework. See www.qresearch.org for further details.

Introduction Stroke accounts for about 10% of deaths internationally and for more than 4% of direct healthcare costs in developed countries.1 If other resources, such as lost productivity, benefit payments, and informal care costs, are taken into account, the total costs double. For example, in the United Kingdom, annual care costs are around £4.4bn (€5.7bn; $6.4bn), but total costs are £9bn a year.2 More than 20% of strokes are recurrent events,3 and if one also takes into account previous history of transient ischaemic attack (TIA), this figure rises to about 30%.1 Therefore, secondary prevention has a major potential role to play in reducing both morbidity and costs of stroke care. Hypertension is a key risk factor for stroke. A 20 mm Hg difference in usual systolic blood pressure is associated with a 60% lower risk of death from stroke in someone aged 50-70 years and a 50% lower risk in someone aged 70-79.4

n has a major potential role to play in reducing both morbidity and costs of stroke care. Hypertension is a key risk factor for stroke. A 20 mm Hg difference in usual systolic blood pressure is associated with a 60% lower risk of death from stroke in someone aged 50-70 years and a 50% lower risk in someone aged 70-79.4 The PROGRESS trial showed that treatment to lower blood pressure in people who have had a stroke or TIA reduces the risk of further stroke.5 However, the best way to apply this evidence in clinical practice is debated.6 7 In particular, uncertainty exists about how intensively to lower blood pressure in people who have had a stroke or TIA.8 A post hoc observational analysis of the PROFESS trial found that people with recent ischaemic stroke whose systolic blood pressure was less than 130 mm Hg had a higher risk of vascular events than those with a blood pressure between 1300 and 140 mm Hg.9 Conversely, participants in PROGRESS whose baseline systolic blood pressure was 120-140 mm Hg and who were randomised to combination therapy had a significantly reduced risk of stroke.10 The SPS3 trial of different blood pressure targets in younger (mean age 63 years) patients with recent lacunar stroke found a non-significant 19% reduction in risk of stroke after one year in people treated with a systolic blood pressure target of less than 130 mm Hg compared with a 130-149 mm Hg target.11 Recent guidelines have drawn different conclusions from the evidence base; the European guidelines recommend a target systolic blood pressure of 140 mm Hg (or higher),12 and British guidelines recommend a target of 130 mm Hg.13

ystolic blood pressure target of less than 130 mm Hg compared with a 130-149 mm Hg target.11 Recent guidelines have drawn different conclusions from the evidence base; the European guidelines recommend a target systolic blood pressure of 140 mm Hg (or higher),12 and British guidelines recommend a target of 130 mm Hg.13 In view of these controversies, the Prevention After Stroke—Blood Pressure (PAST-BP) study compared two different targets for blood pressure lowering after stroke or TIA in people recruited from a prevalent primary care population.14 The aim was to determine whether setting a more intensive target in primary care would lead to a lower blood pressure, as a prelude to a trial powered to detect whether such a strategy would lead to a reduction in recurrence of stroke.

after stroke or TIA in people recruited from a prevalent primary care population.14 The aim was to determine whether setting a more intensive target in primary care would lead to a lower blood pressure, as a prelude to a trial powered to detect whether such a strategy would lead to a reduction in recurrence of stroke. Methods Participants The methods used in PAST-BP have been reported in detail elsewhere.14 PAST-BP was an individually randomised trial in which participants were allocated to either an intensive blood pressure target (<130 mm Hg or a 10 mm Hg reduction if baseline pressure was <140 mm Hg) or a standard target (<140 mm Hg). Patients were recruited from 106 general practices (of which 99 contributed at least one patient) in England during 2009-11. Patients were considered for inclusion if they were on the practice’s TIA/stroke register. They were excluded if their baseline systolic blood pressure was less than 125 mm Hg, they were already taking three or more antihypertensive agents, they had a greater than 20 mm Hg postural change in systolic blood pressure on standing, they were already being treated to a 130 mm Hg systolic blood pressure target, they were unable to provide informed consent, or there was insufficient corroborative evidence that they had had a stroke or TIA. Potentially eligible participants were identified using a search of the general practice’s clinical computer system. A general practitioner reviewed this list to exclude patients for whom a study invitation would be inappropriate. The remainder were sent a letter inviting them to attend a study clinic appointment held at their general practice by a research nurse, where written informed consent was obtained.

clinical computer system. A general practitioner reviewed this list to exclude patients for whom a study invitation would be inappropriate. The remainder were sent a letter inviting them to attend a study clinic appointment held at their general practice by a research nurse, where written informed consent was obtained. Randomisation and masking The central study team at the University of Birmingham randomised patients, with minimisation based on age, sex, diabetes mellitus, atrial fibrillation, baseline systolic blood pressure, and general practice. The research nurse ascertained treatment allocation either by telephone or online. Neither participants nor clinicians were blinded to treatment allocation. A research nurse who was not otherwise involved in the patient’s care obtained the primary outcome measure (blood pressure) by using an automated sphygmomanometer.

Randomisation and masking The central study team at the University of Birmingham randomised patients, with minimisation based on age, sex, diabetes mellitus, atrial fibrillation, baseline systolic blood pressure, and general practice. The research nurse ascertained treatment allocation either by telephone or online. Neither participants nor clinicians were blinded to treatment allocation. A research nurse who was not otherwise involved in the patient’s care obtained the primary outcome measure (blood pressure) by using an automated sphygmomanometer. Procedures Patients randomised to the intensive arm were given a target systolic blood pressure of below 130 mm Hg or a target reduction of 10 mm Hg if their baseline blood pressure was between 125 and 140 mm Hg. The target in the standard arm was less than 140 mm Hg, irrespective of baseline blood pressure. Apart from the different blood pressure targets, the management of blood pressure was the same in both groups and was carried out by a practice nurse (to monitor blood pressure) and a general practitioner (responsible for modifying blood pressure treatment). Patients whose systolic blood pressure at baseline was above target (everyone in the intensive arm and those patients in the standard arm whose blood pressure was ≥140 mm Hg) had their antihypertensive treatment reviewed by their general practitioner. A practice nurse would see all patients at three month intervals (if their blood pressure was below target when previously measured) or after one month (if previous blood pressure was above target) and refer to the general practitioner if the blood pressure was above target. The protocol required no formal down-titration of treatment if blood pressure was below target, but general practitioners had discretion to change or reduce treatment in the light of symptoms attributable to blood pressure drugs. We provided general practitioners with treatment protocols that reflected the national guidelines for blood pressure lowering in operation at the time of the trial.15 In both arms of the trial, the general practitioners had access to a computer based algorithm that actively suggested drugs and dosage if the participant was above target. Follow-up ceased if the participant had a major cardiovascular event.

uidelines for blood pressure lowering in operation at the time of the trial.15 In both arms of the trial, the general practitioners had access to a computer based algorithm that actively suggested drugs and dosage if the participant was above target. Follow-up ceased if the participant had a major cardiovascular event. The primary outcome was change in systolic blood pressure between baseline and one year. Participants had blood pressure measured by a research nurse (separate from the practice nurse’s measurements described above) at baseline and at six and 12 months. Blood pressure was measured using a British Hypertension Society validated automated electronic monitor supplied and validated for the study.16 Blood pressure was measured in a standardised way, with the patient seated for five minutes and then six measurements taken at one minute intervals. The primary outcome was the average of the second and third measurements. Secondary measures of blood pressure included diastolic blood pressure at six and 12 months, systolic blood pressure at six months, and proportion achieving target blood pressures at 12 months. For the systolic blood pressure, we also calculated the means of readings 2-6 and 5-6 to look for any differential effects with regard to habituation to blood pressure measurement.

c blood pressure at six and 12 months, systolic blood pressure at six months, and proportion achieving target blood pressures at 12 months. For the systolic blood pressure, we also calculated the means of readings 2-6 and 5-6 to look for any differential effects with regard to habituation to blood pressure measurement. We identified clinical events through review of the general practice record at 12 months. These comprised major cardiovascular events (composite of fatal and non-fatal stroke, myocardial infarction, fatal coronary heart disease, or other cardiovascular death), emergency hospital admissions, and deaths. Participants were flagged for mortality at the NHS Central Register. Side effects were assessed through the use of standard questionnaires.14

r events (composite of fatal and non-fatal stroke, myocardial infarction, fatal coronary heart disease, or other cardiovascular death), emergency hospital admissions, and deaths. Participants were flagged for mortality at the NHS Central Register. Side effects were assessed through the use of standard questionnaires.14 Statistical analysis We estimated that a sample size of 305 patients in each group would detect a 5 mm Hg difference in systolic blood pressure between groups with 90% power at a significance level of 5% assuming a standard deviation of 17.5 mm Hg, 10% loss to follow-up, 5% mortality, and 10% major vascular events.5 7 We used mixed models for the primary analysis, adjusting for baseline blood pressure, age group (<80 years, ≥80 years), sex, diabetes mellitus, atrial fibrillation, and practice (as a random effect). The principal analysis was a complete case analysis. We also explored the potential effects of missing values by the use of three approaches: multiple imputation, group mean, and last available value. Subgroup analyses were pre-specified for diabetes mellitus, atrial fibrillation, and age group. In addition, we did a subgroup analysis by baseline systolic blood pressure (<140 mm Hg, ≥140 mm Hg). We compared the number of consultations, treatment changes, and side effects by using generalised mixed modelling, adjusting for the same variables as in the primary outcome. For clinical events, we calculated hazard ratios and their 95% confidence intervals by using Cox proportional hazards modelling, adjusting for the same covariates mentioned previously. We checked the proportional hazard assumption with Schoenfeld residual plots and by including interaction terms in the model (for each term by time). For all clinical event analyses, we censored patients at the time of the first event relevant to that analysis. Thus, if a patient had more than one emergency hospital admission, only the first one would be counted. We used SAS 9.2 and Stata 12 for analyses.

uding interaction terms in the model (for each term by time). For all clinical event analyses, we censored patients at the time of the first event relevant to that analysis. Thus, if a patient had more than one emergency hospital admission, only the first one would be counted. We used SAS 9.2 and Stata 12 for analyses. Patient involvement The study was discussed by a stroke survivor group who agreed that it was an important research question and that blood pressure was an important outcome for them. Patients were involved in developing plans for recruitment and design of the study through representation on the Trial Steering Committee. No patients were asked to advise on interpretation or writing up of the results. We plan to disseminate the results of the research to the relevant patient community through local and nationally organised stroke groups.

lans for recruitment and design of the study through representation on the Trial Steering Committee. No patients were asked to advise on interpretation or writing up of the results. We plan to disseminate the results of the research to the relevant patient community through local and nationally organised stroke groups. Results Figure 1 shows the trial profile; 529 patients from 99 general practices (range 1-16 per practice) entered the trial, and 84 patients withdrew from the trial in the 12 months after randomisation (52 (20%) in the intensive target arm and 32 (12%) in the standard target arm; P=0.02). Primary outcome data were available for 379 participants at one year follow-up (182 (68%) in the intensive target arm and 197 (75%) in the standard target arm). All patients were followed up for clinical events and deaths.Table 1 shows patients’ characteristics at baseline. About a quarter of participants were not taking any blood pressure lowering treatment at randomisation (76 in intensive arm; 63 in standard arm). For half of the participants, the index event was a TIA. Just under 20% of participants reported at least moderate disability (modified Rankin score of three or more). There were no important differences in characteristics between participants who did and did not have blood pressure recorded at 12 months (table 1).

m). For half of the participants, the index event was a TIA. Just under 20% of participants reported at least moderate disability (modified Rankin score of three or more). There were no important differences in characteristics between participants who did and did not have blood pressure recorded at 12 months (table 1). Fig 1 Trial profile. *Reasons given: patient was housebound or in nursing home (957; 33%); would be unable to provide consent (338; 12%); comorbidity (216; 7%); blood pressure too low (199; 7%); at risk of falling (164; 6%); insufficient evidence of stroke/transient ischaemic attack (98; 3%); already being treated to 130 mm Hg target (71; 2%); other patient related factors (69; 2%); patient choice (54; 2%); terminally ill (48; 2%); deceased or left practice (41; 1%); participating in another trial (9); no reason given (618; 21%). †Blood pressure <125 mm Hg (447); lack of corroborative evidence of stroke/transient ischaemic attack (60); taking ≥3 antihypertensives (51); orthostatic hypotension (22); already being treated to lower blood pressure target (4); unable to provide informed consent (2). SBP=systolic blood pressure Table 1 Baseline characteristics. Values are numbers (percentages) unless stated otherwise

Fig 1 Trial profile. *Reasons given: patient was housebound or in nursing home (957; 33%); would be unable to provide consent (338; 12%); comorbidity (216; 7%); blood pressure too low (199; 7%); at risk of falling (164; 6%); insufficient evidence of stroke/transient ischaemic attack (98; 3%); already being treated to 130 mm Hg target (71; 2%); other patient related factors (69; 2%); patient choice (54; 2%); terminally ill (48; 2%); deceased or left practice (41; 1%); participating in another trial (9); no reason given (618; 21%). †Blood pressure <125 mm Hg (447); lack of corroborative evidence of stroke/transient ischaemic attack (60); taking ≥3 antihypertensives (51); orthostatic hypotension (22); already being treated to lower blood pressure target (4); unable to provide informed consent (2). SBP=systolic blood pressure Table 1 Baseline characteristics. Values are numbers (percentages) unless stated otherwise Characteristics All participants Participants with systolic blood pressure recorded at 12 months Intensive target (n=266) Standard target (n=263) Intensive target (n=182) Standard target (n=197) Mean (SD) age, years 71.9 (9.1) 71.7 (9.4) 72.6 (8.3) 71.9 (9.5) Male sex 157 (59) 156 (59) 104 (57) 125 (63) White ethnicity 260 (98) 259 (98) 180 (99) 194 (98) Current smoker 25 (9) 33 (13) 15 (8) 27 (14) Mean (SD) SBP, mm Hg 142.9 (14.0) 142.2 (13.4) 143.5 (13.5) 142.2 (12.9) SBP <140 mm Hg 128 (48) 129 (49) 79 (43) 98 (50) SBP ≥140 mm Hg 138 (52) 134 (51) 103 (57) 99 (50) Mean (SD) diastolic blood pressure, mm Hg 79.9 (10.0) 80.4 (9.8) 78.8 (9.3) 80.7 (10.1) Diabetes mellitus 26 (10) 25 (10) 19 (10) 21 (11) Atrial fibrillation 28 (11) 27 (10) 21 (12) 22 (11) Coronary heart disease 41 (15) 46 (17) 28 (15) 35 (18) Chronic kidney disease 26 (10) 30 (11) 19 (10) 23 (12) Heart failure 2 (1) 7 (3) 1 (1) 6 (3) Peripheral vascular disease 11 (4) 11 (4) 7 (4) 6 (3) Stroke 130 (49) 122 (46) 85 (47) 95 (48) Transient ischaemic attack only 135 (51) 141 (54) 97 (53) 102 (52) Mean (SD) No of antihypertensive drugs 1.0 (0.8) 1.1 (0.8) 1.1 (0.8) 1.1 (0.8) Mean (SD) No of other drugs 4.5 (2.5) 4.6 (2.6) 4.5 (2.5) 4.6 (2.6) Mean (SD) total No of drugs 5.6 (2.8) 5.7 (2.7) 5.6 (2.7) 5.7 (2.7) Modified Rankin scale*: 0 or 1 135 (518) 125 (48) 98 (54) 84 (43) 2 65 (24) 69 (26) 42 (23) 57 (29) 3 or 4 47 (18) 51 (19) 29 (16) 42 (21) SBP=systolic blood pressure.

(0.8) 1.1 (0.8) Mean (SD) No of other drugs 4.5 (2.5) 4.6 (2.6) 4.5 (2.5) 4.6 (2.6) Mean (SD) total No of drugs 5.6 (2.8) 5.7 (2.7) 5.6 (2.7) 5.7 (2.7) Modified Rankin scale*: 0 or 1 135 (518) 125 (48) 98 (54) 84 (43) 2 65 (24) 69 (26) 42 (23) 57 (29) 3 or 4 47 (18) 51 (19) 29 (16) 42 (21) SBP=systolic blood pressure. *Data missing for 19 patients in intensive arm and 18 in standard arm (all participants) and for 13 patients in intensive arm and 14 in standard arm (participants with 12 month systolic blood pressure). The intensive target arm was associated with significantly more consultations for blood pressure control with the general practitioner (median visits 2 v 1; P<0.001) and practice nurse (median 3 v 2; P=0.002) than the standard target arm. This higher consultation rate led to more intensifications of blood pressure treatment (458 v 278; P<0.001) and more changes due to side effects (77 v 30; P<0.001). However, patients were also less likely to have their blood pressure treatment increased after review by the general practitioner when the blood pressure was above target in the intensive arm (109 v 57; P=0.005) (table 2). The three factors that contributed most to this difference were symptoms attributed to blood pressure drugs, blood pressure only just above target, and patient not wanting treatment intensified. At the end of the study, the number of antihypertensive drugs that patients were taking had increased by a similar amount in both arms (mean number of antihypertensive drugs 2.1 in intensive arm and 1.9 in standard arm; P=0.13).

rugs, blood pressure only just above target, and patient not wanting treatment intensified. At the end of the study, the number of antihypertensive drugs that patients were taking had increased by a similar amount in both arms (mean number of antihypertensive drugs 2.1 in intensive arm and 1.9 in standard arm; P=0.13). Table 2 Reasons given by general practitioner for not increasing blood pressure treatment after patient referred by practice nurse with blood pressure above target Reason Intensive target (n=109) Standard target (n=57) Other blood pressure readings (eg, home readings) taken into account 17 20 Patient did not want treatment intensified 22 13 Decision taken to re-measure blood pressure at future time 19 12 Symptoms attributed to blood pressure treatment 24 5 Blood pressure only just above target 14 2 Patient had not been taking pills 9 5 Blood pressure reading attributed to anxiety of patient 3 8 Changes to drug treatment already made 4 2 Postural hypotension 3 2 Awaiting specialist advice/test results 5 – Intercurrent illness 3 – Patient too old for further increases in treatment 1 2 Change in lifestyle advocated rather than change in drugs – 1 Reason was given for 164/166 non-intensification decisions. Numbers add up to more than 164, as in some cases two reasons were given.

Awaiting specialist advice/test results 5 – Intercurrent illness 3 – Patient too old for further increases in treatment 1 2 Change in lifestyle advocated rather than change in drugs – 1 Reason was given for 164/166 non-intensification decisions. Numbers add up to more than 164, as in some cases two reasons were given. Treatment to a more intensive target was associated with a significantly greater reduction in systolic blood pressure at 12 months (primary outcome) (table 3). Systolic blood pressure was reduced by 16 mm Hg in the intensive target arm and by 13 mm Hg in the standard target arm. This difference persisted when we calculated it by using the mean of the fifth and sixth readings (−3.2 (95% confidence interval −5.8 to −0.64) mm Hg) or the mean of the second to sixth readings (−3.3 (−5.8 to −0.67) mm Hg) (supplementary table A). When we took account of the missing values by using multiple imputation, the effect size was −3.2 (−5.7 to −0.65) mm Hg (see supplementary table B for results of other methods). The blood pressure target (that is, <130 mm Hg or a 10 mm Hg reduction for those with a baseline systolic blood pressure <140 mm Hg) at one year was achieved in 93 (51%) patients in the intensive arm. Proportions achieving a systolic blood pressure below 140 mm Hg were similar in the two arms (150/182 (82%) v 161/197 (82%); P=0.59), as were those achieving a systolic blood pressure below 130 mm Hg (103/182 (57%) v 107/197 (54%); P=0.36). We found no evidence of a significant difference in effectiveness of using an intensive blood pressure target in any subgroup of patients (fig 2).

e similar in the two arms (150/182 (82%) v 161/197 (82%); P=0.59), as were those achieving a systolic blood pressure below 130 mm Hg (103/182 (57%) v 107/197 (54%); P=0.36). We found no evidence of a significant difference in effectiveness of using an intensive blood pressure target in any subgroup of patients (fig 2). Table 3 Systolic and diastolic blood pressure in intensive target and standard target groups Mean (SD) blood pressure (mm Hg) Mean (SD) difference from baseline (mm Hg) Effect size: mm Hg (95% CI)* Baseline 6 months 12 months 6 months 12 months 6 months 12 months Systolic blood pressure Intensive target† 143.5 (13.5) 125.7 (14.5) 127.4 (14.8) −17.3 (16.7) −16.1 (15.0) −4.12 (−6.84 to -1.40) −2.94 (−5.68 to −0.21) Standard target‡ 142.2 (12.9) 129.3 (14.6) 129.4 (14.8) −12.7 (16.7) −12.8 (17.2) – – Diastolic blood pressure Intensive target† 78.8 (9.3) 73.1 (10.3) 72.0 (9.0) −6.5 (10.7) −6.8 (9.1) −1.14 (−2.86 to 0.58) −1.63 (−3.10 to −0.15) Standard target‡ 80.7 (10.1) 74.6 (9.8) 74.4 (8.9) −6.1 (9.7) −6.3 (9.4) – – *Adjusted for baseline blood pressure, age group (<80, ≥80 years), sex, diabetes mellitus, atrial fibrillation, and general practice (random effect). †Blood pressure data for 193 intensive target patients at six months and 182 at 12 months. ‡Blood pressure data for 198 standard target patients at six months and 197 at 12 months.

Mean (SD) blood pressure (mm Hg) Mean (SD) difference from baseline (mm Hg) Effect size: mm Hg (95% CI)* Baseline 6 months 12 months 6 months 12 months 6 months 12 months Systolic blood pressure Intensive target† 143.5 (13.5) 125.7 (14.5) 127.4 (14.8) −17.3 (16.7) −16.1 (15.0) −4.12 (−6.84 to -1.40) −2.94 (−5.68 to −0.21) Standard target‡ 142.2 (12.9) 129.3 (14.6) 129.4 (14.8) −12.7 (16.7) −12.8 (17.2) – – Diastolic blood pressure Intensive target† 78.8 (9.3) 73.1 (10.3) 72.0 (9.0) −6.5 (10.7) −6.8 (9.1) −1.14 (−2.86 to 0.58) −1.63 (−3.10 to −0.15) Standard target‡ 80.7 (10.1) 74.6 (9.8) 74.4 (8.9) −6.1 (9.7) −6.3 (9.4) – – *Adjusted for baseline blood pressure, age group (<80, ≥80 years), sex, diabetes mellitus, atrial fibrillation, and general practice (random effect). †Blood pressure data for 193 intensive target patients at six months and 182 at 12 months. ‡Blood pressure data for 198 standard target patients at six months and 197 at 12 months. Fig 2 Effect of intensive versus standard target on systolic blood pressure at 12 months for different patient subgroups, adjusted for baseline blood pressure, age group (<80, ≥80 years), sex, diabetes mellitus, atrial fibrillation, and general practice (random effect)

‡Blood pressure data for 198 standard target patients at six months and 197 at 12 months. Fig 2 Effect of intensive versus standard target on systolic blood pressure at 12 months for different patient subgroups, adjusted for baseline blood pressure, age group (<80, ≥80 years), sex, diabetes mellitus, atrial fibrillation, and general practice (random effect) One major cardiovascular event occurred in the intensive target arm (a non-fatal myocardial infarction) and five in the standard care arm (three strokes, one non-fatal myocardial infarction, and one cardiovascular death) (hazard ratio 0.19, 95% confidence interval 0.02 to 1.87; P=0.16). Two deaths occurred in the intensive target arm and one in the standard target arm. The risk of emergency admission was 12.8% per year in the intensive target arm and 7.8% per year in the standard target arm (hazard ratio 1.56, 0.84 to 2.93; P=0.16). Two admissions in each arm were related to falls. Apart from TIA (responsible for five admissions in the standard target arm and three admissions in the intensive target arm) and stroke, no single diagnosis accounted for more than two admissions. Table 4 shows the most common symptoms at 12 months by treatment allocation. There were no significant differences between the two groups. Table 4 Most frequent symptoms at 12 months

One major cardiovascular event occurred in the intensive target arm (a non-fatal myocardial infarction) and five in the standard care arm (three strokes, one non-fatal myocardial infarction, and one cardiovascular death) (hazard ratio 0.19, 95% confidence interval 0.02 to 1.87; P=0.16). Two deaths occurred in the intensive target arm and one in the standard target arm. The risk of emergency admission was 12.8% per year in the intensive target arm and 7.8% per year in the standard target arm (hazard ratio 1.56, 0.84 to 2.93; P=0.16). Two admissions in each arm were related to falls. Apart from TIA (responsible for five admissions in the standard target arm and three admissions in the intensive target arm) and stroke, no single diagnosis accounted for more than two admissions. Table 4 shows the most common symptoms at 12 months by treatment allocation. There were no significant differences between the two groups. Table 4 Most frequent symptoms at 12 months Symptom No (%) Effect size: odds ratio* (95% CI) P value Intensive target arm Standard target arm Pain 93/163 (57) 89/173 (51) 1.17 (0.75 to 1.84) 0.48 Breathlessness 53/148 (36) 49/158 (31) 1.17 (0.72 to 1.92) 0.53 Fatigue 75/149 (50) 88/163 (54) 0.81 (0.51 to 1.28) 0.36 Stiff joints 93/162 (57) 99/176 (56) 0.94 (0.59 to 1.49) 0.80 Sore eyes 35/148 (24) 24/158 (15) 1.68 (0.93 to 3.04) 0.08 Wheeziness 32/163 (20) 28/175 (16) 1.24 (0.70 to 2.21) 0.46 Headaches 27/151 (18) 36/165 (22) 0.69 (0.38 to 1.24) 0.22 Sleep difficulties 56/150 (37) 66/163 (40) 0.81 (0.51 to 1.31) 0.39 Dizziness 45/164 (27) 39/173 (23) 1.24 (0.74 to 2.08) 0.42 Loss of strength 44/148 (30) 51/162 (31) 0.85 (0.51 to 1.40) 0.52 Loss of libido 47/160 (29) 50/171 (29) 1.06 (0.65 to 1.72) 0.83 Impotence 29/129 (22) 31/145 (21) 1.22 (0.65 to 2.30) 0.54 Pins and needles 54/163 (33) 44/176 (25) 1.48 (0.91 to 2.41) 0.11 Cough 40/144 (28) 49/160 (31) 0.86 (0.51 to 1.44) 0.57 Swelling of legs/ankles 51/162 (31) 49/177 (28) 1.10 (0.67 to 1.81) 0.70 Dry mouth 34/147 (23) 36/161 (22) 0.98 (0.57 to 1.70) 0.95 *Adjusted for baseline systolic blood pressure, age group (<80, ≥80 years), sex, diabetes mellitus, atrial fibrillation, and general practice (random effect).

(31) 0.86 (0.51 to 1.44) 0.57 Swelling of legs/ankles 51/162 (31) 49/177 (28) 1.10 (0.67 to 1.81) 0.70 Dry mouth 34/147 (23) 36/161 (22) 0.98 (0.57 to 1.70) 0.95 *Adjusted for baseline systolic blood pressure, age group (<80, ≥80 years), sex, diabetes mellitus, atrial fibrillation, and general practice (random effect). Discussion We found that aiming for a target systolic blood pressure of below 130 mm Hg or a 10 mm Hg reduction from baseline if this was below 140 mm Hg in a primary care population with prevalent cerebrovascular disease led to a lower systolic blood pressure than if a target of below 140 mm Hg target was aimed for. However, the difference was small (about 3 mm Hg) and was associated with increased workload (one extra consultation a year each for general practitioners and nurses). The intensive target arm was not associated with more side effects as measured at follow-up, but more changes to treatment occurred because of side effects during the trial. More people from the intensive target arm withdrew consent for the trial, and this might have reflected unwillingness to persevere with the increased treatment regimen. Perhaps the most important finding was the greater than 10 mm Hg reductions in mean systolic blood pressure in both arms of the study, so that more than 80% of participants in each arm had achieved a blood pressure of below 140 mm Hg by the end of the trial, compared with less than 50% at baseline.

d treatment regimen. Perhaps the most important finding was the greater than 10 mm Hg reductions in mean systolic blood pressure in both arms of the study, so that more than 80% of participants in each arm had achieved a blood pressure of below 140 mm Hg by the end of the trial, compared with less than 50% at baseline. Strengths and weaknesses of study Blood pressure at 12 months was not available for 28% of patients randomised. This reflected a high number of withdrawals from the study, with some differential loss to follow-up in the intensive target arm. However, when we imputed missing values by using multiple imputation (the most robust method), the difference in achieved blood pressure between arms at one year was very similar to that observed. Although we did not achieve our sample size, in the event our trial was adequately powered, as the observed standard deviation in blood pressure was less than we had anticipated in our sample size calculation. This is reflected in the statistical significance of the small difference in observed blood pressure between arms. Nevertheless, the upper limit of the confidence interval around the difference between arms at one year was 5.68 mm Hg, which would be regarded as a clinically important effect. Only 4% of patients on general practice stroke/TIA registers participated in the trial. Participants had a low prevalence of disability for a prevalent cerebrovascular disease population, were younger than typical patients in primary care with a history of cerebrovascular disease, and over-represented people with a history of TIA only.7 It is likely, therefore, that the more intensive target would have been even harder to achieve if the trial population was more representative of people with prevalent cerebrovascular disease. The trial represents a post-stroke primary care population managed by generalists rather than a selective hospital/outpatient population managed by specialists. The outcome measure was not blinded, but a nurse not directly involved in the participant’s care obtained it by using an automated sphygmomanometer, so systematic recording bias is unlikely.

oke primary care population managed by generalists rather than a selective hospital/outpatient population managed by specialists. The outcome measure was not blinded, but a nurse not directly involved in the participant’s care obtained it by using an automated sphygmomanometer, so systematic recording bias is unlikely. The standard target arm in PAST-BP was actively managed, with the support of a computer based algorithm that suggested drug changes, rather than simply receiving “usual care.” If we had used a more passive management strategy in the comparison group, we may have achieved a greater separation in systolic blood pressure between arms. In another blood pressure lowering study of patients with increased cardiovascular risk carried out by our group in the same timeframe, the standard care control arm dropped by 6 mm Hg from a similar baseline compared with 13 mm Hg in the study reported here.17 We used an active control because we wanted to ascertain the effect of setting different blood pressure targets and to avoid confounding that would be introduced by having different management strategies in the two arms. The target in the intensive arm was more complicated than that in the standard care arm, but we minimised the effect of this on adherence to the protocol by ensuring that the primary care staff managed all trial participants in the same way, with prompts to review treatment if blood pressure was above the individualised target.

t in the intensive arm was more complicated than that in the standard care arm, but we minimised the effect of this on adherence to the protocol by ensuring that the primary care staff managed all trial participants in the same way, with prompts to review treatment if blood pressure was above the individualised target. Comparison with other studies and interpretation The change in mean blood pressure that we observed in the intensive target arm was very similar to that observed in the below 130 mm Hg target arm of the SPS3 trial, with both PAST-BP and SPS3 achieving a mean systolic blood pressure in the intensive arm of 127 mm Hg after one year.11 However, the comparison arms had different achieved blood pressures (129 mm Hg in PAST-BP versus 138 mm Hg in SPS3). This reflects the more conservative target in the higher target arm of SPS3 (130-149 mm Hg as opposed to <140 mm Hg) and that antihypertensive treatment was reduced if blood pressure fell below target.

1 However, the comparison arms had different achieved blood pressures (129 mm Hg in PAST-BP versus 138 mm Hg in SPS3). This reflects the more conservative target in the higher target arm of SPS3 (130-149 mm Hg as opposed to <140 mm Hg) and that antihypertensive treatment was reduced if blood pressure fell below target. Most of the observed reduction in blood pressure is likely to have been mediated by increased use of antihypertensive drugs, which on average went up from one to two drugs per person over the year of the study in both arms of the trial. Alternative explanations are that habituation to blood pressure measurement occurred, leading to reduced white coat effect, or that there was regression dilution bias. However, in a blood pressure monitoring trial in a similar post-stroke population with similar mean baseline systolic blood pressure, no fall in blood pressure was observed in the control group over a 12 month period,18 and in the SPS3 trial (also with similar mean baseline systolic blood pressure to PAST-BP) a fall of just 4 mm Hg was seen in the 140 mm Hg target arm over the study period.11 This suggests that the fall of 13 mm Hg observed in the standard target arm of PAST-BP is unlikely to be primarily due to effects of regression dilution or habituation to measurement. Given that we had a relatively low systolic blood pressure inclusion criterion of 125 mm Hg or above, important regression dilution bias would not be anticipated in this study.

observed in the standard target arm of PAST-BP is unlikely to be primarily due to effects of regression dilution or habituation to measurement. Given that we had a relatively low systolic blood pressure inclusion criterion of 125 mm Hg or above, important regression dilution bias would not be anticipated in this study. Only 51% of patients in the intensive target arm of PAST-BP achieved their target blood pressure. Both patients’ wishes and general practitioners’ decision making led to treatment not being intensified when blood pressure was above target (table 2). Greater reluctance to lower blood pressure when near target, higher attribution of symptoms to blood pressure treatment (table 2) despite an absence of objective evidence of increased symptoms (table 4) in the intensive target arm, and greater reluctance of patients to increase treatment hint at the difficulties faced in achieving lower blood pressure targets in clinical practice.19 This impression of practical difficulty is reinforced by the significantly higher proportion of participants that withdrew from the trial in the intensive arm. Although reported side effects and symptoms were similar in the two arms, and serious adverse events were infrequent (two admissions for falls in each arm), significantly more changes to treatment needed to be made because of side effects in the intensive target arm.

s that withdrew from the trial in the intensive arm. Although reported side effects and symptoms were similar in the two arms, and serious adverse events were infrequent (two admissions for falls in each arm), significantly more changes to treatment needed to be made because of side effects in the intensive target arm. Implications Recent evidence from SPRINT and a systematic review highlight the benefits of intensive blood pressure lowering.20 21 In some blood pressure target trials such as SPRINT and SPS3, the trial design maximised the achieved difference in blood pressure between the two arms, with the less intensive arm having a target range rather than simply a below 140 mm Hg systolic target, and with treatment being reduced if blood pressure fell below the target range. This is an appropriate design for an explanatory trial designed to test the question does lowering blood pressure reduce risk of cardiovascular events? In our pragmatic trial, which sought to test the effect of different blood pressure targets as they would be used in clinical practice, the protocol did not stipulate a reduction in blood pressure treatment if the blood pressure was below target and the control arm was actively managed to achieve a target blood pressure below 140 mm Hg. As a result of this, and of reluctance on the part of both clinicians and patients to instigate all increases in blood pressure treatment in the intensive group, the achieved difference in blood pressure between the two arms was small. Nevertheless, we found that active management was associated with clinically important reductions in blood pressure in both arms—the 13 mm Hg reduction achieved in the below 140 mm Hg arm equates to more than 40% and 20% reduction in the risk of stroke and coronary heart disease respectively.22 The reduction in blood pressure in our less intensive arm was similar to that achieved in the active arms of other blood pressure lowering trials and more than in their control groups.11 17 The additional resources needed to achieve the additional 3 mm Hg lower blood pressure in the intensive target arm might be better spent in increasing the proportion of people with stroke in primary care who have a systolic blood pressure below 140 mm Hg. Given this conclusion, we did not feel that a pragmatic trial powered to detect a difference in cardiovascular endpoints achieved using an intensive target in primary care was warranted.

e better spent in increasing the proportion of people with stroke in primary care who have a systolic blood pressure below 140 mm Hg. Given this conclusion, we did not feel that a pragmatic trial powered to detect a difference in cardiovascular endpoints achieved using an intensive target in primary care was warranted. Furthermore, the ongoing ESH-CHL SHOT trial will provide important data on whether intensive blood pressure lowering reduces cardiovascular events in people with stroke (who were excluded from the SPRINT trial).23 The explanatory trial design is likely to lead to clear differences in achieved blood pressure in the treatment arms and confirm whether intensive blood pressure lowering reduces cardiovascular endpoints in the post-stroke population. What is already know on this topic Decreasing blood pressure after stroke is associated with a lower risk of stroke recurrence, but uncertainty exists about what the target blood pressure should be One trial in people with recent lacunar stroke found that a systolic blood pressure target of <130 mm Hg was associated with a non-significant reduction in stroke compared with a target of 130-149 mm Hg No trials of different blood pressure targets after stroke have been carried out in primary care settings What this study adds Patients set a target of <130 mm Hg or a 10 mm Hg reduction if initial blood pressure was <140 mm Hg achieved lower systolic blood pressures than those set a target of <140 mm Hg However, the difference was small (3 mm Hg) in the context of the reduction in blood pressure observed in both arms (13 mm Hg and 16 mm Hg)

What this study adds Patients set a target of <130 mm Hg or a 10 mm Hg reduction if initial blood pressure was <140 mm Hg achieved lower systolic blood pressures than those set a target of <140 mm Hg However, the difference was small (3 mm Hg) in the context of the reduction in blood pressure observed in both arms (13 mm Hg and 16 mm Hg) Active management of blood pressure after stroke/transient ischaemic attack is more important than the target that is set Web Extra Extra material supplied by the author Supplementary tables Click here for additional data file. Contributors: JM, RJMcM, SG, and FDRH had the original idea and obtained the funding. KF, JM, RJMcM, CJT, UM, SV, SG, and FDRH contributed to the protocol. AR did the primary data analysis. KF and SV were responsible for the data collection. JM wrote the first draft of the paper. All authors subsequently refined the manuscript and approved the final version. JM is the guarantor.

nd obtained the funding. KF, JM, RJMcM, CJT, UM, SV, SG, and FDRH contributed to the protocol. AR did the primary data analysis. KF and SV were responsible for the data collection. JM wrote the first draft of the paper. All authors subsequently refined the manuscript and approved the final version. JM is the guarantor. Funding: This report is of independent research funded by the National Institute for Health Research (NIHR; Stroke Prevention in Primary Care, Programme Grant for Applied Research, RP-PG-0606-1153) and by an NIHR Professorship (RJMcM). FDRH is part funded as director of the NIHR School for Primary Care Research, theme leader of the NIHR Oxford Biomedical Research Centre, and director of the NIHR Collaboration for Leadership in Applied Health Research and Care, Oxford. The views expressed in this publication are those of the author(s) and not necessarily those of the NHS. The study sponsor was the University of Birmingham. The study funder and sponsor had no role in the study design; the collection, analysis, or interpretation of data; the writing of the report; or the decision to submit for publication. The researchers are independent of the funders.

d not necessarily those of the NHS. The study sponsor was the University of Birmingham. The study funder and sponsor had no role in the study design; the collection, analysis, or interpretation of data; the writing of the report; or the decision to submit for publication. The researchers are independent of the funders. Competing interests: All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf and declare: JM has received grants from Ferrer and the NIHR; RJMcM has received grants from Ferrer during the conduct of the study and grants and personal fees from Omron, grants from Lloyds Pharmacy, personal fees from the Japanese Society of Hypertension, and personal fees from the American Society of Nephrology outside the submitted work; AR has received grants from the University of Birmingham during the conduct of the study; FDRH has received grants from the NIHR and non-financial support from Omron and Microlife during the conduct of the study; no other relationships or activities that could appear to have influenced the submitted work. Ethical approval: Ethical approval was provided by the Warwickshire Research Ethics Committee (reference 08/H1211/121). All participants gave written informed consent. Transparency: The lead author (the manuscript’s guarantor) affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained.

rency: The lead author (the manuscript’s guarantor) affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained. Data sharing: No additional data available.

Introduction Associations between insufficient or excessive gestational weight gain (GWG) and short and long term maternal and child health outcomes are well described.1 Insufficient weight gain has been linked with increased risks of low birth weight, small for gestational age, and preterm birth, while excessive gain has been associated with large for gestational age, gestational diabetes, preterm birth, caesarean section, infant mortality, postpartum weight retention, and childhood obesity.2 3 4 5 6 7 8 9 Pregnant women are therefore routinely weighed in clinical settings. The benefits of doing so, however, are debatable in the absence of appropriate guidelines or even agreement on what constitutes adequate weight gain.10 11

an section, infant mortality, postpartum weight retention, and childhood obesity.2 3 4 5 6 7 8 9 Pregnant women are therefore routinely weighed in clinical settings. The benefits of doing so, however, are debatable in the absence of appropriate guidelines or even agreement on what constitutes adequate weight gain.10 11 In 1970, the Institute of Medicine/National Research Council reviewed the available evidence on GWG that resulted in good pregnancy outcomes, with subsequent revisions in 1990 and 2009.1 12 13 The latest guidelines evaluated the trade offs between maternal and child health outcomes and weight gain during pregnancy, including the risks of small for gestational age and preterm birth with inadequate GWG and the increased rates of caesarean section and postpartum weight retention with excessive GWG. Based on a recent systematic review, however, these guidelines were all derived from country specific studies that varied in sample selection, study design, and methods of data collection and statistical analysis.14 In the United Kingdom, “routine weighing during pregnancy should be confined to circumstances in which clinical management is likely to be influenced.”15 In countries where routine weighing is recommended, most current guidelines are based on relating observed GWG to pregnancy outcomes and then determining the range of weight gain with the lowest perinatal risk,1 16 17 18 although other authors have attempted to select populations with good perinatal outcomes and then retrospectively determine the associated GWG range.19 20 21 22

ent guidelines are based on relating observed GWG to pregnancy outcomes and then determining the range of weight gain with the lowest perinatal risk,1 16 17 18 although other authors have attempted to select populations with good perinatal outcomes and then retrospectively determine the associated GWG range.19 20 21 22 The World Health Organization recommends that a reference for GWG be based on prospective longitudinal studies of selected populations with a low incidence of maternal and fetal complications, where anthropometric measures are collected before and during pregnancy and postpartum.23 The same “prescriptive” approach was adopted by WHO in producing international growth standards for children aged 0-5 years that have now been adopted by more than 125 countries worldwide,24 and by the International Fetal and Newborn Growth (INTERGROWTH-21st) Consortium for the 21st Century in producing standards for early pregnancy dating,25 fetal growth,26 newborn size,27 and postnatal growth for preterm infants.28 We examined data on GWG obtained, according to WHO recommendations, from healthy pregnant women who were free from identifiable major medical, nutritional, or social and major environmental risk factors.26 29 30 The women had pregnancies with good maternal and perinatal outcomes.31 Based on these data, we report GWG patterns from normal weight women.

, according to WHO recommendations, from healthy pregnant women who were free from identifiable major medical, nutritional, or social and major environmental risk factors.26 29 30 The women had pregnancies with good maternal and perinatal outcomes.31 Based on these data, we report GWG patterns from normal weight women. Methods Study site and population selection INTERGROWTH-21st was a multicentre multiethnic population based project conducted between April 2009 and March 2014 in eight well defined urban sites: Pelotas (Brazil), Turin (Italy), Muscat (Oman), Oxford (UK), Seattle (US), Shunyi County in Beijing (China), the central area of Nagpur (India), and the Parklands suburb of Nairobi (Kenya). The primary aim was to produce international standards for fetal, newborn, and preterm growth using the same conceptual framework as the WHO Multicentre Growth Reference Study24 30 32 to complement the existing WHO Child Growth Standards. We recruited women who started antenatal care before 14 weeks’ gestation with reliable menstrual dates and a confirmatory ultrasound dating scan who met the entry criteria of optimal health, nutrition, education, and socioeconomic status and were not exposed during pregnancy to environmental hazards.25 29 30 These low risk women constituted the population of the Fetal Growth Longitudinal Study (FGLS) component of the INTERGROWTH-21st Project.26

ound dating scan who met the entry criteria of optimal health, nutrition, education, and socioeconomic status and were not exposed during pregnancy to environmental hazards.25 29 30 These low risk women constituted the population of the Fetal Growth Longitudinal Study (FGLS) component of the INTERGROWTH-21st Project.26 Measurements A detailed manual with instructions for all adult measurement techniques, the methods for multicentre standardisation of those measures, and the procedures for the calibration and maintenance of equipment have been published elsewhere.33 34 35 All documentation, protocols, data collection forms, and electronic transfer strategies are available at www.intergrowth21.org. Briefly, the women’s height and weight were measured in duplicate with a Seca 264 stadiometer and Seca 877 scale (Seca, Germany), respectively, on study entry between 9 and 13+6 weeks’ gestation. A first trimester body mass index (BMI) was calculated and categorised as normal weight (18.50-24.99) or overweight (25.00-29.99), according to the WHO definition.36 The same standardised methods and clinical procedures were used to measure maternal weight every five weeks (plus/minus one week) until delivery, so that the possible ranges after recruitment in which weight was measured were 14-18, 19-23, 24-28, 29-33, 34-38, and 39-42 weeks’ gestation.35

ding to the WHO definition.36 The same standardised methods and clinical procedures were used to measure maternal weight every five weeks (plus/minus one week) until delivery, so that the possible ranges after recruitment in which weight was measured were 14-18, 19-23, 24-28, 29-33, 34-38, and 39-42 weeks’ gestation.35 Statistical analyses GWG was calculated as the measured weight at each antenatal visit minus the measured weight in the first trimester. According to prespecified criteria, we excluded pregnancies complicated by fetal death or congenital abnormality, catastrophic or severe medical conditions (such as cancer or HIV), those with severe unanticipated conditions related to pregnancy that required admission to hospital (such as eclampsia or severe pre-eclampsia), and those identified during the study who no longer fulfilled the entry criteria (such as women who started smoking during pregnancy or had an episode of malaria).

er or HIV), those with severe unanticipated conditions related to pregnancy that required admission to hospital (such as eclampsia or severe pre-eclampsia), and those identified during the study who no longer fulfilled the entry criteria (such as women who started smoking during pregnancy or had an episode of malaria). The first step was to assess variation in GWG across sites and whether we could pool the data. A detailed analysis of the methods used to assess the similarity of fetal and newborn data from all eight INTERGROWTH-21st sites to permit pooling has been reported elsewhere.31 37 We applied the same methods to the GWG data by using variance component analysis (analysis of variance (ANOVA)) to calculate the percentage of variance in the longitudinal weight measurements from variance between sites adjusted for gestational age (fixed effects) while sites and individuals were treated as random effects, and a standardised site difference (SSD), similar to a z score, calculated as the difference between the mean of one site and the mean of all sites together. Each difference was then expressed as a proportion of the all sites’ standard deviation (SD) (that is, SD of the data pooled across all sites) at each corresponding gestational age. The SSD allows for direct comparisons across gestational age windows, and we prespecified a value of ≤0.5 as adequate for combining data from all sites. This is similar to the cut off used in the WHO Multicentre Growth Reference Study to create international standards for infant and child growth.38

ing gestational age. The SSD allows for direct comparisons across gestational age windows, and we prespecified a value of ≤0.5 as adequate for combining data from all sites. This is similar to the cut off used in the WHO Multicentre Growth Reference Study to create international standards for infant and child growth.38 In a second step we constructed smoothed centiles of GWG according to gestational age. The statistical methods we used were informed by the recommendations of Altman and Royston39 40 and recent literature reviews.41 42

ing gestational age. The SSD allows for direct comparisons across gestational age windows, and we prespecified a value of ≤0.5 as adequate for combining data from all sites. This is similar to the cut off used in the WHO Multicentre Growth Reference Study to create international standards for infant and child growth.38 In a second step we constructed smoothed centiles of GWG according to gestational age. The statistical methods we used were informed by the recommendations of Altman and Royston39 40 and recent literature reviews.41 42 We applied a multilevel linear regression analysis accounting for repeated measures, adjusting for gestational age, which we treated as a fixed effect, whereas sites and individuals were treated as random effects.38 As weight gain exhibited a non-normal distribution, we log transformed (natural log) data to stabilise variance and transform the data to normality. We added a constant 8.5 for normal weight women to all values to shift the minimum value of the distribution to 1 to ensure no negative values when we modelled on the log scale. The best fitting powers for the mean weight gain were provided by second degree fractional polynomials and further modelled in a multilevel framework to account for the longitudinal design of the study (repeated measures). The data structure comprises two levels—that is, measurements within and between women. Therefore, we fitted a random effects model (two level hierarchical structure) to the longitudinal GWG measurements as a function of gestational age using the runmlwin package in STATA.43 To obtain an equation for the SD, we modelled the resulting variance components from the multilevel model that accounts for the correlations between and within women using fractional polynomials. The SD was modelled on the log scale to stabilise variance. Assessment of goodness of fit incorporated a visual inspection of the overall model fit by comparing empirical centiles (calculated per completed week of gestation—for example, 38 weeks’ gestation=38-38+6 weeks’ gestation) to the fitted centiles; a quantile-quantile (q-q) plot of the residuals; and a plot of fitted z scores across gestational ages.

incorporated a visual inspection of the overall model fit by comparing empirical centiles (calculated per completed week of gestation—for example, 38 weeks’ gestation=38-38+6 weeks’ gestation) to the fitted centiles; a quantile-quantile (q-q) plot of the residuals; and a plot of fitted z scores across gestational ages. As the first weight measurement was taken between 9 and 13+6 weeks’ gestation, we performed a sensitivity analysis to explore the likelihood of potential bias that might arise as a result of this classification. Based on a reported range in weight gain of 0.5-2 kg in the first trimester,1 we performed a post hoc analysis to estimate the proportion of women who were within 2 kg of the lower limit in the normal weight group (and so could have been underweight before conception) and, similarly, those within 2 kg of the lower cut off for overweight women, as they might actually have been normal weight before conception. The data were modelled with the same analytical strategy and the resultant centiles compared with those obtained from our original classification of normal weight (that is, based on the first trimester BMI). All analyses were performed in STATA, version 11.2, software (StataCorp LP, College Station, TX, US). Furthermore, to rule out potential bias from caesarean section, we performed a sensitivity analysis excluding all births by caesarean section and refitting the final model to the remaining data and compared this with the model using all the data.

in STATA, version 11.2, software (StataCorp LP, College Station, TX, US). Furthermore, to rule out potential bias from caesarean section, we performed a sensitivity analysis excluding all births by caesarean section and refitting the final model to the remaining data and compared this with the model using all the data. Patient involvement No patients were involved in setting the research question or the outcome measures, nor were they involved in the design and implementation of the study. There are no plans to involve patients in dissemination. Results A total of 13 108 pregnant women at <14 weeks’ gestation were screened (fig 1), and 4607 met the eligibility criteria, provided consent, and were enrolled in the Fetal Growth Longitudinal Study. The contribution from each site to the total enrolled sample population ranged from 311 (7%) for the US to 640 (14%) for the UK. Fig 1 Flowchart for selecting women included in study of gestational weight gain

Results A total of 13 108 pregnant women at <14 weeks’ gestation were screened (fig 1), and 4607 met the eligibility criteria, provided consent, and were enrolled in the Fetal Growth Longitudinal Study. The contribution from each site to the total enrolled sample population ranged from 311 (7%) for the US to 640 (14%) for the UK. Fig 1 Flowchart for selecting women included in study of gestational weight gain The most common reasons for ineligibility were maternal height <153 cm (1022/8501; 12%), BMI ≥30 (1009/8501; 12%), and age <18 or >35 years (915/8501; 11%) at screening. During the pregnancy, 71 women were lost to follow-up or withdrew consent and 36 were excluded (during pregnancy 29 had severe medical conditions, six took up smoking, and one used recreational drugs). After exclusion of 78 miscarriages, terminations, or stillbirths, there were 4422 live singleton births of which a further 101 were excluded because of congenital malformations. Of the 4321 remaining women, eight were excluded from the analysis (four with only one weight measurement during pregnancy and four who were obvious outliers because of illogical values that could not be corrected during data cleaning). We excluded nine observations with extreme weight changes (defined as a gain or loss of >5 kg/week). Our final sample therefore consisted of 4313 women who contributed 24 977 weight measurements. Of these, 3097 (72%) women had normal weight in the first trimester. Here we report the analyses pertaining to these normal weight women whose data were used to construct the international GWG standard.

USA 220 2.68 (2.14) 0.07 Italy 370 2.55 (1.40) −0.01 All 3078 2.56 (1.61) 0.00 *Calculated by: (site mean of GWG−all sites mean of GWG at each gestational age interval)/all sites’ SD of GWG. SSD adjusted at median gestational age for all sites at each gestational age interval with estimates from final regression model. The mean GWGs were 1.64 kg, 2.86 kg, 2.86 kg, 2.59 kg, and 2.56 kg for the gestational age windows 14-18+6 weeks, 19-23+6 weeks, 24-28+6 weeks, 29-33+6 weeks, and 34-40+0 weeks, respectively (table 2). Of all the weight measurements, 71.7% (10 639/14 846) and 94.9% (14 085/14 846) fell within the expected 1 SD and 2 SD thresholds, respectively, which compares well with 68% and 95% theoretically expected under normality assumptions. On average, across all gestational ages, the absolute magnitude of differences between the observed (empirical) and smoothed centiles was 0.18 kg for the median, 0.37 kg for the 3rd centile, and 0.06 kg for the 97th centile (fig 4). Fig 4 Fitted 3rd (bottom dashed line), 50th (middle dashed line), and 97th (top dashed line) smoothed centile curves for gestational weight gain among normal weight women. Large blue circles show empirical values for each week of gestation, and small grey circles show actual observations

The most common reasons for ineligibility were maternal height <153 cm (1022/8501; 12%), BMI ≥30 (1009/8501; 12%), and age <18 or >35 years (915/8501; 11%) at screening. During the pregnancy, 71 women were lost to follow-up or withdrew consent and 36 were excluded (during pregnancy 29 had severe medical conditions, six took up smoking, and one used recreational drugs). After exclusion of 78 miscarriages, terminations, or stillbirths, there were 4422 live singleton births of which a further 101 were excluded because of congenital malformations. Of the 4321 remaining women, eight were excluded from the analysis (four with only one weight measurement during pregnancy and four who were obvious outliers because of illogical values that could not be corrected during data cleaning). We excluded nine observations with extreme weight changes (defined as a gain or loss of >5 kg/week). Our final sample therefore consisted of 4313 women who contributed 24 977 weight measurements. Of these, 3097 (72%) women had normal weight in the first trimester. Here we report the analyses pertaining to these normal weight women whose data were used to construct the international GWG standard. The demographic characteristics of the study cohort were similar across the eight sites and have been reported elsewhere.31 Women had a median of six weight measurements (range 2-7); median gestational age at first antenatal visit was 11.9 weeks (SD 1.4 weeks); mean maternal age was 28.2 (SD 3.8) years; 97% (3020/3097) were married or living with a partner, and 72% (2230/3097) were nulliparous. Table 1 shows sociodemographic information and pregnancy and perinatal events. Fig 2 shows an example of the crude weight gain trajectories of a simple random sample of 100 normal weight women, illustrating the longitudinal design of the study.

ere married or living with a partner, and 72% (2230/3097) were nulliparous. Table 1 shows sociodemographic information and pregnancy and perinatal events. Fig 2 shows an example of the crude weight gain trajectories of a simple random sample of 100 normal weight women, illustrating the longitudinal design of the study. Table 1 Baseline characteristics and pregnancy outcomes of the normal-weight women. Values are mean (SD) for continuous variables, and number (percentage) for categorical variables

ere married or living with a partner, and 72% (2230/3097) were nulliparous. Table 1 shows sociodemographic information and pregnancy and perinatal events. Fig 2 shows an example of the crude weight gain trajectories of a simple random sample of 100 normal weight women, illustrating the longitudinal design of the study. Table 1 Baseline characteristics and pregnancy outcomes of the normal-weight women. Values are mean (SD) for continuous variables, and number (percentage) for categorical variables Normal BMI (n=3097) Parents Maternal age (years) (SD) 28.2 (3.8) Maternal height (cm) (SD) 162.3 (5.9) Maternal weight (kg) (SD) 57.2 (6.5) Paternal height (cm) (SD) 174.2 (7.3) Body mass index (SD) 21.7 (1.8) Gestational age at first visit (weeks) (SD) 11.9 (1.4) Years of formal education (years) (SD) 15.1 (2.9) Haemoglobin level before 15 weeks’ gestation (g/dL) (SD) 12.5 (1.1) Married/cohabiting (%) 3020 (97.3) Nulliparous (%) 2230 (71.8) Pre-eclampsia (%) 12 (0.4) Pyelonephritis (%) 9 (0.3) Any sexually transmitted infection (%) 1 (0.0) Spontaneous initiation of labour (%) 2127 (68.5) PPROM (<37 weeks’ gestation) (%) 46 (1.5) Caesarean section (%) 1036 (33.4) Mother admitted to intensive care unit (%) 9 (0.3) Infants NICU admission >1 day (%) 160 (5.2) Preterm (<37 weeks’ gestation) (%) 125 (4.0) Preterm and spontaneous onset of labour (%) 82 (2.6) Term LBW (<2500 g; ≥37 weeks) (%) 99 (3.2) Birth weight >4.0 kg (%) 144 (4.7) Neonatal mortality (%) 4 (0.1) Male sex (%) 1534 (49.4) Exclusive breastfeeding at discharge (%) 2698 (86.9) Mean (SD) weight (kg) (≥37 weeks’ gestation) 3.2 (0.4) Mean (SD) length (cm) (≥37 weeks’ gestation) 49.3 (1.9) Mean (SD) head circumference (cm) (≥37 weeks’ gestation) 33.8 (1.3) PPROM=preterm pre-labour rupture of membranes; NICU=neonatal intensive care unit; LBW=low birthweight.

stfeeding at discharge (%) 2698 (86.9) Mean (SD) weight (kg) (≥37 weeks’ gestation) 3.2 (0.4) Mean (SD) length (cm) (≥37 weeks’ gestation) 49.3 (1.9) Mean (SD) head circumference (cm) (≥37 weeks’ gestation) 33.8 (1.3) PPROM=preterm pre-labour rupture of membranes; NICU=neonatal intensive care unit; LBW=low birthweight. Fig 2 Trajectories of gestational weight gain of 100 randomly selected normal weight healthy women with uncomplicated live singleton births We explored the variation in GWG among the sites; the variance within sites (59.6%) was six times higher than the variance between sites (9.6%). The all sites’ SD for GWG ranged from 1.45 kg at 14-19+6 weeks’ gestation to 1.61 kg at 34-40 weeks’ gestation. Within five gestational age windows from 14 weeks to 40+0 weeks, representing 40 comparisons, 37 had standardised site differences (SSDs) ≤0.5 (as prespecified in the protocol) of the SD of all sites combined (fig 3, table 2). The three comparisons that were higher than 0.5 SSD were from China, but the difference was <0.5 at 14-18+6 weeks’ gestation and at 34-40 weeks’ gestation (0.34 and 0.21, respectively).

, 37 had standardised site differences (SSDs) ≤0.5 (as prespecified in the protocol) of the SD of all sites combined (fig 3, table 2). The three comparisons that were higher than 0.5 SSD were from China, but the difference was <0.5 at 14-18+6 weeks’ gestation and at 34-40 weeks’ gestation (0.34 and 0.21, respectively). Fig 3 Standardised site difference (SSD) for gestational weight gain in the Fetal Growth Longitudinal Study of the INTERGROWTH-21st Project. SSD calculated by (site mean of gestational weight gain at each gestational age interval minus all sites mean of gestational weight gain at the same gestational age interval)/all sites’ SD of gestational weight gain at the same gestational age interval. SSD was adjusted at median gestational age for all sites at each gestational age interval Table 2 All sites and individual site means (SD) for gestational weight gain (kg) of normal weight women

Fig 3 Standardised site difference (SSD) for gestational weight gain in the Fetal Growth Longitudinal Study of the INTERGROWTH-21st Project. SSD calculated by (site mean of gestational weight gain at each gestational age interval minus all sites mean of gestational weight gain at the same gestational age interval)/all sites’ SD of gestational weight gain at the same gestational age interval. SSD was adjusted at median gestational age for all sites at each gestational age interval Table 2 All sites and individual site means (SD) for gestational weight gain (kg) of normal weight women Gestational age and Country No of measures Adjusted mean (SD) GWG (kg) Standardised site difference (SSD)* 14-18+6 weeks Brazil 252 1.84 (1.34) 0.14 China 476 2.14 (1.28) 0.34 India 441 1.36 (1.30) −0.20 Kenya 355 1.52 (1.65) −0.09 Oman 360 0.96 (1.53) −0.47 UK 408 1.83 (1.44) 0.13 USA 221 1.79 (1.40) 0.10 Italy 392 1.71 (1.29) 0.05 All 2905 1.64 (1.45) 0.00 19-23+6 weeks Brazil 252 2.62 (1.11) −0.16 China 485 4.00 (1.57) 0.79 India 448 2.22 (1.33) −0.43 Kenya 356 2.57 (1.46) −0.20 Oman 366 2.73 (1.31) −0.08 UK 413 2.80 (1.26) −0.04 USA 211 2.84 (1.40) −0.01 Italy 384 2.75 (1.24) −0.07 All 2915 2.86 (1.46) 0.00 24-28+6 weeks Brazil 251 2.81 (1.46) −0.04 China 448 3.77 (1.68) 0.62 India 455 2.37 (1.40) −0.34 Kenya 360 2.48 (1.29) −0.26 Oman 370 2.91 (1.39) 0.03 UK 412 2.81 (1.27) −0.04 USA 204 2.97 (1.30) 0.07 Italy 375 2.75 (1.35) −0.08 All 2875 2.86 (1.47) 0.00 29-33+6 weeks Brazil 261 2.53 (1.43) −0.04 China 545 3.41 (1.61) 0.55 India 428 2.28 (1.48) −0.21 Kenya 355 2.18 (1.39) −0.27 Oman 363 2.43 (1.39) −0.11 UK 417 2.59 (1.23) 0.00 USA 216 2.52 (1.73) −0.04 Italy 373 2.37 (1.35) −0.15 All 2958 2.59 (1.51) 0.00 34-40+0 weeks Brazil 269 2.48 (1.57) −0.05 China 485 2.91 (1.40) 0.22 India 406 2.19 (1.30) −0.23 Kenya 388 2.42 (2.41) −0.09 Oman 406 2.60 (1.36) 0.02 UK 534 2.61 (1.26) 0.03 USA 220 2.68 (2.14) 0.07 Italy 370 2.55 (1.40) −0.01 All 3078 2.56 (1.61) 0.00 *Calculated by: (site mean of GWG−all sites mean of GWG at each gestational age interval)/all sites’ SD of GWG. SSD adjusted at median gestational age for all sites at each gestational age interval with estimates from final regression model.

The mean GWGs were 1.64 kg, 2.86 kg, 2.86 kg, 2.59 kg, and 2.56 kg for the gestational age windows 14-18+6 weeks, 19-23+6 weeks, 24-28+6 weeks, 29-33+6 weeks, and 34-40+0 weeks, respectively (table 2). Of all the weight measurements, 71.7% (10 639/14 846) and 94.9% (14 085/14 846) fell within the expected 1 SD and 2 SD thresholds, respectively, which compares well with 68% and 95% theoretically expected under normality assumptions. On average, across all gestational ages, the absolute magnitude of differences between the observed (empirical) and smoothed centiles was 0.18 kg for the median, 0.37 kg for the 3rd centile, and 0.06 kg for the 97th centile (fig 4). Fig 4 Fitted 3rd (bottom dashed line), 50th (middle dashed line), and 97th (top dashed line) smoothed centile curves for gestational weight gain among normal weight women. Large blue circles show empirical values for each week of gestation, and small grey circles show actual observations Table 3 provides the values of the smoothed week specific GWG according to gestational age of selected centiles (that is, 3rd, 10th, 25th, 50th, 75th, 90th, and 97th), which are shown graphically in figure 5. We have also provided the corresponding equations for the mean and SD from the multilevel regression model (table 4), allowing for calculation of any desired centiles according to gestational age in exact weeks. For example, centiles can be calculated as mean ±z×SD, where z is −1.88, −1.645, −1.28, 0, 1.28, 1.645, and 1.88 for the 3rd, 5th, 10th, 50th, 90th, 95th, and 97th centiles, respectively.

multilevel regression model (table 4), allowing for calculation of any desired centiles according to gestational age in exact weeks. For example, centiles can be calculated as mean ±z×SD, where z is −1.88, −1.645, −1.28, 0, 1.28, 1.645, and 1.88 for the 3rd, 5th, 10th, 50th, 90th, 95th, and 97th centiles, respectively. Table 3 Smoothed centiles for gestational weight gain (GWG) for women of normal weight (BMI 18.50-24.99) according to gestational age

multilevel regression model (table 4), allowing for calculation of any desired centiles according to gestational age in exact weeks. For example, centiles can be calculated as mean ±z×SD, where z is −1.88, −1.645, −1.28, 0, 1.28, 1.645, and 1.88 for the 3rd, 5th, 10th, 50th, 90th, 95th, and 97th centiles, respectively. Table 3 Smoothed centiles for gestational weight gain (GWG) for women of normal weight (BMI 18.50-24.99) according to gestational age Gestational age (weeks) No of measures Centiles for GWG (kg) 3rd 10th 25th 50th 75th 90th 97th 14 260 −2.34 −1.73 −1.07 −0.25 0.65 1.54 2.50 15 473 −1.77 −1.14 −0.45 0.39 1.32 2.24 3.23 16 705 −1.26 −0.60 0.13 1.01 1.99 2.95 3.98 17 851 −0.80 −0.09 0.67 1.61 2.64 3.65 4.75 18 639 −0.37 0.38 1.19 2.19 3.29 4.36 5.53 19 324 0.03 0.82 1.69 2.75 3.92 5.07 6.31 20 532 0.41 1.25 2.17 3.30 4.55 5.78 7.11 21 627 0.77 1.66 2.64 3.84 5.17 6.49 7.91 22 715 1.11 2.05 3.10 4.37 5.79 7.19 8.72 23 717 1.45 2.44 3.54 4.90 6.41 7.90 9.52 24 399 1.77 2.82 3.98 5.42 7.02 8.61 10.34 25 500 2.09 3.19 4.42 5.94 7.63 9.31 11.15 26 599 2.40 3.56 4.85 6.45 8.24 10.02 11.97 27 675 2.71 3.93 5.28 6.96 8.85 10.73 12.79 28 702 3.02 4.29 5.71 7.47 9.45 11.43 13.61 29 493 3.33 4.65 6.14 7.98 10.06 12.14 14.44 30 526 3.63 5.01 6.56 8.49 10.67 12.86 15.27 31 533 3.94 5.37 6.99 9.00 11.28 13.57 16.10 32 691 4.24 5.73 7.41 9.52 11.89 14.29 16.94 33 715 4.55 6.10 7.84 10.03 12.51 15.01 17.78 34 498 4.85 6.46 8.27 10.55 13.12 15.73 18.62 35 533 5.16 6.82 8.70 11.06 13.74 16.46 19.47 36 514 5.47 7.19 9.14 11.58 14.37 17.19 20.32 37 858 5.78 7.56 9.57 12.11 14.99 17.92 21.18 38 402 6.10 7.93 10.01 12.63 15.62 18.66 22.04 39 230 6.41 8.30 10.45 13.16 16.25 19.40 22.91 40 82 6.73 8.68 10.89 13.69 16.89 20.15 23.79 Total No of measures 14 793 — — — — — — — Fig 5 Smoothed centile curves at 3rd, 10th, 50th, 90th, and 97th centiles for gestational weight gain among healthy normal weight women with uncomplicated live singleton births

66 22.04 39 230 6.41 8.30 10.45 13.16 16.25 19.40 22.91 40 82 6.73 8.68 10.89 13.69 16.89 20.15 23.79 Total No of measures 14 793 — — — — — — — Fig 5 Smoothed centile curves at 3rd, 10th, 50th, 90th, and 97th centiles for gestational weight gain among healthy normal weight women with uncomplicated live singleton births Table 4 Equations for estimating mean and standard deviation (SD) of gestational weight gain in normal weight women according to exact gestational age (weeks)* Estimate Regression equation Mean; log (GWG) 1.382972−56.14743×GA−2+0.2787683×GA0.5 SD; log (GWG) 0.2501993731+142.4297879×GA−2−61.45345×GA−2×LN(GA) GA=exact gestational age in weeks. *All logarithms are natural logarithms. Using equations of mean and SD one can easily compute any desired centiles using relation Pth centile=mean+KSD where K is normal equivalent deviate (z score) corresponding to particular centile—for example, K=1.88 for 97th centile and −1.88 for 3rd centile, and SDs in this equation are predicted estimates from the regression analysis. For example to calculate gestational weight gain at 34 weeks; C50 for GWG=exp((1.382972−56.14743×34−2+0.2787683×340.5)+(0×(0.2501993731+142.4297879×34−2 − 61.45345*34−2*LN(34)))) – 8.75 C3 for GWG=exp((1.382972−56.14743×34−2+0.2787683×340.5)+(−1.88×(0.2501993731+142.4297879×34−2 − 61.45345*34−2×LN(34)))) – 8.75 C97 for GWG=exp((1.382972− 56.14743×34−2+0.2787683×340.5)+(1.88×(0.2501993731+142.4297879×34−2 −61.45345*34−2×LN(34)))) – 8.75

C50 for GWG=exp((1.382972−56.14743×34−2+0.2787683×340.5)+(0×(0.2501993731+142.4297879×34−2 − 61.45345*34−2*LN(34)))) – 8.75 C3 for GWG=exp((1.382972−56.14743×34−2+0.2787683×340.5)+(−1.88×(0.2501993731+142.4297879×34−2 − 61.45345*34−2×LN(34)))) – 8.75 C97 for GWG=exp((1.382972− 56.14743×34−2+0.2787683×340.5)+(1.88×(0.2501993731+142.4297879×34−2 −61.45345*34−2×LN(34)))) – 8.75 The sensitivity analyses performed to assess the impact of potential misclassification of BMI status resulted in 30% (932/3097) of women being excluded from the normal weight group and 53% (n=639/1216) of women being added from the overweight group to form a reclassified group of normal weight women (n=2804 women, 17 579 observations). The resultant centile values were remarkably similar and indistinguishable when we superimposed them on the normal weight GWG chart. Sensitivity results when we excluded women with caesarean sections had minimal effect compared with results using all the data (data not shown). Tables containing the mean and SD, centile values, and z scores by gestational age, expressed in completed weeks’ gestation (as recommended by WHO ICD1044), and printable charts are available at www.intergrowth21.org.

The sensitivity analyses performed to assess the impact of potential misclassification of BMI status resulted in 30% (932/3097) of women being excluded from the normal weight group and 53% (n=639/1216) of women being added from the overweight group to form a reclassified group of normal weight women (n=2804 women, 17 579 observations). The resultant centile values were remarkably similar and indistinguishable when we superimposed them on the normal weight GWG chart. Sensitivity results when we excluded women with caesarean sections had minimal effect compared with results using all the data (data not shown). Tables containing the mean and SD, centile values, and z scores by gestational age, expressed in completed weeks’ gestation (as recommended by WHO ICD1044), and printable charts are available at www.intergrowth21.org. Discussion Principal findings Despite the range of cultures, behaviours, clinical practices, and traditions, which can strongly influence gestational weight gain (GWG), we observed strikingly similar patterns of weight gain in the populations studied, reflecting their overall good health and living conditions, nutritional status, and access to adequate standardised healthcare. The proportion of total variance explained by population differences was <10% of the total variance. This finding indicates not only that separate GWG charts for women from different ethnic/racial groups are not required, as is the case for growth standards from early pregnancy to 5 years of age,24 26 27 but that the observed differences by race/ethnicity reported in some studies45 46 47 48 are more likely caused by socioeconomic, medical, cultural, and nutritional factors than true biological differences in the process of nutrient absorption or fat deposition among healthy women. We adopted a prescriptive approach, employed highly trained anthropometrists to measure maternal weight prospectively in duplicate, and used uniform and standardised measurement equipment and protocols. We used the patterns in weight gain in women with a normal BMI in early pregnancy to produce international standards, using statistical techniques that account for repeated measurements within women at one site and between women across sites. We developed a standard, as well as the accompanying centile chart and simple formulae, to allow any desired centiles or z scores to be calculated. These tools complement the already published fetal growth, neonatal size, and postnatal growth of preterm infant standards from the INTERGROWTH-21st Project.26 27 28

sites. We developed a standard, as well as the accompanying centile chart and simple formulae, to allow any desired centiles or z scores to be calculated. These tools complement the already published fetal growth, neonatal size, and postnatal growth of preterm infant standards from the INTERGROWTH-21st Project.26 27 28 Comparison with other studies Comparisons with previous studies on this subject are difficult because of wide variations in study designs, methods, and populations selected. In particular, some studies based GWG on maternally recalled weight before pregnancy, while we measured weight using standardised methods at the first trimester visit. Nonetheless, the weight gain at term of women in the Fetal Growth Longitudinal Study (13.7 kg) was comparable with the range recommended in 2009 by the Institute of Medicine/National Research Council for normal weight women (11.5-16.0 kg) and optimal GWG reported for a multiethnic Singaporean population (13.7 kg), but about 2-3 kg less than that for low risk urban populations in Leuven, Belgium (15.9 kg) and Pittsburgh, USA (16.4 kg).1 16 21 22 Other prospective longitudinal studies of healthy women in Mexico City, urban regions of Argentina, and rural Malawi reported GWG at term of 12.1 kg, 10.7 kg, and 3.7-6.4 kg,49 50 51 respectively, and large cross sectional studies of low risk Japanese women, well nourished women in Switzerland, and Swedish birth registry records have reported singleton term GWG of 10.0 kg, 15.5 kg, and 13.8 kg, respectively.17 47 52 All these studies were based on country specific populations and used various classifications of BMI status. Furthermore, most of them relied on recalled or routinely recorded weight measurements from medical records or weight data from large population databases with questionable measurement sources, validity, and reliability.

studies were based on country specific populations and used various classifications of BMI status. Furthermore, most of them relied on recalled or routinely recorded weight measurements from medical records or weight data from large population databases with questionable measurement sources, validity, and reliability. Strengths and limitations of study We recognise that our study has some limitations. As the first weight measurement was taken between 9 and 13+6 weeks’ gestation, the BMI classification of women as normal weight was not based on a value from before pregnancy. The results of the post hoc sensitivity analysis, however, were reassuring, and we believe that the effect of any possible misclassification is therefore small. Measurements before pregnancy are seldom available in clinical practice or research studies, especially in low risk women.53 54 Recruitment of women who intend to conceive is also challenging and might be culturally unacceptable in some populations, which would introduce selection bias; this could explain why there are few studies with measured pre-pregnancy weight, which should ideally be used to construct GWG references or standards. Consequently, clinicians and researchers have often relied on self reported pre-pregnancy weight to estimate BMI and monitor GWG,55 despite the considerable limitations of error and recall bias.56 Another limitation is that it was not possible to infer the most appropriate GWG pattern for women who are underweight or obese as our population consisted only of healthy women with a BMI range of 18.5-<30. Underweight women are at increased risk for several adverse outcomes, including fetal growth restriction, so adequate GWG is especially important for this group.5 57 58 59 Conversely, the growing problem of maternal obesity throughout the world has led to great interest in whether limiting GWG can reduce the risk of the associated adverse outcomes.1 It could be argued that the sample size is relatively small compared with epidemiological studies that have reported data from large populations—for example, the Danish National Birth Cohort of more than 60 000 women.60 It is always difficult to reach a balance between sample size and data quality, particularly when larger samples require the use of routinely collected clinical information.

epidemiological studies that have reported data from large populations—for example, the Danish National Birth Cohort of more than 60 000 women.60 It is always difficult to reach a balance between sample size and data quality, particularly when larger samples require the use of routinely collected clinical information. We decided when designing the study that it was more important to have a sufficiently large sample, collected prospectively in a scientifically robust manner, with standardised methods, quality control, and equipment, than a larger sample using data that have been routinely collected with less rigour and precision.

epidemiological studies that have reported data from large populations—for example, the Danish National Birth Cohort of more than 60 000 women.60 It is always difficult to reach a balance between sample size and data quality, particularly when larger samples require the use of routinely collected clinical information. We decided when designing the study that it was more important to have a sufficiently large sample, collected prospectively in a scientifically robust manner, with standardised methods, quality control, and equipment, than a larger sample using data that have been routinely collected with less rigour and precision. Policy implications Our results have several practical implications. Firstly, we are aware that in some settings, such as the UK, routine weight monitoring is not recommended.15 In most countries worldwide and in particular those with large populations at risk of under-nutrition, however, weight monitoring at antenatal visits is common practice. Our aim was to contribute to the standardisation of weight monitoring and the more systematic use of the data obtained. Overall, we suggest that the standards (as part of first level nutritional screening) can be used to alert clinicians to deviations in weight, triggering clinical inquiries as to whether such deviations are associated with complications related to pregnancy, medical conditions, or eating disorders. We would discourage clinicians, however, from telling women that deviations are due to pregnancy complications or recommending immediate behaviour changes as our data do not provide sufficient evidence for the standards to be interpreted in this way. Secondly, we believe that consideration should be given to referring women who are underweight before pregnancy for nutritional advice and treatment if necessary and that it is safe to suggest that during pregnancy such women should have GWG at least compatible with those of normal weight women. Finally, our data cannot be used to make recommendations to underweight, overweight, or obese women beyond those already provided by NICE.61

utritional advice and treatment if necessary and that it is safe to suggest that during pregnancy such women should have GWG at least compatible with those of normal weight women. Finally, our data cannot be used to make recommendations to underweight, overweight, or obese women beyond those already provided by NICE.61 Conclusions In summary, we have described patterns of GWG among normal weight women that are compatible with desirable healthy pregnancy outcomes, which provide a basis to guide clinical recommendations on weight gain. To facilitate the use of such recommendations in clinical settings, epidemiological studies with data on important long term maternal and childhood outcomes are needed to identify optimal centile (that is, outcome based cut off points) categories associated with the best health outcomes. Towards that end, the INTERGROWTH-21st Project is currently collecting one and two year follow-up data, including postpartum maternal weight patterns. We anticipate that the publication of this GWG standard will prompt debate among epidemiologists, nutritionists, obstetricians, and midwives about what the optimal thresholds should be. We believe that this standard is more robust than any other available charts and adds to the set of international standards from the INTERGROWTH-21st Project, which aims to improve pregnancy care practices and outcomes by establishing benchmarks against which all women, their unborn babies, and newborns can be compared.26 27 28

e that this standard is more robust than any other available charts and adds to the set of international standards from the INTERGROWTH-21st Project, which aims to improve pregnancy care practices and outcomes by establishing benchmarks against which all women, their unborn babies, and newborns can be compared.26 27 28 What is already known on this topic Guidelines and charts for gestational weight gain (GWG) that are currently in use around the world were derived from country specific studies A recent systematic review assessing the quality of these studies has shown considerable heterogeneity in methods, in particular in terms of sample selection, study design, and methods of data collection and statistical analysis This could explain the variation in recommendations and the lack of consensus regarding what constitutes adequate weight gain What this study adds This multi-country study of GWG adopted a prescriptive and highly standardised approach to describing the GWG patterns of normal weight women at low risk of adverse maternal and perinatal outcomes The generated standards could be used to alert clinicians to deviations in weight, which should then initiate a series of questions to determine whether the changes are associated with complications related to pregnancy, medical conditions, or eating disorders These standards are more scientifically robust than other published charts and add to the set of international standards from the INTERGROWTH-21st Project

The generated standards could be used to alert clinicians to deviations in weight, which should then initiate a series of questions to determine whether the changes are associated with complications related to pregnancy, medical conditions, or eating disorders These standards are more scientifically robust than other published charts and add to the set of international standards from the INTERGROWTH-21st Project We thank the health authorities in Pelotas, Brazil; Beijing, China; Nagpur, India; Turin, Italy; Nairobi, Kenya; Muscat, Oman; Oxford, UK; and Seattle, US, who facilitated the project by allowing participation of these study sites as collaborating centres. We are extremely grateful to Philips Medical Systems who provided the ultrasound equipment and technical assistance throughout the project. We also thank MedSciNet UK for setting up the INTERGROWTH-21st website and for the development, maintenance, and support of the online data management system. We thank the parents and infants who participated in the studies and the more than 200 members of the research teams who made the implementation of this project possible. The participating hospitals included: Brazil, Pelotas (Hospital Miguel Piltcher, Hospital São Francisco de Paula, Santa Casa de Misericórdia de Pelotas, and Hospital Escola da Universidade Federal de Pelotas); China, Beijing (Beijing Obstetrics and Gynaecology Hospital, Shunyi Maternal and Child Health Centre, and Shunyi General Hospital); India, Nagpur (Ketkar Hospital, Avanti Institute of Cardiology Private Limited, Avantika Hospital, Gurukrupa Maternity Hospital, Mulik Hospital and Research Centre, Nandlok Hospital, Om Women’s Hospital, Renuka Hospital and Maternity Home, Saboo Hospital, Brajmonhan Taori Memorial Hospital, and Somani Nursing Home); Kenya, Nairobi (Aga Khan University Hospital, MP Shah Hospital and Avenue Hospital); Italy, Turin (Ospedale Infantile Regina Margherita Sant’ Anna and Azienda Ospedaliera Ordine Mauriziano); Oman, Muscat (Khoula Hospital, Royal Hospital, Wattayah Obstetrics and Gynaecology Poly Clinic, Wattayah Health Centre, Ruwi Health Centre, Al-Ghoubra Health Centre and Al-Khuwair Health Centre); UK, Oxford (John Radcliffe Hospital); and US, Seattle (University of Washington Hospital, Swedish Hospital, and Providence Everett Hospital).

ula Hospital, Royal Hospital, Wattayah Obstetrics and Gynaecology Poly Clinic, Wattayah Health Centre, Ruwi Health Centre, Al-Ghoubra Health Centre and Al-Khuwair Health Centre); UK, Oxford (John Radcliffe Hospital); and US, Seattle (University of Washington Hospital, Swedish Hospital, and Providence Everett Hospital). Full acknowledgment of all those who contributed to the development of the INTERGROWTH-21st Project protocol appears at www.intergrowth21.org.uk Members of the International Fetal and Newborn Growth Consortium for the 21st Century (INTERGROWTH-21st) and its committees Scientific advisory committee M Katz (chair from January 2011), MK Bhan, C Garza, S Zaidi, A Langer, PM Rothwell (from February 2011), D Weatherall (chair until December 2010). Steering committee ZA Bhutta (chair), J Villar (principal investigator), S Kennedy (project director), DG Altman, FC Barros, E Bertino, F Burton, M Carvalho, L Cheikh Ismail, WC Chumlea, MG Gravett, YA Jaffer, A Lambert, P Lumbiganon, JA Noble, RY Pang, AT Papageorghiou, M Purwar, J Rivera, CG Victora. Executive committee J Villar (chair), DG Altman, ZA Bhutta, L Cheikh Ismail, S Kennedy, A Lambert, JA Noble, AT Papageorghiou. Project coordinating unit J Villar (head), S Kennedy, L Cheikh Ismail, A Lambert, AT Papageorghiou, M Shorten, L Hoch (until May 2011), HE Knight (until August 2011), EO Ohuma (from September 2010), C Cosgrove (from July 2011), I Blakey (from March 2011), D Bishop (from February 2014). Data analysis group DG Altman (head), EO Ohuma, J Villar.

Project coordinating unit J Villar (head), S Kennedy, L Cheikh Ismail, A Lambert, AT Papageorghiou, M Shorten, L Hoch (until May 2011), HE Knight (until August 2011), EO Ohuma (from September 2010), C Cosgrove (from July 2011), I Blakey (from March 2011), D Bishop (from February 2014). Data analysis group DG Altman (head), EO Ohuma, J Villar. Data management group DG Altman (head), F Roseman, N Kunnawar, SH Gu, JH Wang, MH Wu, M Domingues, P Gilli, L Juodvirsiene, L Hoch (until May 2011), N Musee (until June 2011), H Al-Jabri (until October 2010), S Waller (until June 2011), C Cosgrove (from July 2011), D Muninzwa (from October 2011), EO Ohuma (from September 2010), D Yellappan (from November 2010), A Carter (from July 2011), S Ash (from August 2011), D Reade (from June 2012), R Miller (from June 2012). Ultrasound group AT Papageorghiou (head), L Salomon (senior external advisor), A Leston, A Mitidieri, F Al-Aamri, W Paulsene, J Sande, WKS Al-Zadjali, C Batiuk, S Bornemeier, M Carvalho, M Dighe, P Gaglioti, N Jacinta, S Jaiswal, JA Noble, K Oas, M Oberto, E Olearo, MG Owende, J Shah, S Sohoni, T Todros, M Venkataraman, S Vinayak, L Wang, D Wilson, QQ Wu, S Zaidi, Y Zhang, P Chamberlain (until September 2012), D Danelon (until July 2010), I Sarris (until June 2010), J Dhami (until July 2011), C Ioannou (until February 2012), CL Knight (from October 2010), R Napolitano (from July 2011), S Wanyonyi (from May 2012), C Pace (from January 2011), V Mkrtychyan (from June 2012).

Zaidi, Y Zhang, P Chamberlain (until September 2012), D Danelon (until July 2010), I Sarris (until June 2010), J Dhami (until July 2011), C Ioannou (until February 2012), CL Knight (from October 2010), R Napolitano (from July 2011), S Wanyonyi (from May 2012), C Pace (from January 2011), V Mkrtychyan (from June 2012). Anthropometry group L Cheikh Ismail (head), WC Chumlea (senior external advisor), F Al-Habsi, ZA Bhutta, A Carter, M Alija, JM Jimenez-Bustos, J Kizidio, F Puglia, N Kunnawar, H Liu, S Lloyd, D Mota, R Ochieng, C Rossi, M Sanchez Luna, YJ Shen, HE Knight (until August 2011), DA Rocco (from June 2012), IO Frederick (from June 2012). Neonatal group ZA Bhutta (head), E Albernaz, M Batra, BA Bhat, E Bertino, P Di Nicola, F Giuliani, I Rovelli, K McCormick, R Ochieng, RY Pang, V Paul, V Rajan, A Wilkinson, A Varalda (from September 2012). Environmental Health Group B Eskenazi (head), LA Corra, H Dolk, J Golding, A Matijasevich, T de Wet, JJ Zhang, A Bradman, D Finkton, O Burnham, F Farhi. Participating countries and local investigators Brazil: FC Barros (principal investigator), M Domingues, S Fonseca, A Leston, A Mitidieri, D Mota, IK Sclowitz, MF da Silveira. China: RY Pang (principal investigator), YP He, Y Pan, YJ Shen, MH Wu, QQ Wu, JH Wang, Y Yuan, Y Zhang. India: M Purwar (principal investigator), A Choudhary, S Choudhary, S Deshmukh, D Dongaonkar, M Ketkar, V Khedikar, N Kunnawar, C Mahorkar, I Mulik, K Saboo, C Shembekar, A Singh, V Taori, K Tayade, A Somani.

Brazil: FC Barros (principal investigator), M Domingues, S Fonseca, A Leston, A Mitidieri, D Mota, IK Sclowitz, MF da Silveira. China: RY Pang (principal investigator), YP He, Y Pan, YJ Shen, MH Wu, QQ Wu, JH Wang, Y Yuan, Y Zhang. India: M Purwar (principal investigator), A Choudhary, S Choudhary, S Deshmukh, D Dongaonkar, M Ketkar, V Khedikar, N Kunnawar, C Mahorkar, I Mulik, K Saboo, C Shembekar, A Singh, V Taori, K Tayade, A Somani. Italy: E Bertino (principal investigator), P Di Nicola, M Frigerio, G Gilli, P Gilli, M Giolito, F Giuliani, M Oberto, L Occhi, C Rossi, I Rovelli, F Signorile, T Todros. Kenya: W Stones and M Carvalho (co- principal investigators), J Kizidio, R Ochieng, J Shah,, S Vinayak, N Musee (until June 2011), C Kisiang’ani (until July 2011), D Muninzwa (from August 2011). Oman: YA Jaffer (principal investigator), J Al-Abri, J Al-Abduwani, FM Al-Habsi, H Al-Lawatiya, B Al-Rashidiya, WKS Al-Zadjali, FR Juangco, M Venkataraman, H Al-Jabri (until October 2010), D Yellappan (from November 2010). UK: S Kennedy (principal investigator), L Cheikh Ismail, AT Papageorghiou, F Roseman, A Lambert, EO Ohuma, S Lloyd, R Napolitano (from July 2011), C Ioannou (until February 2012), I Sarris (until June 2010). US: MG Gravett (principal investigator), C Batiuk, M Batra, S Bornemeier, M Dighe, K Oas, W Paulsene, D Wilson, IO Frederick, HF Andersen, SE Abbott, AA Carter, H Algren, DA Rocco, TK Sorensen, D Enquobahrie, S Waller (until June 2011).

UK: S Kennedy (principal investigator), L Cheikh Ismail, AT Papageorghiou, F Roseman, A Lambert, EO Ohuma, S Lloyd, R Napolitano (from July 2011), C Ioannou (until February 2012), I Sarris (until June 2010). US: MG Gravett (principal investigator), C Batiuk, M Batra, S Bornemeier, M Dighe, K Oas, W Paulsene, D Wilson, IO Frederick, HF Andersen, SE Abbott, AA Carter, H Algren, DA Rocco, TK Sorensen, D Enquobahrie, S Waller (until June 2011). Contributors: JV and SHK were responsible for conceiving the INTERGROWTH-21st Project. JV, SHK, DGA, and JAN prepared the original protocol with later input from ATP, LCI, FCB, and ZAB. JV, ATP, LCI, AL, and ZAB supervised and coordinated the project’s overall undertaking. EOO performed the statistical analysis in collaboration with DGA. EOO, and DGA were responsible for data management and analysis in collaboration with JV. RP, FCB, WS, YAJ, MGG, and MP were collaborators and implemented the project in their respective countries. LCI, DCB, and EOO wrote the paper in collaboration with BFA and KR, with input from all co-authors. All co-authors read the report and made suggestions about its content. JV is guarantor. Funding: This project was supported by the Bill and Melinda Gates Foundation to the University of Oxford. The sponsor played no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

ported by the Bill and Melinda Gates Foundation to the University of Oxford. The sponsor played no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication. Competing interests: All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf and declare: no support from any organisation for the submitted work; no financial relationships with any organisations that might have an interest in the submitted work in the previous three years; no other relationships or activities that could appear to have influenced the submitted work. Ethical approval: The INTERGROWTH-21st Project was approved by the Oxfordshire research ethics committee ‘C’ (reference: 08/H0606/139), the research ethics committees of the individual participating institutions, and the corresponding regional health authorities where the project was implemented. All participants gave informed consent before taking part. Transparency: The manuscript’s guarantor affirms that the manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained. Data sharing: The relevant anonymised patient level data are available on reasonable request from the corresponding author.

Introduction Schizophrenia and other psychotic disorders lead to lifelong health and social adversities, culminating in a reduction in life expectancy of 10-25 years.1 Immigrants and their descendants are, on average, 2.5 times more likely to have a psychotic disorder than the majority ethnic group in a given setting,2 3 although the exact risk varies by ethnicity and setting. For example, in Europe, incidence rates for people of black Caribbean or African descent are approximately five times higher than those for the white European population.2 4 These marked differences persist after adjustment for age, sex, and socioeconomic position,5 are maintained in the descendants of first generation migrants,2 and do not seem to be attributable to higher incidence rates in people’s country of origin or selective migration.6 7 8 9 Possible explanations centre on various social determinants of health, including severe or repeated exposure to psychosocial adversities such as trauma, abuse, socioeconomic disadvantage, discrimination, and social isolation. If this is the case, people granted refugee status may be particularly vulnerable to psychosis, given their increased likelihood of having experienced conflict, persecution, violence, or other forms of psychosocial adversity.10 11

such as trauma, abuse, socioeconomic disadvantage, discrimination, and social isolation. If this is the case, people granted refugee status may be particularly vulnerable to psychosis, given their increased likelihood of having experienced conflict, persecution, violence, or other forms of psychosocial adversity.10 11 Although refugees have more mental health problems than their non-refugee counterparts,11 12 including post-traumatic stress disorder and common mental disorders,13 14 little is known about the risk of psychosis in refugees. One previous longitudinal study from Denmark observed that refugees were at elevated risk of psychosis compared with the native-born Danish population.15 However, the risk in refugees was not compared with that in other non-refugee migrants (henceforth referred to as migrants), who are known to be at increased risk,16 making attribution of this excess directly to a refugee effect impossible. More recently, a Canadian cohort study found that refugees had a modestly increased risk of schizophrenia compared with other migrants,17 but neither group was at elevated risk compared with an ethnically diverse Canadian-born background population, making this finding difficult to interpret and contrary to a large literature on immigration and psychosis.2

found that refugees had a modestly increased risk of schizophrenia compared with other migrants,17 but neither group was at elevated risk compared with an ethnically diverse Canadian-born background population, making this finding difficult to interpret and contrary to a large literature on immigration and psychosis.2 Here, we clarify the risk of non-affective psychotic disorders, including schizophrenia, in refugees compared with other migrants and the native-born Swedish population in a national population based cohort of 1.3 million people. Sweden has a total population size of 9.7 million inhabitants, of whom 1.6 million were born abroad. In 2011 refugees constituted 12% of the total immigrant population. Sweden experienced high levels of labour immigration between 1940 and 1970, followed by substantial refugee immigration.18 On a per capita basis, Sweden grants more refugee applications than any other high income country,19 which, combined with national linked register data, makes it an excellent setting in which to conduct this research. We hypothesised that refugees would have a higher risk of non-affective psychotic disorders than migrants and that risk for both groups would be elevated compared with the Swedish-born population. We also hypothesised that the risk in refugees compared with migrants would vary by region of origin, given putative differences in the pre-migratory experiences of migrants from different regions and differences in how they might adjust to a new society.

ups would be elevated compared with the Swedish-born population. We also hypothesised that the risk in refugees compared with migrants would vary by region of origin, given putative differences in the pre-migratory experiences of migrants from different regions and differences in how they might adjust to a new society. Methods Study design and population We established a retrospective cohort of 1 347 790 people born after 1 January 1984, who were born in Sweden to two Swedish-born parents (n=1 191 004; 88.4%) or were refugees (n=24 123; 1.8%) or non-refugee first generation migrants (n=132 663; 9.8%) granted residency in Sweden. To permit valid comparisons between refugees and migrants, we restricted the immigrant sample to people born in geographical regions with at least 1000 refugees in our cohort (see below). We excluded people without an official residence permit in Sweden—that is, undocumented migrants or people with an official asylum decision pending. We followed participants from their 14th birthday, or date of arrival in Sweden if later, until diagnosis of an ICD-10 (international classification of diseases, 10th revision) non-affective psychotic disorder (F20-29), emigration, death, or 31 December 2011, whichever was sooner. We could not include people who immigrated to Sweden before 1 January 1998 (n=53 855), because refugee status was not sufficiently recorded in the Swedish national registers before this date. We also excluded 812 (0.06%) participants with missing data on municipality of residence in Sweden at cohort entry, needed for estimation of urban residency as a covariate (see below). Excluded participants did not differ from immigrants included in the cohort by sex (51.0% (27 471/53 855) versus 50.7% (79 863/157 531) men; χ2 P=0.21) but had a higher disposable income (11.0% (5924/53 855) versus 5.4% (8533/157 531) were in the highest income quarter; χ2 P<0.001) and were more likely to come from the former Yugoslavia (32.4% (17 457/53 855) versus 8.4% (13 275/157 531); χ2 P<0.001) than other regions. Crude incidence rates were similar between excluded (77.7 (95% confidence interval 70.4 to 85.8) per 100 000 person years) and included immigrants (86.6 (79.1 to 94.7) per 100 000 person years).

ikely to come from the former Yugoslavia (32.4% (17 457/53 855) versus 8.4% (13 275/157 531); χ2 P<0.001) than other regions. Crude incidence rates were similar between excluded (77.7 (95% confidence interval 70.4 to 85.8) per 100 000 person years) and included immigrants (86.6 (79.1 to 94.7) per 100 000 person years). Data sources We extracted data from a large, longitudinal database of linked national registers, known as Psychiatry Sweden, which included data on all people officially resident in Sweden after 1 January 1932, linked via a unique personal identity number and anonymised by Statistics Sweden for research purposes. We obtained relevant outcome, exposure, and covariate data from the following registers: the register of the total population to identify cohort participants and obtain basic demographic data (birth date, sex, country of birth); the multi-generation register to link participants to their parents for identification of the native-born Swedish population; the longitudinal integration database for health insurance and labour market studies (LISA) to obtain data on disposable income; the immigration and emigration database (STATIV) to obtain migration and refugee data; the national patient register to obtain outcome data; and the causes of death register for data pertaining to mortality.

integration database for health insurance and labour market studies (LISA) to obtain data on disposable income; the immigration and emigration database (STATIV) to obtain migration and refugee data; the national patient register to obtain outcome data; and the causes of death register for data pertaining to mortality. Patient involvement No patients were involved in setting the research question or the outcome measures, nor were they involved in developing plans for design or implementation of the study. No patients were asked to advise on interpretation or writing up of results. However, we will disseminate the results of our research to agencies responsible for the healthcare of refugee and migrant groups in Sweden. Outcome Our primary outcome was an ICD-10 clinical diagnosis of non-affective psychotic disorder (F20-29), which included schizophrenia (F20) and all other non-affective psychotic disorders (F21-29). We defined cases as cohort participants with a first recorded diagnosis between 1 January 1998 and 31 December 2011 in the national patient register, which records diagnoses following inpatient and outpatient admissions in Sweden (including privately run public healthcare settings). Inpatient records are complete since 1987, and complete recording from outpatient settings began in 2001. We excluded anyone with a recorded diagnosis of non-affective psychotic disorder made before the age of 14 years (n=156).

nt and outpatient admissions in Sweden (including privately run public healthcare settings). Inpatient records are complete since 1987, and complete recording from outpatient settings began in 2001. We excluded anyone with a recorded diagnosis of non-affective psychotic disorder made before the age of 14 years (n=156). Exposures Our primary exposure was refugee status, defined as refugee, other migrant, or person born in Sweden to two Swedish-born parents, obtained from the STATIV database, which records the reason why a residence permit was granted. Permanent residency for asylum in Sweden is based on the Swedish Migration Agency’s definition of refugee status,18 made in accordance with Swedish law and the UN Refugee Convention, as someone who, “owing to a well-founded fear of being persecuted . . . is unable to, or owing to such fear, is unwilling to avail himself of the protection of that country.”20 All other immigrants granted official residency were classified as migrants. We identified people born in Sweden to two Swedish-born parents (henceforth the “Swedish-born” group) via linkage to the multi-generation register.

to, or owing to such fear, is unwilling to avail himself of the protection of that country.”20 All other immigrants granted official residency were classified as migrants. We identified people born in Sweden to two Swedish-born parents (henceforth the “Swedish-born” group) via linkage to the multi-generation register. As a secondary exposure, we classified people according to region of origin, as defined by country of birth. Although Statistics Sweden records data on specific country of birth, information is released for research purposes according to 13 larger geographical regions to ensure confidentiality. From this variable, we derived a broader region of origin variable for analysis, which included Sweden (Swedish-born only) and four other regions from which at least 1000 refugees in our cohort originated—sub-Saharan Africa, Asia, eastern Europe and Russia, and the Middle East and north Africa (see supplementary table A). Confounders We included age at risk and sex as two a priori confounder variables in all analyses. We also included individual disposable income in Sweden and population density at cohort entry as covariates, to adjust for possible differences between refugees, migrants, and the Swedish-born population.

As a secondary exposure, we classified people according to region of origin, as defined by country of birth. Although Statistics Sweden records data on specific country of birth, information is released for research purposes according to 13 larger geographical regions to ensure confidentiality. From this variable, we derived a broader region of origin variable for analysis, which included Sweden (Swedish-born only) and four other regions from which at least 1000 refugees in our cohort originated—sub-Saharan Africa, Asia, eastern Europe and Russia, and the Middle East and north Africa (see supplementary table A). Confounders We included age at risk and sex as two a priori confounder variables in all analyses. We also included individual disposable income in Sweden and population density at cohort entry as covariates, to adjust for possible differences between refugees, migrants, and the Swedish-born population. We defined disposable income as annual disposable income, based on total family income from all registered sources, including wages, welfare benefits, other social subsidies, and pensions. Statistics Sweden estimated individual disposable income, weighting total family income according to household size and composition, with younger children given lower weights than older household members. We measured disposable income at the earliest point during follow-up (available in LISA at 16 years old or arrival in Sweden, if later). To account for inflation, we categorised individual disposable income into quarters, relative to all other cohort members assigned a disposable income score in the same year.

members. We measured disposable income at the earliest point during follow-up (available in LISA at 16 years old or arrival in Sweden, if later). To account for inflation, we categorised individual disposable income into quarters, relative to all other cohort members assigned a disposable income score in the same year. We defined urban residency according to the population density of each participant’s municipality of residence at cohort entry, expressed as the total population per square kilometre (ppkm2). Sweden consists of 290 municipalities (median population density 26.3 (interquartile range 12.2-75.7) ppkm2). For descriptive purposes, we classified participants into three population density categories: 0-26.2 ppkm2 (very rural areas, below Swedish median), 26.3-260 ppkm2 (rural and semi-rural areas), and 260.1-4617.2 ppkm2 (metropolitan, suburban, and urban areas). To adjust more effectively for population density, we used a continuous measure in our analyses, first transformed on to the natural logarithm scale to account for its positive skewed distribution across municipalities.

pkm2 (rural and semi-rural areas), and 260.1-4617.2 ppkm2 (metropolitan, suburban, and urban areas). To adjust more effectively for population density, we used a continuous measure in our analyses, first transformed on to the natural logarithm scale to account for its positive skewed distribution across municipalities. Statistical analyses We recorded basic descriptive statistics and crude incidence rates for refugees, migrants, and the Swedish-born group. Next, we fitted Cox proportional hazard models to estimate hazard ratios and 95% confidence intervals according to each exposure variable. Follow-up time was based on the earliest date of entry into the risk period (date of 14th birthday or, for all immigrants older than 14 years on arrival, date of immigration) until exit from the cohort. We modelled age at risk as a time varying covariate, using Lexis expansion to stratify each participant into N observations, taking into account differing ages at risk over the follow-up period (14-16, 17-19, 20-22, 23-25, 26-27; Nmax=5).

older than 14 years on arrival, date of immigration) until exit from the cohort. We modelled age at risk as a time varying covariate, using Lexis expansion to stratify each participant into N observations, taking into account differing ages at risk over the follow-up period (14-16, 17-19, 20-22, 23-25, 26-27; Nmax=5). We initially examined the effect of refugee status on risk of non-affective psychotic disorder, after adjustment for age at risk, sex, and their interaction, if statistically significant. In a second adjustment, we added disposable income and population density. We tested whether the relation between refugee status and non-affective disorder differed between men and women by fitting an interaction term between refugee status and sex, with results presented separately for men and women, where appropriate. We repeated these analyses for our secondary exposure variable, region of origin. Next, to determine whether risk of non-affective psychotic disorder in refugees relative to migrants differed by region of origin, we fitted a Cox regression model to a subset of the cohort, excluding the Swedish-born group who did not contribute information to these analyses. Given the small sample of female refugees diagnosed as having psychosis (n=27), we did these analyses for both sexes combined and, separately, for men only. We assessed all statistical interactions by using likelihood ratio tests against a model without the relevant interaction term.

bute information to these analyses. Given the small sample of female refugees diagnosed as having psychosis (n=27), we did these analyses for both sexes combined and, separately, for men only. We assessed all statistical interactions by using likelihood ratio tests against a model without the relevant interaction term. To minimise the possibility that any immigrants diagnosed as having non-affective psychotic disorder may have been prevalent (that is, existing) cases on arrival in Sweden, we did sensitivity analyses on all models, excluding any refugee or non-refugee migrant given a diagnosis within 12 months of immigration. Finally, we checked our main models (via likelihood ratio tests) for departure from proportional hazards. We used Stata v13 to analyse the data. Results We identified 3704 cases during more than 8.9 million person years of follow-up (table 1). Median age at first diagnosis in the Swedish-born population was 20.1 (interquartile range 18.3-22.3) years, younger than for refugees (21.0 (19.2-23.7) years; Mann-Whitney P<0.001) and non-refugees (20.9 (18.7-23.6) years; P<0.001), for whom age at first diagnosis was similar (P=0.30). Following arrival in Sweden, time to first diagnosis was shorter for refugees (median 2.8 (0.7-5.6) years) than for migrants (3.9 (1.2-7.0) years; Mann-Whitney P=0.02). Table 1 Cohort characteristics by migrant status—refugees, non-refugee migrants, and Swedish-born population. Values are numbers (percentages)

Results We identified 3704 cases during more than 8.9 million person years of follow-up (table 1). Median age at first diagnosis in the Swedish-born population was 20.1 (interquartile range 18.3-22.3) years, younger than for refugees (21.0 (19.2-23.7) years; Mann-Whitney P<0.001) and non-refugees (20.9 (18.7-23.6) years; P<0.001), for whom age at first diagnosis was similar (P=0.30). Following arrival in Sweden, time to first diagnosis was shorter for refugees (median 2.8 (0.7-5.6) years) than for migrants (3.9 (1.2-7.0) years; Mann-Whitney P=0.02). Table 1 Cohort characteristics by migrant status—refugees, non-refugee migrants, and Swedish-born population. Values are numbers (percentages) Characteristics Swedish-born population Non-refugee migrants Refugee migrants Cases (n=3232) Person years* (n=8 384 891) Cases (n=379) Person years* (n=471 308) Cases (n=93) Person years* (n=73 604) Sex: Men 1778 (55.0) 4 310 990 (51.4) 234 (62) 232 118 (49.2) 66 (71) 41 069 (55.8) Women 1454 (45.0) 4 073 901 (48.6) 145 (38) 239 190 (50.8) 27 (29) 32 535 (44.2) Birth year: 1984-86 1279 (39.6) 2 928 401 (34.9) 175 (46) 185 052 (39.3) 35 (38) 23 820 (32.4) 1987-89 1111 (34.4) 2 510 835 (29.9) 107 (28) 125 770 (26.7) 28 (30) 19 093 (25.9) 1990-92 649 (20.1) 1 896 903 (22.6) 74 (20) 91 965 (19.5) 22 (24) 16 837 (22.9) 1993-95 174 (5.4) 903 840 (10.8) 19 (5) 56 237 (11.9) 8 (9) 11 728 (15.9) 1996-97 19 (0.6) 144 911 (1.7) 4 (1) 12 283 (2.6) 0 (0) 2127 (2.9) Region of origin: Sweden 3232 (100.0) 8 345 891 (100.0) - - - - Sub-Saharan Africa - - 111 (29) 59 447 (12.6) 31 (33) 18 670 (25.4) Asia - - 66 (17) 105 647 (22.4) 15 (16) 12 929 (17.6) Eastern Europe - - 80 (21) 134 094 (28.5) 7 (8) 6546 (8.9) Middle East - - 122 (32) 172 120 (36.5) 40 (43) 35 459 (48.2) Income: Lowest quarter 1156 (35.8) 2 161 330 (25.8) 264 (70) 339 062 (71.9) 63 (68) 51 953 (70.6) Second quarter 830 (25.7) 2 185 386 (26.1) 52 (14) 63 153 (13.4) 12 (13) 10 486 (14.2) Third quarter 679 (21.0) 2 073 841 (24.7) 45 (12) 35 919 (7.6) 13 (14) 6768 (9.2) Highest quarter 567 (17.5) 1 964 334 (23.4) 18 (5) 33 174 (7.0) 5 (5) 4398 (6.0) Population density†: 0-26.2 875 (27.1) 2 303 728 (27.5) 50 (1) 55 129 (11.7) 25 (27) 21 746 (29.5) 26.3-260 1698 (52.5) 4 472 698 (53.3) 168 (44) 216 155 (45.9) 49 (53) 35 031 (47.6) 260.1-4617.2 659 (20.4) 1 608 466 (19.2) 161 (42) 200 024 (42.4) 19 (20) 16 827 (22.9) *Rounded to nearest integer.

7.0) 5 (5) 4398 (6.0) Population density†: 0-26.2 875 (27.1) 2 303 728 (27.5) 50 (1) 55 129 (11.7) 25 (27) 21 746 (29.5) 26.3-260 1698 (52.5) 4 472 698 (53.3) 168 (44) 216 155 (45.9) 49 (53) 35 031 (47.6) 260.1-4617.2 659 (20.4) 1 608 466 (19.2) 161 (42) 200 024 (42.4) 19 (20) 16 827 (22.9) *Rounded to nearest integer. †People per km2. The crude incidence rate of non-affective psychotic disorders was 38.5 (95% confidence interval 37.2 to 39.9) per 100 000 person years in the Swedish-born population, 80.4 (72.7 to 88.9) per 100 000 person years in migrants, and 126.4 (103.1 to 154.8) per 100 000 person years in refugees. This corresponded to an absolute rate difference of 45.9 (19.0 to 72.9) per 100 000 person years in refugees compared with migrants, in addition to an extra 41.9 (33.7 to 50.1) cases per 100 000 person years in migrants compared with the Swedish-born population. Compared with the Swedish-born population, hazard ratios were 2.90 (95% confidence interval 2.31 to 3.64) in refugees and 1.75 (1.51 to 2.02) in migrants, after adjustment for age, sex, their interaction, disposable income, and population density (table 2). Refugees were 1.66 (1.32 to 2.09) times more likely to be diagnosed as having non-affective psychotic disorders than were migrants. These associations were more pronounced in men than women (likelihood ratio test P for interaction=0.001; table 2 and fig 1). Table 2 Risk of non-affective psychoses by migrant status after adjustment for confounders. Values are hazard ratios (95% CIs) Category All Men Women Model 1 Model 2 Model 1 Model 2 Model 1 Model 2

The crude incidence rate of non-affective psychotic disorders was 38.5 (95% confidence interval 37.2 to 39.9) per 100 000 person years in the Swedish-born population, 80.4 (72.7 to 88.9) per 100 000 person years in migrants, and 126.4 (103.1 to 154.8) per 100 000 person years in refugees. This corresponded to an absolute rate difference of 45.9 (19.0 to 72.9) per 100 000 person years in refugees compared with migrants, in addition to an extra 41.9 (33.7 to 50.1) cases per 100 000 person years in migrants compared with the Swedish-born population. Compared with the Swedish-born population, hazard ratios were 2.90 (95% confidence interval 2.31 to 3.64) in refugees and 1.75 (1.51 to 2.02) in migrants, after adjustment for age, sex, their interaction, disposable income, and population density (table 2). Refugees were 1.66 (1.32 to 2.09) times more likely to be diagnosed as having non-affective psychotic disorders than were migrants. These associations were more pronounced in men than women (likelihood ratio test P for interaction=0.001; table 2 and fig 1). Table 2 Risk of non-affective psychoses by migrant status after adjustment for confounders. Values are hazard ratios (95% CIs) Category All Men Women Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Swedish-born as reference Non-refugee migrant 2.28 (1.99 to 2.62) 1.75 (1.51 to 2.02) 2.61 (2.22 to 3.07) 2.01 (1.70 to 2.38) 1.91 (1.58 to 2.31) 1.44 (1.19 to 1.76) Refugee migrant 3.61 (2.87 to 4.53) 2.90 (2.31 to 3.64) 4.28 (3.28 to 5.58) 3.49 (2.67 to 4.55) 2.65 (1.80 to 3.92) 2.07 (1.40 to 3.06) Non-refugee migrant as reference

Swedish-born as reference Non-refugee migrant 2.28 (1.99 to 2.62) 1.75 (1.51 to 2.02) 2.61 (2.22 to 3.07) 2.01 (1.70 to 2.38) 1.91 (1.58 to 2.31) 1.44 (1.19 to 1.76) Refugee migrant 3.61 (2.87 to 4.53) 2.90 (2.31 to 3.64) 4.28 (3.28 to 5.58) 3.49 (2.67 to 4.55) 2.65 (1.80 to 3.92) 2.07 (1.40 to 3.06) Non-refugee migrant as reference Refugee migrant 1.58 (1.26 to 1.99) 1.66 (1.32 to 2.09) 1.64 (1.25 to 2.15) 1.74 (1.32 to 2.28) 1.39 (0.92 to 2.10) 1.43 (0.95 to 2.16) Model 1 was adjusted for age at risk, sex, and their interaction. Model 2 was also adjusted for disposable income and population density. A likelihood ratio test confirmed statistical interaction between sex and age at risk in model 1 (χ2 (df=4) 71.5; P<0.001) and model 2 (χ2 (df=4) 73.0; P<0.001), as well as between sex and refugee status in model 1 (χ2 (df=2) 11.7; P=0.003) and model 2 (χ2 (df=2) 13.5; P=0.001). Hazard ratios by refugee status are therefore presented separately for men and women. Fig 1 Hazard ratios for schizophrenia and other non-affective psychotic disorders by refugee status and sex. Model 1 was adjusted for age at risk, sex, and their interaction (where appropriate). Model 2 was additionally adjusted for disposable income and population density. Swedish-born group provides reference category, except for fourth (white) bar in each group, which shows hazard ratio for refugees relative to non-refugee migrants. Error bars represent 95% confidence intervals

eraction (where appropriate). Model 2 was additionally adjusted for disposable income and population density. Swedish-born group provides reference category, except for fourth (white) bar in each group, which shows hazard ratio for refugees relative to non-refugee migrants. Error bars represent 95% confidence intervals Taking refugees and migrants together, immigrants from all regions of origin had increased rates of disorder relative to the Swedish-born population, after adjustment for age at risk and sex (supplementary table B). Hazard ratios were most pronounced for all immigrants from sub-Saharan Africa (hazard ratio 5.23, 4.32 to 6.34), which was also observed for both men (6.68, 5.33 to 8.37) and women (3.64, 2.68 to 4.94) separately. These patterns persisted after adjustment for disposable income and population density, ranging from 1.41 (1.11 to 1.78) in people from eastern Europe and Russia to 4.10 (3.38 to 4.98) in people from sub-Saharan Africa, relative to the Swedish-born population.

8, 5.33 to 8.37) and women (3.64, 2.68 to 4.94) separately. These patterns persisted after adjustment for disposable income and population density, ranging from 1.41 (1.11 to 1.78) in people from eastern Europe and Russia to 4.10 (3.38 to 4.98) in people from sub-Saharan Africa, relative to the Swedish-born population. We next investigated whether the elevated rates of non-affective psychotic disorders in refugees compared with migrants differed by region of origin, excluding the Swedish-born population who did not contribute to these analyses. For men and women combined, we found evidence that the rate of non-affective psychosis in refugees compared with migrants varied by region of origin (table 3; likelihood ratio test P=0.05). This finding was even more pronounced in men (likelihood ratio test P=0.007), such that rates of non-affective psychotic disorder were elevated in refugees compared with migrants from all regions of origin, except sub-Saharan Africa (hazard ratio 0.68, 0.40 to 1.16), after adjustment for age at risk, sex, disposable income, and population density (table 3). Male refugees from eastern Europe and Russia were at greatest risk compared with their migrant counterparts (hazard ratio 2.88, 1.22 to 6.82). In general, the rate of psychotic disorders in refugees relative to migrants became smaller as the crude incidence rate in non-refugees from each region of origin increased (table 3). We made no attempt to examine this effect in women, given insufficient numbers of refugees (n=27).

ts (hazard ratio 2.88, 1.22 to 6.82). In general, the rate of psychotic disorders in refugees relative to migrants became smaller as the crude incidence rate in non-refugees from each region of origin increased (table 3). We made no attempt to examine this effect in women, given insufficient numbers of refugees (n=27). Table 3 Risk of non-affective psychoses in refugees relative to non-refugees, by region of origin

ts (hazard ratio 2.88, 1.22 to 6.82). In general, the rate of psychotic disorders in refugees relative to migrants became smaller as the crude incidence rate in non-refugees from each region of origin increased (table 3). We made no attempt to examine this effect in women, given insufficient numbers of refugees (n=27). Table 3 Risk of non-affective psychoses in refugees relative to non-refugees, by region of origin Category All Men Crude incidence rate (95% CI) per 100 000 PYAR Hazard ratio (95% CI): model 2 Crude incidence rate (95% CI) per 100 000 PYAR Hazard ratio (95% CI): model 2 Swedish-born 38.5 (37.2 to 39.9) - 41.2 (39.4 to 43.2) - Eastern Europe: Non-refugees 59.7 (47.9 to 74.3) 1 62.5 (45.9 to 85.2) 1 Refugees 106.9 (51.0 to 224.3) 1.76 (0.81 to 3.82) 184.1 (82.7 to 409.8) 2.88 (1.22 to 6.82) Asia: Non-refugees 62.5 (49.1 to 79.5) 1 67.0 (48.3 to 92.9) 1 Refugees 116.0 (69.9 to 192.4) 1.78 (1.01 to 3.14) 146.1 (83.0 to 257.3) 2.20 (1.13 to 4.25) Middle East and north Africa: Non-refugees 70.9 (59.4 to 84.6) 1 94.4 (75.9 to 117.4) 1 Refugees 112.8 (82.7 to 153.8) 1.56 (1.08 to 2.23) 143.5 (100.3 to 205.2) 1.55 (1.01 to 2.36) Sub-Saharan Africa: Non-refugees 186.7 (155.0 to 224.9) 1 269.0 (215.1 to 336.3) 1 Refugees 166.0 (116.8 to 236.1) 0.81 (0.54 to 1.23) 207.1 (130.5 to 328.8) 0.68 (0.40 to 1.16) Estimates from model 1 and model 2 were similar; only data from model 2, adjusted for age at risk, sex, their interaction (for both sexes combined), disposable income, and population density, are reported. Likelihood ratio test χ2 (df=3) and P values, for statistical interaction between refugee status and region of origin were 8.0 and 0.05 for full sample and 12.0 and 0.007 in analysis restricted to men. Given small number of refugee women with outcome (n=27), no attempt was made to inspect risk by region of origin separately for women.

test χ2 (df=3) and P values, for statistical interaction between refugee status and region of origin were 8.0 and 0.05 for full sample and 12.0 and 0.007 in analysis restricted to men. Given small number of refugee women with outcome (n=27), no attempt was made to inspect risk by region of origin separately for women. PYAR=person years at-risk. Sensitivity analyses excluding potentially prevalent cases among immigrants did not appreciably alter estimates of associations for our main exposures (supplementary tables C and D). The assumption of proportional hazards was not violated (P=0.84 and P=0.13 for analyses of refugee status and region of origin, respectively). Discussion In this cohort study, we found that refugees granted asylum in a high income setting were, on average, 66% more likely to develop schizophrenia or another non-affective psychotic disorder than non-refugee migrants from the same regions of origin and up to 3.6 times more likely to do so than the Swedish-born population.

In this cohort study, we found that refugees granted asylum in a high income setting were, on average, 66% more likely to develop schizophrenia or another non-affective psychotic disorder than non-refugee migrants from the same regions of origin and up to 3.6 times more likely to do so than the Swedish-born population. Strengths and weaknesses of study This study has several methodological strengths. It was based on a large, national population based cohort of more than 1.3 million people, followed for more than 8.9 million person years by using linked Swedish register data. This research has not previously been possible owing to a lack of information on the reason for migration in official Swedish registers; one earlier attempt to investigate this question in Sweden could not distinguish between refugees and non-refugees from the same region.21 Swedish register data are known to be reliable for research purposes,22 23 and diagnosis of psychotic disorders recorded in the national patient register has good validity and positive predictive value.24 25 26 This register is highly complete, recording all psychiatric contacts from inpatient settings from 1987 onwards and from outpatient settings since 2001. Although this may have led to slight under-ascertainment from outpatient settings between 1998 and 2000, we have no reason to believe that this would have introduced differential bias by refugee status or region of origin. We cannot exclude the possibility that we underestimated the true incidence of non-affective psychoses in Sweden, particularly for certain groups, such as recent immigrants or refugees, who may have been unfamiliar with the Swedish healthcare system, have faced greater language barriers, or had poor health literacy.27 If these accessibility factors differed according to sex, the true incidence among migrant and refugee women may have been underestimated in the Swedish patient register, making our hazard ratios conservative.

iliar with the Swedish healthcare system, have faced greater language barriers, or had poor health literacy.27 If these accessibility factors differed according to sex, the true incidence among migrant and refugee women may have been underestimated in the Swedish patient register, making our hazard ratios conservative. Sensitivity analyses suggested that our results were not attributable to prevalent cases among refugees and migrants. In our study, migrants and refugees from sub-Saharan Africa were at increased risk of having a psychotic disorder, compared with the Swedish-born group. This finding is consistent with many other European and worldwide studies.2 Although diagnostic bias has been proposed to explain excess rates of psychotic disorders observed in ethnic minorities,8 little evidence supports this possibility in general.28 Studies in which psychiatrists were blinded to participants’ ethnicity during the diagnostic process have confirmed rates of psychotic disorders in ethnic minority groups,29 including people of black Caribbean and black African origin. In Sweden, by law, interpreters have to aid clinical consultations when necessary. Furthermore, any diagnostic biases are less likely to have accounted for observed differences in risk between refugees and migrants from the same regions of origin observed in our study. Refugees are also at elevated risk of post-traumatic stress disorder,13 which can present with psychotic features; however, our findings are unlikely to be attributable to misdiagnosed cases of this disorder among refugees, as it often presents comorbidly in people exposed to potentially traumatic events and experiences.30

gees are also at elevated risk of post-traumatic stress disorder,13 which can present with psychotic features; however, our findings are unlikely to be attributable to misdiagnosed cases of this disorder among refugees, as it often presents comorbidly in people exposed to potentially traumatic events and experiences.30 We were unable to include immigrants who arrived in Sweden before 1998 in our study, because data on refugee status were unavailable before that year. These groups were more likely to come from the former Yugoslavia, reflecting geopolitical conflicts of the time. This may have reduced our power to detect differences between refugees and other migrants from eastern Europe, but we have no reason to believe their exclusion would have otherwise biased our estimates; the crude incidence in this group was comparable to that for included immigrants, despite their higher post-migratory disposable income. Finally, notwithstanding our large cohort size, the number of cases in refugees was small, which limited our power to detect effects in certain groups, most notably women, for whom risk of non-affective psychotic disorders is, on average, half that of men.31

rants, despite their higher post-migratory disposable income. Finally, notwithstanding our large cohort size, the number of cases in refugees was small, which limited our power to detect effects in certain groups, most notably women, for whom risk of non-affective psychotic disorders is, on average, half that of men.31 As our study was based on routine register data, information on potentially relevant experiences before migration was unavailable. Such pre-migratory experiences remain an important area for future research. Our cohort included migrants and refugees exposed to various humanitarian crises resulting from conflict (such as Iraq, Iran, Afghanistan, the Balkans, central Africa) as well as famine (such as east Africa). Although it is too early to determine whether people currently seeking refuge in Europe following contemporary humanitarian crises (in Syria, Iraq, Afghanistan, parts of north Africa, Kosovo, Albania) would also be at greater risk of psychotic disorder, we assume that our findings will generalise to these groups for two reasons. Firstly, a degree of geographical overlap exists between the regions we included and those generating current humanitarian crises.32 Secondly, we presume that exposure to war, persecution, and exposure to other psychosocial adversity would have a universal effect on individual risk of psychosis, independent of other risk factors.

degree of geographical overlap exists between the regions we included and those generating current humanitarian crises.32 Secondly, we presume that exposure to war, persecution, and exposure to other psychosocial adversity would have a universal effect on individual risk of psychosis, independent of other risk factors. We adjusted for possible differences between refugees, migrants, and the Swedish-born population with regard to age, sex, disposable income, and population density at cohort entry. We did not include other post-migratory markers of potential social disadvantage; such factors may lie on the causal pathway between immigration and risk psychosis, thus making adjustment difficult to interpret. We were unable to examine risk of psychosis in so-called second generation refugees or migrants, because our study population was born after 1984, making their children too young to have entered the risk period for psychosis before the end of our follow-up period in 2011.

thus making adjustment difficult to interpret. We were unable to examine risk of psychosis in so-called second generation refugees or migrants, because our study population was born after 1984, making their children too young to have entered the risk period for psychosis before the end of our follow-up period in 2011. Clinical and public health implications of study Contemporary humanitarian crises in Europe, the Middle East, north Africa, and central Asia have contributed to more displaced people, asylum seekers, and refugees worldwide than at any time since the second world war.33 The severe social, economic, and health inequalities faced by displaced populations arising from these crises are often compounded by national immigration policies and structural constraints in receiving countries. In turn, exposure to these ongoing adversities seems likely to contribute to the increased risk of post-traumatic stress disorder and common mental disorders among refugees.11 12 13 Our data highlight further mental health inequalities facing such groups.34 Clinicians and service planners in high income settings should be aware of the early signs of psychosis in refugees, for whom median presentation to services after arrival to Sweden was more than a year sooner than for other migrant groups. Just as for the general population, refugees and their families will benefit from timely and early intervention and care, particularly in those exposed to severe psychosocial adversity.

refugees, for whom median presentation to services after arrival to Sweden was more than a year sooner than for other migrant groups. Just as for the general population, refugees and their families will benefit from timely and early intervention and care, particularly in those exposed to severe psychosocial adversity. Our findings are consistent with the hypothesis that increased risk of non-affective psychotic disorders among immigrants is due to a higher frequency of exposure to social adversity before migration,35 including the effects of war, violence, or persecution. Further studies will be needed to confirm this possibility. Violence experienced by children and adults who flee persecution has been linked to worse subsequent mental health in general.11 36 Intriguingly, our study suggested that risk of psychosis in refugees relative to other migrants varied by region of origin in our data. Although this finding needs to be replicated in larger samples, it suggests that in addition to refugee status, context matters. For example, we observed no differences in risk of psychosis between refugees and non-refugee migrants among immigrants from sub-Saharan Africa, perhaps because both groups had highly increased rates of disorder (more than 165 new cases per 100 000 person years).

sts that in addition to refugee status, context matters. For example, we observed no differences in risk of psychosis between refugees and non-refugee migrants among immigrants from sub-Saharan Africa, perhaps because both groups had highly increased rates of disorder (more than 165 new cases per 100 000 person years). One parsimonious explanation for this finding is that a larger proportion of sub-Saharan Africa immigrants will have been exposed to deleterious psychosocial adversities before emigration, irrespective of refugee status. By contrast, pre-migratory psychosocial adversities experienced by refugees from eastern Europe and Russia may differ substantially compared with non-refugee migrants from these countries, thus confining excess risk to refugees from such regions. It is also possible that post-migratory factors, such as discrimination, racism, and social exclusion, may explain the high rates of psychotic disorder in migrants and refugees from sub-Saharan Africa, given the absence of a “refugee effect” in this group. Visible minority status may lead to more post-migratory psychosocial adversity. In general population samples, some evidence suggests that perceived discrimination and ethnic density (proximity to one’s own ethnic group) are, respectively, risk and protective factors for psychosis.37 38 Although we controlled for income and post-migratory urban residency, we were unable to investigate other post-migratory factors, including racism, discrimination, and ethnic density, in the available data; further exploration of such factors presents an important avenue for future research. Other factors, including difficulties in the asylum process, also warrant further investigation. For example, women seeking asylum are less likely to be granted refugee status than men, given greater structural and cultural barriers in the asylum process.39 In our study, such an effect would have led to a higher proportion of women being classified as migrants, which may have partially explained why differences in incidence between female refugees and non-refugees were less pronounced than for their male counterparts. A recent study by Oram et al has further highlighted high levels of severe mental illness faced by trafficked migrants, who represent another vulnerable group of migrants.40

e partially explained why differences in incidence between female refugees and non-refugees were less pronounced than for their male counterparts. A recent study by Oram et al has further highlighted high levels of severe mental illness faced by trafficked migrants, who represent another vulnerable group of migrants.40 Conclusion Our study shows that, on average, refugees in a high income setting face substantially elevated rates of schizophrenia and other non-affective psychoses, in addition to the array of other mental, physical, and social inequalities that already disproportionately affect these vulnerable populations. This risk exceeded the well established excess burden of psychosis experienced in immigrant and ethnic minority groups more generally and thus emphasises the need to take the early signs and symptoms of psychosis into account in refugee populations, as part of any clinical mental health service responses to the current global humanitarian crises. More broadly, our findings support the possibility that exposure to psychosocial adversity increases the risk of psychosis. What is already known on this topic Immigrant populations are at elevated risk of schizophrenia and other non-affective psychotic disorders Whether refugees have rates of these disorders over and above those typically observed in non-refugee immigrant groups is unclear

Conclusion Our study shows that, on average, refugees in a high income setting face substantially elevated rates of schizophrenia and other non-affective psychoses, in addition to the array of other mental, physical, and social inequalities that already disproportionately affect these vulnerable populations. This risk exceeded the well established excess burden of psychosis experienced in immigrant and ethnic minority groups more generally and thus emphasises the need to take the early signs and symptoms of psychosis into account in refugee populations, as part of any clinical mental health service responses to the current global humanitarian crises. More broadly, our findings support the possibility that exposure to psychosocial adversity increases the risk of psychosis. What is already known on this topic Immigrant populations are at elevated risk of schizophrenia and other non-affective psychotic disorders Whether refugees have rates of these disorders over and above those typically observed in non-refugee immigrant groups is unclear What this study adds The incidence rate of a non-affective psychotic disorder was 66% higher among refugees than among non-refugee migrants from similar regions of origin, and nearly three times greater than in the native-born Swedish population These patterns were apparent for men and women, although they were stronger in men Refugees from all regions of origin had higher rates of psychotic disorder than non-refugee migrants, except for people from sub-Saharan Africa, for whom rates in both groups were similarly high relative to the Swedish-born population

What this study adds The incidence rate of a non-affective psychotic disorder was 66% higher among refugees than among non-refugee migrants from similar regions of origin, and nearly three times greater than in the native-born Swedish population These patterns were apparent for men and women, although they were stronger in men Refugees from all regions of origin had higher rates of psychotic disorder than non-refugee migrants, except for people from sub-Saharan Africa, for whom rates in both groups were similarly high relative to the Swedish-born population Clinicians and health service planners should be aware of early signs of psychosis in vulnerable migrant populations, who may benefit from timely and early interventions Web Extra Extra material supplied by the author Supplementary tables Click here for additional data file. Contributors: A-CH and CD conceived of the study. A-CH, JK, and CD designed the study and obtained funding. A-CH, CD, and HD acquired the migration data. CD and HD acquired all other cohort data. A-CH, JBK, and HD prepared the data. A-CH, JK, CD, CM, and GL interpreted statistical analyses. HD coordinated the data management, and A-CH wrote the study protocol. A-CH and JBK did the statistical analyses, drafted the data tables, and co-wrote the manuscript. All authors critically revised the paper for important intellectual content and approved the final version. JBK and CD were co-senior authors of the manuscript. A-CH and JBK are the guarantors.

-CH wrote the study protocol. A-CH and JBK did the statistical analyses, drafted the data tables, and co-wrote the manuscript. All authors critically revised the paper for important intellectual content and approved the final version. JBK and CD were co-senior authors of the manuscript. A-CH and JBK are the guarantors. Funding: A-CH is supported by FORTE (grant number 2014-1430; 2014-2678). JBK is supported by a Sir Henry Dale fellowship jointly funded by the Wellcome Trust and the Royal Society (grant number 101272/Z/13/Z). CD is supported by the Swedish Research Council (grant number 523-2010-1052). The funders had no involvement in any aspect of the design of this study, preparation of results, or decision to submit for publication. Competing interests: All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf (available on request from the corresponding author) and declare: financial support to A-CH, JBK, and CD as described above; no other financial relationships with any organisations that might have an interest in the submitted work in the previous three years; no other relationships or activities that could appear to have influenced the submitted work. Ethical approval: This research has ethical approval as part of Psychiatry Sweden “Psykisk ohälsa, psykiatrisk sjukdom: förekomst och etiologi,” approved by the Stockholm Regional Ethical Review Board (number 2010/1185-31/5). Data sharing: The statistical code is available from the corresponding author. Under Swedish law and ethical approval, patient level data cannot be made available.

Ethical approval: This research has ethical approval as part of Psychiatry Sweden “Psykisk ohälsa, psykiatrisk sjukdom: förekomst och etiologi,” approved by the Stockholm Regional Ethical Review Board (number 2010/1185-31/5). Data sharing: The statistical code is available from the corresponding author. Under Swedish law and ethical approval, patient level data cannot be made available. Transparency declaration: The lead authors (the manuscript's guarantors) affirm that the manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained.

Introduction Antimicrobial resistance is an internationally recognised threat to health. The contribution of primary healthcare is particularly important as this is where almost 80% of all antibiotics used within the health service are prescribed.1 Bacterial infections resistant to antibiotics can limit the availability of effective treatment options, rendering some commonly encountered bacterial infections difficult to treat, including those of the urinary tract. Antibiotic resistant infections are also twice as likely to be associated with greater morbidity and mortality and are associated with increased healthcare costs.2 In low income countries, affordability of second line drugs and reduced access to healthcare can restrict the use of newer broad spectrum antibiotics, resulting in growing concerns for increased morbidity and mortality from antibiotic resistant infections in these countries.3

ssociated with increased healthcare costs.2 In low income countries, affordability of second line drugs and reduced access to healthcare can restrict the use of newer broad spectrum antibiotics, resulting in growing concerns for increased morbidity and mortality from antibiotic resistant infections in these countries.3 Children receive a lot of primary healthcare services and, as such, receive a disproportionately high number of antibiotics compared with middle aged populations.4 Children are also key drivers of infection within communities and can contribute to the spread of bacteria from person to person. Despite this, little research has been published describing the prevalence of bacterial resistance in children or the risk factors of importance in this group. In 2010, Costelloe and colleagues conducted a systematic review that reported strong associations between previous exposure to routinely prescribed antibiotics in primary care and antimicrobial resistance persisting for up to 12 months.5 Most of the contributing studies, however, were conducted in adults.

n this group. In 2010, Costelloe and colleagues conducted a systematic review that reported strong associations between previous exposure to routinely prescribed antibiotics in primary care and antimicrobial resistance persisting for up to 12 months.5 Most of the contributing studies, however, were conducted in adults. Urinary tract infections are one of the most common bacterial infections seen in primary care.6 In children with a suspected urinary tract infection, the most common management strategy is to treat empirically with an antibiotic while results of culture and sensitivity testing are awaited. Young children are more vulnerable to immediate and long term complications, including renal scarring and renal failure,7 and therefore require prompt appropriate treatment. Escherichia coli is responsible for over 80% of all urinary tract infections8 and is also the most common cause of bacteraemia and foodborne infections and a cause of meningitis in neonates.9

te and long term complications, including renal scarring and renal failure,7 and therefore require prompt appropriate treatment. Escherichia coli is responsible for over 80% of all urinary tract infections8 and is also the most common cause of bacteraemia and foodborne infections and a cause of meningitis in neonates.9 We conducted a systematic review to investigate the prevalence of resistance in community acquired E coli urinary tract infection to the most commonly prescribed antibiotics given to children in primary care and to quantify the relation between previous exposure to antibiotics in primary care and bacterial resistance. We stratified results by OECD (Organisation for Economic Co-operation and Development) status of the study countries as antibiotics tend to be used differently in these groups. In the more developed OECD countries antibiotics are obtained mostly only by prescription, whereas in “developing” non-OECD countries many antibiotics, including those commonly used to treat urinary tract infection, can be obtained over the counter, without the need for a prescription.10 11 12 13 14

ly in these groups. In the more developed OECD countries antibiotics are obtained mostly only by prescription, whereas in “developing” non-OECD countries many antibiotics, including those commonly used to treat urinary tract infection, can be obtained over the counter, without the need for a prescription.10 11 12 13 14 Methods Search strategy and selection criteria We searched Medline, Embase, and Cochrane for articles published in any language between 1955 and October 2015. MeSH terms for these databases included “drug resistance”, “antimicrobial resistance”, “bacterial resistance”, “primary health care”, “urinary tract infections”, and “children”. MeSH terms were combined with text word searches that included “antibiotic(s)”, “primary care”, “family practice”, “ambulatory care”, “community”, “UTI”, and “urinary bacteria”. Grey and unpublished literature was searched for with ISI Web of Knowledge software and included journal articles, patents, websites, conference proceedings, government and national reports, and open access material. We screened reference lists of selected key papers and contacted authors who appeared multiple times to request details of further published and unpublished work. All full text papers were subject to citation searches. Appendix 1 details the full search strategy. Our review protocol was published on PROSPERO (www.crd.york.ac.uk/PROSPERO/).

lists of selected key papers and contacted authors who appeared multiple times to request details of further published and unpublished work. All full text papers were subject to citation searches. Appendix 1 details the full search strategy. Our review protocol was published on PROSPERO (www.crd.york.ac.uk/PROSPERO/). Two independent reviewers (AB and HT) screened all titles and abstracts independently for eligibility. Studies were eligible for inclusion if they met the following criteria: investigated and reported patterns of resistance in laboratory diagnosed E coli positive isolates from children with urinary tract infection from primary care, defined as the first point of contact in the healthcare system; or investigated associations between previous antibiotic exposure and bacterial resistance; and study participants were children and young people aged 0-17 presenting with symptoms of urinary tract infection who had provided a urine sample. We included hospital based studies when it was clear that the investigation was for community acquired urinary tract infection, defined as a laboratory diagnosed infection from urine samples taken within 48 hours of admission.

ged 0-17 presenting with symptoms of urinary tract infection who had provided a urine sample. We included hospital based studies when it was clear that the investigation was for community acquired urinary tract infection, defined as a laboratory diagnosed infection from urine samples taken within 48 hours of admission. Data extraction and quality assessment Full text papers for all eligible studies were obtained, and three reviewers (AB, CC, and IL) extracted data independently using a purpose built spreadsheet. The following information was extracted from each paper, when provided: author, journal, year of publication, study design, study country, economic status, participants and recruitment location, recruitment time period, age range, method of urine sample collection and testing, method of antimicrobial sensitivity testing, bacteria cultured and reported antibiotic sensitivities, previously prescribed antibiotics, and time between antibiotic exposure and urine sample collection. Level of development was measured with the OECD status of the country in which the study was conducted.15 The OECD is an international economic organisation first established in 1948, now made up of 34 countries, which aims to work together and with emerging and developing economies to reduce poverty through economic growth and financial stability.15 Member countries tend to be “developed” countries, whereas non-member countries tend to be “developing.” We used OECD status as a general measure of country level development and primary care infrastructure and a proxy marker for use of over-the-counter antibiotics. For antimicrobial exposure, time was generally recorded as a period of days, weeks, or months before the urine sample was taken and resistance was measured with standard local laboratory methods. When any information was unclear in the paper, we contacted authors for clarification.

use of over-the-counter antibiotics. For antimicrobial exposure, time was generally recorded as a period of days, weeks, or months before the urine sample was taken and resistance was measured with standard local laboratory methods. When any information was unclear in the paper, we contacted authors for clarification. We extracted and reported resistance to antibiotics commonly prescribed and reported for urinary tract infection in children in primary care: ampicillin, co-amoxiclav (amoxicillin-clavulanic acid), co-trimoxazole (trimethoprim-sulfamethoxazole), trimethoprim, nitrofurantoin, ciprofloxacin, and ceftazidime (as a marker for cephalosporin resistance). Ampicillin was reported in place of amoxicillin because of more frequent reporting and its equivalence in spectrum of antimicrobial activity.16

illin-clavulanic acid), co-trimoxazole (trimethoprim-sulfamethoxazole), trimethoprim, nitrofurantoin, ciprofloxacin, and ceftazidime (as a marker for cephalosporin resistance). Ampicillin was reported in place of amoxicillin because of more frequent reporting and its equivalence in spectrum of antimicrobial activity.16 We used the Cochrane collaboration’s risk of bias tool to assess papers for quality.17 Selection bias was assessed with the Critical Appraisal Skills Programme (CASP) checklist for cohort and case-control studies (www.casp-uk.net). We produced quality assessment charts based on a traffic light system of “good,” “adequate,” and “poor” reporting (see appendix 2), as recommended by Cochrane.17 Our key quality criteria for eligible studies were a reliable measure of antibiotic resistance; clear reporting of bacterial resistance in children and young people aged up to 17; and clear reporting of urinary bacteria isolated as community acquired. The same key quality indicators applied for papers that included information on previous antibiotic exposure, with the addition of adjustment for confounders including age, sex, previous admission to hospital, and comorbidities. Data synthesis and analysis All statistical analyses were conducted with Stata version 13 software, and all methods undertaken according to PRISMA guidelines.18

We used the Cochrane collaboration’s risk of bias tool to assess papers for quality.17 Selection bias was assessed with the Critical Appraisal Skills Programme (CASP) checklist for cohort and case-control studies (www.casp-uk.net). We produced quality assessment charts based on a traffic light system of “good,” “adequate,” and “poor” reporting (see appendix 2), as recommended by Cochrane.17 Our key quality criteria for eligible studies were a reliable measure of antibiotic resistance; clear reporting of bacterial resistance in children and young people aged up to 17; and clear reporting of urinary bacteria isolated as community acquired. The same key quality indicators applied for papers that included information on previous antibiotic exposure, with the addition of adjustment for confounders including age, sex, previous admission to hospital, and comorbidities. Data synthesis and analysis All statistical analyses were conducted with Stata version 13 software, and all methods undertaken according to PRISMA guidelines.18 We calculated estimates of pooled prevalence of resistance by generating a forest plot for each antibiotic, stratified by OECD status. Forest plots illustrated the proportion of resistant E coli for each country, along with 95% confidence intervals, and the pooled prevalence of resistance per antibiotic per economic country group (OECD v non-OECD). We calculated pooled estimates for each country and for OECD and non-OECD groups using the pooled country estimates. Pooled prevalence estimates were generated for children/young people of all age groups (ages 0-17) and children aged 0-5, for comparison. When we could identify the first line antibiotics, these were indicated in the forest plot. I2 of 25%, 50%, and 75% were used to signify low level, moderate level, and high level heterogeneity, in line with Cochrane recommendations.17 All pooled estimates and 95% confidence intervals were generated with double arcsine transformation to adjust for variance instability. This avoids implausible 95% confidence intervals for prevalence estimates when generated under the normal approximation.19

level heterogeneity, in line with Cochrane recommendations.17 All pooled estimates and 95% confidence intervals were generated with double arcsine transformation to adjust for variance instability. This avoids implausible 95% confidence intervals for prevalence estimates when generated under the normal approximation.19 For studies investigating the association between previous antibiotic exposure and bacterial resistance, the outcome measure was the odds ratio of bacterial resistance in children previously exposed to antibiotics compared with those children previously unexposed. The crude estimates from these studies were grouped according to the reported preceding exposure time period (0-1 month, 0-3 months, and 0-6 months). One study investigated exposure at discrete time intervals up to 12 months or more before urine sampling and was reported separately. We carried out a random effects meta-analysis and generated a pooled odds ratio for each exposure time period measured. These were compared with adjusted odds ratios for each time period, when reported. We assessed heterogeneity using the I2 statistic. Meta-regression was used to investigate differences in the odds ratios between antibiotic exposure and resistance across different time periods. Finally, we generated funnel plots to explore the possibility of small study effects, which can be caused by publication bias.

e assessed heterogeneity using the I2 statistic. Meta-regression was used to investigate differences in the odds ratios between antibiotic exposure and resistance across different time periods. Finally, we generated funnel plots to explore the possibility of small study effects, which can be caused by publication bias. Patient involvement No patients were involved in setting the research question or the outcome measures, nor were they involved in developing plans for design or implementation of the review. No patients were asked to advise on interpretation or writing up of results. There are no plans to disseminate the results of the research to study participants or the relevant patient community.

rch question or the outcome measures, nor were they involved in developing plans for design or implementation of the review. No patients were asked to advise on interpretation or writing up of results. There are no plans to disseminate the results of the research to study participants or the relevant patient community. Results Study characteristics We identified 4246 articles through database searches. Of these, we assessed 3115 non-duplicated papers and excluded 2491 on basis of title (fig 1). The 624 remaining papers were assessed by abstract screening; 540 did not meet our eligibility criteria. We obtained and assessed 84 full text papers, with 26 papers not meeting our eligibility criteria for the following reasons: 12 had no primary care data, 11 did not report antibiotic susceptibilities for E coli urinary tract infection bacteria, two studies were in adults, and one paper reported duplicate data from another included paper. We therefore included 58 papers in our review,8 20 21 22 23 24 25 26 27 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 65 66 67 68 69 70 71 72 73 74 75 76 of which five papers (all from OECD countries) reported information on previous antibiotic exposure and were included in our meta-analysis. No grey literature or national reports were eligible for inclusion in the review. Fig 1 Data search and extraction (PRISMA flow chart)

Results Study characteristics We identified 4246 articles through database searches. Of these, we assessed 3115 non-duplicated papers and excluded 2491 on basis of title (fig 1). The 624 remaining papers were assessed by abstract screening; 540 did not meet our eligibility criteria. We obtained and assessed 84 full text papers, with 26 papers not meeting our eligibility criteria for the following reasons: 12 had no primary care data, 11 did not report antibiotic susceptibilities for E coli urinary tract infection bacteria, two studies were in adults, and one paper reported duplicate data from another included paper. We therefore included 58 papers in our review,8 20 21 22 23 24 25 26 27 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 65 66 67 68 69 70 71 72 73 74 75 76 of which five papers (all from OECD countries) reported information on previous antibiotic exposure and were included in our meta-analysis. No grey literature or national reports were eligible for inclusion in the review. Fig 1 Data search and extraction (PRISMA flow chart) Table 1 summarises the characteristics of the 58 included studies (full details are in appendix 3). Thirty three studies from OECD countries reported resistance in 73 375 E coli isolates from the same number of children, with the exception of one UK study that included multiple urine isolates per child. As data reported in the UK study were analysed with a two level model of samples nested within patients, we reported it separately in our meta-analysis.24 All studies were observational; 25 were retrospective, six prospective, and two case-control. Thirty reported information on prevalence of resistance in E coli urinary tract infection isolates, with the three remaining reporting the association between previous antibiotic exposure and E coli resistance only.22 38 51Table 1 also summarises the 25 studies included from non-OECD studies that reported bacterial resistance in 4408 E coli isolates from the same number of children. All were observational; 10 were retrospective, 11 prospective, one case-control, and three cross sectional. All 25 non-OECD studies reported information on prevalence of resistance in urinary E coli. No non-OECD studies reported information on previous antibiotic exposure. Figure 2 shows the number of studies per country included in the review. Most studies were conducted in OECD countries, and there were relatively few studies from each country.

s reported information on prevalence of resistance in urinary E coli. No non-OECD studies reported information on previous antibiotic exposure. Figure 2 shows the number of studies per country included in the review. Most studies were conducted in OECD countries, and there were relatively few studies from each country. Table 1 Characteristics of included papers on antibiotic resistance in paediatric E coli urinary tract infections by OECD (Organisation for Economic Co-operation and Development) status of study country

s reported information on prevalence of resistance in urinary E coli. No non-OECD studies reported information on previous antibiotic exposure. Figure 2 shows the number of studies per country included in the review. Most studies were conducted in OECD countries, and there were relatively few studies from each country. Table 1 Characteristics of included papers on antibiotic resistance in paediatric E coli urinary tract infections by OECD (Organisation for Economic Co-operation and Development) status of study country Study characteristics No of papers from OECD countries (n=33) No of papers from non-OECD countries (n=25) Study design: Retrospective observational 258 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 1052 53 54 55 56 57 58 59 60 61 Prospective observational 644 45 46 47 48 49 1162 63 64 65 66 67 68 69 70 71 72 Case-control 250 51 173 Cross-sectional 0 374 75 76 No of children in study: 0-100 235 40 756 62 64 66 67 68 76 101-500 1223 25 29 32 36 38 44 45 46 47 48 49 1353 54 55 58 60 61 65 69 71 72 73 74 75 501-1000 628 30 33 39 43 50 252 57 1001-10 000 721 24 27 31 37 42 51 259 70 ≥10 001 68 20 22 26 34 41 163 Age range (years)*: 0-5 98 29 31 33 34 36 43 45 50 662 65 69 74 75 76 6-17 58 31 34 36 50 0 0-17 308 20 21 22 23 24 25 26 27 28 30 32 33 34 35 36 37 38 39 40 41 42 44 46 47 48 49 50 51 1952 53 54 55 56 57 58 59 60 61 63 64 66 67 68 70 71 72 73 Recruitment location: GP practice/paediatric office 1221 22 24 25 26 28 30 32 34 37 41 50 553 58 61 67 70 Outpatient/clinic 108 27 29 36 40 42 47 48 49 51 955 56 57 59 63 64 68 71 73 Emergency department 720 35 38 39 43 45 46 160 Hospital admission 423 31 33 44 952 54 65 66 69 72 74 75 76 Not reported 0 162 Method of urine sampling: At least one of clean catch, catheter, or suprapubic aspiration 2022 23 25 27 28 29 30 31 32 33 34 35 38 39 43 44 46 47 49 51 1153 54 61 63 65 67 72 74 75 76 79 Clean catch only 326 40 41 459 64 70 73 Catheter only 145 0 Suprapubic aspiration only 0 360 62 69 Not reported 98 20 21 24 36 37 42 48 50 752 55 56 57 66 68 71 Antibiotic susceptibilities reported: Ampicillin 258 20 21 23 26 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 1552 53 56 57 58 61 62 65 68 69 70 71 72 73 74 Co-amoxiclav 218 20 21 23 25 26 27 28 29 30 31 33 34 37 41 42 44 46 47 48 49 852 61 62 69 70 72 73 74 Co-trimoxazole 248 20 21 23 25 27 28 29 30 31 32 33 34 36 38 39 42 44 45 46 47 48 49 50 1852 53 54 56 57 58 59 60 61 64 65 66 67 69 72 73 74 76 Trimethoprim 724 26 33 35 37 41 43 170 Nitrofurantoin 218 20 25 26 27 28 32 33 35 37 38 39 41 42 43 44 45 46 47 49 50 1853 54 57 58 59 60 61 62 64 66 67 68 69 70 72 73 75 76 Ciprofloxacin 178 20 25 26 27 28 29 30 31 32 33 35 37 41

36 38 39 42 44 45 46 47 48 49 50 1852 53 54 56 57 58 59 60 61 64 65 66 67 69 72 73 74 76 Trimethoprim 724 26 33 35 37 41 43 170 Nitrofurantoin 218 20 25 26 27 28 32 33 35 37 38 39 41 42 43 44 45 46 47 49 50 1853 54 57 58 59 60 61 62 64 66 67 68 69 70 72 73 75 76 Ciprofloxacin 178 20 25 26 27 28 29 30 31 32 33 35 37 41 42 45 46 47 48 1152 55 56 58 59 61 62 63 64 66 68 Ceftazidime 1020 25 28 29 31 39 41 45 46 48 852 53 55 56 62 70 73 75 Method of antimicrobial susceptibility testing: Disk diffusion 2320 21 23 25 27 29 31 32 33 35 36 37 39 40 41 42 43 44 45 47 48 49 51 2152 53 54 55 56 57 58 59 60 63 65 66 67 68 69 70 71 72 73 74 75 76 Minimum inhibitory concentration 28 34 0 Vitek 326 28 50 0 Not reported 522 24 30 38 46 455 61 62 64 Guidelines used to interpret antimicrobial sensitivities: CLSI 258 20 21 22 23 25 26 27 28 29 31 32 34 35 36 38 39 41 43 44 47 48 49 50 51 1853 54 56 57 58 60 63 64 65 66 67 68 69 71 72 74 75 76 BSAC 137 0 Not reported 724 30 33 40 42 45 46 752 55 59 61 62 70 73 Previous antibiotic exposure information† 522 24 38 50 51 0 CLSI=Clinical Laboratory Standards Institute; BSAC=British Standard for Antimicrobial Chemotherapy. *Age 0-5: papers that report data specifically for this age group; 6-17: papers that report data specifically for this age group; 0-17: papers that which report data for children/young people within 0-17 and do not fit into previous reported age groups. Papers can appear more than once depending on how results are reported.

0-5: papers that report data specifically for this age group; 6-17: papers that report data specifically for this age group; 0-17: papers that which report data for children/young people within 0-17 and do not fit into previous reported age groups. Papers can appear more than once depending on how results are reported. †No studies from non-OECD countries collected previous antibiotic exposure data and were not included in meta-analysis. Fig 2 Geographical distribution of urinary E coli resistance prevalence to ampicillin (%) by OECD and non-OECD countries,15 with number of included studies per country in parentheses)

0-5: papers that report data specifically for this age group; 6-17: papers that report data specifically for this age group; 0-17: papers that which report data for children/young people within 0-17 and do not fit into previous reported age groups. Papers can appear more than once depending on how results are reported. †No studies from non-OECD countries collected previous antibiotic exposure data and were not included in meta-analysis. Fig 2 Geographical distribution of urinary E coli resistance prevalence to ampicillin (%) by OECD and non-OECD countries,15 with number of included studies per country in parentheses) Thirty one (20 OECD v 11 non-OECD) studies used mixed methods for urine collection, including clean catch, catheter, or suprapubic aspiration, with the remaining studies using a single method. There were no differences in rates of resistance detected between the different methods of urine sampling that studies used. Antimicrobial sensitivity testing was carried out with standard disk diffusion methods in 44 studies, with one interpreted and reported according to British Standard for Antimicrobial Chemotherapy (BSAC), 43 with Clinical and Laboratory Standards Institute (CLSI) guidelines,77 78 and 14 not reported. There were no differences in resistance between studies that did and did not report the use of antimicrobial sensitivity guidelines. All children had presented to primary care facilities (18 studies), outpatient clinics (19 studies), or emergency departments (eight studies) with symptoms of a urinary tract infection, with some children sent to a secondary or tertiary care hospital for urine tests on admission (12 studies). Of the 12 inpatient studies, nine stated that they included only community acquired E coli isolates; the three remaining confirmed urine samples were collected within 48 hours of admission.

inary tract infection, with some children sent to a secondary or tertiary care hospital for urine tests on admission (12 studies). Of the 12 inpatient studies, nine stated that they included only community acquired E coli isolates; the three remaining confirmed urine samples were collected within 48 hours of admission. The quality assessment “traffic light” charts for the included studies showed that, for the five studies reporting information on antibiotic exposure, reporting was generally good for our all our key quality indicators (appendix 2). For studies reporting prevalence of resistance only, overall quality was good with the exception of adjustment for confounding. Prevalence of resistance in urinary E coli Table 2 shows the prevalence of E coli urinary isolates resistance to antibiotics. These data were obtained from forest plots generated for each antibiotic (appendix 4). Table 2 Pooled percentage prevalence (95% confidence interval) of resistance to antibiotics in primary care used to treat urinary E coli infection in children (see appendix 4 for corresponding forest plots) by OECD (Organisation for Economic Co-operation and Development) status of study country

Prevalence of resistance in urinary E coli Table 2 shows the prevalence of E coli urinary isolates resistance to antibiotics. These data were obtained from forest plots generated for each antibiotic (appendix 4). Table 2 Pooled percentage prevalence (95% confidence interval) of resistance to antibiotics in primary care used to treat urinary E coli infection in children (see appendix 4 for corresponding forest plots) by OECD (Organisation for Economic Co-operation and Development) status of study country Antibiotic OECD Non-OECD Pooled prevalence (%) No of isolates tested No of reporting studies I2 (%) Pooled prevalence (%) No of isolates tested No of reporting studies I2 (%) Ampicillin 53.4 (46.0 to 60.8) 66 503 25 (11 countries)8 20 21 23 26 28 29 30 31 32 33 34 35 37 38 39 40 41 42 44 46 47 48 49 50 7 79.8 (73.0 to 87.7) 2265 15 (11 countries)52 53 56 57 58 61 62 65 68 69 70 71 72 73 74 25 Co-amoxiclav 8.2 (7.9 to 9.6) 65 076 21 (9 countries)8 20 21 23 25 26 27 28 29 30 31 33 34 37 41 42 44 46 47 48 49 45 60.3 (40.9 to 79.0) 1256 8 (8 countries)52 61 62 69 70 72 73 74 62 Co-trimoxazole 30.2 (20.5 to 39.3) 50 230 24 (9 countries)8 20 21 23 25 27 28 29 30 31 32 33 34 36 38 39 42 44 45 46 47 48 49 50 28 69.6 (59.8 to 81.5) 2590 18 (10 countries)52 53 54 56 57 58 59 60 61 64 65 66 67 69 72 73 74 76 37 Trimethoprim 23.6 (13.9 to 32.3) 18 977 7 (5 countries)24 26 33 35 37 41 43 16 Too few data* 596 1 (1 country)70 Too few data* Nitrofurantoin 1.3 (0.8 to 1.7) 50 994 21 (13 countries)8 20 25 26 27 28 32 33 35 37 38 39 40 41 42 43 44 45 46 47 49 50 0 17.0 (9.8 to 24.2) 3020 18 (10 countries)53 54 57 58 59 60 61 62 64 66 67 68 69 70 72 73 75 76 42 Ciprofloxacin 2.1 (0.8 to 4.4) 52 209 17 (9 countries)8 20 25 26 27 28 31 32 33 35 37 41 42 45 46 47 48 59 26.8 (11.1 to 43.0) 1723 11 (7 countries)52 55 56 58 59 61 62 63 64 66 68 35 Ceftazidime† 2.4 (0.9 to 3.3) 25 805 10 (8 countries)20 25 28 29 31 39 41 45 46 48 58 26.1 (14.6 to 37.5) 1136 8 (5 countries)52 53 55 56 62 70 73 75 54 *Only one study from non-OECD countries (Saudi-Arabia).

1 32 33 35 37 41 42 45 46 47 48 59 26.8 (11.1 to 43.0) 1723 11 (7 countries)52 55 56 58 59 61 62 63 64 66 68 35 Ceftazidime† 2.4 (0.9 to 3.3) 25 805 10 (8 countries)20 25 28 29 31 39 41 45 46 48 58 26.1 (14.6 to 37.5) 1136 8 (5 countries)52 53 55 56 62 70 73 75 54 *Only one study from non-OECD countries (Saudi-Arabia). †Marker for cephalosporin resistance. For all antibiotics tested, the prevalence of antibiotic resistance was higher in non-OECD than in OECD countries. For all countries the prevalence of resistance was highest for ampicillin. Figure 2 shows the pooled prevalence (or single study reported prevalence if n=1) of ampicillin resistance by country. Switzerland had the lowest prevalence at 41%, with Ghana and Nigeria highest at 100%. Pooled prevalences of resistance to co-trimoxazole and trimethoprim were high in OECD countries, with co-trimoxazole resistance at 30%. Resistance to co-trimoxazole was more than twice as high in non-OECD compared with OECD countries. Trimethoprim resistance was reported in only one non-OECD study, conducted by Al-Mugeiren and colleagues in Saudi-Arabia, which reported 67% resistance from 596 isolates.70 Nitrofurantoin resistance was the lowest of all reported for all countries. Pooled prevalences of resistance to ciprofloxacin and ceftazidime in children’s E coli urinary isolates were both around 2% in OECD countries; however, resistance to both antibiotics was over 10 times higher in non-OECD countries, both over 26% (table 2).

Pooled prevalences of resistance to co-trimoxazole and trimethoprim were high in OECD countries, with co-trimoxazole resistance at 30%. Resistance to co-trimoxazole was more than twice as high in non-OECD compared with OECD countries. Trimethoprim resistance was reported in only one non-OECD study, conducted by Al-Mugeiren and colleagues in Saudi-Arabia, which reported 67% resistance from 596 isolates.70 Nitrofurantoin resistance was the lowest of all reported for all countries. Pooled prevalences of resistance to ciprofloxacin and ceftazidime in children’s E coli urinary isolates were both around 2% in OECD countries; however, resistance to both antibiotics was over 10 times higher in non-OECD countries, both over 26% (table 2). When we stratified by “reported first line” antibiotic versus “first line not specified” for each country, estimates of prevalence of resistance for OECD countries were similar to overall OECD estimates reported in table 2, with little difference in estimates when first line treatment was specified or not. In non-OECD countries, however, pooled estimates of resistance were generally higher for those countries that specified the antibiotic as first line. The difference was particularly large for co-trimoxazole (first line pooled resistance 76.2% (95% confidence interval 64.1% to 87.2%) versus non-first line resistance 55.6% (26.6% to 84.7%)) and ciprofloxacin (first line pooled resistance 58.1% (51.5% to 64.7%) versus non-first-line pooled resistance 15.8% (4.7% to 26.8%) (appendix 5).

s particularly large for co-trimoxazole (first line pooled resistance 76.2% (95% confidence interval 64.1% to 87.2%) versus non-first line resistance 55.6% (26.6% to 84.7%)) and ciprofloxacin (first line pooled resistance 58.1% (51.5% to 64.7%) versus non-first-line pooled resistance 15.8% (4.7% to 26.8%) (appendix 5). Prevalence of resistance in children aged 0-5 Twelve studies reported resistance in urinary E coli specifically for children aged 0-5, seven from OECD countries and five from non-OECD countries (table 3). As there were insufficient data to compare with children and young people aged 6-17, we compared these data with data from all children (table 2). As with all children, the prevalence of antibiotic resistance in children aged 0-5 was higher in non-OECD than OECD countries. Compared with the data for all children, in OECD countries the pooled prevalence of resistance in urinary E coli in children aged 0-5 was higher for ampicillin and ceftazidime, and lower for co-amoxiclav, co-trimoxazole, and nitrofurantoin (table 2). In non-OECD countries, resistance was higher against all reported antibiotics for children aged 0-5 compared with all children. Table 3 Pooled prevalence (%) of resistance to antibiotics in primary care used to treat urinary E coli infection in children aged 0-5 by OECD (Organisation for Economic Co-operation and Development) status of study country

Prevalence of resistance in children aged 0-5 Twelve studies reported resistance in urinary E coli specifically for children aged 0-5, seven from OECD countries and five from non-OECD countries (table 3). As there were insufficient data to compare with children and young people aged 6-17, we compared these data with data from all children (table 2). As with all children, the prevalence of antibiotic resistance in children aged 0-5 was higher in non-OECD than OECD countries. Compared with the data for all children, in OECD countries the pooled prevalence of resistance in urinary E coli in children aged 0-5 was higher for ampicillin and ceftazidime, and lower for co-amoxiclav, co-trimoxazole, and nitrofurantoin (table 2). In non-OECD countries, resistance was higher against all reported antibiotics for children aged 0-5 compared with all children. Table 3 Pooled prevalence (%) of resistance to antibiotics in primary care used to treat urinary E coli infection in children aged 0-5 by OECD (Organisation for Economic Co-operation and Development) status of study country Antibiotic OECD Non-OECD Pooled prevalence (%) No of isolates tested No of reporting studies I2 (%) Pooled prevalence (%) No of isolates tested No of reporting studies I2 (%) Ampicillin 55.0 (48.6 to 61.4) 5273 5 (4 countries)8 29 31 33 34 10 90.3 (73.4 to 100) 176 3 (3 countries)65 69 74 0 Co-amoxiclav 9.6 (5.7 to 13.5) 5273 5 (4 countries)8 29 31 33 34 51 71.9 (40.7 to 100) 89 3 (3 countries)62 69 74 66 Co-trimoxazole 29.8 (21.0 to 38.5) 5405 7 (5 countries)8 29 31 33 34 36 45 39 71.0 (44.9 to 97.0) 257 5 (4 countries)65 69 74 75 76 0 Trimethoprim Too few data* 188 1 (1 country)33 Too few data* No data† 0 0 — Nitrofurantoin 0.4 (0.0 to 0.7) 3089 5 (5 countries)8 33 29 43 45 45 35.2 (31.6 to 38.8) 145 3 (3 countries)62 69 75 0 Ciprofloxacin 6.2 (3.2 to 9.3) 4544 4 (4 countries)8 31 33 45 33 Too few data‡ 49 1 (1 country)62 Too few datac Ceftazidime§ 4.9 (0.3 to 9.5) 1535 4 (4 countries)29 31 33 45 28 43.6 (9.0 to 78.2) 130 2 (2 countries)62 75 0 *Only one study from OECD countries (Austria).

35.2 (31.6 to 38.8) 145 3 (3 countries)62 69 75 0 Ciprofloxacin 6.2 (3.2 to 9.3) 4544 4 (4 countries)8 31 33 45 33 Too few data‡ 49 1 (1 country)62 Too few datac Ceftazidime§ 4.9 (0.3 to 9.5) 1535 4 (4 countries)29 31 33 45 28 43.6 (9.0 to 78.2) 130 2 (2 countries)62 75 0 *Only one study from OECD countries (Austria). †No studies from non-OECD countries reported resistance to trimethoprim in children aged 0-5. ‡Only one study from non-OECD countries (India). §Marker for cephalosporin resistance.

35.2 (31.6 to 38.8) 145 3 (3 countries)62 69 75 0 Ciprofloxacin 6.2 (3.2 to 9.3) 4544 4 (4 countries)8 31 33 45 33 Too few data‡ 49 1 (1 country)62 Too few datac Ceftazidime§ 4.9 (0.3 to 9.5) 1535 4 (4 countries)29 31 33 45 28 43.6 (9.0 to 78.2) 130 2 (2 countries)62 75 0 *Only one study from OECD countries (Austria). †No studies from non-OECD countries reported resistance to trimethoprim in children aged 0-5. ‡Only one study from non-OECD countries (India). §Marker for cephalosporin resistance. Association between previous antibiotic exposure and bacterial resistance Figure 3 shows a forest plot of five studies investigating the relation between previous exposure to any versus no antibiotics and bacterial resistance. The studies varied in the combinations of antibiotic resistance and exposure investigated, some reporting resistance and exposure to any antibiotic, while others reported resistance to trimethoprim, co-trimoxazole, or third generation cephalosporins. In figure 3, for all time periods of exposure the crude odds of resistance were greater in children exposed to antibiotics than in those who were unexposed. The effect sizes show a decline between exposure time periods of 0-1 month (pooled odds ratio 8.38, 95% confidence interval 2.84 to 24.77) and 0-3 months (3.38, 2.05 to 5.55), then an increase at 0-6 months (13.23, 7.84 to 22.31). The study by Allen and colleagues, which explored the association between exposure to any antibiotic in the previous six months and resistance to co-trimoxazole, measured previous exposure to antibiotics for four weeks or more within the six months before the urine sample was taken. The three other studies shown in figure 3 measured exposure to any antibiotic for an unspecified total prescription time period. Given the overlap in the exposure time periods reported by the included studies, we did not conduct a meta-regression analysis for the data presented in figure 3.

the urine sample was taken. The three other studies shown in figure 3 measured exposure to any antibiotic for an unspecified total prescription time period. Given the overlap in the exposure time periods reported by the included studies, we did not conduct a meta-regression analysis for the data presented in figure 3. Fig 3 Pooled crude odds ratios (log scale) for resistance in children’s urinary bacteria and previous exposure to any antibiotic. Studies grouped according to time period after antibiotic use during which exposure was measured and ordered within each time period by increasing standard error There was no evidence of heterogeneity within groups in the 0-1 month time period, with too few studies in the 0-3 and 0-6 months time periods to calculate heterogeneity. Of the four studies included in figure 3, three reported odds ratios adjusted for sex, age, race, renal abnormalities, and previous admission to hospital. The adjusted odds ratio did not differ substantially from the crude pooled estimates.

few studies in the 0-3 and 0-6 months time periods to calculate heterogeneity. Of the four studies included in figure 3, three reported odds ratios adjusted for sex, age, race, renal abnormalities, and previous admission to hospital. The adjusted odds ratio did not differ substantially from the crude pooled estimates. The study by Duffy and colleagues was the only one of those measuring the association between antibiotic exposure and resistance to report results based on multiple urinary isolates per child and with a more accurate measure of antibiotic exposure24; therefore we chose to report this study separately. Figure 4 shows the crude multilevel odds ratios for resistance to trimethoprim relative to the number of days since exposure to trimethoprim. Duffy and colleagues reported multilevel crude odds ratios for the association between exposure and resistance to trimethoprim, based on the number of urinary isolates reported in the paper not individual patients, along with the number of isolates with reported exposure to trimethoprim only for each time period. The sample level variables included in the model were age at test, time since most recent trimethoprim prescription, and year of test; patient level variables included sex, socioeconomic status, rurality, and total number of E coli isolates in the study period. The crude odds ratios in figure 4 show a decrease in resistance to trimethoprim with increasing time since exposure to trimethoprim. We conducted a meta-regression analysis on the crude odds ratios calculated from the paper by Duffy and colleagues, which showed a β coefficient of −0.4 (95% confidence interval −0.61 to −0.19), indicating an important time trend.

crease in resistance to trimethoprim with increasing time since exposure to trimethoprim. We conducted a meta-regression analysis on the crude odds ratios calculated from the paper by Duffy and colleagues, which showed a β coefficient of −0.4 (95% confidence interval −0.61 to −0.19), indicating an important time trend. Fig 4 Individual crude multilevel odds ratios for trimethoprim resistance in urinary isolates of children from Duffy and colleagues24 and previous trimethoprim prescribing Publication bias There were too few studies to assess publication bias. Discussion Principal findings The 58 studies from both OECD and non-OECD countries provide evidence of high rates of bacterial resistance in E coli isolates from children with urinary tract infection to some of the most commonly prescribed antibiotics in primary care. Worldwide, rates of resistance to ampicillin were the highest and nitrofurantoin rates lowest, irrespective of OECD status. Resistance to all reported antibiotics was higher in non-OECD than OECD countries, with resistance rates higher to first line than to non-first line antibiotics. Resistance could render several first line antibiotics ineffective in some countries. Prescribing of antibiotics to individual children in primary care is an important contributor to bacterial resistance, which can persist for up to six months after prescription.

higher to first line than to non-first line antibiotics. Resistance could render several first line antibiotics ineffective in some countries. Prescribing of antibiotics to individual children in primary care is an important contributor to bacterial resistance, which can persist for up to six months after prescription. Strengths and limitations To our knowledge, this is the first systematic review and meta-analysis to explore and report global evidence regarding the prevalence of bacterial resistance in children’s urinary tract infection and associations with the routine use of antibiotics in primary care. WHO recently published their “global action plan” on antimicrobial resistance, which described data relating to the prevalence of resistance, including geographical patterns, as a key gap in our current knowledge,80 which this systematic review in part fills. Our review was rigorously conducted according to the Cochrane guidelines.17 We chose to stratify our results by OECD status to reflect both national development and likely availability of over-the-counter antibiotics.3 81

cal patterns, as a key gap in our current knowledge,80 which this systematic review in part fills. Our review was rigorously conducted according to the Cochrane guidelines.17 We chose to stratify our results by OECD status to reflect both national development and likely availability of over-the-counter antibiotics.3 81 We are, however, aware of several limitations. Firstly, antibiotics are used differently within OECD and non-OECD countries,82 83 84 85 and over-the-counter use is difficult to measure. A 2011 systematic review reported high worldwide variability in non-prescription antibiotics,81 with some evidence of less than 100% agreement between OECD status and over-the-counter availability. To our knowledge there is no better country level alternative, and none of the included studies reported or measured availability of over-the-counter antibiotics. We also acknowledge that factors other than antibiotic use and over-the-counter availability can account for differences in resistance prevalence between OECD and non-OECD countries, including poor sanitation, unstable governance, and lower levels of regulation of medicines. Although it is useful to explore changes in resistance over time, (for example, to understand the impact of vaccination), we were also unable to explore this as the data collected overlapped in terms of recruitment periods.

es, including poor sanitation, unstable governance, and lower levels of regulation of medicines. Although it is useful to explore changes in resistance over time, (for example, to understand the impact of vaccination), we were also unable to explore this as the data collected overlapped in terms of recruitment periods. Of the five studies we included in our meta-analysis, most reported the association between previous antibiotic exposure and resistance within overlapping time periods. This implies that the associations with longer time periods (such as 0-6 months) could reflect a combination of long and short term associations. The odds ratios were highest in the 0-6 months time period; probably because this individual study measured exposure to antibiotics for a combined total of four weeks or more within the previous six months compared with no exposure within the six month time period.68 The other studies measured exposure to antibiotics within an undefined combined length of prescription time within the measured time period versus no antibiotic prescriptions. None of the studies we included in our meta-analysis reported antibiotic doses, so we were unable to evaluate any dose-response effects.

riod.68 The other studies measured exposure to antibiotics within an undefined combined length of prescription time within the measured time period versus no antibiotic prescriptions. None of the studies we included in our meta-analysis reported antibiotic doses, so we were unable to evaluate any dose-response effects. In most countries it is standard practice to treat empirically with an antibiotic when a child presents to primary care with a suspected urinary tract infection. Sometimes a urine sample is taken only if the illness does not respond to first line antibiotic treatment. This can lead to falsely high reported resistance rates to first line antibiotics. This problem would be removed if only incident cases were included or systematic urine sampling was used, but studies did not present this information. That said, there were no obvious differences in resistance rates according to the timing of the urine sampling. Furthermore, variation in sampling strategies could explain some of the variation in pooled prevalence of resistance between OECD and non-OECD countries, though this could not be explored from the data available. Reverse causality and other confounding associations could also have introduced bias to our findings; including previous hospital admissions, comorbidities, age, and sex. Studies that attempted to adjust for confounding factors, however, did not show differences between crude and adjusted estimates of association.

ble. Reverse causality and other confounding associations could also have introduced bias to our findings; including previous hospital admissions, comorbidities, age, and sex. Studies that attempted to adjust for confounding factors, however, did not show differences between crude and adjusted estimates of association. Results in the context of existing research Prevalence of urinary bacterial resistance We believe our rates of prevalence of resistance are accurate as they are consistent with other data sources. The highest reported resistance to ampicillin in our review was similar to the reported aminopenicillin group resistance in the European EARS-Net database and US Centre for Disease Dynamics, Economics and Policy (CDDEP) databases.86 87 Resistance to ampicillin in other studies from the US ranged between 36% and 54%, suggesting that resistance to antibiotics in young children is similar to that of the general population. The similarities observed here could be a result of transmission between age groups of genetic resistance factors such as plasmids, facilitated through frequent interaction between children and adults. Trimethoprim resistance was reported by three studies from the UK, all with large sample sizes (>1700 isolates); all reported resistance in excess of 20%. These are similar to levels of trimethoprim resistance reported by other UK based studies; Bean and colleagues reported trimethoprim resistance in community acquired urinary isolates from adults and children at 39%.88 Additionally, Farrell and colleagues reported 27% resistance in E coli urinary isolates from all age groups.89 In total, seven OECD studies from five countries (UK, Ireland, Austria, Australia, and Sweden) reported susceptibility to trimethoprim; these were also the only countries to report trimethoprim as a first line antibiotic treatment for urinary tract infection (appendix 5). Trimethoprim resistance was infrequently tested for in many studies from OECD countries, which probably because it is not a first line treatment in their country. Co-trimoxazole was the most common first line treatment for urinary tract infection worldwide (15 countries and 37 studies). Resistance to co-trimoxazole was relatively high worldwide, particularly in non-OECD countries at 64%. Resistance to nitrofurantoin, an antibiotic used almost exclusively for urinary tract infections, was low worldwide, supporting its continued effectiveness as a first line treatment for uncomplicated infections.90 91 92

sistance to co-trimoxazole was relatively high worldwide, particularly in non-OECD countries at 64%. Resistance to nitrofurantoin, an antibiotic used almost exclusively for urinary tract infections, was low worldwide, supporting its continued effectiveness as a first line treatment for uncomplicated infections.90 91 92 For many of the antibiotics reported in this review, the pooled prevalence of resistance was higher in children aged 0-5 than in all children and young people (0-17). It has been previously suggested that resistance levels are likely to be higher in those communities with a higher proportion of young children because of their high consumption of antibiotics.93 A study conducted in France found that children aged under 7 consumed three times more antibiotics than older populations.94 The findings in our review support this theory as resistance to all commonly prescribed antibiotics worldwide was higher in younger children than in children of predominantly older age. Our findings also suggest there could be a reversible element of antibiotic resistance, whereby reduced use of antibiotics (in older children) reduces the selective pressure that favours antibiotic resistant strains.

ribed antibiotics worldwide was higher in younger children than in children of predominantly older age. Our findings also suggest there could be a reversible element of antibiotic resistance, whereby reduced use of antibiotics (in older children) reduces the selective pressure that favours antibiotic resistant strains. Association between previous antibiotic exposure and bacterial resistance Our meta-analysis showing an association between exposure to antibiotics in the previous six months and isolation of resistant urinary isolates is similar to our previous 2010 review, which explored the effect of antibiotic prescribing in primary care on the development of resistance in individual patients of all ages.5 Consistent with our previous review, we found some evidence from Duffy and colleagues of decreasing resistance for increasing time from antibiotic prescribing.24

revious 2010 review, which explored the effect of antibiotic prescribing in primary care on the development of resistance in individual patients of all ages.5 Consistent with our previous review, we found some evidence from Duffy and colleagues of decreasing resistance for increasing time from antibiotic prescribing.24 Policy, clinical, and research implications Our findings detail global high level resistance to some of the most commonly prescribed antibiotics for children primary care, which could result in several drugs becoming ineffective first line treatments in many countries. The Infectious Diseases Society of America (IDSA) in collaboration with the European Society for Microbiology and Infectious Diseases (ESCMID)95 recommend that an antibiotic should be selected for first line empirical treatment of urinary tract infection only if the local prevalence of resistance is less than 20%. According to these guidelines, our review suggests ampicillin, co-trimoxazole, and trimethoprim are no longer suitable first line treatment options for urinary tract infection in many OECD countries and that as a result many guidelines, such as those published by the National Institute for Health and Care Excellence (NICE), might need updating. In non-OECD countries, resistance to all first line antibiotics specified for urinary tract infections was in excess of 20% (appendix 5), suggesting that choices of first line treatment might need to be re-evaluated in less well developed countries. Our results also support the need for prescribing guidelines to reflect patterns of local resistance and that, for many areas, nitrofurantoin might be the most appropriate first line treatment for lower urinary tract infection. That said, care is needed because ruling out the use of some first line antibiotics could lead clinicians to prescribe broad spectrum second line antibiotics, such as co-amoxiclav, cephalosporins, and quinolones, resulting in a vicious cycle of increasing use of broad spectrum antibiotics and bacterial resistance.

ion. That said, care is needed because ruling out the use of some first line antibiotics could lead clinicians to prescribe broad spectrum second line antibiotics, such as co-amoxiclav, cephalosporins, and quinolones, resulting in a vicious cycle of increasing use of broad spectrum antibiotics and bacterial resistance. Prevalence of resistance to common antibiotics in primary care was higher in non-OECD countries than OECD countries, which could be because of weaker infrastructure of primary care, weaker regulation of antibiotic use, and the need for higher use of antibiotic because of higher risks of serious bacterial infection in children living in non-OECD countries. Improved infrastructure of primary care, access to healthcare, and antibiotic regulation might be necessary to reduce the burden of antimicrobial resistance in these settings.

use, and the need for higher use of antibiotic because of higher risks of serious bacterial infection in children living in non-OECD countries. Improved infrastructure of primary care, access to healthcare, and antibiotic regulation might be necessary to reduce the burden of antimicrobial resistance in these settings. Furthermore, the results indicate that bacterial resistance to antibiotics can persist for up to six months after antibiotic exposure in individual children. The study conducted by Duffy and colleagues is an exemplar of how future studies should measure associations between resistance and time since exposure to antibiotics.24 In addition, future studies should also consider inclusion of incident data whenever possible to facilitate better comparison with other studies. Primary care clinicians should consider the impact of any antibiotic use on subsequent antimicrobial resistance and avoid their unnecessary use by following local and national guidance whenever possible. When antibiotic treatment is needed, our findings suggest that clinicians should consider a child’s antibiotic use in the past six months when selecting further treatment, avoiding the use of broad spectrum antibiotics whenever possible.96 Our findings also support other evidence for the continued availability of nitrofurantoin as an effective treatment for uncomplicated urinary tract infections in primary care.91 97

antibiotic use in the past six months when selecting further treatment, avoiding the use of broad spectrum antibiotics whenever possible.96 Our findings also support other evidence for the continued availability of nitrofurantoin as an effective treatment for uncomplicated urinary tract infections in primary care.91 97 Conclusions Prevalence of resistance to commonly prescribed primary care antibiotics in E coli urinary tract infections in children is high, particularly in non-OECD countries, where one possible explanation is availability of antibiotics over the counter. This could render some drugs ineffective as first line treatments for urinary tract infection. Routine use of antibiotics in primary care contributes to antimicrobial resistance in children, which can persist for up to six months after antibiotic prescription. What is already known on this topic Throughout the world, children are prescribed a lot of antibiotics in primary care Such routine use increases the probability of antibiotic resistance in adults with urinary tract infections Substantial variations in antibiotic use exist globally, with over-the-counter availability common in many countries What this study adds Prevalence of antibiotic resistance in urinary tract infection in children caused by E coli is high globally, including to some first line treatments such as trimethoprim Several antibiotics for children commonly used in primary care, including ampicillin (amoxicillin) and trimethoprim, could be ineffective first line treatment options

What this study adds Prevalence of antibiotic resistance in urinary tract infection in children caused by E coli is high globally, including to some first line treatments such as trimethoprim Several antibiotics for children commonly used in primary care, including ampicillin (amoxicillin) and trimethoprim, could be ineffective first line treatment options Urinary tract bacterial isolates from individual children with previous primary care prescriptions for antibiotic were more likely to be resistant to treatment, and this increased risk can persist for up to six months Web Extra Extra material supplied by the author Appendix 1: Medline and Embase search strategy Click here for additional data file. Appendix 2: Data quality charts (by studies reporting prevalence of resistance only and prevalence plus antibiotic exposure) Click here for additional data file. Appendix 3: Study characteristics Click here for additional data file. Appendix 4: Supplementary forest plots Click here for additional data file. Appendix 5: Prevalence of resistance by country level reported first line antibiotic treatment for urinary tract infection in OECD and non-OECD countries Click here for additional data file. CC is affiliated with the National Institute for Health Research Health Protection Research Unit (NIHR HPRU) in Healthcare Associated Infection and Antimicrobial Resistance at Imperial College London in partnership with Public Health England (PHE). The views expressed are those of the author(s) and not necessarily those of the NHS, the NIHR, the Department of Health, or Public Health England.

tection Research Unit (NIHR HPRU) in Healthcare Associated Infection and Antimicrobial Resistance at Imperial College London in partnership with Public Health England (PHE). The views expressed are those of the author(s) and not necessarily those of the NHS, the NIHR, the Department of Health, or Public Health England. Contributors: ADH and CC conceived and secured funding for the study. AB performed the searches. AB and HVT identified eligible studies. AB, CC, HVT, and ADH appraised study quality; data were extracted by AB, CC, and IFL. AB and CC transformed data and performed the meta-analyses. MW made substantial contributions to the overall study design and the presentation of results. AB, CC, and ADH drafted first sections of the text. All authors contributed to, reviewed, and approved the final draft. All authors received access to all the data (including statistical reports and tables) in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. AB is guarantor. Funding: AB was supported to conduct this study through a doctoral fellowship from the National Institute for Health Research School for Primary Care Research (NIHR-SPCR). ADH is supported by a NIHR research professorship (NIHR RP-R2-12-012). The views expressed in this publication are those of the authors and not necessarily those of the NIHR. The funder had no role in the study design; data collection, data analysis, and interpretation of data; in the writing of the report; or in the decision to submit the article for publication.

hip (NIHR RP-R2-12-012). The views expressed in this publication are those of the authors and not necessarily those of the NIHR. The funder had no role in the study design; data collection, data analysis, and interpretation of data; in the writing of the report; or in the decision to submit the article for publication. Competing interests: All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf and declare: no support from any organisation for the submitted work; no financial relationships with any organisations that might have an interest in the submitted work in the previous three years; no other relationships or activities that could appear to have influenced the submitted work. Ethical approval: Not required. Transparency: AB affirms that the manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted, and that any discrepancies from the study as planned have been explained. Data sharing: No additional data available.

Introduction The growing prevalence of long term conditions means that new and more efficient approaches to healthcare delivery are needed that support people to manage their own care, with less reliance on consultations with expensively trained healthcare professionals. Effective self management, as part of a shift in the management of long term conditions, can help improve health outcomes and reduce costs.1 2 Many countries are exploring a greater use of technologies, such as the internet, remote monitoring, and telephone support as a way of expanding provision and increasing access to care for a large number of people at relatively low cost. In the United Kingdom, current policy envisages these “telehealth” approaches as having potential to transform the delivery of healthcare to make the national health service sustainable for the future.3 In the United States, the Veterans Health Administration has enrolled more than 50 000 people in a home telehealth programme,4 5 and in Europe the Renewing Health Consortium is evaluating telehealth programmes in nine countries.6

he delivery of healthcare to make the national health service sustainable for the future.3 In the United States, the Veterans Health Administration has enrolled more than 50 000 people in a home telehealth programme,4 5 and in Europe the Renewing Health Consortium is evaluating telehealth programmes in nine countries.6 The volume of literature on the effectiveness of specific telehealth interventions is burgeoning, with promising effects for some applications. However, recent reviews have highlighted that much of the evidence is of poor quality; results are inconsistent; there is a lack of theoretical underpinning, which makes it difficult to interpret the mixed results; and there is some evidence of publication bias in favour of positive results.7 8 9 10 Furthermore, focusing on specific applications or technologies in isolation is of limited value since they need to be considered in the context of their implementation within the healthcare system. In practice, large scale healthcare programmes based on telehealth involve the combined use of technologies—for example, online programmes or remote monitoring with telephone support from advisors following computerised algorithms. In the recent five year strategic plan for the NHS, it is argued that evaluation is needed of “combinatorial innovation,” in which a range of technologies are provided in combination with new ways of working.11 12 Few rigorous pragmatic studies have been done on implementation of this approach in the real world.8 Furthermore, a key aspect of the argument for telehealth is increased efficiency, but there are few studies incorporating economic analyses, and the limited evidence available suggests that many telehealth interventions are not cost effective.13

dies have been done on implementation of this approach in the real world.8 Furthermore, a key aspect of the argument for telehealth is increased efficiency, but there are few studies incorporating economic analyses, and the limited evidence available suggests that many telehealth interventions are not cost effective.13 We conducted a research programme to develop a conceptual model for the effective use of telehealth in long term conditions, based on literature reviews,14 15 qualitative research,16 and surveys of patients’ views.17 Designated the telehealth in chronic disease (TECH) model, this builds on existing approaches such as the chronic care model by creating a framework for improving the management of chronic diseases through telehealth.18 We used this model to design the Healthlines service for the management of long term conditions, based on the combined use of internet based health applications that had evidence of effectiveness supported by non-clinically qualified staff working using tailored computerised algorithms.19

diseases through telehealth.18 We used this model to design the Healthlines service for the management of long term conditions, based on the combined use of internet based health applications that had evidence of effectiveness supported by non-clinically qualified staff working using tailored computerised algorithms.19 We evaluated the Healthlines service through linked pragmatic multicentre randomised controlled trials with nested process and economic evaluations in two exemplar conditions: depression and increased cardiovascular risk. This paper reports the findings for patients with an increased cardiovascular risk. Although hypertension, obesity, and hyperlipidaemia are often considered as long term conditions, it is more appropriate to consider them as risk factors, with their combined effect determining overall cardiovascular risk.20 This was considered an appropriate exemplar because of the high number of people affected (10% of adults aged 35-74 in England and Wales have a 10 year cardiovascular risk ≥20%),21 which has serious health consequences as a result of heart attacks, strokes, kidney disease, and other problems. Cardiovascular disease causes 28% of deaths in England, accounts for 10% of all hospital admissions, and involves an annual expenditure in England of almost £7bn.22 A low cost intervention that could be made widely available to large numbers of people could have a beneficial impact at a population level even if the effect for an individual was small.

% of deaths in England, accounts for 10% of all hospital admissions, and involves an annual expenditure in England of almost £7bn.22 A low cost intervention that could be made widely available to large numbers of people could have a beneficial impact at a population level even if the effect for an individual was small. Evidence exists for the effectiveness of specific relevant technological approaches, such as home blood pressure monitoring,23 mobile phone applications to support smoking cessation,24 and online interventions for weight loss.25 This evidence provided a good basis for the hypothesis that combining these “active ingredients” and implementing them within a new telehealth model of care would be effective and cost effective. Furthermore, the introduction in 2008 of the NHS Health Check programme was likely to identify a large number of people at high cardiovascular risk, and there was a need to explore ways to expand provision of care to manage them once they had been identified.26 Our hypothesis was that the Healthlines service for patients with high cardiovascular risk would be more clinically effective and cost effective than usual care, while also improving participant’s quality of life, risk behaviours, and experience of care.

Evidence exists for the effectiveness of specific relevant technological approaches, such as home blood pressure monitoring,23 mobile phone applications to support smoking cessation,24 and online interventions for weight loss.25 This evidence provided a good basis for the hypothesis that combining these “active ingredients” and implementing them within a new telehealth model of care would be effective and cost effective. Furthermore, the introduction in 2008 of the NHS Health Check programme was likely to identify a large number of people at high cardiovascular risk, and there was a need to explore ways to expand provision of care to manage them once they had been identified.26 Our hypothesis was that the Healthlines service for patients with high cardiovascular risk would be more clinically effective and cost effective than usual care, while also improving participant’s quality of life, risk behaviours, and experience of care. Methods Design This was a pragmatic, multicentre, randomised controlled trial comparing the Healthlines service in addition to usual care versus usual care alone in adults with a high risk of cardiovascular disease. The study was registered before recruitment of the first participant, and the study protocol has been published.19 After the trial commenced there were no important changes to the methods, apart from the addition of a nested substudy of different forms of information in the patient invitation to assess the impact on participant recruitment rates. This did not alter the design or outcomes for the main trial; results of this substudy are published elsewhere.27

here were no important changes to the methods, apart from the addition of a nested substudy of different forms of information in the patient invitation to assess the impact on participant recruitment rates. This did not alter the design or outcomes for the main trial; results of this substudy are published elsewhere.27 Participants Patients eligible for the trial were aged between 40 and 74 years, had a risk of a cardiovascular event in the next 10 years of 20% or more calculated using the QRISK2 score,21 and had one or more of the following modifiable risk factors (systolic blood pressure ≥140 mm Hg, body mass index ≥30, being a current smoker, or any combination of these). Participants required access to a telephone, the internet, and an email address. We excluded people who had a previous cardiovascular event; were pregnant or planning pregnancy; had a serious mental health problem, dementia, severe learning disability, or substance dependency; were receiving palliative care; or were unable to communicate verbally in English.

one, the internet, and an email address. We excluded people who had a previous cardiovascular event; were pregnant or planning pregnancy; had a serious mental health problem, dementia, severe learning disability, or substance dependency; were receiving palliative care; or were unable to communicate verbally in English. Participants were recruited from 42 general practices covering populations with a range of sociodemographic characteristics in and around Bristol, Sheffield, and Southampton, England. We used patients’ medical records to identify those who had at least one modifiable risk factor and estimated 10 year cardiovascular risk of 18% or more (we were over-inclusive at the initial screening stage because QRISK2 scores based on historical records may not reflect current risk and we wanted to invite potentially eligible people to have an updated risk assessment). A random sample of these potentially eligible patients in each practice was sent information about the study by post, after general practitioners screened the list for patients with known exclusion criteria. We sent information to between 250 and 285 patients in each practice, altering the sampling fraction over time to achieve our recruitment targets. A researcher telephoned patients who expressed an interest in the study to conduct initial eligibility screening and then invited them for an assessment of cardiovascular risk status by a practice nurse at their participating general practice. The nurse measured the patients’ blood pressure, weight and height, smoking status, and total cholesterol to high density lipoprotein cholesterol ratio, and collected all other relevant information needed to calculate the patient’s QRISK2 score (see supplementary appendix 1). Patients who had a QRISK2 score of 20% or more and had one or more of the specified modifiable risk factors completed a baseline questionnaire and consent form, either online or by post.

ol ratio, and collected all other relevant information needed to calculate the patient’s QRISK2 score (see supplementary appendix 1). Patients who had a QRISK2 score of 20% or more and had one or more of the specified modifiable risk factors completed a baseline questionnaire and consent form, either online or by post. Intervention and control Participants in the control group could continue to receive all care normally provided by the NHS, but had no contact with the Healthlines service. Usual care involved management of cardiovascular risk factors by primary care clinicians, including, in some cases, referral to community services for advice about smoking cessation and weight management.

p could continue to receive all care normally provided by the NHS, but had no contact with the Healthlines service. Usual care involved management of cardiovascular risk factors by primary care clinicians, including, in some cases, referral to community services for advice about smoking cessation and weight management. Participants in the intervention arm received support from the Healthlines service in addition to usual NHS care. The Healthlines service is a multifaceted intervention, incorporating a range of strategies to address the various components of the TECH model (see box 1).18 The intervention is based around regular telephone calls from a health advisor, supported by patient specific tailored algorithms and standardised scripts generated through a computerised behavioural management programme. This programme was originally developed and successfully evaluated in the United States by Bosworth et al and includes a series of modules on subjects such as drug adherence, diet, and smoking cessation.28 29 The standardised scripts generated by the software were based on recognised principles for behaviour change, such as stimulus control, problem solving, cognitive restructuring, and goal setting.30 We modified the programme to reflect English management guidelines and referral options, wrote additional modules with new content, and adapted the language to suit an English population.

recognised principles for behaviour change, such as stimulus control, problem solving, cognitive restructuring, and goal setting.30 We modified the programme to reflect English management guidelines and referral options, wrote additional modules with new content, and adapted the language to suit an English population. Box 1: Components of the Healthlines cardiovascular disease risk intervention, reflecting the TECH conceptual model, with examples of strategies Computerised behaviour management programme, providing interactive scripts used by health advisors Modules include: o Knowledge about cardiovascular risk and healthy lifestyles o Review of drugs and side effects o Optimisation of drugs for blood pressure lowering o Home blood pressure monitoring o Review of statins o Support for drug adherence o Smoking and nicotine replacement therapy o Healthy eating o Weight loss and Orlistat o Alcohol use o Exercise Motivational interviewing. All health advisors were trained in motivational interviewing Self monitoring and feedback—for example, blood pressure online self monitoring programme with automated feedback Treatment optimisation and intensification—health advisors monitor treatment response, and send emails to clinicians to intensify treatment when necessary, along with reminders of treatment guidelines Addressing drug adherence—monthly review, scripts incorporated evidence based strategies to promote adherence Improving care coordination—sharing all information sent to clinicians with patients Supporting primary care—all aspects of the intervention designed to support rather than duplicate primary care

Treatment optimisation and intensification—health advisors monitor treatment response, and send emails to clinicians to intensify treatment when necessary, along with reminders of treatment guidelines Addressing drug adherence—monthly review, scripts incorporated evidence based strategies to promote adherence Improving care coordination—sharing all information sent to clinicians with patients Supporting primary care—all aspects of the intervention designed to support rather than duplicate primary care Strategies to promote engagement of patients—through continuity of care with the same advisor; providing technical support with getting online Strategies to promote engagement of primary care clinicians—emphasising the evidence based nature of intervention components and how it can support their work in primary care

Supporting primary care—all aspects of the intervention designed to support rather than duplicate primary care Strategies to promote engagement of patients—through continuity of care with the same advisor; providing technical support with getting online Strategies to promote engagement of primary care clinicians—emphasising the evidence based nature of intervention components and how it can support their work in primary care For the initial assessment, health advisors contacted each participant by telephone to discuss their health needs and to agree on specific goals. After the initial call, the advisors telephoned each participant approximately every month for one year. The software was interactive and provided different computerised scripts so that the content of each call was tailored to meet each participant’s particular needs and goals. The software provided health advisors with links to relevant and quality assured online resources and applications to support self management (for example, to help with losing weight or stopping smoking), and the advisors sent these links to participants by email or post. To avoid an anonymous “call centre” approach, the same advisor telephoned each participant on each occasion when possible, since our earlier qualitative research had identified a relationship with the advisor as an important factor in engaging prospective participants.15 The Healthlines service was designed to improve access to care and was available until 8 pm on weekdays and 2 pm on Saturdays.

participant on each occasion when possible, since our earlier qualitative research had identified a relationship with the advisor as an important factor in engaging prospective participants.15 The Healthlines service was designed to improve access to care and was available until 8 pm on weekdays and 2 pm on Saturdays. Participants were also provided with access to a Healthlines web portal where they could obtain further information about cardiovascular disease, access other online resources, request a call-back from Healthlines staff, see copies of letters to their general practitioner, and use a blood pressure self monitoring system. Participants with a baseline systolic blood pressure of 140 mm Hg or more were offered a validated home blood pressure monitor (Omron, M3) by their practice nurse, requested to take their blood pressure twice daily for the first week and weekly thereafter, and to upload their readings to the Healthlines portal. The portal calculated average readings over the previous six days initially and thereafter over the previous six weeks. Using these readings, participants were automatically advised by the portal whether their blood pressure was within their target, when to take their blood pressure again, and what to do if their blood pressure was too high or too low. Target blood pressure was based on UK guidelines,31 although an individual’s target could be modified by his or her general practitioner. At each telephone contact, health advisors reviewed average blood pressure readings, and participants with above target readings were asked to see their doctor to review their treatment. Advisors sent an email to the general practice, attaching details of the patient’s recent blood pressure readings and a summary of guidelines from the National Institute for Health and Care Excellence about recommended steps for intensifying treatment.

readings were asked to see their doctor to review their treatment. Advisors sent an email to the general practice, attaching details of the patient’s recent blood pressure readings and a summary of guidelines from the National Institute for Health and Care Excellence about recommended steps for intensifying treatment. The Healthlines advisors were not clinically qualified but had experience of working as health advisors for NHS Direct and had a further three weeks of training in health coaching, motivational interviewing, treatment options (including drugs) for hypertension, smoking and obesity, and use of the Healthlines computerised management programme. The Healthlines service was originally hosted by NHS Direct, which provided a range of telehealth services through a network of call centres and a nationally recognised website. When NHS Direct closed in March 2014, delivery of the intervention was paused for two months while the staff and computer systems were transferred to a new provider (Solent NHS Trust). Although the Healthlines service resumed unaltered after this hiatus, about two thirds of participants experienced some disruption, and some participants could not receive the full number of telephone calls during their 12 month follow-up period.

d computer systems were transferred to a new provider (Solent NHS Trust). Although the Healthlines service resumed unaltered after this hiatus, about two thirds of participants experienced some disruption, and some participants could not receive the full number of telephone calls during their 12 month follow-up period. Outcomes The primary outcome was the proportion of participants responding positively to treatment, defined as maintaining or reducing their 10 year cardiovascular risk 12 months after randomisation, estimated using the QRISK2 score. Since cardiovascular risk increases with age, maintaining or reducing risk over 12 months requires an improvement in at least one modifiable risk factor. We treated the QRISK2 score (continuous) as a secondary outcome. The estimate of risk was based on data collected at an assessment by a nurse or healthcare assistant at the participant’s general practice at six and 12 months after recruitment using the same procedures as used at baseline (see supplementary appendix 1). We calculated follow-up QRISK2 scores by updating age and values for modifiable risk factors only. Other variables such as diagnoses of atrial fibrillation or diabetes were not altered to avoid bias from the greater attention paid to participants in the intervention arm.

used at baseline (see supplementary appendix 1). We calculated follow-up QRISK2 scores by updating age and values for modifiable risk factors only. Other variables such as diagnoses of atrial fibrillation or diabetes were not altered to avoid bias from the greater attention paid to participants in the intervention arm. Cardiovascular risk is a composite outcome, and the individual risk factors of blood pressure, weight (and body mass index), smoking, and cholesterol level were important secondary outcomes. Other secondary outcomes were quality of life, exercise, diet, satisfaction with treatment received and with amount of support received, perceived access to care, self management skills and self efficacy, drug adherence, health literacy, use of telehealth, and perceptions of care coordination. Table 6 lists the specific measurement instruments used. Secondary outcome measures were collected through patient questionnaires, completed online or by post at baseline and six and 12 months after randomisation. We obtained data about prescriptions and primary care consultations from general practice records and details on use of the intervention from Healthlines records. Potential serious adverse events were identified through reports from participants or health professionals, further inquiry about hospital admissions reported in outcome questionnaires, or admissions, deaths, or other potential serious adverse events identified through review of primary care notes at the end of the trial. We logged all such events with a description of the event and an assessment of expectedness, relatedness, and seriousness and we reported to the trial monitoring committee, sponsor, and ethics committee as appropriate.

ntial serious adverse events identified through review of primary care notes at the end of the trial. We logged all such events with a description of the event and an assessment of expectedness, relatedness, and seriousness and we reported to the trial monitoring committee, sponsor, and ethics committee as appropriate. Sample size The sample size was chosen pragmatically, taking into account the size of effect that would be likely to influence practice and might be feasible to detect in a trial of reasonable size. Based on a previous study we assumed that 35% of participants in the control arm would maintain or reduce their cardiovascular risk over 12 months.32 Including 240 participants in each trial arm for analysis would provide 80% power (5% α) and 90% power (1% α) to detect differences of 13 and 18 percentage points, respectively. Assuming 20% attrition, we therefore aimed to recruit 600 participants, 300 in each trial arm.

uce their cardiovascular risk over 12 months.32 Including 240 participants in each trial arm for analysis would provide 80% power (5% α) and 90% power (1% α) to detect differences of 13 and 18 percentage points, respectively. Assuming 20% attrition, we therefore aimed to recruit 600 participants, 300 in each trial arm. Randomisation and masking Participants who provided consent were randomly allocated in 1:1 ratio to the intervention or usual care group. Allocation was made using a web randomisation system hosted by the Bristol Randomised Controlled Trials Collaboration, and automated to ensure concealment. Randomisation was stratified by location of recruitment (Bristol, Sheffield, or Southampton) and minimised by general practice and baseline QRISK2 score. Researchers notified the participants of their allocation by email. Participants were not masked to treatment allocation. Practice nurses or healthcare assistants collected data for the QRISK2 score and may have been aware of treatment allocation at follow-up, but the variables of relevance on smoking (validated using a carbon monoxide monitor), blood pressure, and cholesterol level were all based on objective quantitative data. All other outcome data were collected by participant self report or electronic download from medical records and were entered and analysed blinded to treatment allocation.

levance on smoking (validated using a carbon monoxide monitor), blood pressure, and cholesterol level were all based on objective quantitative data. All other outcome data were collected by participant self report or electronic download from medical records and were entered and analysed blinded to treatment allocation. Statistical analysis Analysis was conducted according to CONSORT guidelines, following an analysis plan agreed in advance with the independent trial steering committee and data monitoring committee. We used descriptive statistics to compare baseline characteristics of trial participants by allocated arm. The primary analysis of response to treatment after 12 months was conducted using a mixed effects logistic regression model adjusted for site, baseline QRISK2 score, and general practice (as a random effect). Participants were analysed according to allocated arm. We conducted sensitivity analyses of the primary outcome using: the assumption that all participants were exactly one year older at 12 months’ follow-up, simple imputation of missing outcome data that assumed no treatment response, multiple imputation of missing data, exclusion of general practitioner’s practice as a random effect, and adjustment by time between randomisation and follow-up. By fitting interaction terms between trial arm and subgroup variables, we investigated whether any effect of the Healthlines intervention on the primary outcome differed according to subgroups defined by sex, age, baseline QRISK2 score, and presence or absence of each of the modifiable risk factors (hypertension, obesity, smoking) at baseline.

rms between trial arm and subgroup variables, we investigated whether any effect of the Healthlines intervention on the primary outcome differed according to subgroups defined by sex, age, baseline QRISK2 score, and presence or absence of each of the modifiable risk factors (hypertension, obesity, smoking) at baseline. In secondary analysis of the primary outcome, we estimated the complier average causal effect of the Healthlines intervention when received as intended. We described compliance as little or none (two or fewer telephone calls), partial (three to 11 calls), or full (12 or 13 calls). We estimated the complier average causal effect at 12 months using principal stratification in two ways: classifying partial compliers as either non-compliers or full compliers.33 Secondary outcomes were analysed in a similar manner to the primary outcome. We estimated between group effects using linear, logistic, or binomial mixed effects regression models, adjusted for stratification and minimisation variables and value of the outcome at baseline. Participants were analysed as randomised without imputation of missing data. To reduce the number of statistical comparisons, we estimated between group differences for secondary outcomes (other than cardiovascular risk factors) only at the final 12 month follow-up time point. We described serious adverse events by study arm.

pants were analysed as randomised without imputation of missing data. To reduce the number of statistical comparisons, we estimated between group differences for secondary outcomes (other than cardiovascular risk factors) only at the final 12 month follow-up time point. We described serious adverse events by study arm. We assessed the cost effectiveness of the Healthlines intervention from an NHS perspective at 12 months from randomisation. Cost effectiveness was not listed as a secondary outcome in the trial registry because we viewed it as an approach to analysis rather than as an outcome; however, assessment of cost effectiveness was specified a priori as an aim in the registry and described in the published protocol. The methods and results of the economic evaluation will be described in detail elsewhere. In brief, we compared health system costs with incremental quality adjusted life years, measured using the EQ-5D-5L generic quality of life questionnaire34 at baseline and six and 12 months post-randomisation, to produce an estimate of net monetary benefit. We also developed a cohort simulation model in order to estimate the cost effectiveness of the intervention over the estimated remaining lifetime of trial participants. All analyses were conducted using Stata version 13 MP2. The trial was registered prospectively with Current Controlled Trials (ISRCTN 27508731).

We assessed the cost effectiveness of the Healthlines intervention from an NHS perspective at 12 months from randomisation. Cost effectiveness was not listed as a secondary outcome in the trial registry because we viewed it as an approach to analysis rather than as an outcome; however, assessment of cost effectiveness was specified a priori as an aim in the registry and described in the published protocol. The methods and results of the economic evaluation will be described in detail elsewhere. In brief, we compared health system costs with incremental quality adjusted life years, measured using the EQ-5D-5L generic quality of life questionnaire34 at baseline and six and 12 months post-randomisation, to produce an estimate of net monetary benefit. We also developed a cohort simulation model in order to estimate the cost effectiveness of the intervention over the estimated remaining lifetime of trial participants. All analyses were conducted using Stata version 13 MP2. The trial was registered prospectively with Current Controlled Trials (ISRCTN 27508731). Patient involvement There was strong and valuable patient and public involvement throughout the Healthlines research programme. Two service user groups (Mental Health Research Network and NHS Direct user group) provided feedback on the initial questionnaire about patients’ preferences and needs in relation to telehealth, which helped to inform the intervention design.17 Two representatives of these groups became members of the management group for the five year research programme. They contributed to the design of the patient survey,17 participated in a workshop to develop the TECH model that underlies the intervention,18 and became members of the trial steering committee for the randomised trial.19 They commented on the acceptability of the intervention to potential participants and obtained feedback from their user groups on the outcome measures. At the end of the trial they contributed to a workshop of key stakeholders, which was held to discuss interpretation and dissemination of the findings. They also provided useful feedback on the final report of the programme, and in particular the lay summary. We have thanked all participants for their involvement and given them details of the website where all published results will be made publically available (www.bristol.ac.uk/healthlines/).

nation of the findings. They also provided useful feedback on the final report of the programme, and in particular the lay summary. We have thanked all participants for their involvement and given them details of the website where all published results will be made publically available (www.bristol.ac.uk/healthlines/). Results Participants were recruited between 3 December 2012 and 23 July 2013. Of 7582 people sent information about the study, 1205 (16%) expressed interest and were assessed. Of these, 641 were eligible and randomly allocated to the Healthlines intervention (n=325) or usual care (n=316) arms (fig 1). In total, 597 (93%) of the participants provided follow-up data after six months’ follow-up and 586 (91%) after 12 months’ follow-up (the primary outcome). Fig 1 Flow of participants through trial comparing Healthlines cardiovascular disease intervention with usual care

Results Participants were recruited between 3 December 2012 and 23 July 2013. Of 7582 people sent information about the study, 1205 (16%) expressed interest and were assessed. Of these, 641 were eligible and randomly allocated to the Healthlines intervention (n=325) or usual care (n=316) arms (fig 1). In total, 597 (93%) of the participants provided follow-up data after six months’ follow-up and 586 (91%) after 12 months’ follow-up (the primary outcome). Fig 1 Flow of participants through trial comparing Healthlines cardiovascular disease intervention with usual care Table 1 shows the characteristics of participants in the trial. Overall, the participants were at high risk of a cardiovascular event (mean 10 year risk 31%) owing to combinations of modifiable and non-modifiable risk factors. The participants were predominantly white men aged more than 60, and at baseline 356 (56%) were obese (body mass index ≥30), 450 (70%) had a blood pressure of 140 mm Hg or more, and 528 (18%) were current smokers. The two trial arms were well balanced except that there were fewer smokers and more participants with diabetes in the intervention arm. These factors both contribute to the baseline QRISK2 score, which was included as a covariate in all analyses, so we did not conduct additional statistical adjustment for these imbalances. Table 1 Baseline characteristics of participants allocated to usual care or Healthlines intervention. Values are percentages (numbers) unless otherwise stated

Table 1 shows the characteristics of participants in the trial. Overall, the participants were at high risk of a cardiovascular event (mean 10 year risk 31%) owing to combinations of modifiable and non-modifiable risk factors. The participants were predominantly white men aged more than 60, and at baseline 356 (56%) were obese (body mass index ≥30), 450 (70%) had a blood pressure of 140 mm Hg or more, and 528 (18%) were current smokers. The two trial arms were well balanced except that there were fewer smokers and more participants with diabetes in the intervention arm. These factors both contribute to the baseline QRISK2 score, which was included as a covariate in all analyses, so we did not conduct additional statistical adjustment for these imbalances. Table 1 Baseline characteristics of participants allocated to usual care or Healthlines intervention. Values are percentages (numbers) unless otherwise stated Characteristics Usual care (n=316) Intervention (n=325) Mean (SD) age at CVD assessment (years) 67.3 (4.7) 67.5 (4.9) Women 21 (66) 18 (60) White ethnicity 99 (313) 99 (321) Current employment situation: n=311 n=316 Full time 13 (39) 17 (54) Part time 14 (43) 9 (29) Unemployed 1 (4) 1 (2) Unable to work: long term illness/disability 2 (7) 1 (3) Unable to work: carer 1 (3) 1 (2) Retired 63 (196) 66 (210) Homemaker 1 (3) 1 (4) Other 5 (16) 4 (12) Occupation (most recent or current): n=294 n=294 Administrative or secretarial 11 (31) 10 (29) Associate professional or technical 15 (45) 12 (35) Elementary* 10 (28) 5 (16) Managers or senior officials 19 (55) 22 (65) Personal services† 2 (5) 3 (9) Process, plant, and machine operatives 5 (15) 6 (17) Professionals 19 (57) 22 (64) Sales and customer services 4 (11) 4 (13) Skilled trades 16 (47) 16 (46) Highest education qualification: n=307 n=318 Degree or higher degree 21 (65) 23 (72) A levels or equivalent 19 (58) 17 (53) GCSEs/O levels or equivalent 45 (137) 43 (136) No qualifications 15 (47) 18 (57) Accommodation: n=315 n=323 Own accommodation or buying with mortgage 84 (264) 87 (281) Part rented or rented 15 (46) 12 (40) Live rent-free 2 (5) 1 (2) Mean (SD) index of multiple deprivation 16.7 (12.6) 15.5 (11.3) Clinical data: Mean (SD) QRISK2 score 30.8 (9.5) 31.1 (10.2) Mean (SD) systolic blood pressure (mm Hg) 148.1 (17.6) 147.6 (16.2) Mean (SD) diastolic blood pressure (mm Hg) 80.0 (10.4) 81.2 (9.6) Mean weight (SD) 91.9 (18.9) 93.2 (17.3) Mean (SD) body mass index 30.9 (5.7) 31.2 (5.4) Mean (SD) total cholesterol level (mmol/L) 4.9 (1.2); n=315 4.9 (1.2);n=324 Mean (SD) total cholesterol:HDL ratio 4.2 (1.4); n=315 4.2 (1.5); n=323 Smoking status: Non-smoker 33 (103) 35 (114) Former smoker 47 (148) 50 (163) Light smoker 9 (30) 8 (25) Moderate smoker 5 (17) 5 (16) Heavy smoker 6 (18) 2 (7) Taking antihypertensive 61 (193) 64 (209) Taking lipid lowering drug 49 (153/312) 49 (158/322) Diabetes 20 (62) 24 (77) Chronic kidney disease 11 (34) 6 (20) Atrial fibrillation 6 (20) 7 (23) Rheumatoid arthritis 3 (8) 2 (6) CVD=cardiovascular disease; GCSE=general certificate of secondary education; HDL=high density lipoprotein c

ypertensive 61 (193) 64 (209) Taking lipid lowering drug 49 (153/312) 49 (158/322) Diabetes 20 (62) 24 (77) Chronic kidney disease 11 (34) 6 (20) Atrial fibrillation 6 (20) 7 (23) Rheumatoid arthritis 3 (8) 2 (6) CVD=cardiovascular disease; GCSE=general certificate of secondary education; HDL=high density lipoprotein c holesterol. *Jobs not needing qualifications, such as cleaners. †For example, care worker, teaching assistant. Primary outcome After 12 months a slightly higher proportion of participants in the intervention arm had improved or maintained their cardiovascular risk compared with those in the usual care arm (50% v 42%, respectively; number to treat=13), although this apparent difference had wide confidence intervals and could be due to chance (adjusted odds ratio 1.3, 95% confidence interval 1.0 to 1.9; P=0.08). This conclusion was largely unchanged in our sensitivity analyses (table 2). There was no evidence that the intervention was differentially effective for any of the prespecified subgroups defined by baseline characteristics, although the study was not powered to detect these interaction effects (table 3). Table 2 Improving or maintaining cardiovascular risk as a binary outcome. Values are percentages (No/total No) unless stated otherwise

Primary outcome After 12 months a slightly higher proportion of participants in the intervention arm had improved or maintained their cardiovascular risk compared with those in the usual care arm (50% v 42%, respectively; number to treat=13), although this apparent difference had wide confidence intervals and could be due to chance (adjusted odds ratio 1.3, 95% confidence interval 1.0 to 1.9; P=0.08). This conclusion was largely unchanged in our sensitivity analyses (table 2). There was no evidence that the intervention was differentially effective for any of the prespecified subgroups defined by baseline characteristics, although the study was not powered to detect these interaction effects (table 3). Table 2 Improving or maintaining cardiovascular risk as a binary outcome. Values are percentages (No/total No) unless stated otherwise Primary outcome Usual care* Intervention† Adjusted odds ratio (95% CI) P value Primary analysis: Improved/maintained QRISK2 after 12 months 43 (124/291) 50 (148/295) 1.3 (1.0 to 1.9) 0.08 Secondary analysis: Improved/maintained QRISK2 after 6 months 46 (137/296) 48 (145/301) 1.1 (0.8 to 1.5) 0.65 Sensitivity analyses of improved/maintained QRISK2 after 12 months: Assuming all participants were one year older 45 (130/291) 52 (153/295) 1.3 (1.0 to 1.9) 0.01 Simple imputation, assuming missing binary outcome is non-response 40 (124/316) 46 (148/325) 1.3 (0.9 to 1.8) 0.11 Multiple imputation 44 (139/316) 50 (163/325) 1.3 (0.9 to 1.8) 0.11 Not including general practice as random effect 43 (124/291) 50 (148/295) 1.3 (1.0 to 1.9) 0.08 Adjusted by days since randomisation to primary outcome assessment 43 (124/291) 50 (148/295) 1.3 (1.0 to 1.9) 0.09 All analyses adjusted by site (Bristol, Sheffield, or Southampton) and baseline QRISK2 score. Analyses are further adjusted by other covariates if specified. General practice included as random effect unless specified otherwise.

randomisation to primary outcome assessment 43 (124/291) 50 (148/295) 1.3 (1.0 to 1.9) 0.09 All analyses adjusted by site (Bristol, Sheffield, or Southampton) and baseline QRISK2 score. Analyses are further adjusted by other covariates if specified. General practice included as random effect unless specified otherwise. *n=296 at six months; n=291 at 12 months; n=316 for imputed data. †n=301 at six months; n=295 at 12 months; n=325 for imputed data. Table 3 Subgroup analyses of primary outcome

randomisation to primary outcome assessment 43 (124/291) 50 (148/295) 1.3 (1.0 to 1.9) 0.09 All analyses adjusted by site (Bristol, Sheffield, or Southampton) and baseline QRISK2 score. Analyses are further adjusted by other covariates if specified. General practice included as random effect unless specified otherwise. *n=296 at six months; n=291 at 12 months; n=316 for imputed data. †n=301 at six months; n=295 at 12 months; n=325 for imputed data. Table 3 Subgroup analyses of primary outcome Subgroups Improving or maintaining QRISK2 at 12 month follow-up Adjusted odds ratio* (95% CI) Interaction P value Usual care (n=291) Intervention (n=295) Baseline CVD assessment age group: 40-59 54 (7/13) 61 (11/18) 1.5 (0.3 to 6.6) 60-69 44 (78/177) 49 (75/152) 1.2 (0.8 to 1.9) ≥70 39 (39/101) 50 (62/125) 1.6 (0.9 to 2.8) 0.71 Men 46 (105/227) 51 (125/243) 1.2 (0.9 to 1.8) Women 30 (19/64) 44 (23/52) 1.8 (0.8 to 4.0) 0.37 Baseline QRISK2 score: 17.3-24.9 37 (37/101) 45 (44/98) 1.4 (0.8 to 2.5) 25.0-29.9 38 (26/68) 44 (35/79) 1.2 (0.6 to 2.4) ≥30.0 50 (61/122) 58 (69/118) 1.4 (0.8 to 2.4) 0.95 Baseline modifiable risk factor: Systolic blood pressure <140 mm Hg 33 (30/90) 41 (35/85) 1.5 (0.8 to 2.8) Systolic blood pressure ≥140 mm Hg 47 (94/201) 54 (113/210) 1.3 (0.9 to 1.9) 0.73 Body mass index <30.0 50 (65/131) 52 (67/129) 1.1 (0.6 to 1.8) Body mass index ≥30.0 37 (59/160) 49 (81/166) 1.7 (1.1 to 2.6) 0.20 Current smoker 51 (29/57) 53 (23/43) 1.1 (0.5 to 2.5) Not current smoker 41 (95/234) 50 (125/252) 1.4 (1.0 to 2.1) 0.55 All analyses adjusted by site (Bristol, Sheffield, or Southampton), baseline outcome, and baseline QRISK2 score. General practice included as random effect.

(59/160) 49 (81/166) 1.7 (1.1 to 2.6) 0.20 Current smoker 51 (29/57) 53 (23/43) 1.1 (0.5 to 2.5) Not current smoker 41 (95/234) 50 (125/252) 1.4 (1.0 to 2.1) 0.55 All analyses adjusted by site (Bristol, Sheffield, or Southampton), baseline outcome, and baseline QRISK2 score. General practice included as random effect. *Odds ratio comparing intervention with usual care. Secondary outcomes Evidence was lacking of any between group difference in the proportion of participants who improved or maintained their cardiovascular risk after six months’ follow-up (table 2). There was also no evidence of a difference between groups in QRISK2 score when treated as a continuous measure (table 4). However, the intervention was associated with improvements in some of the individual modifiable risk factors that contribute to cardiovascular risk, including reductions in systolic and diastolic blood pressure and weight and body mass index after 12 months’ follow up (table 4). The intervention did not lead to improvements in cholesterol levels (table 4) or smoking rates (table 5). Table 4 Secondary outcomes: QRISK2 score as a continuous outcome and individual modifiable cardiovascular risk factors of blood pressure, cholesterol level, weight, and body mass index

Secondary outcomes Evidence was lacking of any between group difference in the proportion of participants who improved or maintained their cardiovascular risk after six months’ follow-up (table 2). There was also no evidence of a difference between groups in QRISK2 score when treated as a continuous measure (table 4). However, the intervention was associated with improvements in some of the individual modifiable risk factors that contribute to cardiovascular risk, including reductions in systolic and diastolic blood pressure and weight and body mass index after 12 months’ follow up (table 4). The intervention did not lead to improvements in cholesterol levels (table 4) or smoking rates (table 5). Table 4 Secondary outcomes: QRISK2 score as a continuous outcome and individual modifiable cardiovascular risk factors of blood pressure, cholesterol level, weight, and body mass index Secondary outcome Usual care* Intervention† Adjusted difference in means (95% CI) P value Unadjusted mean (SD) No Unadjusted mean (SD) No QRISK2 score as continuous variable: 6 months 31.0 (9.5) 296 31.4 (10.3) 301 0.1 (−0.2 to 0.4) 0.49 12 months 31.2 (10.3) 291 31.3 (10.7) 295 −0.4 (−1.2 to 0.3) 0.27 Systolic blood pressure (mm Hg): 6 months 141.4 (15.4) 296 141.0 (15.1) 301 0.0 (−1.9 to 1.9) 0.10 12 months 142.2 (16.1) 291 139.6 (14.0) 295 −2.7 (−4.7 to −0.6) 0.01 Diastolic blood pressure (mm Hg): 6 months 78.0 (9.7) 296 78.2 (9.9) 301 −0.6 (−1.8 to 0.6) 0.34 12 months 78.7 (9.9) 291 76.6 (9.2) 295 −2.8 (−4.0 to −1.6) <0.001 Total cholesterol level (mmol/L)‡: 12 months 4.7 (1.1) 288 4.6 (1.2) 295 −0.1 (−0.2 to 0.0) 0.17 Total cholesterol:HDL ratio: 12 months 4.0 (1.5) 287 4.0 (1.7) 294 −0.1 (−0.2 to 0.1) 0.45 Weight (kg): 6 months 91.1 (18.4) 296 91.7 (17.7) 301 −0.9 (−1.5 to −0.2) 0.006 12 months 91.2 (19.1) 291 91.3 (17.5) 293 −1.0 (−1.8 to −0.3) 0.008 Body mass index (kg/m2): 6 months 30.6 (5.4) 296 30.7 (5.5) 301 −0.3 (−0.5 to −0.1) 0.006 12 months 30.8 (5.7) 291 30.5 (5.4) 293 −0.4 (−0.6 to −0.1) 0.008 HDL=high density lipoprotein cholesterol.

1 (18.4) 296 91.7 (17.7) 301 −0.9 (−1.5 to −0.2) 0.006 12 months 91.2 (19.1) 291 91.3 (17.5) 293 −1.0 (−1.8 to −0.3) 0.008 Body mass index (kg/m2): 6 months 30.6 (5.4) 296 30.7 (5.5) 301 −0.3 (−0.5 to −0.1) 0.006 12 months 30.8 (5.7) 291 30.5 (5.4) 293 −0.4 (−0.6 to −0.1) 0.008 HDL=high density lipoprotein cholesterol. All analyses adjusted by site (Bristol, Sheffield, or Southampton), baseline QRISK2 score, and baseline outcome. General practice included as random effect. *n=296 at six months; n=291 at 12 months. †n=301 at six months; n=295 at 12 months. ‡Cholesterol was not remeasured after six months. Baseline cholesterol measurement was used to calculate QRISK2 at six months. Table 5 Secondary outcome: smoking. Values are percentages (numbers) unless stated otherwise Smoker status Usual care* Intervention† Adjusted odds ratio (95% CI) P value Smoker at 6 months: Yes 18 (52/296) 15 (45/301) No 82 (244/296) 85 (256/301) 0.3 (0.1 to 1.2) 0.01 Smoker at 12 months: Yes 18 (52/291) 17 (49/295) No 82 (239/291) 83 (246/295) 0.4 (0.2 to 1.0) 0.06 All analyses adjusted by site (Bristol, Sheffield, or Southampton), baseline QRISK2 score, and baseline smoking category (non-smoker, former smoker, light smoker, moderate smoker, heavy smoker). General practice included as random effect. *n=296 at six months; n=291 at 12 months. †n=301 at six months; n=295 at 12 months.

Smoker status Usual care* Intervention† Adjusted odds ratio (95% CI) P value Smoker at 6 months: Yes 18 (52/296) 15 (45/301) No 82 (244/296) 85 (256/301) 0.3 (0.1 to 1.2) 0.01 Smoker at 12 months: Yes 18 (52/291) 17 (49/295) No 82 (239/291) 83 (246/295) 0.4 (0.2 to 1.0) 0.06 All analyses adjusted by site (Bristol, Sheffield, or Southampton), baseline QRISK2 score, and baseline smoking category (non-smoker, former smoker, light smoker, moderate smoker, heavy smoker). General practice included as random effect. *n=296 at six months; n=291 at 12 months. †n=301 at six months; n=295 at 12 months. Table 6 shows that the intervention was associated with improvements in several of the secondary outcomes. Participants in the intervention arm reported that they improved their diets and increased their level of exercise. They were more likely to adhere to their treatment with statins and antihypertensive drugs. Participants in the intervention arm reported improved access to care and expressed greater satisfaction with the amount of support they received and their overall treatment. They also reported that their care was better organised and coordinated. For ease of presentation, table 6 only shows data on secondary outcomes after 12 months’ follow-up. Findings after six months are available from the authors. Table 6 Secondary outcomes at 12 month follow-up

Table 6 shows that the intervention was associated with improvements in several of the secondary outcomes. Participants in the intervention arm reported that they improved their diets and increased their level of exercise. They were more likely to adhere to their treatment with statins and antihypertensive drugs. Participants in the intervention arm reported improved access to care and expressed greater satisfaction with the amount of support they received and their overall treatment. They also reported that their care was better organised and coordinated. For ease of presentation, table 6 only shows data on secondary outcomes after 12 months’ follow-up. Findings after six months are available from the authors. Table 6 Secondary outcomes at 12 month follow-up Secondary outcome Usual care (n=300) Intervention (n=300) Adjusted difference in means (95% CI) P value Unadjusted mean (SD) No Unadjusted mean (SD) No Quality of life (EQ-5D-5L)34 0.78 (0.2) 297 0.81 (0.2) 295 0.01 (−0.01 to 0.03) 0.41 Patient behaviours: Exercise behaviour (heiQ subscale: health directed behaviour’)*44 2.9 (0.8) 294 3.0 (0.8) 297 0.1 (0.0 to 0.2) 0.003 Diet (starting the conversation questionnaire)*45 10.3 (2.1) 299 10.9 (2.1) 300 0.6 (0.4 to 0.9) <0.001 Patient experience: Satisfaction with treatment*† 3.7 (0.8) 215 3.9 (0.7) 244 0.1 (0.0 to 0.3) 0.03 Satisfaction with amount of support received*† 2.8 (0.6) 207 3.1 (0.5) 260 0.3 (0.2 to 0.4) <0.001 Perceived access to care*† 5.5 (1.7) 293 5.8 (1.3) 287 0.3 (0.0 to 0.5) 0.02 Self management skills and self efficacy (heiQ):44 Self monitoring and insight* 3.2 (0.4) 295 3.3 (0.4) 295 0.1 (0.0 to 0.1) 0.07 Constructive attitudes and approaches* 3.3 (0.5) 296 3.4 (0.5) 295 0.0 (0.0 to 0.1) 0.63 Skill and technique acquisition* 3.1 (0.5) 297 3.2 (0.5) 295 0.1 (0.1 to 0.2) <0.001 Health services navigation* 3.1 (0.6) 296 3.2 (0.5) 297 0.0 (0.0 to 0.1) 0.27 Drug adherence (Morisky)*: 46

95 3.3 (0.4) 295 0.1 (0.0 to 0.1) 0.07 Constructive attitudes and approaches* 3.3 (0.5) 296 3.4 (0.5) 295 0.0 (0.0 to 0.1) 0.63 Skill and technique acquisition* 3.1 (0.5) 297 3.2 (0.5) 295 0.1 (0.1 to 0.2) <0.001 Health services navigation* 3.1 (0.6) 296 3.2 (0.5) 297 0.0 (0.0 to 0.1) 0.27 Drug adherence (Morisky)*: 46 Anti-hypertensives‡ 3.8 (0.5) 194 3.9 (0.3) 203 0.1 (0.0 to 0.2) 0.01 Statins‡ 3.6 (0.8) 165 3.8 (0.5) 169 0.2 (0.1 to 0.3) 0.005 Health literacy (eHEALs)*47 3.9 (0.7) 296 4.0 (0.7) 295 0.1 (0.0 to 0.2) 0.13 Care coordination (Haggerty): 48 Role clarity and coordination* 2.9 (0.5) 247 3.0 (0.3) 263 0.1 (0.0 to 0.1) 0.02 Evidence of care plan* 3.8 (2.1) 209 4.9 (2.0) 236 1.2 (0.8 to 1.5) <0.001 Overall experience of organisation of healthcare* 3.6 (0.9) 296 3.8 (0.7) 296 0.1 (0.0 to 0.2) 0.04 Self organisation of healthcare* 3.9 (1.1) 283 3.8 (1.0) 287 −0.1 (−0.2 to 0.1) 0.37 Use of telehealth*†‡: Online searching 1.6 (0.7) 297 1.6 (0.7) 296 0.1 (−0.0 to 0.2) 0.10 Online forum or group 1.1 (0.3) 295 1.1 (0.4) 298 0.0 (−0.0 to 0.1) 0.29 All analyses adjusted by site (Bristol, Sheffield, or Southampton), baseline outcome (if measured), and baseline QRISK2 score. General practice included as random effect. *Higher score is more positive (less access difficulties, greater satisfaction). †Based on scales generated before main trial analysis using principal components analysis and incorporating questions taken from existing validated questionnaires or constructed for this research. ‡Only applicable to those taking antihypertensives or statins.

*Higher score is more positive (less access difficulties, greater satisfaction). †Based on scales generated before main trial analysis using principal components analysis and incorporating questions taken from existing validated questionnaires or constructed for this research. ‡Only applicable to those taking antihypertensives or statins. §Five level ordered categorical variable (never/almost never to daily/almost daily). After 12 months the incremental cost effectiveness ratio was estimated to be £10 859 ($15 600; €13 800) in 2012/13 prices (incremental cost £138, 95% confidence interval £66 to £211; incremental gain in quality adjusted life years 0.012, 95% confidence interval −0.001 to 0.026). The net monetary benefit at a cost effectiveness threshold of £20 000 was estimated to be £116 (−£58 to £291). The intervention was likely to be cost effective at this threshold after 12 months, with a probability of 0.77. The cohort simulation study showed that the lifetime cost effectiveness of the intervention increased the greater the duration of effect of the intervention on cardiovascular disease risk beyond the follow-up period of the trial. Further details will be published elsewhere.

after 12 months, with a probability of 0.77. The cohort simulation study showed that the lifetime cost effectiveness of the intervention increased the greater the duration of effect of the intervention on cardiovascular disease risk beyond the follow-up period of the trial. Further details will be published elsewhere. Engagement with the intervention Participants in the intervention arm received a median of 10 (interquartile range 8-12) encounters with the Healthlines service out of a possible maximum of 13 encounters. The mean duration of each encounter was 18 (SD 9.5) minutes. Using a complier average causal effect analysis, we explored whether the number of encounters received in the intervention arm was associated with the primary outcome. The results suggest an increase in effect of the intervention among participants who received all or most of the planned number of encounters (table 7). Table 7 Complier average causal effect analysis of primary outcome. Values are unadjusted odds ratios (95% confidence intervals) unless stated otherwise

Engagement with the intervention Participants in the intervention arm received a median of 10 (interquartile range 8-12) encounters with the Healthlines service out of a possible maximum of 13 encounters. The mean duration of each encounter was 18 (SD 9.5) minutes. Using a complier average causal effect analysis, we explored whether the number of encounters received in the intervention arm was associated with the primary outcome. The results suggest an increase in effect of the intervention among participants who received all or most of the planned number of encounters (table 7). Table 7 Complier average causal effect analysis of primary outcome. Values are unadjusted odds ratios (95% confidence intervals) unless stated otherwise Amount of intervention received (No of encounters) Maintenance/reduction in QRISK2 at 12 month follow-up Partial compliers classified as non-compliers Partial compliers classified as full compliers Usual care (n=291) Intervention (n=293) Intervention v usual care Intervention v usual care None (0-2) 43 (124/291) 29 (4/14) Partial (3-11) 44 (77/177) Full (12-13) 65 (66/102) 2.4 (1.4 to 4.3) 1.4 (1.0 to 1.9) Three participants who never received Healthlines intervention and two participants who only received unscheduled non-encounter calls are categorised as receiving none of the intervention. Two intervention arm participants had missing data on encounters.

7/177) Full (12-13) 65 (66/102) 2.4 (1.4 to 4.3) 1.4 (1.0 to 1.9) Three participants who never received Healthlines intervention and two participants who only received unscheduled non-encounter calls are categorised as receiving none of the intervention. Two intervention arm participants had missing data on encounters. Participants in the intervention arm logged in to the Healthlines website on a median of 14 (interquartile range 3-47) occasions, more than once a month. Overall, 296 (91%) of the participants were given a blood pressure monitor, of whom 200 entered at least one reading, uploading a median of 70 (48-102) blood pressure readings.

ts in the intervention arm logged in to the Healthlines website on a median of 14 (interquartile range 3-47) occasions, more than once a month. Overall, 296 (91%) of the participants were given a blood pressure monitor, of whom 200 entered at least one reading, uploading a median of 70 (48-102) blood pressure readings. Healthlines advisors sent a median of 5 (2-9) letters by email to participants’ doctors. Of these, 138/310 of the participants’ doctors were sent letters advising commencement or review of blood pressure drugs, 32 (10%) were asked to consider statin treatment, 7 (2%) were asked to prescribe orlistat for obesity, and 3 (1%) were asked to prescribe drugs to aid smoking cessation. However, based on data from the medical records, the intervention and control groups did not differ in the number of changes in drugs (starting new treatments or changing dose) for hypertension or lipid lowering, with a median of 0 (0-1) changes for both types of treatment. Similarly, there was no evidence of a difference between the arms in the proportion of participants who reported taking statins or drugs for hypertension, the proportion who had a change in treatment prescribed, or the types of drug prescribed (table 8). These data were not specified as outcomes, but we have presented them to explore the mechanism of effect of the intervention. Table 8 Treatment optimisation: cardiovascular risk related drug prescriptions over trial. Values are percentages (numbers) unless stated otherwise

Healthlines advisors sent a median of 5 (2-9) letters by email to participants’ doctors. Of these, 138/310 of the participants’ doctors were sent letters advising commencement or review of blood pressure drugs, 32 (10%) were asked to consider statin treatment, 7 (2%) were asked to prescribe orlistat for obesity, and 3 (1%) were asked to prescribe drugs to aid smoking cessation. However, based on data from the medical records, the intervention and control groups did not differ in the number of changes in drugs (starting new treatments or changing dose) for hypertension or lipid lowering, with a median of 0 (0-1) changes for both types of treatment. Similarly, there was no evidence of a difference between the arms in the proportion of participants who reported taking statins or drugs for hypertension, the proportion who had a change in treatment prescribed, or the types of drug prescribed (table 8). These data were not specified as outcomes, but we have presented them to explore the mechanism of effect of the intervention. Table 8 Treatment optimisation: cardiovascular risk related drug prescriptions over trial. Values are percentages (numbers) unless stated otherwise Measure of treatment optimisation Usual care (n=316) Intervention (n=325) Adjusted odds ratio (95% CI): intervention v usual care P value Experienced at least one change in drugs over 12 month period*: Antihypertensive 32 (100) 38 (123) 1.3† (0.9 to 1.8) 0.12 Cholesterol drugs, including statins 22 (71) 26 (84) 1.2† (0.8 to 1.7) 0.33 Self reported use of drugs over 12 month period‡: Antihypertensive 68 (196/289) 70 (202/287) Statin 57 (165/297) 57 (166/290) Prescribed at least one drug over trial period*: Antiplatelet 18 (57) 19 (62) Cholesterol drugs, including statins 61 (192) 62 (201) Smoking cessation 1 (3) 2 (5) Obesity drugs 1 (2) 1 (4) Antihypertensive 70 (222) 73 (236) Prescribed antihypertensive by drug class over trial period*: ACE inhibitors or ARBs 50 (159) 52 (170) β blockers 18 (58) 16 (52) Calcium channel blockers 36 (114) 40 (129) Diuretics 29 (90) 29 (93) Other 8 (26) 8 (26) ACE=angiotensin converting enzyme; ARBs=angiotensin receptor blockers.

hypertensive 70 (222) 73 (236) Prescribed antihypertensive by drug class over trial period*: ACE inhibitors or ARBs 50 (159) 52 (170) β blockers 18 (58) 16 (52) Calcium channel blockers 36 (114) 40 (129) Diuretics 29 (90) 29 (93) Other 8 (26) 8 (26) ACE=angiotensin converting enzyme; ARBs=angiotensin receptor blockers. *Medical records data. †Only these between treatment group comparisons are analysed because treatment optimisation was a key aspect of the intervention. Analyses are adjusted by site (Bristol, Sheffield, or Southampton) and baseline QRISK2 score. General practice included as random effect. ‡Questionnaire data. Over the 12 month period, there was no evidence of a difference between the intervention and control arms in the number of times participants consulted in primary care (mean 11.28 (SD 8.8, n=313) and 11.42 (SD=7.9 n=325), respectively (adjusted incidence ratio 0.99, 0.89 to 1.09, P=0.80). Patient safety Over the course of the trial, 76 adverse events were reported by participants, 38 in each trial arm. Twenty four serious and unexpected events occurred in the usual care arm and 22 in the intervention arm (see supplementary appendix 2). Only one serious event in the intervention arm was likely to be related: a participant was admitted to hospital with low blood pressure, which could have been due to antihypertensive drugs not being reduced after weight loss.

occurred in the usual care arm and 22 in the intervention arm (see supplementary appendix 2). Only one serious event in the intervention arm was likely to be related: a participant was admitted to hospital with low blood pressure, which could have been due to antihypertensive drugs not being reduced after weight loss. Discussion This study suggests a modest benefit from the Healthlines service in terms of the proportion of people reducing or maintaining their cardiovascular disease risk over 12 months. Despite the large sample size, the estimate of effect had wide confidence intervals and could be consistent with no effect or a 90% increase in the odds of reducing or maintaining risk. The results for the primary outcome were not statistically significant either in the complete case analysis or after multiple imputation of missing data. Furthermore, there was no evidence of a difference between the trial arms in average risk, treating QRISK2 as a continuous measure (a secondary outcome). Cardiovascular risk is a composite measure, based on several underlying risk factors. The Healthlines intervention was associated with small but meaningful improvements in several of these factors, including reductions in blood pressure and weight but not in cholesterol level or smoking. It was also associated with improvements in self management behaviours such as diet and physical activity, better adherence to drugs, and greater participant satisfaction with support, access to care, and treatment received. It is important to note that these improvements in self management behaviours would reduce cardiovascular risk beyond the benefit captured in the QRISK2 score and are also likely to reduce risk for many other common and serious diseases, so our focus on cardiovascular risk measured using QRISK2 is likely to be conservative in terms of estimating overall benefit.

nts in self management behaviours would reduce cardiovascular risk beyond the benefit captured in the QRISK2 score and are also likely to reduce risk for many other common and serious diseases, so our focus on cardiovascular risk measured using QRISK2 is likely to be conservative in terms of estimating overall benefit. The intervention was not successful at promoting optimisation of drug treatment in line with current guidelines, which was a key intended mechanism for reducing blood pressure and cholesterol levels. This is consistent with previous research highlighting the problem of clinical inertia—that treatment is not necessarily intensified in people who fail to reach treatment targets even when regular monitoring shows inadequate control.35 Although the observed reduction in cardiovascular risk was small (and could be due to chance), the likely reduction in cardiovascular events in the longer term means that the Healthlines service was likely to be cost effective. Strengths and limitations of this study This is a large and pragmatic trial of a telehealth intervention to reduce cardiovascular risk. It is a complex intervention combining a range of telehealth approaches and has a strong theoretical foundation based on the underlying telehealth in chronic disease (TECH) conceptual model.18 The large sample size and high level of participant retention enhance internal validity, whereas the multicentre recruitment and broad inclusion criteria enhance external validity.

ehealth approaches and has a strong theoretical foundation based on the underlying telehealth in chronic disease (TECH) conceptual model.18 The large sample size and high level of participant retention enhance internal validity, whereas the multicentre recruitment and broad inclusion criteria enhance external validity. The Healthlines intervention incorporates the use of several telehealth approaches, which have reasonable evidence of effectiveness, such as home blood pressure monitoring, and we sought to implement them on a wide scale. Most research studies of telehealth interventions relate to specific technological innovations and can be characterised as efficacy trials, in that they demonstrate the effect of a well defined intervention in people with tightly defined inclusion and exclusion criteria, and who are motivated to use the particular application. These studies may lead to estimates of effect that are exaggerated when compared with the effects observed with wider implementation of the application. By contrast, this trial was pragmatic, testing an intervention as delivered by a mainstream NHS provider in a way that could be rolled out quickly on a wide scale.

on. These studies may lead to estimates of effect that are exaggerated when compared with the effects observed with wider implementation of the application. By contrast, this trial was pragmatic, testing an intervention as delivered by a mainstream NHS provider in a way that could be rolled out quickly on a wide scale. This study has several limitations. Firstly, only 16% of those sent information about the study expressed an interest in it. This response rate is not unusual in primary care based trials in which people who may not have an expressed health need are invited to take part in research. Indeed the response rate in this trial was higher than in several other influential trials of related interventions.36 37 38 However, if non-respondents differ from respondents because of disinterest in research this could reduce the generalisability of the trial findings. Based on information from 2741 people who gave a reason for non-participation, the most common reasons were related to technology rather than to research: 1491 (54%) had no internet access and 1225 (45%) did not feel confident using computers (people could provide more than one reason).39 Many people (n=1135, 41%) did not feel they needed additional support with health problems. It is important to note that less than half of those invited for an NHS Health Check actually attend, and not everyone who smokes or is overweight is motivated to change. We also recognise that telehealth interventions are not necessarily of interest to everyone, and take-up in routine service use may be low. However, healthcare is likely to be increasingly personalised, with different forms of care being chosen by different groups in the population. Telehealth interventions may be useful for a minority of potential participants if (as in the case of increased cardiovascular risk) the total number of people at risk is large.

owever, healthcare is likely to be increasingly personalised, with different forms of care being chosen by different groups in the population. Telehealth interventions may be useful for a minority of potential participants if (as in the case of increased cardiovascular risk) the total number of people at risk is large. Secondly, the closure of NHS Direct towards the end of the trial meant that delivery of the intervention was disrupted and many participants received less than the full course of intervention encounters. However, that we were able to move the service quickly to another provider demonstrates the transferability of the approach. Thirdly, we analysed a large number of secondary outcomes in order to capture the range of potential effects from this complex intervention, but this raises the possibility of some apparent differences being due to chance because of multiple testing. Fourthly, the sample size was chosen pragmatically and assumed that 35% of participants in the control arm would maintain or reduce their cardiovascular risk over 12 months. In the trial, a higher than anticipated proportion of those in the control group achieved this, perhaps because of the impact of the NHS Health Checks programme.26 This reduced the power of the study to detect differences between the intervention and control groups, but this will have been mitigated to some extent by the fact that we recruited and successfully followed up more patients than anticipated. Fifthly, the study was limited to patients aged less than 75 years (because this is the age range in which QRISK2 has been validated and is also the age group targeted by NHS Health Checks), but this intervention could potentially also help older people. The study also excluded people without access to the internet; however, the proportion of the population with access is increasing rapidly.

e this is the age range in which QRISK2 has been validated and is also the age group targeted by NHS Health Checks), but this intervention could potentially also help older people. The study also excluded people without access to the internet; however, the proportion of the population with access is increasing rapidly. Finally, the use of cardiovascular risk as a composite outcome has limitations because the QRISK2 score is strongly dominated by non-modifiable factors such as age and sex. We chose to analyse the QRISK2 as a binary measure of “response to treatment” for the primary outcome because this approach is sensitive to changes in modifiable risk factors. The number of patients needed to treat to gain benefit from the intervention was 13. However, because only a minority of participants benefited, there was no statistically significant change in QRISK2 averaged across all participants when analysed as a continuous variable (a secondary outcome). Nevertheless, the small changes in modifiable risk factors observed in this trial are likely to be associated with meaningful benefits. Based on the systematic review by Law et al,40 the reductions in blood pressure observed in this trial would lead to a 23% reduction in the relative risk of stroke and a 15% reduction in the relative risk of a heart attack. The combined effect (along with the reduction in weight) suggests that these small changes in modifiable risk factors are likely to be worthwhile, particularly at a population level when applied to the large number of people at high risk of cardiovascular disease.

a 15% reduction in the relative risk of a heart attack. The combined effect (along with the reduction in weight) suggests that these small changes in modifiable risk factors are likely to be worthwhile, particularly at a population level when applied to the large number of people at high risk of cardiovascular disease. Comparison with other studies This was a trial of the implementation of the combined use of a range of telehealth interventions to deal with cardiovascular risk factors. The results are broadly consistent with earlier trials, which have studied different components of the intervention in isolation to reduce individual risk factors. A systematic review of trials of blood pressure self monitoring showed that this was associated with small reductions in both systolic and diastolic blood pressure of a similar size to those achieved in the Healthlines cardiovascular disease risk trial.23 A Cochrane review found that computer based interactive interventions for weight loss were associated with a mean weight loss of 1.5 kg (95% confidence interval 0.9 to 2.1 kg) compared with no or minimal intervention, an effect which is also consistent with our findings.25 Systematic reviews on internet based telehealth interventions for smoking cessation show mixed effects, although mobile phone based interventions are effective and telephone quitlines can improve cessation rates in those people who proactively contact them.24 41 42 It is important to note that the above reviews were all based on people who had the risk factor of interest, and many trials only included those who were motivated to change the specific risk factor. In the Healthlines cardiovascular disease risk trial only a proportion of participants had raised blood pressure, were obese, or were smokers at baseline, and they were not necessarily motivated to change the main factor contributing to their risk, so effects are likely to be smaller than in studies on specific risk factors.

Healthlines cardiovascular disease risk trial only a proportion of participants had raised blood pressure, were obese, or were smokers at baseline, and they were not necessarily motivated to change the main factor contributing to their risk, so effects are likely to be smaller than in studies on specific risk factors. The Healthlines intervention tested in this trial had a similar impact on blood pressure reduction as the earlier trials by Bosworth et al, which used a similar behavioural management system (but provided by nurses rather than lay staff and without incorporating the use of internet resources).28 29 However, the Healthlines trial had less impact than two trials from the United States, which involved blood pressure self management with pharmacist management of drugs by phone or over the internet.36 37 The involvement of pharmacists to directly alter drugs without the intermediate step of sending advice to primary care doctors may be associated with more effective optimisation of treatment but could be problematic in a routine primary care context, when patients often have comorbidities and other factors need to be considered in treatment decisions.

directly alter drugs without the intermediate step of sending advice to primary care doctors may be associated with more effective optimisation of treatment but could be problematic in a routine primary care context, when patients often have comorbidities and other factors need to be considered in treatment decisions. Two systematic reviews of telehealth interventions to reduce overall cardiovascular risk have recently been published.14 43 Several studies demonstrated small improvements in blood pressure and weight, findings for cholesterol were equivocal, and there was no evidence of increased rates of smoking cessation. Our results are consistent with these findings but provide much stronger evidence from a large, rigorous and pragmatic trial.

ed.14 43 Several studies demonstrated small improvements in blood pressure and weight, findings for cholesterol were equivocal, and there was no evidence of increased rates of smoking cessation. Our results are consistent with these findings but provide much stronger evidence from a large, rigorous and pragmatic trial. Implications for clinicians and for policy The development of the Healthlines service reflected a conceptual framework that was based on promoting self management, improving drug adherence and optimisation of drug treatment, coordination of care, and the active engagement of patients and primary care clinicians.18 This randomised controlled trial shows modest but cost effective benefit in cardiovascular risk reduction. Delineating how components of a multifaceted intervention work, alone or in combination; their effect on doctor practice in terms of optimisation or intensification of medicines and their effect on behaviour modification by patients is complex. What is clear is that patients who engaged with the intervention seem to gain the most in terms of cardiovascular risk, but some components of the intervention, particularly optimisation or intensification of drugs, were ineffective. To improve the effectiveness of the intervention it will be important to target it at those who are motivated to change their risk behaviours, and to improve communication with primary care prescribers about drug treatment recommendations.

ention, particularly optimisation or intensification of drugs, were ineffective. To improve the effectiveness of the intervention it will be important to target it at those who are motivated to change their risk behaviours, and to improve communication with primary care prescribers about drug treatment recommendations. Conclusions Optimism about the potential of telehealth approaches to improve the accessibility, convenience, and efficiency of healthcare has been considerable. This study adds to the growing evidence base, which suggests that healthcare delivery systems based on telehealth may be associated with some benefits, although these should not be assumed. However, this study has demonstrated the feasibility of delivering an intervention on a wide scale at relatively low cost and using non-clinically trained health advisors supported by computerised algorithms. This increases the capacity of the healthcare system to provide an intervention to large numbers of people. Further development of this type of intervention is justified to increase the effectiveness of the Healthlines service approach. What is already known on this topic Given the increasing prevalence of long term health conditions, it is necessary to explore new ways to deliver healthcare and to support self management to expand provision of care at low cost There is considerable optimism among policy makers that greater use of digital health technologies (“telehealth”) in combination with new ways of working could transform healthcare delivery, helping the UK national health service to be sustainable

What is already known on this topic Given the increasing prevalence of long term health conditions, it is necessary to explore new ways to deliver healthcare and to support self management to expand provision of care at low cost There is considerable optimism among policy makers that greater use of digital health technologies (“telehealth”) in combination with new ways of working could transform healthcare delivery, helping the UK national health service to be sustainable Evidence about the effectiveness of telehealth interventions is equivocal, with some benefits from specific technologies but little evidence of effectiveness in real world implementation What this study adds Some evidence suggests that an intervention combining the use of a range of digital health technologies with telephone support from trained lay health advisors, leads to a modest improvement in overall cardiovascular risk for a minority of participants The intervention had no impact on average cardiovascular risk but was associated with improvements in specific cardiovascular risk factors and health behaviours and patient perceptions of support and access to care Web extra Extra material supplied by authors Web appendix: supplementary information Click here for additional data file.

The intervention had no impact on average cardiovascular risk but was associated with improvements in specific cardiovascular risk factors and health behaviours and patient perceptions of support and access to care Web extra Extra material supplied by authors Web appendix: supplementary information Click here for additional data file. We thank the patients, healthcare professionals, health information advisors, and other NHS Direct staff who contributed time and effort to make this trial possible; Sarah Williams at Solent NHS Trust who took over responsibility for hosting the intervention when NHS Direct closed; Steve Bellerby who managed implementation of the Healthlines software at both trusts; members of the trial steering and data monitoring committees, particularly Michelle McPhail and Anne Jacob who were patient and public representatives; Hayden Bosworth for permission to adapt his behavioural intervention, and both he and Felicia McCann for their help with adaptation; Roberta Ara for input into the design and implementation of the lifetime cost effectiveness cohort simulation model; Glyn Lewis and Simon Brownsell who were applicants on the research programme grant but not directly involved in this trial; the Primary Care Research Network (now, NIHR Clinical Research Network) for assisting us with recruitment of general practitioners; and Frederika Collihole, Richard Campbell, Ben Davies, Lorna Duncan, Diane Beck, and Janet Cooke who supported participant recruitment, data entry, and trial administration. This study was designed and delivered in collaboration with Bristol Randomised Trials Collaboration (BRTC), a UKCRC Registered Clinical Trials Unit in receipt of National Institute for Health Research CTU support funding.

Beck, and Janet Cooke who supported participant recruitment, data entry, and trial administration. This study was designed and delivered in collaboration with Bristol Randomised Trials Collaboration (BRTC), a UKCRC Registered Clinical Trials Unit in receipt of National Institute for Health Research CTU support funding. Contributors: CS, AO’C, SH, JN, SL, LY, TF, AR, CP, and AAM developed the protocol for the study, obtained funding, provided methodological advice, and supervised the conduct of the trial. CS led protocol development and the funding application, acted as chief investigator with overall responsibility for the conduct of the trial, and led the drafting of the article. AO’C supervised the conduct of the trial in Sheffield. CT, MSM, and LE acted as trial managers, coordinating the conduct of the trial across the centres. LE, AF, KG, and KH undertook participant recruitment and follow-up, data collection, and data entry. DG developed the statistical analysis plan and undertook the statistical analysis. PD undertook the economic analysis. SL coordinated development and delivery of the intervention with NHS Direct. AAM supervised the statistical analysis. All authors, external and internal, had full access to all of the data (including statistical reports and tables) in the study and can take responsibility for the integrity of the data and the accuracy of the data analysis. All authors approved the final version of this manuscript. CS is guarantor.

tatistical analysis. All authors, external and internal, had full access to all of the data (including statistical reports and tables) in the study and can take responsibility for the integrity of the data and the accuracy of the data analysis. All authors approved the final version of this manuscript. CS is guarantor. Funding: This report summarises independent research funded by the National Institute for Health Research (NIHR) under its programme grant for applied research (grant reference No RP-PG-0108-10011). CS’s time is supported by the NIHR Collaboration for Leadership in Applied Health Research and Care West (CLAHRC West) at University Hospitals Bristol NHS Foundation Trust. The views and opinions expressed in this report are those of the authors and do not necessarily reflect those of the NIHR the NHS, or the Department of Health. The funder had no role in the conduct of the study, the writing of the manuscript, or the decision to submit it for publication. The corresponding author (CS) had full access to all the data in the study and had final responsibility for the decision to submit for publication. CS, AOC, and JN act as members of boards for NIHR but were not on the board which commissioned this project.

f the manuscript, or the decision to submit it for publication. The corresponding author (CS) had full access to all the data in the study and had final responsibility for the decision to submit for publication. CS, AOC, and JN act as members of boards for NIHR but were not on the board which commissioned this project. Competing interests: All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf and declare: support from NIHR in grant funding but no other support for the submitted work; no financial relationships with any organisations that might have an interest in the submitted work in the previous three years; their spouses, partners, or children have no financial relationships that may be relevant to the submitted work; and no non-financial interests that may be relevant to the submitted work. The Healthlines patient portal is the intellectual property of Solent NHS Trust. The telephone algorithms were adapted with permission from a patient assessment system, which is the intellectual property of Duke University. Interested readers should refer to the website www.bristol.ac.uk/healthlines/ and contact the author for further information. Ethical approval: This study was approved by the National Research Ethics Service Committee South West–Frenchay (reference 12/SW/0009). All participants provided informed consent to take part in the trial.

Competing interests: All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf and declare: support from NIHR in grant funding but no other support for the submitted work; no financial relationships with any organisations that might have an interest in the submitted work in the previous three years; their spouses, partners, or children have no financial relationships that may be relevant to the submitted work; and no non-financial interests that may be relevant to the submitted work. The Healthlines patient portal is the intellectual property of Solent NHS Trust. The telephone algorithms were adapted with permission from a patient assessment system, which is the intellectual property of Duke University. Interested readers should refer to the website www.bristol.ac.uk/healthlines/ and contact the author for further information. Ethical approval: This study was approved by the National Research Ethics Service Committee South West–Frenchay (reference 12/SW/0009). All participants provided informed consent to take part in the trial. Data sharing: The research team will consider reasonable requests for sharing of patient level data. Requests should be made to CS. Consent for data sharing was not obtained but the presented data are anonymised and risk of identification is low.

Ethical approval: This study was approved by the National Research Ethics Service Committee South West–Frenchay (reference 12/SW/0009). All participants provided informed consent to take part in the trial. Data sharing: The research team will consider reasonable requests for sharing of patient level data. Requests should be made to CS. Consent for data sharing was not obtained but the presented data are anonymised and risk of identification is low. Transparency: The lead author (manuscript’s guarantor) affirms that the manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned and registered have been explained.

Introduction Cardiovascular disease (CVD) is the leading cause of death worldwide.1 Furthermore, substantial socioeconomic inequalities have been observed in CVD mortality in England and elsewhere.2 3 These inequalities powerfully reflect much greater premature mortality, and hence shorter life expectancy, among the most deprived groups. In England, the current governmental action plan to tackle the burden of CVD includes a programme known as NHS (National Health Service) Health Checks. Introduced in 2009, this programme promotes the screening of all healthy adults aged 40 to 74 for CVD risk stratification, and treatment of those at high risk.4 5 Recently, the debate about the programme’s scientific foundation, effectiveness, and cost effectiveness, however, has been heated.6 7 8 9 10 Despite the controversy, the programme remains policy.

motes the screening of all healthy adults aged 40 to 74 for CVD risk stratification, and treatment of those at high risk.4 5 Recently, the debate about the programme’s scientific foundation, effectiveness, and cost effectiveness, however, has been heated.6 7 8 9 10 Despite the controversy, the programme remains policy. Beyond the obvious importance of the debate to national public health, the programme’s relevance extends internationally. Choices about public health policy in the United Kingdom influence policy worldwide; the UK policies on tobacco control and salt reduction are two recent examples.11 12 In essence, the debate about NHS Health Checks originates from the archetypal debate of targeted “high risk” versus “population-wide” preventive interventions that was first articulated by Geoffrey Rose.13 Rose argued that population-wide interventions are more effective than ones aimed at high risk groups because the majority of incident cases occur in the multitudinous group of people at low and intermediate risk. In Rose terminology, NHS Health Checks is a typical “high risk intervention,” as it targets people at high risk rather than lowering risk in the whole population.

effective than ones aimed at high risk groups because the majority of incident cases occur in the multitudinous group of people at low and intermediate risk. In Rose terminology, NHS Health Checks is a typical “high risk intervention,” as it targets people at high risk rather than lowering risk in the whole population. The effectiveness of high risk interventions for CVD prevention has been previously challenged.14 More recently, a Cochrane systematic review and the Inter99 trial found no benefits of health checks on CVD morbidity or mortality.15 16 There were, however, major limitations to these studies: Inter99 trialled a counselling intervention not supported by additional drug treatment, and in the Cochrane review nine out of 14 trials were conducted before 1980, when the treatment options for high risk people were limited. In addition, high risk interventions may be more effective in populations with high clustering of risk factors, resulting in a high concentration of the risk to certain groups in the population.17 In fact, the English population has such characteristics, with the risk of CVD being higher among those in the most socioeconomically deprived groups.18

entions may be more effective in populations with high clustering of risk factors, resulting in a high concentration of the risk to certain groups in the population.17 In fact, the English population has such characteristics, with the risk of CVD being higher among those in the most socioeconomically deprived groups.18 High risk interventions may generate health inequalities because they require active participation of people in both screening and treatment of those at high risk, favouring those with more resources.14 19 20 21 The particular effect of NHS Health Checks on socioeconomic health inequalities remains unclear however. A national study reported no difference in the coverage of the intervention by deprivation,22 whereas several smaller, but more detailed, studies showed substantially lower uptake in deprived areas.23 24 25 We estimated the potential impact of universal screening for primary prevention of CVD on disease burden and socioeconomic health inequalities in England. Available data on the effectiveness of the NHS Health Check programme have been used to model this scenario. We further compared universal CVD screening with an alternative approach targeting only deprived areas, a feasible population-wide intervention, and a combination of both.

conomic health inequalities in England. Available data on the effectiveness of the NHS Health Check programme have been used to model this scenario. We further compared universal CVD screening with an alternative approach targeting only deprived areas, a feasible population-wide intervention, and a combination of both. Methods Building on experience from the original, validated IMPACT model26 and the more recent IMPACTSEC27 and IMPACT2 models,28 we created IMPACTNCD, a discrete time dynamic stochastic microsimulation model. IMPACTNCD simulates the life course of synthetic individuals under different counterfactual scenarios, up to 2030 (the projection horizon). During the simulation, CVD incidence and CVD and non-CVD mortality are recorded. The results are stratified by year, five year age group, sex, and fifths of index of multiple deprivation. The last is a relative measure of area deprivation that is widely used by public health authorities in England, and it has been used as the measure of socioeconomic classification for this study.29 A more detailed description of the model is provided in the supplementary material and the source code is available at https://github.com/ChristK/IMPACTncd/tree/CVD-policy-options. Scenarios We considered five scenarios.

Methods Building on experience from the original, validated IMPACT model26 and the more recent IMPACTSEC27 and IMPACT2 models,28 we created IMPACTNCD, a discrete time dynamic stochastic microsimulation model. IMPACTNCD simulates the life course of synthetic individuals under different counterfactual scenarios, up to 2030 (the projection horizon). During the simulation, CVD incidence and CVD and non-CVD mortality are recorded. The results are stratified by year, five year age group, sex, and fifths of index of multiple deprivation. The last is a relative measure of area deprivation that is widely used by public health authorities in England, and it has been used as the measure of socioeconomic classification for this study.29 A more detailed description of the model is provided in the supplementary material and the source code is available at https://github.com/ChristK/IMPACTncd/tree/CVD-policy-options. Scenarios We considered five scenarios. Baseline (current trends) In the baseline scenario, we assumed that the recent observed trends in CVD risk factor trajectories by age, sex, and socioeconomic status will continue in the near future. We extracted the trends from the health survey for England 2001-12, a nationally representative series of health surveys conducted in England annually.30 31 32 33 34 35 36 37 38 39 40 41 42

the recent observed trends in CVD risk factor trajectories by age, sex, and socioeconomic status will continue in the near future. We extracted the trends from the health survey for England 2001-12, a nationally representative series of health surveys conducted in England annually.30 31 32 33 34 35 36 37 38 39 40 41 42 Universal screening This scenario modelled the potential health effects of universal screening to identify and treat people at high risk for CVD. Input variables were informed from current implementation of the NHS Health Check programme. Eligible people were defined as adults aged between 40 and 74, excluding those with a known history of CVD, atrial fibrillation, diabetes mellitus, rheumatoid arthritis, or renal disease; closely resembling real life eligibility criteria. Based on existing evidence we assumed an uptake of 50% for screening,43 and we calibrated the distribution of the estimated 10 year risk of developing CVD among those participating: 70% with a less than 10% risk, 25% with between 10% and 20%, and 5% with more than 20%.22 In addition, we calibrated the age distribution so that around 30% of those screened were older than 60.22 Participants with a higher than 10% estimated 10 year risk of developing CVD were considered at high risk and eligible for treatment. We used the QRISK2 score to estimate the 10 year risk of developing CVD, as perceived from healthcare.44

d the age distribution so that around 30% of those screened were older than 60.22 Participants with a higher than 10% estimated 10 year risk of developing CVD were considered at high risk and eligible for treatment. We used the QRISK2 score to estimate the 10 year risk of developing CVD, as perceived from healthcare.44 Based on published evidence, we assumed that about 24% with an estimated risk of 20% or more and total cholesterol of 5 mmol/L or more will be prescribed atorvastatin 20 mg and about 27% with an estimated risk of 20% or more and a systolic blood pressure of 135 mm Hg or more will be prescribed antihypertensive drugs. For those with a risk between 10% and 20% we assumed that about 17% and 20% will be prescribed treatment, respectively.45 We assumed an 80% persistence with treatment and a mean adherence of approximately 70%, roughly based on evidence from Denmark.46 Moreover, we modelled high risk participants with a body mass index of more than 50 kg/m2 to undergo bariatric surgery and reduce their body mass index to 30 kg/m2. We assumed that with lifestyle counselling half of the high risk participants consuming fewer than five fruit and vegetable portions daily will increase their consumption by a portion daily. Half of those being active for less than five days a week will increase their physical activity by an active day each week, and all high risk participants will decrease their body mass index by around 1%.45 47 Finally, we modelled 10% of high risk smokers to achieve cessation for a year and have a probability of relapse equal to that of the general population by sex, fifth of multiple deprivation, and years since cessation.48 49

each week, and all high risk participants will decrease their body mass index by around 1%.45 47 Finally, we modelled 10% of high risk smokers to achieve cessation for a year and have a probability of relapse equal to that of the general population by sex, fifth of multiple deprivation, and years since cessation.48 49 Concentrated screening In the concentrated screening scenario, we simulated a hypothetical strategy where screening had only been implemented in the most deprived fifths (groups 4 and 5), the groups with the greatest concentration of CVD risk. We assumed that the uptake of the intervention was 50% and the risk and age distribution in the participants was similar to that in the eligible population. Otherwise, the strategy is similar to the previous universal screening scenario. Given the recent criticism about the cost and cost effectiveness of the intervention,9 offering the intervention where the risk is more concentrated may reduce costs.

tion in the participants was similar to that in the eligible population. Otherwise, the strategy is similar to the previous universal screening scenario. Given the recent criticism about the cost and cost effectiveness of the intervention,9 offering the intervention where the risk is more concentrated may reduce costs. Population-wide intervention This scenario modelled the effects of a feasible population-wide structural intervention targeting unhealthy diet and smoking. Several studies have found that a tax on sugar sweetened beverages may reduce the prevalence of obesity.50 51 52 For this scenario we assumed that such a tax may reduce the mean increase in body mass index by about 5% annually. Moreover, the United Kingdom has had one of the world’s most successful salt reduction strategies, including public awareness campaigns, food labelling, and voluntary reformulation of processed foods.53 Modelling studies suggested that the addition of mandatory reformulation of processed foods may further reduce mean systolic blood pressure by 0.8 mm Hg54; we modelled this decrease. A large randomised trial in the United States showed that subsidies on fruits and vegetables may increase consumption by about half a portion daily, and a modelling study in the UK found that subsidising fruits and vegetables combined with taxation of unhealthy foods may increase fruit and vegetable annual consumption by about 10%.55 56 We modelled an increase of a portion of fruit and vegetable each day in 50% of the population. Finally, a SimSmoke modelling study estimated that full compliance with the framework convention on tobacco control may reduce smoking prevalence by 13% (relative) in five years57; we modelled this decrease.

y about 10%.55 56 We modelled an increase of a portion of fruit and vegetable each day in 50% of the population. Finally, a SimSmoke modelling study estimated that full compliance with the framework convention on tobacco control may reduce smoking prevalence by 13% (relative) in five years57; we modelled this decrease. Population-wide intervention and concentrated screening This scenario is the combination of the population-wide intervention and concentrated screening strategies. We modelled the implementation of a population-wide strategy identical to the previous scenario, complemented by concentrated screening for people at high risk of CVD in the most deprived fifths (groups 4 and 5). Common scenario assumptions All interventions begun in 2011 and were linearly diffused into the population over a five year period. Trends in population risk factors were assumed to be the same as those of the baseline scenario for all but the population-wide intervention. All of the scenarios assumed that CVD case fatality will keep improving by 3% (relative) annually. In addition, we assumed a socioeconomic gradient in CVD case fatality, forcing the more deprived people to experience worse outcomes. Both case fatality assumptions were based on recent trends and are supported by the British Heart Foundation’s statistics on coronary heart disease.2 Finally, a five year lag time was assumed between exposure to cardiovascular risk factors and disease.

ity, forcing the more deprived people to experience worse outcomes. Both case fatality assumptions were based on recent trends and are supported by the British Heart Foundation’s statistics on coronary heart disease.2 Finally, a five year lag time was assumed between exposure to cardiovascular risk factors and disease. Model description Inputs and logic IMPACTNCD synthesises information from the Office for National Statistics and the health surveys for England on the English population’s demographics and its exposure to CVD associated risk factors, to generate a close-to-reality synthetic population.58 Well established causal pathways between CVD and the associated risk factors are used to translate exposure into CVD incidence and mortality, in a competing risk framework. We obtained effect sizes for exposures from published meta-analyses and longitudinal studies (see supplementary table S1).

synthetic population.58 Well established causal pathways between CVD and the associated risk factors are used to translate exposure into CVD incidence and mortality, in a competing risk framework. We obtained effect sizes for exposures from published meta-analyses and longitudinal studies (see supplementary table S1). The risk factors we considered for this study were age, sex, fifth of deprivation, body mass index, systolic blood pressure, total cholesterol level, diabetes mellitus (diagnosis or increased glycated haemoglobin level/no diabetes), smoking status (current, former, or never smoker), environmental tobacco exposure (binary variable), fruit and vegetable consumption (portions daily), and physical activity (days with at least 30 minutes of moderate or vigorous physical activity each week). CVD was defined as the sum of coronary heart disease and stroke (any type) cases. As this study focuses on primary prevention, we considered only the first ever episode of coronary heart disease or stroke. The competing risk framework allowed people to develop coronary heart disease and/or stroke separately, and to die from these two diseases or any other cause.

disease and stroke (any type) cases. As this study focuses on primary prevention, we considered only the first ever episode of coronary heart disease or stroke. The competing risk framework allowed people to develop coronary heart disease and/or stroke separately, and to die from these two diseases or any other cause. Model outputs We report the cumulative estimates of cases and deaths prevented or postponed as measures of overall effectiveness of the modelled interventions. To measure the impact of the modelled interventions on absolute and relative socioeconomic health inequalities, we developed and used two regression based metrics inspired by the slope index of inequality59; the absolute equity slope index and the relative equity slope index. The absolute equity slope index measures the impact of an intervention on absolute inequality; for example, a value of 100 means 100 more cases were prevented or postponed in most deprived areas compared with least deprived areas, resulting in a decrease in absolute inequality. The relative equity slope index takes into account the pre-existing socioeconomic gradient of disease burden and measures the impact of an intervention on relative inequality. Positive values mean the intervention tackles relative inequalities and negative values that the intervention generates relative inequality. Finally, we summarised the overall impact of each scenario on CVD burden and equity in the equity summary chart.

and measures the impact of an intervention on relative inequality. Positive values mean the intervention tackles relative inequalities and negative values that the intervention generates relative inequality. Finally, we summarised the overall impact of each scenario on CVD burden and equity in the equity summary chart. Uncertainty and sensitivity analysis IMPACTNCD implements a second order Monte Carlo design that allows uncertainty to be quantified from the outputs. We used distributions to model the uncertainty around all scenario specific inputs and the sampling error of the risk associated with the CVD related risk factors. The probabilistic sensitivity analysis has been incorporated in our estimates. We summarise the distributions by reporting medians and interquartile ranges in the form of first and third fourths. The supplementary file provides a more detailed description of the sources of uncertainty and the relevant distributions. We ran three further scenarios offering slight variations on the two primary ones of universal screening and population-wide intervention: a universal screening variation, where we assumed a treatment threshold recommendation of 20% risk instead of 10%; another variation on universal screening, where we assumed a socioeconomic differential in screening uptake, with the most deprived of the population to be 10% less likely to participate; and a variation on the population-wide intervention, where we only modelled dietary interventions, excluding smoking interventions. The supplementary file provides detailed information on the extra scenarios.

ifferential in screening uptake, with the most deprived of the population to be 10% less likely to participate; and a variation on the population-wide intervention, where we only modelled dietary interventions, excluding smoking interventions. The supplementary file provides detailed information on the extra scenarios. Validation We assessed the predictive validity of the IMPACTNCD model by comparing the estimated number of deaths from CVD with the observed number of deaths from the same causes for 2006 to 2013 in England.60 We further compared the IMPACTNCD output with CVD mortality forecasts from a bayesian age-period-cohort model.61 Patient involvement No patients were involved in setting the research question or the outcome measures, nor were they involved in developing plans for design or implementation of the study. No patients were asked to advise on interpretation or writing up of results. There are no plans to disseminate the results of the research to study participants or the relevant patient community. Results IMPACTNCD outputs for CVD burden and inequality are summarised for ages 30 to 84. Because of the assumed five year time lag, the interventions affect the population from 2016 up to the projection horizon of 2030. The impact of the five scenarios on risk factor trajectories are further illustrated in additional graphs in the supplementary file.

burden and inequality are summarised for ages 30 to 84. Because of the assumed five year time lag, the interventions affect the population from 2016 up to the projection horizon of 2030. The impact of the five scenarios on risk factor trajectories are further illustrated in additional graphs in the supplementary file. Overall effectiveness Under the baseline scenario, IMPACTNCD estimated about 1.4 million (interquartile range 1.3-1.5) cases of CVD and 540 000 deaths (interquartile range 520 000 to 550 000) between 2016 and 2030. The most effective intervention was the combination of the population-wide intervention and concentrated screening. The population-wide intervention alone had the second highest effectiveness, whereas the universal and the concentrated screening scenarios were considerably less effective (table 1). Despite the improvement of most CVD related risk factors, the proportion of high risk people in the eligible population is slowly increasing over time, because of population aging (fig 1). Table 1 Estimated cases and deaths prevented or postponed under each scenario, by 2030 Scenarios No (interquartile range) prevented or postponed Cases Deaths Universal screening 19 000 (11 000-28 000) 3000 (−1000-6000) Concentrated screening 17 000 (9000-26 000) 2000 (−1000-5000) Population-wide intervention 67 000 (57 000-77 000) 8000 (4000-11 000) Population-wide intervention and concentrated screening 82 000 (73 000-93 000) 9000 (6000-13 000) Results rounded to nearest 1000.

ersal screening 19 000 (11 000-28 000) 3000 (−1000-6000) Concentrated screening 17 000 (9000-26 000) 2000 (−1000-5000) Population-wide intervention 67 000 (57 000-77 000) 8000 (4000-11 000) Population-wide intervention and concentrated screening 82 000 (73 000-93 000) 9000 (6000-13 000) Results rounded to nearest 1000. Fig 1 Proportion of high risk people eligible for universal screening population projections, by age group and sex. 10 year risk of cardiovascular disease (CVD) was estimated from QRISK2 score. Error bars represent interquartile ranges Socioeconomic inequalities When socioeconomic inequalities were considered, the patterns for reductions in absolute and relative inequalities were similar. The combination of the population-wide intervention and concentrated screening seemed the most powerful among the simulated interventions (tables 2 and 3). Concentrated screening alone was the second most powerful intervention in tackling inequalities, followed by the population-wide intervention. Finally, universal screening of CVD is likely to have a small, if any, effect on socioeconomic inequalities. Table 2 Cases prevented or postponed according to fifth of deprivation by 2030, along with absolute equity slope index for each scenario

Socioeconomic inequalities When socioeconomic inequalities were considered, the patterns for reductions in absolute and relative inequalities were similar. The combination of the population-wide intervention and concentrated screening seemed the most powerful among the simulated interventions (tables 2 and 3). Concentrated screening alone was the second most powerful intervention in tackling inequalities, followed by the population-wide intervention. Finally, universal screening of CVD is likely to have a small, if any, effect on socioeconomic inequalities. Table 2 Cases prevented or postponed according to fifth of deprivation by 2030, along with absolute equity slope index for each scenario Deprivation fifth* No (interquartile range) of cases prevented or postponed Universal screening Concentrated screening Population-wide intervention Population-wide intervention+concentrated screening First (least deprived) 3400 (−1400-8300) 0 10 800 (5900-15 500) 10 800 (6200-15 700) Second 2900 (−1500-8400) 0 12 200 (6200-17 200) 11 500 (6600-17 000) Third 4000 (−900-9300) 0 13 100 (8100-18 300) 12 600 (7400-17 700) Fourth 3700 (−1600-8600) 6400 (1500-11 800) 12 500 (7100-18 400) 18 700 (13 900-24 200) Fifth (most deprived) 4900 (−600-10 400) 10 700 (5300-16 300) 18 700 (13 000-24 000) 28 600 (22 800-33 200) Absolute equity slope index 1700 (−6200-9300) 14 100 (5700-23 000) 8400 (−400-16 900) 21 100 (12 800-29 300) Results rounded to nearest 1000. *According to index of multiple deprivation.

Deprivation fifth* No (interquartile range) of cases prevented or postponed Universal screening Concentrated screening Population-wide intervention Population-wide intervention+concentrated screening First (least deprived) 3400 (−1400-8300) 0 10 800 (5900-15 500) 10 800 (6200-15 700) Second 2900 (−1500-8400) 0 12 200 (6200-17 200) 11 500 (6600-17 000) Third 4000 (−900-9300) 0 13 100 (8100-18 300) 12 600 (7400-17 700) Fourth 3700 (−1600-8600) 6400 (1500-11 800) 12 500 (7100-18 400) 18 700 (13 900-24 200) Fifth (most deprived) 4900 (−600-10 400) 10 700 (5300-16 300) 18 700 (13 000-24 000) 28 600 (22 800-33 200) Absolute equity slope index 1700 (−6200-9300) 14 100 (5700-23 000) 8400 (−400-16 900) 21 100 (12 800-29 300) Results rounded to nearest 1000. *According to index of multiple deprivation. Table 3 Relative percentage reduction in cases of cardiovascular disease according to fifth of deprivation by 2030, along with relative equity slope index for each scenario Deprivation fifth* Relative % reduction (interquartile range) Universal screening Concentrated screening Population-wide intervention Population-wide intervention+concentrated screening First (least deprived) 1.3 (−0.5-3.1) 0 4.1 (2.2-5.9) 4.0 (2.4-6.0) Second 1.1 (−0.5-2.9) 0 4.2 (2.2-5.9) 4.0 (2.3-5.9) Third 1.4 (−0.3-3.2) 0 4.6 (2.8-6.3) 4.4 (2.6-6.2) Fourth 1.3 (−0.6-3.1) 2.4 (0.6-4.3) 4.6 (2.7-6.6) 6.9 (5.1-8.9) Fifth (most deprived) 1.6 (−0.2-3.3) 3.6 (1.8-5.3) 6.2 (4.4-8.0) 9.4 (7.6-11.2) Relative equity slope index 0.4 (−2.4-3.2) 4.9 (1.8-7.9) 2.3 (−0.7-5.3) 6.7 (3.8-9.5) Results rounded to one decimal place.

) Third 1.4 (−0.3-3.2) 0 4.6 (2.8-6.3) 4.4 (2.6-6.2) Fourth 1.3 (−0.6-3.1) 2.4 (0.6-4.3) 4.6 (2.7-6.6) 6.9 (5.1-8.9) Fifth (most deprived) 1.6 (−0.2-3.3) 3.6 (1.8-5.3) 6.2 (4.4-8.0) 9.4 (7.6-11.2) Relative equity slope index 0.4 (−2.4-3.2) 4.9 (1.8-7.9) 2.3 (−0.7-5.3) 6.7 (3.8-9.5) Results rounded to one decimal place. *According to index of multiple deprivation. Equity summary chart We summarised our estimates for the effectiveness and equity of the modelled interventions in the equity summary chart (fig 2). The horizontal axis of the chart represents the cases of CVD prevented or postponed and the vertical axis the reduction in absolute inequality. Scenarios above the equity curve (dashed curve in the figure) decrease relative socioeconomic inequality, and scenarios below the curve increase it. The vertical distance from the curve approximates the impact of the scenario on relative inequality. (See the supplementary file for more details about this chart.) The combination of the population-wide intervention and concentrated screening is by far the most effective and equitable intervention. Concentrated screening is also equitable but with few mortality gains.

he impact of the scenario on relative inequality. (See the supplementary file for more details about this chart.) The combination of the population-wide intervention and concentrated screening is by far the most effective and equitable intervention. Concentrated screening is also equitable but with few mortality gains. Fig 2 Equity summary chart of effectiveness and equity of all modelled interventions, compared with baseline scenario (beginning of axes). Dashed line represents “equity” curve. Interventions below the curve increase relative inequality, whereas interventions above it decrease relative inequalities. Smaller coloured dots represent reference points used to fit equity curve. Horizontal and vertical error bars represent interquartile ranges Sensitivity analysis Adding assumptions to extend the scenarios did not displace our main findings. The three most notable results of the sensitivity analysis were: Raising the treatment threshold from 10% to 20% further reduced the effectiveness of universal screening by about 60% in preventing CVD cases. However, in preventing deaths from CVD the effectiveness decreased by only 15% as raising the treatment threshold excludes younger participants at intermediate risk from treatment. Assuming a differential uptake of universal screening by deprivation fifth essentially eliminated the estimated small potential benefit of universal screening in tackling health inequalities.

Raising the treatment threshold from 10% to 20% further reduced the effectiveness of universal screening by about 60% in preventing CVD cases. However, in preventing deaths from CVD the effectiveness decreased by only 15% as raising the treatment threshold excludes younger participants at intermediate risk from treatment. Assuming a differential uptake of universal screening by deprivation fifth essentially eliminated the estimated small potential benefit of universal screening in tackling health inequalities. A population-wide intervention targeting only diet would still be about twice as effective as universal screening and more than twice as effective as population-wide intervention targeting smoking alone—so the relative ranking of scenario effectiveness would remain unaltered. For detailed results see supplementary tables S11-S13. Validation We assessed the predictive validity of the IMPACTNCD model by comparing the estimated number of deaths from CVD with the observed number of deaths from the same cause for 2006 to 2013 in England (fig 3). See the supplementary file for detailed graphs by age group, sex, deprivation fifth, and disease. Fig 3 Number of deaths from cardiovascular disease (CVD) in England, by year for ages 30 to 84. Office for National Statistics reported deaths (observed) versus IMPACTNCD estimated. Observed deaths after 2010 were adjusted to account for changes in ICD-10 version used by the Office for National Statistics from 2011 onwards. Error bars represent interquartile ranges

ase (CVD) in England, by year for ages 30 to 84. Office for National Statistics reported deaths (observed) versus IMPACTNCD estimated. Observed deaths after 2010 were adjusted to account for changes in ICD-10 version used by the Office for National Statistics from 2011 onwards. Error bars represent interquartile ranges Discussion Our results strongly suggest that universal screening and treatment of people at high risk is not the most effective option for primary prevention of cardiovascular disease (CVD) overall, nor for reducing socioeconomic inequalities. In contrast, prevention strategies that include population-wide structural interventions seem to be the consistently better options for reducing overall CVD burden and inequalities. This echoes and quantifies findings from other, mostly theoretical, studies supporting that structural population-wide interventions are powerful, while reducing socioeconomic health inequalities.13 14 62 63 Indeed, the impact of the population-wide intervention scenario on reduction in estimated mortality and inequalities seems compatible with previous estimates, considering the different methodologies.64 Furthermore, the effectiveness and equity of population-wide structural interventions can be further improved by the addition of targeted interventions in the most deprived groups, as highlighted in the combined scenario of the population-wide intervention and concentrated screening.

ring the different methodologies.64 Furthermore, the effectiveness and equity of population-wide structural interventions can be further improved by the addition of targeted interventions in the most deprived groups, as highlighted in the combined scenario of the population-wide intervention and concentrated screening. Compared with other modelling approaches, our IMPACTNCD model estimated that NHS Health Checks might prevent approximately 1000 non-fatal and 200 fatal cases of CVD annually. This is comparable with the Department of Health estimates of 1600 non-fatal CVD cases and 650 deaths prevented annually.4 Furthermore, the Department of Health modelling approach assumed an intervention uptake of 75%; higher than the current observed levels. Using the Archimedes model, Schuetz et al estimated that health checks in the UK could prevent some 12 CVD cases per 1000 population screened after 30 years’ follow-up65 (7500 CVD cases prevented each year extrapolating to the eligible English population). That higher estimate reflects the researchers’ apparently unrealistic assumption of 100% screening uptake and 50% overall uptake of treatment.

K could prevent some 12 CVD cases per 1000 population screened after 30 years’ follow-up65 (7500 CVD cases prevented each year extrapolating to the eligible English population). That higher estimate reflects the researchers’ apparently unrealistic assumption of 100% screening uptake and 50% overall uptake of treatment. The scenarios We modelled the universal screening scenario to closely resemble the current implementation of the NHS Health Check programme, based on published evidence. Therefore, we maintain that our estimates on the effectiveness of this scenario are not far from the real world effectiveness of NHS Health Checks. However, our output suggesting that universal screening might reduce socioeconomic inequalities seems to contradict existing empirical and modelling evidence.14 19 20 21 This is because we generously assumed identical screening uptake and treatment adherence for all socioeconomic groups. In fact, any potential reduction in socioeconomic health inequalities was essentially eliminated when we considered a small socioeconomic differential in uptake in the sensitivity analysis. Furthermore, additional health inequalities may arise from differential persistence and adherence to treatment by deprivation status.46

ct, any potential reduction in socioeconomic health inequalities was essentially eliminated when we considered a small socioeconomic differential in uptake in the sensitivity analysis. Furthermore, additional health inequalities may arise from differential persistence and adherence to treatment by deprivation status.46 The population-wide intervention scenario on the other hand, is based mostly on structural policies targeting price and availability. This scenario potential effectiveness was mostly based on natural experiments,66 67 and on previous modelling studies from the UK and elsewhere. The size of the changes in the population risk factors that we modelled were modest, and actually smaller than the reductions observed in countries such as France, Finland, and the USA during recent decades.68 69 70 This scenario estimated the reduction in mortality conservatively, because it ignored the beneficial effect of the policies on survival from CVD. Similarly, it underestimated the reduction of the gap in inequalities, because it did not fully consider the current disproportionate burden of poor diet among the most deprived of the population,71 and hence the potential for improvement through population-wide policies.

al effect of the policies on survival from CVD. Similarly, it underestimated the reduction of the gap in inequalities, because it did not fully consider the current disproportionate burden of poor diet among the most deprived of the population,71 and hence the potential for improvement through population-wide policies. Finally, the concentrated screening strategy was the weakest in terms of overall effectiveness, yet more powerful in tackling inequalities. Its increased impact on socioeconomic health inequalities is a direct consequence of the concentrated prevention only to the more deprived quantiles of the population. However, the scenario assumptions may not fully hold in real world implementation. Hence, concentrated screening represents a challenge for public health practitioners and policymakers to exploit the opportunity of a smaller and more homogeneous eligible population and to implement better recruitment and tactics for treatment adherence. Yet, cost effectiveness might also fall because of loss of economies of scale.

rated screening represents a challenge for public health practitioners and policymakers to exploit the opportunity of a smaller and more homogeneous eligible population and to implement better recruitment and tactics for treatment adherence. Yet, cost effectiveness might also fall because of loss of economies of scale. Public health implications This IMPACTNCD modelling may help stakeholders to understand better the interplay between preventive policies, risk factors, disease, and inequalities, and thus potentially inform health policy and strategy. Hence, when compared with the alternative feasible interventions, universal screening seemed inferior both in primary prevention and in reducing socioeconomic health inequalities. Additionally, we estimated that the proportion of young people at high risk aged less than 60 in the eligible population will decrease in future (fig 1). This will render universal screening less effective and less cost effective for this age group, because a larger number will need to be screened to identify each high risk individual.

estimated that the proportion of young people at high risk aged less than 60 in the eligible population will decrease in future (fig 1). This will render universal screening less effective and less cost effective for this age group, because a larger number will need to be screened to identify each high risk individual. Our study suggests that despite the high clustering of risk factors in the most deprived parts of the population, structural population-wide approaches remain more effective than high risk ones for the prevention of CVD. Population-wide approaches also seem to be more effective in reducing absolute and relative socioeconomic health inequalities, generally cost much less than a universal screening programme, and may even be cost saving.72 73 In this study, we did not model the full potential of these policies, as we focused only on diet and smoking interventions; we did not, for example, incorporate alcohol consumption or physical activity. In addition, we did not simulate the likely wider benefits of improved diet and smoking cessation on the plethora of relevant non-communicable diseases. Despite this restricted scope, for CVD prevention we estimated that structural policies targeting diet could be twice as effective as those targeting smoking. Yet, structural interventions for a healthier diet are currently underutilised compared with tobacco control. Several countries have now introduced taxes on sugary drinks or sugar, including Finland, France, Latvia, and Mexico. The UK has recently followed their example. Hungary is the only European country currently taxing unhealthy “junk” food.74 However, fiscal interventions may face opposition from commercial vested interests.75 Interestingly, an increasing body of evidence from empirical studies and modelling analyses suggest that the maximum health impact with a neutral effect on poverty may occur when food or drinks taxes are combined with subsidies for healthy foods.56 76 77

rventions may face opposition from commercial vested interests.75 Interestingly, an increasing body of evidence from empirical studies and modelling analyses suggest that the maximum health impact with a neutral effect on poverty may occur when food or drinks taxes are combined with subsidies for healthy foods.56 76 77 Moreover, the combination of a population-wide intervention with an intervention targeting the most deprived members, may further improve effectiveness and equity. This approach is in the spirit of proportionate universalism that was identified in the Marmot review as the best approach to tackle socioeconomic inequalities in health.78 Our study provides evidence that in CVD prevention proportionate universalism may be the best option not only for tackling inequalities but also for overall effectiveness.

spirit of proportionate universalism that was identified in the Marmot review as the best approach to tackle socioeconomic inequalities in health.78 Our study provides evidence that in CVD prevention proportionate universalism may be the best option not only for tackling inequalities but also for overall effectiveness. Strengths and limitations of this study IMPACTNCD is the first microsimulation model to synthesise core principles of social and CVD epidemiology, vital demographics, published literature, and recent health surveys for England to create a synthetic population of England, including socioeconomic structure, at the individual level. The microsimulation approach allows for the simulation of detailed scenarios and explores the distributional nature of their impact on the population, in a competing risks framework. Microsimulation allows for greater flexibility and more detailed simulation, demanding more statistical and computational resources than older approaches; we utilised the Farr Institute’s statistical high performance computing facilities.79 Many assumptions must be made with such models. Yet, despite the potential frailty of such assumptions, this model validated well against observed CVD mortality, even when multiply stratified. Finally, to ensure transparency, we have made the IMPACTNCD source code open under GNU GPLv3 license.

e computing facilities.79 Many assumptions must be made with such models. Yet, despite the potential frailty of such assumptions, this model validated well against observed CVD mortality, even when multiply stratified. Finally, to ensure transparency, we have made the IMPACTNCD source code open under GNU GPLv3 license. Models are simplifications of reality and thus possess inherent limitations. At least four items were not included in the current model. Firstly, the multiplicative risk assumption is considered the status quo in comparative risk assessments80; however, this may oversimplify the complex nature of interactions between multiple risk factors and disease outcome over the life course. Secondly, IMPACTNCD currently ignores the effect of risk factors on CVD case fatality, although in this study we considered only primary prevention scenarios. Thirdly, complex population dynamics such as migration, social mobility, and the socioeconomic consequences of disease were not modelled. We consider this bias would be relatively small for projections with a short horizon. Fourthly, the model ignores the impact of universal screening in recognising previously undiagnosed cases of atrial fibrillation and other opportunistic diagnoses. Reassuringly, most of these biases apply across all scenarios; their effects would thus be reduced in comparisons between scenarios.

s with a short horizon. Fourthly, the model ignores the impact of universal screening in recognising previously undiagnosed cases of atrial fibrillation and other opportunistic diagnoses. Reassuringly, most of these biases apply across all scenarios; their effects would thus be reduced in comparisons between scenarios. Conclusions When comparing primary prevention strategies for reducing CVD burden and inequalities, universal screening seems less effective than alternative strategies that incorporate population-wide approaches. Further research is needed to identify the best mix of population-wide and risk targeted CVD strategies to maximise cost effectiveness and minimise inequalities. What is already known on this topic Two main strategies for the primary prevention of cardiovascular disease (CVD) is to screen the population, find those individuals at high risk, and treat them or to reduce the CVD risk of the whole population irrespective of individuals’ baseline risk Evidence suggests that the second approach is more effective and likely more equitable, yet this depends on the distribution of CVD risk throughout the population In England, the Department of Health adopted the first approach, although this decision has recently attracted some criticism What this study adds In England, despite the observed higher concentration of CVD risk in more deprived areas, structural population-wide interventions targeting unhealthy diet and tobacco might be three times more effective than the existing screening policy

In England, the Department of Health adopted the first approach, although this decision has recently attracted some criticism What this study adds In England, despite the observed higher concentration of CVD risk in more deprived areas, structural population-wide interventions targeting unhealthy diet and tobacco might be three times more effective than the existing screening policy Structural population-wide interventions are also likely to be more equitable than screening A comprehensive strategy, combining structural population-wide interventions with screening in the most deprived areas (where CVD risk is concentrated) is most likely to maximise both effectiveness and equity of primary CVD prevention Web extra Extra material supplied by authors Appendix: supplementary information Click here for additional data file. We thank Julia Critchley for her valuable advice about modelling physical activity, and the reviewers and editorial committee for their comments and suggestions, which have greatly improved the original manuscript. Contributors: All authors made substantial contributions to the conception and design of the study. CK and MOF had the original idea. CK prepared and conducted data analysis. All authors drafted and critically revised the manuscript. All authors, external and internal, had full access to all of the data (including statistical reports and tables) in the study and can take responsibility for the integrity of the data and the accuracy of the data analysis. CK is the guarantor of this study.

s. All authors drafted and critically revised the manuscript. All authors, external and internal, had full access to all of the data (including statistical reports and tables) in the study and can take responsibility for the integrity of the data and the accuracy of the data analysis. CK is the guarantor of this study. Funding: The health survey for England was funded by the Department of Health until 2004 and the Health and Social Care Information Centre from 2005. IB and CK were supported by Medical Research Council Health eResearch Centre grant MR/K006665/1. SC, MOF, MGC, KA, and PB were supported by the National Institute for Health Research through a grant (SPHR-LIL-PH1-MCD) to the LiLaC collaboration between the University of Liverpool and Lancaster University. The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: All authors have completed the ICMJE uniform disclosure form at http://www.icmje.org/coi_disclosure.pdf and declare: no support from any organisation for the submitted work; no financial relationships with any organisations that might have an interest in the submitted work in the previous three years, no other relationships or activities that could appear to have influenced the submitted work. Ethical approval: Not required as this study is an analysis of previously collected data. Ethical approval for each survey was obtained by the health survey for England team.

Competing interests: All authors have completed the ICMJE uniform disclosure form at http://www.icmje.org/coi_disclosure.pdf and declare: no support from any organisation for the submitted work; no financial relationships with any organisations that might have an interest in the submitted work in the previous three years, no other relationships or activities that could appear to have influenced the submitted work. Ethical approval: Not required as this study is an analysis of previously collected data. Ethical approval for each survey was obtained by the health survey for England team. Data sharing: Anonymised, non-identifiable participant level cross sectional survey data are freely available for academic researchers and public health staff to download from the UK data service (www.ukdataservice.ac.uk). The source code for IMPACTNCD is available at https://github.com/christk/impactncd/tree/cvd-policy-options. Transparency: The lead author (the manuscript’s guarantor) affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned have been explained.

Introduction Cardiovascular disease and cancer remain the two most common causes of death and in 2013 accounted for 25.5 million deaths worldwide.1 Grains are one of the major staple foods consumed around the world and provide 56% of the energy and 50% of the protein intake. They constitute the largest component of recommended daily intake in all dietary guidelines.2 Because of their important role in most diets around the world, interest in the health effects of grain consumption, and in particular whole grains, is increasing.3 4 A high intake of whole grains has been associated with a reduced risk of type 2 diabetes,5 coronary heart disease,6 and obesity.6 Whole grains contain endosperm, germ, and bran, in contrast with refined grains, which have the germ and bran removed during the milling process. Whole grains are good sources of fibre, B vitamins, and some trace minerals such as iron, magnesium, and zinc.7 These nutrients are found in the outer layer of the grains or the bran that function as a protective shell for the germ and endosperm inside. The germ is nourishment for the seed and contains antioxidants, vitamin E, and some B vitamins, while the endosperm provides carbohydrates, protein, and energy.7 Consumption of whole grains differs considerably between populations,8 with the main source being whole grain bread in Scandinavian countries,9 whole grain bread and breakfast cereals in the United States,10 brown rice, unrefined maize and sorghum in some African countries,11 and brown rice in Asia,12 although most of the rice consumed in Asia is white rice.13 14

bly between populations,8 with the main source being whole grain bread in Scandinavian countries,9 whole grain bread and breakfast cereals in the United States,10 brown rice, unrefined maize and sorghum in some African countries,11 and brown rice in Asia,12 although most of the rice consumed in Asia is white rice.13 14 Several previous prospective studies have found a lower risk of coronary heart disease,4 9 15 16 17 stroke,16 18 cardiovascular disease,4 19 20 and all cause mortality4 9 16 20 21 22 associated with a high intake of whole grains, though not all studies reported a clear association.23 24 25 26 27 We have previously reported an inverse association between dietary fibre and whole grain intake and risk of colorectal cancer,28 and a previous review of mostly case-control studies reported a lower risk of several individual cancers, mainly of the digestive system, with higher intake of whole grains,3 but data from cohort studies are limited. Whether whole grain consumption is associated with risk of total cancer is not clear, and clarification of this question would be important from a public health point of view. Epidemiological studies on whole grains and total cancer, however, have reported mixed results, with some studies suggesting a possible inverse association,4 9 22 29 while others have shown no clear association.20 27 Of the cohort studies on whole grains and cardiovascular disease or all cause mortality, some16 20 22 but not all4 15 17 21 29 studies reported a possible plateau effect, with most of the benefit observed at relatively low levels of intake. Although two previous meta-analyses suggested an inverse association between high versus low intake of whole grains and coronary heart disease,6 30 no dose-response analyses were conducted, thus questions remain about the strength and shape of the dose-response relation between whole grains and coronary heart disease and the amount of whole grains that need to be eaten to reduce risk of coronary heart disease and other chronic diseases. Whole grain intake has also been inversely associated with other less common causes of death including deaths from infection,4 20 22 respiratory disease,4 9 20 22 diabetes,9 20 22 and kidney disease20 in some studies, but the available data are limited.

reduce risk of coronary heart disease and other chronic diseases. Whole grain intake has also been inversely associated with other less common causes of death including deaths from infection,4 20 22 respiratory disease,4 9 20 22 diabetes,9 20 22 and kidney disease20 in some studies, but the available data are limited. Despite a growing body of epidemiological evidence for the health benefits of whole grain consumption, dietary recommendations have often been unclear or inconsistent with regard to the amount of whole grains that should be eaten to reduce the risk of chronic disease.

reduce risk of coronary heart disease and other chronic diseases. Whole grain intake has also been inversely associated with other less common causes of death including deaths from infection,4 20 22 respiratory disease,4 9 20 22 diabetes,9 20 22 and kidney disease20 in some studies, but the available data are limited. Despite a growing body of epidemiological evidence for the health benefits of whole grain consumption, dietary recommendations have often been unclear or inconsistent with regard to the amount of whole grains that should be eaten to reduce the risk of chronic disease. For example the World Cancer Research Fund 2007 report recommended that people should “eat relatively unprocessed cereals (grains) and/or pulses with every meal,”31 while in the United Kingdom there is no specific recommendation other than “choosing whole grain, brown or high fibre varieties wherever you can,” but no specific quantities of whole grains were recommended.32 In the US and Canada the recommendation is that “all adults eat at least half their grains as whole grains” so at least three servings of whole grains should be consumed each day,33 while in Scandinavian countries intake of at least 75 g per day of whole grain (dry weight), which equals about 250 g a day (eight servings/day) of whole grain products (fresh weight), is recommended.34 There might be several reasons for the inconsistent dietary guidelines for whole grain intake, including difficulties in measuring intake, differences in the consumption patterns between populations, or lack of data on intake in some populations, but it might also be because most previous meta-analyses considered only selected disease endpoints and did not conduct dose-response analyses.6 30 We found a reduced risk of incidence of type 2 diabetes associated with up to two to three servings a day (60-90 g/day) of whole grain but no further reductions in risk with higher intakes,5 while in a second meta-analysis of whole grain intake and colorectal cancer a linear inverse association was observed with intakes of up to 180 g/day.28 Whether the association is linear or reaches a plateau for other chronic disease outcomes and all cause mortality, or whether only specific types of whole grains are associated with chronic disease and all cause mortality, would be important to clarify to provide more detailed and consistent dietary recommendations with regard to the amount and types of whole grains that should be consumed to reduce the risk of chronic disease and premature mortality. Answering this question would also clarify whether there are additional benefits with high intakes such as those recommended in the Scandinavian guidelines34 and whether such high recommendations are justified.

types of whole grains that should be consumed to reduce the risk of chronic disease and premature mortality. Answering this question would also clarify whether there are additional benefits with high intakes such as those recommended in the Scandinavian guidelines34 and whether such high recommendations are justified. Several large cohort studies including more than 22 000 cases of cardiovascular disease and more than 662 000 participants9 20 22 27 35 have been published since or were missed36 by the previous meta-analyses of whole grains and cardiovascular disease.6 30 To provide a more comprehensive, up to date, and detailed assessment of whole grain intake and several health outcomes we conducted a systematic review and meta-analysis of whole grain consumption in relation to coronary heart disease, stroke, cardiovascular disease, and total cancer and all cause mortality, as well as less common causes of mortality including respiratory disease, infectious disease, diabetes, neurological disease, and all non-cardiovascular, non-cancer causes combined. We aimed to clarify the strength and the shape of the dose-response relation between whole grain intake and these outcomes. We also summarised data on specific types of whole grains as well as on refined grains and total grains. Because of the limited amount of data, however, the main focus of our current analysis is on whole grains.

rify the strength and the shape of the dose-response relation between whole grain intake and these outcomes. We also summarised data on specific types of whole grains as well as on refined grains and total grains. Because of the limited amount of data, however, the main focus of our current analysis is on whole grains. Methods Search strategy and inclusion criteria We searched the PubMed and Embase databases from their inception (1966 and 1947, respectively) to 31 May 2014 and later updated the search to 3 April 2016. Details of the search terms are provided in table S1 in appendix 1. We included prospective studies of grain intake and incidence or mortality from coronary heart disease, stroke, cardiovascular disease, total cancer, and all cause and cause specific mortality if they reported adjusted relative risk estimates and 95% confidence intervals. For the dose-response analyses a quantitative measure of the intake for at least three categories of grain intake or a risk estimate for grain intake on a continuous scale had to be available. We searched the references of the retrieved reports for any additional studies. A list of the excluded studies is provided in table S2 in appendix 1. We followed standard criteria (PRISMA criteria) for reporting meta-analyses.37 The authors of one study22 were contacted for clarification of the amount of whole grain intake, which was reported in ounces/day (1 ounce=28 g) in the publication but was clarified to be in ounces/1000 kcal/day by the authors. In another study,38 the authors were contacted for clarification of the increment for bread which was clarified to be 100 g/d rather than 10 g/d.

cation of the amount of whole grain intake, which was reported in ounces/day (1 ounce=28 g) in the publication but was clarified to be in ounces/1000 kcal/day by the authors. In another study,38 the authors were contacted for clarification of the increment for bread which was clarified to be 100 g/d rather than 10 g/d. Patient involvement No patients were involved in setting the research question or the outcome measures, nor were they involved in developing plans for design, or implementation of the study. No patients were asked to advise on interpretation or writing up of results. There are no plans to disseminate the results of the research to study participants or the relevant patient community. Data extraction From each study we extracted name of first author, publication year, country, name of the study, follow-up period, sample size and number of cases or deaths, type of outcome, sex, age, type of grains, amount or frequency of intake, relative risks and 95% confidence intervals, and variables adjusted for in the analysis. Data were extracted by one author (DA) and checked by another author (DCG) for accuracy.

study, follow-up period, sample size and number of cases or deaths, type of outcome, sex, age, type of grains, amount or frequency of intake, relative risks and 95% confidence intervals, and variables adjusted for in the analysis. Data were extracted by one author (DA) and checked by another author (DCG) for accuracy. Statistical methods We calculated summary relative risks of cardiovascular disease, total cancer, and mortality for the highest versus the lowest level of intake and for 90 g a day increment (three servings/day; the approximate median range across studies) using the random effects model by DerSimonian and Laird,39 which takes into account variation (heterogeneity) both within and between studies. The average of the natural logarithm of the relative risks was estimated and the relative risk from each study was weighted with random effects weights. When studies reported data separately by sex we pooled the relative risks using a fixed effects model before inclusion in the meta-analysis. A two tailed P<0.05 was considered significant.

tural logarithm of the relative risks was estimated and the relative risk from each study was weighted with random effects weights. When studies reported data separately by sex we pooled the relative risks using a fixed effects model before inclusion in the meta-analysis. A two tailed P<0.05 was considered significant. We conducted linear dose-response analyses using the method by Greenland and Longnecker40 to compute study specific slopes (linear trends) and 95% confidence intervals from the natural logarithm of the relative risks across categories of grain intake. For each category of grain intake we used the mean or median if it was reported in the publication and estimated the midpoint of the upper and lower bound for the remaining studies. When extreme categories were open ended or had extreme upper or lower values we used the width of the adjacent interval to calculate an upper or lower cut-off value. For total grains, whole grains, and refined grains we used 30 g as a serving size (one slice of bread or one bowl of breakfast cereal) to recalculate results from studies reporting data in g/day to servings/day, as in our previous analyses.5 28 For intake of pasta we used 150 g as a serving size while for total rice we used 167.25 g as a serving size (cooked weight) based on a weighted average of the serving size for white rice (158 g/day) and brown rice (195 g/day), weighted by the proportion of rice intake of each type (75% white rice and 25% brown rice),41 unless a serving size was specified in the paper. Separate analyses were conducted for studies reporting on total whole grains and specific subtypes of whole grains. We assessed a potential non-linear dose-response relation between grain intake and cardiovascular disease, cancer, and all cause and cause specific mortality using restricted cubic splines with three knots at 10%, 50%, and 90% centiles of the distribution, which were combined using multivariate meta-analysis.42 43 The 95% confidence intervals were derived from the standard errors of the differences in linear predictors between each given point on the dose-response curve and a stated reference value, computed from the covariate values and the covariance matrix of the estimated coefficients.44 A likelihood ratio test was used to assess the difference between the non-linear and linear models to test for non-linearity.45

inear predictors between each given point on the dose-response curve and a stated reference value, computed from the covariate values and the covariance matrix of the estimated coefficients.44 A likelihood ratio test was used to assess the difference between the non-linear and linear models to test for non-linearity.45 Heterogeneity between studies was evaluated with Q and I2 statistics.46 For the Q statistic a P<0.10 was considered to be significant. I2 is the amount of total variation explained by variation between studies. We carried out subgroup and meta-regression analyses stratified by study characteristics (duration of follow-up, sex, geographical location, number of cases, whether the method of dietary assessment had been validated, study quality, and adjustment for confounding factors) to investigate potential sources of heterogeneity. Influence analyses in which we excluded one study at a time from each analysis were conducted to investigate the robustness of the findings. We assessed publication bias with Egger’s test47 and have provided funnel plots in analyses including 10 or more studies. Study quality was assessed with the Newcastle-Ottawa scale, which awards 0-9 points based on the selection, comparability, and outcome assessment.48 We considered studies with 0-3, 4-6, and 7-9 points to represent low, medium, and high quality studies, respectively. Stata version 12.0 software (StataCorp, TX) was used for the analyses.

assessed with the Newcastle-Ottawa scale, which awards 0-9 points based on the selection, comparability, and outcome assessment.48 We considered studies with 0-3, 4-6, and 7-9 points to represent low, medium, and high quality studies, respectively. Stata version 12.0 software (StataCorp, TX) was used for the analyses. Results We included 45 cohort studies (64 publications)4 9 15 16 17 18 19 20 21 22 23 24 25 26 27 29 35 36 41 38 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 in the analyses of grain intake and coronary heart disease, stroke, cardiovascular disease, total cancer, and all cause mortality and other causes of mortality (table S3-S12 in appendix 1, fig 1). Twenty studies were from Europe, 16 were from the US, and nine were from Asia. The studies included in the analyses of whole grains included 7068 cases of coronary heart disease, 2337 cases of stroke, 26 243 cases of cardiovascular disease, 34 346 deaths from cancer, and 100 726 all cause deaths. The number of participants ranged from 245 012 to 705 253. Tables S3-S12 in appendix 1 provide a summary of the study characteristics. Figure 1 shows a flowchart of the study selection. Figures S1-S20 in appendix 2 show the results from the high versus low analyses and scatter plots from the non-linear dose-response analyses. Figures S21-S102 in appendix 2 show results for specific types of grains, refined grains, and total grains. Fig 1 Flow chart of study selection

Results We included 45 cohort studies (64 publications)4 9 15 16 17 18 19 20 21 22 23 24 25 26 27 29 35 36 41 38 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 in the analyses of grain intake and coronary heart disease, stroke, cardiovascular disease, total cancer, and all cause mortality and other causes of mortality (table S3-S12 in appendix 1, fig 1). Twenty studies were from Europe, 16 were from the US, and nine were from Asia. The studies included in the analyses of whole grains included 7068 cases of coronary heart disease, 2337 cases of stroke, 26 243 cases of cardiovascular disease, 34 346 deaths from cancer, and 100 726 all cause deaths. The number of participants ranged from 245 012 to 705 253. Tables S3-S12 in appendix 1 provide a summary of the study characteristics. Figure 1 shows a flowchart of the study selection. Figures S1-S20 in appendix 2 show the results from the high versus low analyses and scatter plots from the non-linear dose-response analyses. Figures S21-S102 in appendix 2 show results for specific types of grains, refined grains, and total grains. Fig 1 Flow chart of study selection Whole grains and coronary heart disease Seven cohort studies4 9 15 16 17 23 63 investigated the association between whole grain intake and risk of coronary heart disease and included a total of 7068 cases and 316 491 participants (one additional publication was included only in the subgroup analysis of incidence of coronary heart disease64 as it overlapped with another publication9). The summary relative risk for high versus low intake was 0.79 (95% confidence interval 0.73 to 0.86; I2=0%, Pheterogeneity=0.63) (fig S1 in appendix 2, table 1). The summary relative risk per 90 g/day was 0.81 (0.75 to 0.87; I2=9%, Pheterogeneity=0.36) (fig 2, table 1). Although the test for non-linearity was significant for the association between whole grain intake and coronary heart disease (P<0.001), with a slightly steeper reduction in risk up to three servings a day than more than three servings a day, there was a clear dose-response relation, and there were further reductions in risk up to 210 g/day (fig 2, fig S2 in appendix 2, table S13 in appendix 1).

n whole grain intake and coronary heart disease (P<0.001), with a slightly steeper reduction in risk up to three servings a day than more than three servings a day, there was a clear dose-response relation, and there were further reductions in risk up to 210 g/day (fig 2, fig S2 in appendix 2, table S13 in appendix 1). Table 1 Intake of total whole grains and effect on coronary heart disease, stroke, cardiovascular disease, total cancer, all cause mortality, and cause specific mortality. Analysis of low versus high intake and dose-response analysis

n whole grain intake and coronary heart disease (P<0.001), with a slightly steeper reduction in risk up to three servings a day than more than three servings a day, there was a clear dose-response relation, and there were further reductions in risk up to 210 g/day (fig 2, fig S2 in appendix 2, table S13 in appendix 1). Table 1 Intake of total whole grains and effect on coronary heart disease, stroke, cardiovascular disease, total cancer, all cause mortality, and cause specific mortality. Analysis of low versus high intake and dose-response analysis High v low analysis Dose-response analysis No of studies RR* (95% CI) I2 P value† Dose (g/day) No of studies RR* (95% CI) I2 P value† Incidence Coronary heart disease 5 0.80 (0.74 to 0.87) 0 0.62 90 5 0.84 (0.77 to 0.92) 34 0.20 Stroke 3 0.86 (0.60 to 1.20) 65 0.06 90 3 0.84 (0.59 to 1.20) 74 0.02 Cardiovascular disease 2 0.89 (0.81 to 0.99) 0 0.40 90 2 0.87 (0.78 to 0.97) 0 0.85 Incidence or mortality Coronary heart disease 6 0.79 (0.73 to 0.86) 0 0.63 90 7 0.81 (0.75 to 0.87) 9 0.36 Stroke 5 0.87 (0.72 to 1.05) 32 0.21 90 6 0.88 (0.75 to 1.03) 56 0.04 Cardiovascular disease 9 0.84 (0.80 to 0.87) 0 0.48 90 10 0.78 (0.73 to 0.85) 40 0.09 Mortality Coronary heart disease 2 0.65 (0.52 to 0.83) 33 0.22 90 3 0.81 (0.74 to 0.89) 10 0.33 Stroke 2 0.85 (0.64 to 1.13) 0 0.99 90 3 0.86 (0.74 to 0.99) 34 0.20 Cardiovascular disease 7 0.81 (0.75 to 0.87) 37 0.15 90 8 0.71 (0.61 to 0.82) 72 0.001 Total cancer 6 0.89 (0.82 to 0.96) 72 0.003 90 6 0.85 (0.80 to 0.91) 37 0.16 All cause 9 0.82 (0.77 to 0.88) 83 <0.001 90 11 0.83 (0.77 to 0.90) 83 <0.001 Respiratory disease 4 0.81 (0.69 to 0.94) 63 0.05 90 4 0.78 (0.70 to 0.87) 0 0.46 Diabetes 4 0.64 (0.42 to 0.98) 64 0.04 90 4 0.49 (0.23 to 1.05) 85 <0.001 Infectious disease 3 0.80 (0.68 to 0.96) 0 0.68 90 3 0.74 (0.56 to 0.96) 0 0.85 Nervous system disease 2 1.13 (0.89 to 1.43) 29 0.24 90 2 1.15 (0.66 to 2.02) 79 0.03 Non-cardiovascular, non-cancer causes 5 0.79 (0.69 to 0.92) 86 <0.001 90 5 0.78 (0.75 to 0.82) 0 0.99 *RR<1 favours those with higher intake

05) 85 <0.001 Infectious disease 3 0.80 (0.68 to 0.96) 0 0.68 90 3 0.74 (0.56 to 0.96) 0 0.85 Nervous system disease 2 1.13 (0.89 to 1.43) 29 0.24 90 2 1.15 (0.66 to 2.02) 79 0.03 Non-cardiovascular, non-cancer causes 5 0.79 (0.69 to 0.92) 86 <0.001 90 5 0.78 (0.75 to 0.82) 0 0.99 *RR<1 favours those with higher intake †P for heterogeneity. Fig 2 Forest plot for consumption of whole grains (per 90 g/day) and risk of coronary heart disease, with graph illustrating non-linear response Subtypes of whole grains including whole grain bread,9 15 29 51 52 60 62 whole grain breakfast cereals,9 15 51 53 and added bran15 17 54 were inversely associated with coronary heart disease, but no association was observed for germ,15 17 refined grains,4 16 58 63 white bread,51 60 refined grain breakfast cereals,51 53 total rice,41 56 59 or total grains,55 61 62 63 while rye was only inversely associated in the high versus low analysis and not in the dose-response analysis (figs S21-S44 in appendix 2, table 2).9 49 57 Table 2 Intakes of subtypes of grains and effect on coronary heart disease, stroke, cardiovascular disease, cancer, and mortality. Analysis of high versus low intake and dose-response analysis

Subtypes of whole grains including whole grain bread,9 15 29 51 52 60 62 whole grain breakfast cereals,9 15 51 53 and added bran15 17 54 were inversely associated with coronary heart disease, but no association was observed for germ,15 17 refined grains,4 16 58 63 white bread,51 60 refined grain breakfast cereals,51 53 total rice,41 56 59 or total grains,55 61 62 63 while rye was only inversely associated in the high versus low analysis and not in the dose-response analysis (figs S21-S44 in appendix 2, table 2).9 49 57 Table 2 Intakes of subtypes of grains and effect on coronary heart disease, stroke, cardiovascular disease, cancer, and mortality. Analysis of high versus low intake and dose-response analysis Type of grain High v low analysis Dose-response analysis No of studies RR* (95% CI) I2 P value† Dose (g/day) No of studies RR* (95% CI) I2 P value† Coronary heart disease Whole grain bread 7 0.83 (0.75 to 0.92) 0 0.64 90 5 0.83 (0.76 to 0.92) 0 0.53 Whole grain breakfast cereals 4 0.72 (0.64 to 0.82) 0 0.92 30 4 0.81 (0.75 to 0.88) 0 0.69 Rye products 2 0.81 (0.70 to 0.94) 0 0.47 30 2 0.97 (0.91 to 1.05) 54 0.14 Added bran 3 0.78 (0.63 to 0.95) 65 0.06 10 2 0.72 (0.58 to 0.89) 34 0.22 Germ 2 0.73 (0.33 to 1.64) 65 0.09 2 2 0.88 (0.76 to 1.03) 0 0.65 Refined grains 4 1.16 (0.84 to 1.59) 48 0.12 90 5 1.13 (0.90 to 1.42) 57 0.05 White bread 2 1.07 (0.86 to 1.34) 50 0.16 90 2 0.96 (0.53 to 1.76) 86 0.007 Refined grain breakfast cereals 2 1.15 (0.79 to 1.67) 70 0.07 30 2 1.14 (0.75 to 1.73) 72 0.06 Total rice 4 0.98 (0.90 to 1.07) 0 0.44 100 4 0.99 (0.95 to 1.03) 7 0.36 Total grains 3 1.07 (0.91 to 1.25) 0 0.47 90 2 1.07 (0.88 to 1.30) 0 0.40 Stroke Whole grain bread 2 0.88 (0.75 to 1.03) 0 0.89 90 1 0.88 (0.72 to 1.07) — — Whole grain breakfast cereals 2 0.99 (0.53 to 1.86) 77 0.04 30 2 1.07 (0.69 to 1.64) 78 0.03 Refined grains 4 0.95 (0.78 to 1.14) 23 0.28 90 5 0.91 (0.81 to 1.02) 29 0.23 Total rice 4 1.02 (0.94 to 1.11) 0 0.95 100 4 1.00 (0.97 to 1.03) 0 0.87 Total grains 4 0.89 (0.79 to 1.00) 6 0.36 90 5 0.93 (0.85 to 1.02) 62 0.03 Cardiovascular disease Whole grain bread 4 0.83 (0.75 to 0.92) 0 0.78 90 3 0.87 (0.80 to 0.95) 0 0.71 Whole grain breakfast cereals 2 0.74 (0.65 to 0.84) 4 0.31 30 2 0.84 (0.78 to 0.90) 0 0.82 Bran 3 0.82 (0.76 to 0.88) 0 0.64 10 2 0.85 (0.79 to 0.90) 0 0.37 Germ 2 1.06 (0.97 to 1.16) 0 0.41 2 2 1.05 (0.96 to 1.15) 0 0.41 Refined grains 2 1.02 (0.91 to 1.14) 16 0.27 90 3 0.98 (0.90 to 1.06) 56 0.11 Total breakfast cereals 2 0.80 (0.70 to 0.90) 55 0.14 30 3 0.80 (0.68 to 0.93) 73 0.03 Total rice 3 0.96 (0.90 to 1.03) 0 0.54 100 3 0.98 (0.95 to 1.00) 0 0.47 Total grains 3 0.94 (0.84 to 1.06) 0 0.47 90 1 0.83 (0.70 to 1.00) — — Total cancer Whole grain bread 3 0.89 (0.78 to 1.01) 42 0.18 90 3 0.91 (0.85 to 0.96) 0 0.63 Brown rice 3 1.07 (0.91 to 1.26) 27 0.26 100 3 0.98 (0.92 to 1.04) 0 0.61 Refined grains 1 0.98 (0.82 to 1.16) — — 90 2

54 100 3 0.98 (0.95 to 1.00) 0 0.47 Total grains 3 0.94 (0.84 to 1.06) 0 0.47 90 1 0.83 (0.70 to 1.00) — — Total cancer Whole grain bread 3 0.89 (0.78 to 1.01) 42 0.18 90 3 0.91 (0.85 to 0.96) 0 0.63 Brown rice 3 1.07 (0.91 to 1.26) 27 0.26 100 3 0.98 (0.92 to 1.04) 0 0.61 Refined grains 1 0.98 (0.82 to 1.16) — — 90 2 0.94 (0.90 to 0.99) 0 0.60 White rice 3 0.87 (0.76 to 1.01) 53 0.12 100 3 0.98 (0.92 to 1.05) 49 0.14 Total breakfast cereals 1 0.90 (0.86 to 0.95) — — 30 2 0.90 (0.82 to 1.00) 36 0.21 Total rice 4 0.95 (0.88 to 1.02) 65 0.03 100 4 0.98 (0.95 to 1.01) 55 0.08 Total grains 1 0.92 (0.80 to 1.06) — — 90 2 0.97 (0.96 to 0.99) 0 0.51 All cause mortality Whole grain bread 5 0.81 (0.74 to 0.88) 57 0.05 90 2 0.85 (0.82 to 0.89) 0 0.36 Whole grain breakfast cereals 3 0.79 (0.72 to 0.86) 50 0.14 30 2 0.87 (0.84 to 0.90) 0 0.85 Oats or oatmeal 3 0.89 (0.76 to 1.04) 90 <0.001 20 1 0.88 (0.83 to 0.92) — — Refined grains 2 1.02 (0.93 to 1.12) 0 0.64 90 4 0.95 (0.91 to 0.99) 20 0.29 Pasta 1 0.61 (0.26 to 1.45) — — 150 2 0.85 (0.74 to 0.99) 54 0.14 Total bread 3 0.77 (0.72 to 0.81) 0 0.42 90 3 0.83 (0.80 to 0.85) 0 0.41 Total breakfast cereals 2 0.87 (0.81 to 0.93) 47 0.17 30 3 0.89 (0.83 to 0.96) 92 <0.001 Total grains 13 0.91 (0.87 to 0.95) 4 0.41 90 7 0.96 (0.90 to 1.02) 71 0.002 *RR<1 favours those with higher intake. †P for heterogeneity.

0.94 (0.90 to 0.99) 0 0.60 White rice 3 0.87 (0.76 to 1.01) 53 0.12 100 3 0.98 (0.92 to 1.05) 49 0.14 Total breakfast cereals 1 0.90 (0.86 to 0.95) — — 30 2 0.90 (0.82 to 1.00) 36 0.21 Total rice 4 0.95 (0.88 to 1.02) 65 0.03 100 4 0.98 (0.95 to 1.01) 55 0.08 Total grains 1 0.92 (0.80 to 1.06) — — 90 2 0.97 (0.96 to 0.99) 0 0.51 All cause mortality Whole grain bread 5 0.81 (0.74 to 0.88) 57 0.05 90 2 0.85 (0.82 to 0.89) 0 0.36 Whole grain breakfast cereals 3 0.79 (0.72 to 0.86) 50 0.14 30 2 0.87 (0.84 to 0.90) 0 0.85 Oats or oatmeal 3 0.89 (0.76 to 1.04) 90 <0.001 20 1 0.88 (0.83 to 0.92) — — Refined grains 2 1.02 (0.93 to 1.12) 0 0.64 90 4 0.95 (0.91 to 0.99) 20 0.29 Pasta 1 0.61 (0.26 to 1.45) — — 150 2 0.85 (0.74 to 0.99) 54 0.14 Total bread 3 0.77 (0.72 to 0.81) 0 0.42 90 3 0.83 (0.80 to 0.85) 0 0.41 Total breakfast cereals 2 0.87 (0.81 to 0.93) 47 0.17 30 3 0.89 (0.83 to 0.96) 92 <0.001 Total grains 13 0.91 (0.87 to 0.95) 4 0.41 90 7 0.96 (0.90 to 1.02) 71 0.002 *RR<1 favours those with higher intake. †P for heterogeneity. Whole grains and stroke Six cohort studies4 9 16 18 24 63 were included in the analysis of whole grain intake and risk of stroke and included 2337 cases and 245 012 participants. The pooled relative risk for high versus low intake was 0.87 (95% confidence interval 0.72 to 1.05; I2=32%, Pheterogeneity=0.21 (fig S3 in appendix 2, table 1). The summary relative risk per 90 g/day was 0.88 (0.75 to 1.03; I2=56%, Pheterogeneity=0.04) (fig 3, table 1). There was evidence of non-linearity between whole grain and risk of stroke (P<0.001), and there was no further reduction in risk above 120-150 g/day (fig 3, fig S4 in appendix 2, table S13 in appendix 1).

e 1). The summary relative risk per 90 g/day was 0.88 (0.75 to 1.03; I2=56%, Pheterogeneity=0.04) (fig 3, table 1). There was evidence of non-linearity between whole grain and risk of stroke (P<0.001), and there was no further reduction in risk above 120-150 g/day (fig 3, fig S4 in appendix 2, table S13 in appendix 1). Fig 3 Forest plot for consumption of whole grains (per 90 g/day) and risk of stroke, with graph illustrating non-linear response We found no clear association between intake of whole grain bread,9 52 whole grain breakfast cereals,9 53 refined grains,4 16 18 24 63 total rice,41 56 59 68 or total grains18 24 61 63 67 69 and risk of stroke (figs S45-S56 in appendix 2, table 2). Whole grains and cardiovascular disease Ten cohort studies (nine publications)4 9 19 20 21 22 26 27 63 investigated whole grain intake and risk of cardiovascular disease and included 26 243 cases and 704 317 participants. The summary relative risk for high versus low intake was 0.84 (95% confidence interval 0.80 to 0.87; I2=0%, Pheterogeneity=0.48) (fig S5 in appendix 2, table 1). The summary relative risk was 0.78 (0.73 to 0.85; I2=40%, Pheterogeneity=0.09) per 90 g/day (fig 4, table 1). There was evidence of a non-linear association between whole grain intake and risk of cardiovascular disease (P<0.001), with a stronger reduction in risk from no intake up to 50 g/day than with higher intakes, but with slight further reductions in risk with intakes up to 200 g/day (fig 4, fig S6 in appendix 2, table S13 in appendix 1).

e of a non-linear association between whole grain intake and risk of cardiovascular disease (P<0.001), with a stronger reduction in risk from no intake up to 50 g/day than with higher intakes, but with slight further reductions in risk with intakes up to 200 g/day (fig 4, fig S6 in appendix 2, table S13 in appendix 1). Fig 4 Forest plot for consumption of whole grains (per 90 g/day) and risk of cardiovascular disease, with graph illustrating non-linear response Intake of whole grain bread,9 29 36 52 62 whole grain breakfast cereals,9 53 total breakfast cereals,36 53 71 and bran,20 54 but not germ,20 refined grains,4 20 35 63 total rice,41 56 59 or total grains,61 62 70 were inversely associated with the risk of cardiovascular disease (figs S57-S72 in appendix 2, table 2).

Fig 4 Forest plot for consumption of whole grains (per 90 g/day) and risk of cardiovascular disease, with graph illustrating non-linear response Intake of whole grain bread,9 29 36 52 62 whole grain breakfast cereals,9 53 total breakfast cereals,36 53 71 and bran,20 54 but not germ,20 refined grains,4 20 35 63 total rice,41 56 59 or total grains,61 62 70 were inversely associated with the risk of cardiovascular disease (figs S57-S72 in appendix 2, table 2). Whole grains and total cancer Six cohort studies (five publications)4 9 20 22 27 were included in the analysis of whole grain intake and risk of total cancer and included 34 346 deaths from cancer among 640 065 participants. The summary relative risk for the high versus the low intake was 0.89 (95% confidence interval 0.82 to 0.96; I2=72%, Pheterogeneity=0.003) (fig S7 in appendix 2, table 1). The summary relative risk per 90 g/day was 0.85 (0.80 to 0.91; I2=37%, Pheterogeneity=0.16) (fig 5, table 1). The heterogeneity seemed to be explained by one large US study,22 and when this was excluded there was no evidence of heterogeneity (I2=0%, P=0.74) and the association remained similar (summary relative risk 0.87, 0.83 to 0.92). There was no evidence of a non-linear association between whole grain intake and total cancer (P=0.15; fig 5 fig S8 in appendix 2, table S13 in appendix 1). Fig 5 Forest plot for consumption of whole grains (per 90 g/day) and risk of total cancer, with graph illustrating non-linear response

Whole grains and total cancer Six cohort studies (five publications)4 9 20 22 27 were included in the analysis of whole grain intake and risk of total cancer and included 34 346 deaths from cancer among 640 065 participants. The summary relative risk for the high versus the low intake was 0.89 (95% confidence interval 0.82 to 0.96; I2=72%, Pheterogeneity=0.003) (fig S7 in appendix 2, table 1). The summary relative risk per 90 g/day was 0.85 (0.80 to 0.91; I2=37%, Pheterogeneity=0.16) (fig 5, table 1). The heterogeneity seemed to be explained by one large US study,22 and when this was excluded there was no evidence of heterogeneity (I2=0%, P=0.74) and the association remained similar (summary relative risk 0.87, 0.83 to 0.92). There was no evidence of a non-linear association between whole grain intake and total cancer (P=0.15; fig 5 fig S8 in appendix 2, table S13 in appendix 1). Fig 5 Forest plot for consumption of whole grains (per 90 g/day) and risk of total cancer, with graph illustrating non-linear response There was an inverse association between intake of whole grain bread9 29 36 52 and total cancer, and there were weak inverse associations between intake of refined grains,4 20 and total grains74 75 and total cancer in the dose-response analysis, but no association was observed between brown rice,92 white rice,92 total breakfast cereals,36 71 and total rice73 92 and total cancer (figs S73-S86 in appendix 2, table 2).

d there were weak inverse associations between intake of refined grains,4 20 and total grains74 75 and total cancer in the dose-response analysis, but no association was observed between brown rice,92 white rice,92 total breakfast cereals,36 71 and total rice73 92 and total cancer (figs S73-S86 in appendix 2, table 2). Whole grains and all cause mortality Eleven cohort studies (10 publications)4 9 16 19 20 21 22 25 27 investigated the association between whole grain intake and all cause mortality and included 100 726 deaths and 705 253 participants. The pooled relative risk for high versus low intake was 0.82 (95% confidence interval 0.77 to 0.88; I2=83%, Pheterogeneity<0.001) (fig S9 in appendix 2, table 1). The summary relative risk was 0.83 (0.77 to 0.90; I2=83%, Pheterogeneity<0.001) per 90 g/day (fig 6, table 1). The heterogeneity was reduced when we excluded two outlying studies21 25 (I2=66%, P=0.003), but the association was not substantially altered (summary relative risk 0.81, 0.76 to 0.86). Although the test for non-linearity was significant (P<0.001), and steeper reductions in risk were observed at lower intakes, there was a clear dose-response relation, and the lowest risk was observed at 225 g/day (fig 6, fig S10 in appendix 2, table S13 in appendix 1). Fig 6 Forest plot for consumption of whole grains (per 90 g/day) and risk of all cause mortality, with graph illustrating non-linear response

Whole grains and all cause mortality Eleven cohort studies (10 publications)4 9 16 19 20 21 22 25 27 investigated the association between whole grain intake and all cause mortality and included 100 726 deaths and 705 253 participants. The pooled relative risk for high versus low intake was 0.82 (95% confidence interval 0.77 to 0.88; I2=83%, Pheterogeneity<0.001) (fig S9 in appendix 2, table 1). The summary relative risk was 0.83 (0.77 to 0.90; I2=83%, Pheterogeneity<0.001) per 90 g/day (fig 6, table 1). The heterogeneity was reduced when we excluded two outlying studies21 25 (I2=66%, P=0.003), but the association was not substantially altered (summary relative risk 0.81, 0.76 to 0.86). Although the test for non-linearity was significant (P<0.001), and steeper reductions in risk were observed at lower intakes, there was a clear dose-response relation, and the lowest risk was observed at 225 g/day (fig 6, fig S10 in appendix 2, table S13 in appendix 1). Fig 6 Forest plot for consumption of whole grains (per 90 g/day) and risk of all cause mortality, with graph illustrating non-linear response Intakes of whole grain bread,9 29 52 62 90 whole grain breakfast cereals,9 53 87 pasta, 38 80 total bread, 38 76 77 and total breakfast cereals,53 71 38 were inversely associated with all cause mortality, and, in addition, total grain consumption61 62 63 65 66 76 82 83 84 85 87 88 89 91 was inversely associated with mortality in the high versus low analysis, but not in the dose-response analysis, while refined grain intake was weakly inversely associated with mortality in the dose-response analysis, but not in the high versus low analysis4 16 20 63 (table 2). There was no association between intake of oats or oatmeal9 86 90 and mortality (figs S87-S102 in appendix 2, table 2).

not in the dose-response analysis, while refined grain intake was weakly inversely associated with mortality in the dose-response analysis, but not in the high versus low analysis4 16 20 63 (table 2). There was no association between intake of oats or oatmeal9 86 90 and mortality (figs S87-S102 in appendix 2, table 2). Whole grains and other causes of death Inverse associations were also observed for the association between whole grains and mortality from respiratory disease (fig 7, figs S11-S12 in appendix 2, table 1, 6617 deaths, 632 849 participants),4 9 20 22 diabetes (fig 8, figs S13-S14 in appendix 2, table 1, 808 deaths, 632 849 participants),4 9 20 22 infectious diseases (fig 9, fig S15-S16 in appendix 2, table 1, 1386 deaths, 512 839 participants),4 20 22 and non-cardiovascular, non-cancer causes (fig 11, figs S19-S20 in appendix 2, table 1, 25 697 deaths, 640 065 participants),4 9 20 22 27 but not for diseases of the nervous system (fig 10, fig S17-S18 in appendix 2, table 1, 2285 deaths, 145 397 participants).4 20 There was evidence of non-linearity in the analyses of mortality from respiratory disease (P=0.001), diabetes (P<0.001), infectious diseases (P=0.003), and diseases of the nervous system (P<0.001), with most of the reduction in risk observed with intakes up to about 60-90 g/day for diabetes and infectious diseases, but with further reductions in risk with higher intakes for respiratory disease mortality (figs 7-10, figs S12, S14, S16, S18, S20 in appendix 2). The analysis of mortality from diseases of the nervous system showed a slight positive association at low intakes, but no association at intakes of 90 g/day, while the association with all non-cardiovascular, non-cancer causes of death showed little evidence of non-linearity (P=0.06; figs 10-11, table S14 in appendix 1).

e analysis of mortality from diseases of the nervous system showed a slight positive association at low intakes, but no association at intakes of 90 g/day, while the association with all non-cardiovascular, non-cancer causes of death showed little evidence of non-linearity (P=0.06; figs 10-11, table S14 in appendix 1). Fig 7 Forest plot for consumption of whole grains (per 90 g/day) and risk of mortality from respiratory disease, with graph illustrating non-linear response Fig 8 Forest plot for consumption of whole grains (per 90 g/day) and risk of mortality from diabetes, with graph illustrating non-linear response Fig 9 Forest plot for consumption of whole grains (per 90 g/day) and risk of mortality from infectious diseases, with graph illustrating non-linear response Fig 11 Forest plot for consumption of whole grains (per 90 g/day) and risk of mortality from non-cardiovascular, non-cancer causes, with graph illustrating non-linear response Fig 10 Forest plot for consumption of whole grains (per 90 g/day) and risk of mortality from diseases of nervous system, with graph illustrating non-linear response

Fig 11 Forest plot for consumption of whole grains (per 90 g/day) and risk of mortality from non-cardiovascular, non-cancer causes, with graph illustrating non-linear response Fig 10 Forest plot for consumption of whole grains (per 90 g/day) and risk of mortality from diseases of nervous system, with graph illustrating non-linear response Publication bias, subgroup and meta-regression analyses, study quality, and influence analyses There was no evidence of small study bias such as publication bias with Egger’s test for coronary heart disease (P=0.11), cardiovascular disease (P=0.31), total cancer (P=0.44), and all cause mortality (P=0.99) (figs S103-S104 in appendix 2), but some indication for stroke (P=0.01). There were, however, few studies in this analysis, and exclusion of one outlying study18 from the analysis attenuated Egger’s test to non-significance (P=0.13) and made the summary estimate significant (0.82, 95% confidence interval 0.72 to 0.93). There was little evidence of heterogeneity between subgroups in subgroup and meta-regression analyses stratified by study characteristics including duration of follow-up, sex, type of outcome, geographical location, number of cases or deaths, or adjustment for confounding factors (table S15-S16 in appendix 1). The association between whole grain intake and coronary heart disease, cardiovascular disease, and total cancer was consistent among both US and European studies, while the association with stroke and all cause mortality was observed only in US studies. For all cause mortality, however, exclusion of one study,25 in which intake of whole grains was extremely low (and which partly explained the heterogeneity), made the summary estimate significant for the European studies as well (summary relative risk 0.82, 0.79 to 0.85; I2=0%, Pheterogeneity=0.36). In the analysis of whole grain intake and stroke there was a significant association among studies with stroke mortality as the outcome, among US studies, and among studies with a validated dietary assessment, however, there was no significant difference between these subgroup analyses (table S15 in appendix 1). There was also little evidence of heterogeneity in the remaining subgroup analyses (tables S15-S16 in appendix 1).

ke mortality as the outcome, among US studies, and among studies with a validated dietary assessment, however, there was no significant difference between these subgroup analyses (table S15 in appendix 1). There was also little evidence of heterogeneity in the remaining subgroup analyses (tables S15-S16 in appendix 1). Mean (median) study quality scores for the studies on whole grains were 7.9 (8.0) for coronary heart disease, 7.7 (8.0) for stroke, 7.7 (8.0) for cardiovascular disease, 7.8 (8.0) for total cancer, and 7.9 (8.0) for mortality, out of a maximum of 9 points (tables S17-S21 in appendix 1). In sensitivity analyses in which we excluded one study at a time from each analysis the summary estimates were not substantially altered for coronary heart disease, cardiovascular disease, total cancer, and all cause mortality, but for stroke there was one study24 that explained the lack of association (figs S105-S109 in appendix 2).

analyses in which we excluded one study at a time from each analysis the summary estimates were not substantially altered for coronary heart disease, cardiovascular disease, total cancer, and all cause mortality, but for stroke there was one study24 that explained the lack of association (figs S105-S109 in appendix 2). Discussion In this dose-response meta-analysis we found an inverse association between whole grain intake and outcomes of several major chronic diseases, including coronary heart disease, stroke, cardiovascular disease overall, total cancer, and all cause mortality as well as less common causes of death such as from respiratory disease, diabetes, infectious disease, and all non-cardiovascular, non-cancer causes. There were reductions of 21%, 16%, 11%, and 18%, respectively, in the relative risk of coronary heart disease, cardiovascular disease, total cancer, and all cause mortality for the highest versus lowest category of whole grain intake. In the dose-response analyses there were reductions of 19%, 22%, 15%, and 17%, respectively, in the relative risk per 90 g/day (one serving equals 30 g), while the association for stroke was significant only in the non-linear dose-response analysis. There were also reductions of 19%, 36%, 20%, and 21% in the relative risk of mortality from respiratory disease, diabetes, infectious disease, and all non-cardiovascular, non-cancer causes, respectively, with a high versus low intake of whole grains. No evidence of an association was observed for mortality from nervous system disorders in the high versus low or linear dose-response analysis. There was some evidence of a slight positive association at low intakes but not at intakes of 90 g/day or more. There was indication of non-linearity in several of the dose-response analyses, with somewhat steeper reductions in risk at lower levels of intake in most of the analyses. There were, however, further reductions in the risk of coronary heart disease and mortality from cancer, respiratory disease, and all non-cardiovascular, non-cancer causes of death as well from all cause mortality up to intakes as high as 210-225 g/day (seven to seven and a half servings a day).

n most of the analyses. There were, however, further reductions in the risk of coronary heart disease and mortality from cancer, respiratory disease, and all non-cardiovascular, non-cancer causes of death as well from all cause mortality up to intakes as high as 210-225 g/day (seven to seven and a half servings a day). Although current dietary guidelines recommend whole grains rather than refined grains, recommendations have often not quantified the amount of whole grain intake that should be consumed,31 thus the current analysis provides a considerable improvement of the evidence base for the level of whole grains that should be consumed to reduce the risk of chronic diseases and mortality. Relatively few people might have three or more servings a day of whole grains. As indicated by the benefits we observed in the non-linear dose-response analyses at an intake of even one or two servings a day in relation to most of the outcomes, however, even moderate increases in whole grain intake could reduce the risk of premature mortality. Also, as many people might have a total grain intake of three or more servings a day, replacement of most or all of the refined grains with whole grains could increase whole grain intake substantially.

of the outcomes, however, even moderate increases in whole grain intake could reduce the risk of premature mortality. Also, as many people might have a total grain intake of three or more servings a day, replacement of most or all of the refined grains with whole grains could increase whole grain intake substantially. Although there was some evidence of non-linear associations between whole grain intake and coronary heart disease, stroke, cardiovascular disease, and mortality from all causes, respiratory disease, diabetes, and infections, with stronger reductions in risk observed at lower levels of intake, in most of the analyses there was a clear dose-response relation with further reductions with intakes up to seven to seven and a half servings a day (210-225 g/day). In addition, there were inverse associations for some subtypes of whole grains or total grains and coronary heart disease (whole grain bread, whole grain breakfast cereals, added bran), cardiovascular disease (whole grain bread, whole grain breakfast cereals, bran, total breakfast cereals), total cancer (whole grain bread, total grains), and all cause mortality (whole grain bread, whole grain cereals, total grains, total bread, pasta), which supports the findings for whole grain intake overall. There was little evidence of an association between intake of refined grains and any of the outcomes. The number of studies in the analyses of grain subtypes, however, was low. Given that whole grain consumption differs substantially between populations, both with regard to type and amount, and because most of the current data are from US and European studies it is possible that effect sizes might differ in other populations.

ber of studies in the analyses of grain subtypes, however, was low. Given that whole grain consumption differs substantially between populations, both with regard to type and amount, and because most of the current data are from US and European studies it is possible that effect sizes might differ in other populations. Limitations of the study Our meta-analysis has some limitations that should be mentioned. There was high heterogeneity in the analysis of whole grains and all cause mortality. With the exception of one study from the Netherlands,25 which had a small range of whole grain intake (the interquartile range was 0-0 and 10.6-13.5 g/d in men/women, respectively), however, the heterogeneity seemed to be more due to differences in the strength of the association between studies than to differences in the direction of the association. Exclusion of two outlying studies21 25 reduced the heterogeneity in the analysis of all cause mortality but did not substantially alter the summary estimates.

he heterogeneity seemed to be more due to differences in the strength of the association between studies than to differences in the direction of the association. Exclusion of two outlying studies21 25 reduced the heterogeneity in the analysis of all cause mortality but did not substantially alter the summary estimates. Although we took into account the different amounts and ranges of whole grain intake between studies in the dose-response analysis, studies could also have differed by the types of whole grains consumed, by how accurately they measured whole grain intake, or by how they defined whole grains. This could have contributed to heterogeneity between studies. In addition, given the diversity of whole grain products available it is difficult to assess intake accurately in epidemiological studies, and some degree of measurement error is inevitable. A recent review recommended reporting intake as the actual amount of whole grain per dry weight.93 As some studies have classified some whole grain items (breakfast cereals, muesli) as whole grain foods if they have a whole grain content of ≥25% or >50% of the weight of the product, then a grain product could be considered whole grain if its whole grain content varied between 25-100 g or 51-100 g per 100 g of the product. Somebody could consume a product with 24 g or 50 g of whole grain per 100 g of the product and still be considered to eat no whole grain, leading to misclassification of the exposure. Most of the studies seemed to report intake as the amount or frequency of whole grain food or product intake (fresh weight including water content), while only two publications9 20 reported intakes in actual amount of whole grain food (dry weight). One study that reported results for both whole grain products (fresh weight) and actual whole grain intake (dry weight) in relation to mortality, however, found similar associations for the two.9 Most of the associations were similar for different types of whole grains, and, in addition, most of the US studies seemed to define whole grains similarly, while few of the European studies provided a definition of whole grain.

ke (dry weight) in relation to mortality, however, found similar associations for the two.9 Most of the associations were similar for different types of whole grains, and, in addition, most of the US studies seemed to define whole grains similarly, while few of the European studies provided a definition of whole grain. People with a high intake of whole grains might have different lifestyles, diets,20 94 or socioeconomic status94 than those with a low intake, thus confounding by other lifestyle factors is a potential source of bias. In subgroup analyses we found that the associations observed persisted among studies that adjusted for smoking, alcohol, physical activity, BMI, and other dietary factors such as sugar sweetened beverages, red meat, and fruit and vegetables. Though differences in socioeconomic factors or deprivation could also have influenced the findings, both the Nurses’ Health Study and the Health Professionals Follow-up Study, cohorts in which there would be relatively little confounding by socioeconomic status or deprivation, found similar results to the overall analysis, and there was no evidence of heterogeneity in the results stratified by adjustment for education.

th the Nurses’ Health Study and the Health Professionals Follow-up Study, cohorts in which there would be relatively little confounding by socioeconomic status or deprivation, found similar results to the overall analysis, and there was no evidence of heterogeneity in the results stratified by adjustment for education. The number of studies that investigated subtypes of whole grains and total or refined grains was limited. Any further studies should therefore try to clarify associations between specific subtypes of grains and cardiovascular disease, cancer, and mortality, as well as less common causes of mortality. As in any meta-analysis of published studies publication bias could have influenced the results. Though we found some indication of small study effects such as publication bias in the analysis of stroke, there was no evidence of publication bias for the remaining outcomes, although the number of studies was moderate and power to detect such bias is low when there are few studies. Strengths of the study Strengths of the current study include the comprehensive analyses of intake of whole grain and subtypes of grain in relation to a range of chronic disease and mortality outcomes including high versus low analyses; linear and non-linear dose-response analyses; the detailed subgroup, sensitivity, and influence analyses; the large numbers of cases or deaths and participants included; and the high quality of the studies included.

ain in relation to a range of chronic disease and mortality outcomes including high versus low analyses; linear and non-linear dose-response analyses; the detailed subgroup, sensitivity, and influence analyses; the large numbers of cases or deaths and participants included; and the high quality of the studies included. Mechanisms Several mechanisms could explain the beneficial effect observed between whole grain intake and coronary heart disease, cardiovascular disease, cancer, and all cause and cause specific mortality. Whole grains are rich in fibre, which can reduce the postprandial glucose and insulin responses leading to better glycaemic control.95 Epidemiological studies have suggested a lower risk of overweight and obesity6 96 97 and of type 2 diabetes5 6 among people with a high whole grain intake. Though both adiposity and type 2 diabetes are established risk factors for cardiovascular disease, cancer, and mortality, in our analysis all the studies adjusted for BMI, suggesting an association independent of BMI. One study on whole grains and coronary heart disease17 and another study on mortality9 found little difference between hazard ratios adjusted or not adjusted for BMI, so if anything BMI might mediate only a small part of the association.

is all the studies adjusted for BMI, suggesting an association independent of BMI. One study on whole grains and coronary heart disease17 and another study on mortality9 found little difference between hazard ratios adjusted or not adjusted for BMI, so if anything BMI might mediate only a small part of the association. Higher whole grain intake has been associated with a lower prevalence or risk of hypertension or raised blood pressure,95 98 99 hypertriglyceridaemia,95 100 and lower concentrations of total and low density lipoprotein cholesterol,97 100 which are important cardiovascular risk factors. Higher fibre intake has been associated with reduced risk of coronary heart disease,101 stroke,102 some cancers,28 103 and all cause mortality.104 105 106 Fibre intake, in particular soluble fibre, might reduce cholesterol concentrations by inhibiting reabsorption of bile acid and by bacterial fermentation of fibre in the colon, which results in the production of short chain fatty acids, which inhibit cholesterol synthesis in the liver.107 Dietary fibre can reduce the risk of cancer by mechanic removal of damaged cells from the digestive tract,108 increasing stool bulk, diluting carcinogens, decreasing transit time, altering the gut microbiota,28 109 110 111 and binding oestrogens in the colon and increasing the faecal excretion of oestrogens, leading to lower oestrogen concentrations.103 112

ncer by mechanic removal of damaged cells from the digestive tract,108 increasing stool bulk, diluting carcinogens, decreasing transit time, altering the gut microbiota,28 109 110 111 and binding oestrogens in the colon and increasing the faecal excretion of oestrogens, leading to lower oestrogen concentrations.103 112 Whole grain consumption has been found to be inversely associated with mortality from inflammatory diseases,4 and an intervention study found reduced concentrations of fasting serum glucose, measures of lipid peroxidation, and homocysteine concentrations among participants fed a whole grain/legume powder supplement.113 Whole grain intake has been associated with lower levels of inflammatory markers (PAI-1, CRP)114 115 116 117 and liver enzymes (GGT, ASAT),117 higher levels of which have been associated with increased risk of cardiovascular disease, cancer, and mortality.118 119 120 Whole grain intake has also been associated with higher levels of adiponectin,116 which increases insulin sensitivity and reduces inflammation. Whole grains also contain several other potentially beneficial components that could explain some of the current findings.121

ascular disease, cancer, and mortality.118 119 120 Whole grain intake has also been associated with higher levels of adiponectin,116 which increases insulin sensitivity and reduces inflammation. Whole grains also contain several other potentially beneficial components that could explain some of the current findings.121 Further studies are needed to clarify whether there is an underlying mechanism for the non-linear association between whole grain intake and cardiovascular disease, all cause mortality, and mortality from respiratory disease, diabetes, and infectious diseases. A high intake of whole grains could also reduce the risk of chronic disease and mortality indirectly, by displacement of unhealthy foods or drinks. The association for cardiovascular disease and mortality, however, persisted in studies that adjusted for intake of red and processed meat and sugar sweetened beverages.

s. A high intake of whole grains could also reduce the risk of chronic disease and mortality indirectly, by displacement of unhealthy foods or drinks. The association for cardiovascular disease and mortality, however, persisted in studies that adjusted for intake of red and processed meat and sugar sweetened beverages. Policy implications and future research We found that a high whole grain intake was associated with reduced risk of coronary heart disease, cardiovascular disease, total cancer, and all cause mortality as well as mortality from respiratory disease, infections, diabetes, and all non-cardiovascular, non-cancer causes combined. This is strong evidence that a high intake of whole grains is beneficial for several health outcomes in a dose-response manner. In addition, a high intake of whole grains has previously been associated with reduced risk of colorectal cancer,28 type 2 diabetes,5 and overweight or obesity.6 Altogether these findings have important public health implications as whole grain intake can be modified relatively easily by replacing refined grains and could have a large effect on the burden of chronic disease if adopted in the general population. As shown in the current meta-analysis, a high intake of whole grain is not only associated with reduced risk of cardiovascular disease and diabetes but also with mortality from cancer, respiratory disease, infectious disease, and all non-cancer, non-cardiovascular causes combined. The current findings therefore strongly support existing dietary recommendations to increase whole grain consumption in the general population. From a practical angle a whole grain product intake of 90 g/day can be achieved, for example, by eating a portion of whole grain breakfast cereals (30-40 g) at breakfast and a piece of whole grain pita bread for dinner (60 g). The non-linear analyses suggested that the reduction in risk of mortality is steepest at the lowest level of whole grain intake (people who increase from no intake of whole grain to two servings/day) and that perhaps targeting subjects with a very low intake might have a greater impact. Further reductions were observed up to 210-225 g/day (seven to seven and a half servings a day), however, suggesting further benefits with even higher intakes. Most of the studies included in the analyses of whole grains were from the US, and only a few European studies have been published so far.

ave a greater impact. Further reductions were observed up to 210-225 g/day (seven to seven and a half servings a day), however, suggesting further benefits with even higher intakes. Most of the studies included in the analyses of whole grains were from the US, and only a few European studies have been published so far. Whole grain intake is higher in Northern Europe34 122 than in the US, and populations in Northern Europe might therefore be promising for further studies of the association between whole grains and health outcomes, both in terms of examining more extreme intakes and specific types of whole grains. Further studies are needed in other geographical locations, as are studies of specific diseases and less common causes of death and that incorporate biomarkers of whole grain intake.123 In conclusion our results provide further evidence for the beneficial effects of diets high in whole grains on the risk of coronary heart disease, cardiovascular disease, total cancer, and all cause mortality, as well as mortality from respiratory disease, infections, diabetes, and all non-cardiovascular, non-cancer causes combined. Reductions in risk are observed up to 210-225 g/day or seven to seven and a half servings a day, and the current findings support dietary recommendations to increase intake of whole grains and as much as possible to choose whole grains rather than refined grains. What is already known on this topic A high intake of whole grains has been associated with a lower risk of type 2 diabetes, cardiovascular disease, and weight gain

In conclusion our results provide further evidence for the beneficial effects of diets high in whole grains on the risk of coronary heart disease, cardiovascular disease, total cancer, and all cause mortality, as well as mortality from respiratory disease, infections, diabetes, and all non-cardiovascular, non-cancer causes combined. Reductions in risk are observed up to 210-225 g/day or seven to seven and a half servings a day, and the current findings support dietary recommendations to increase intake of whole grains and as much as possible to choose whole grains rather than refined grains. What is already known on this topic A high intake of whole grains has been associated with a lower risk of type 2 diabetes, cardiovascular disease, and weight gain Recommendations for whole grain intake have often been unclear or inconsistent with regard to the amount and types of whole grain foods that should be consumed to reduce chronic disease and risk of mortality What this study adds A high intake of whole grains was associated with reduced risk of coronary heart disease, cardiovascular disease, total cancer, and all cause mortality, as well as mortality from respiratory disease, infectious disease, diabetes, and all non-cardiovascular, non-cancer causes Reductions in risk were observed up to an intake of 210-225 g/day (seven to seven and a half servings/day) and for whole grain bread, whole grain breakfast cereals, and added bran

What this study adds A high intake of whole grains was associated with reduced risk of coronary heart disease, cardiovascular disease, total cancer, and all cause mortality, as well as mortality from respiratory disease, infectious disease, diabetes, and all non-cardiovascular, non-cancer causes Reductions in risk were observed up to an intake of 210-225 g/day (seven to seven and a half servings/day) and for whole grain bread, whole grain breakfast cereals, and added bran The results strongly support dietary recommendations to increase intake of whole grain foods in the general population to reduce risk of chronic diseases and premature mortality Web Extra Extra material supplied by the author Appendix 1: Supplementary tables S1-S21 Click here for additional data file. Appendix 2: Supplementary figures S1-S109 Click here for additional data file. We thank Tao Huang and Lu Qi (Department of Nutrition, Harvard T Chan School of Public Health) for clarification of the data from the NIH-AARP Diet and Health Study, and Diewertje Sluik (Division of Human Nutrition, Wageningen University) for clarification of the data from the European Prospective Investigation into Cancer and Nutrition study.

Qi (Department of Nutrition, Harvard T Chan School of Public Health) for clarification of the data from the NIH-AARP Diet and Health Study, and Diewertje Sluik (Division of Human Nutrition, Wageningen University) for clarification of the data from the European Prospective Investigation into Cancer and Nutrition study. Contributors: DA and TN conceived and designed the study. DA, NK, EG, LTF, PB, TN, DCG, ER, and ST acquired, analysed, and interpreted the data. DCG checked data extractions. DA drafted the manuscript, which was critically revised for important intellectual content by all authors. DA and DCG carried out the statistical analysis. DA, LJV, ST, and ER obtained funding. TN supervised the study. All authors have read and approved the final manuscript. DA is guarantor and had full access to all the data and takes responsibility for the integrity of the data and the accuracy of the data analysis. Funding: This project was funded by Olav og Gerd Meidel Raagholt’s Stiftelse for Medisinsk Forskning, the liaison committee between the Central Norway Regional Health Authority (RHA) and the Norwegian University of Science and Technology (NTNU), and the Imperial College National Institute of Health Research (NIHR) Biomedical Research Centre (BRC). The funders had no role in the study design, data collection, data analysis and interpretation, writing of the report, or the decision to submit the article for publication.

iversity of Science and Technology (NTNU), and the Imperial College National Institute of Health Research (NIHR) Biomedical Research Centre (BRC). The funders had no role in the study design, data collection, data analysis and interpretation, writing of the report, or the decision to submit the article for publication. Competing interests: All authors have completed the ICMJE uniform disclosure form at http://www.icmje.org/coi_disclosure.pdf and declare: no support from any organisation for the submitted work; no financial relationships with any organisation that might have an interest in the submitted work in the previous three years; no other relationships or activities that could appear to have influenced the submitted work. Ethical approval: Not required. Transparency: The lead author affirms that the manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned have been disclosed. Data sharing: No additional data available.

and proportion (%) of RTI consultations with antibiotics prescribed. Rates and proportions were standardised for age and sex using the 2013 European standard population. After excluding practices with insufficient data, because of short periods of contributing to CPRD, we estimated rates for 610 CPRD general practices. Statistical analysis In the final stage of the analysis, we estimated the numbers of infective complications with person years at risk, in relation to general practice specific rates of RTI consultations and antibiotic prescribing. Mixed effects Poisson models were fitted using the hglm package24 in the R program.25 General practice was fitted as a random effect. The log of person years was included as offset. Fixed effects included sex, year, age group, region, and deprivation fifth. We evaluated the association of age standardised RTI consultation rate with rates of infective complications. After adjusting for the RTI consultation rate, we evaluated the association of the antibiotic prescribing rate and the antibiotic prescribing proportion with infective complication rates. Incident rate ratios (95% confidence intervals) were estimated for each fourth of RTI consultation rate, antibiotic prescribing rate, or antibiotic prescribing proportion, using the lowest fourth for reference. The RTI consultation rate, antibiotic prescribing rate, and antibiotic prescribing proportion were also fitted as continuous predictors, and we estimated incident rate ratios for each 10 unit change in the predictor. We evaluated whether the addition of quadratic terms improved goodness of fit. As there were small numbers of events for intracranial abscess, and mixed effects models did not converge, we omitted the random effect for general practice for this outcome. Regression models were not fitted for Lemierre’s syndrome because this condition was rare. The ggplot226 and forestplot27 packages in R were used to present the results.

Introduction Statins reduce the risk of cardiovascular disease1 and are widely recommended as part of the strategy for primary and secondary prevention.2 3 4 5 6 Severe adverse effects associated with statins are extremely rare,7 but concerns over purportedly high rates of side effects such as muscle pain and weakness have been raised in the academic press and reported in the national media. In October 2013, two articles published in The BMJ were perceived as critical of statins, with one suggesting that side effects might outweigh the overall health benefits in patients at low and intermediate risk.8 9 Although the comments on rates of side effects were based on evidence from non-blinded observational data, and the articles were focused on the benefit:risk ratio in those with low risk, they generated extensive and broader discussion in the media about statins. The debate peaked in March 2014, when most national media outlets in the United Kingdom covered the subject.10 11 Media coverage of these articles was probably intensified because of the impending changes in the guidelines proposed by the UK National Institute for Health and Care Excellence (NICE, July 2014), which broadened eligibility for statins from patients with a high (≥20%) 10 year risk of cardiovascular disease to those with intermediate (≥10%) 10 year risk.

probably intensified because of the impending changes in the guidelines proposed by the UK National Institute for Health and Care Excellence (NICE, July 2014), which broadened eligibility for statins from patients with a high (≥20%) 10 year risk of cardiovascular disease to those with intermediate (≥10%) 10 year risk. As a society we are increasingly exposed to numerous and disparate sources of health information, and it has been shown that this bombardment leads to a lack of clarity about which sources patients and others should trust.12 Studies in Denmark, Australia, Turkey, and France have suggested that media debate about side effects of statins has led to measurable effects on certain aspects of use,13 14 15 16 and qualitative work has shown that concerns over side effects and a desire for clearer information regarding the risks and benefits can affect use in patients in the UK.17 18 19 No large studies to date, however, have comprehensively evaluated the effects of media debates about treatment with statins on prescribing for both the primary and secondary prevention of cardiovascular disease and the consequences for public health.

the risks and benefits can affect use in patients in the UK.17 18 19 No large studies to date, however, have comprehensively evaluated the effects of media debates about treatment with statins on prescribing for both the primary and secondary prevention of cardiovascular disease and the consequences for public health. Using prescribing data from routinely collected UK primary care records, we quantified the potential association between the debate in the media about the side effects of statins and initiation and cessation of treatment in UK primary care for both primary and secondary prevention of cardiovascular disease. We investigated whether any potential media effects differed by key patient level characteristics and estimated the public health impact of any changes in patterns of use that might have arisen from the controversy, in particular the resulting number of excess cardiovascular events.

revention of cardiovascular disease. We investigated whether any potential media effects differed by key patient level characteristics and estimated the public health impact of any changes in patterns of use that might have arisen from the controversy, in particular the resulting number of excess cardiovascular events. Methods Study design and setting This ecological interrupted time series study used prospectively collected data from the UK Clinical Practice Research Datalink (CPRD), a primary care database containing anonymised data from about 6.9% of the UK population.20 21 General practitioners play a key role in the UK healthcare system as they are responsible for primary healthcare and specialist referrals. The CPRD includes prescriptions and clinical diagnoses from primary care, as well as diagnoses from secondary care that are typically fed back to general practitioners. Those represented in the database are broadly representative of the UK population in terms of age and sex.21 In the interrupted time series design, population level outcomes (in this case, proportions of people starting/stopping statins) are calculated over time, and then statistical regression techniques are used to investigate how trends in these outcomes are affected by a population level exposure that occurs in a single well defined time period (here, widespread media coverage about statins over a six month period)—that is, the exposure is viewed as a potential “interruption” to the underlying trends in the outcome(s) over time.22

estigate how trends in these outcomes are affected by a population level exposure that occurs in a single well defined time period (here, widespread media coverage about statins over a six month period)—that is, the exposure is viewed as a potential “interruption” to the underlying trends in the outcome(s) over time.22 We produced a code list for statins by identifying all drugs that included the word “statin” in either the product name or the drug substance name. The proportions of patients initiating and stopping statins were calculated for each month from January 2011 to March 2015. Definitions for our study populations and ascertainment of statin initiation and cessation are described below.

at included the word “statin” in either the product name or the drug substance name. The proportions of patients initiating and stopping statins were calculated for each month from January 2011 to March 2015. Definitions for our study populations and ascertainment of statin initiation and cessation are described below. Statin initiation Study population For each calendar month, we identified all individuals aged >40 registered at their general practice for at least a year, with no previous recorded prescriptions for a statin, and no previous cardiovascular disease events, and who had either a newly recorded 10 year cardiovascular risk score (hereafter referred to as simply 10 year risk score) of >20% (appendix part 1) (that is, eligible to start taking a statin for primary prevention of cardiovascular disease) or an incident cardiovascular event (that is, eligible to start taking a statin for secondary prevention). Incident events were defined as a first record of coronary heart disease (myocardial infarction, angina, revascularisation procedures), cerebrovascular disease (stroke, transient ischaemic attack), or peripheral vascular disease (abdominal aortic aneurism and intermittent claudication) dated at least a year after the start of a patient’s CPRD follow-up. These events were identified within patient records by searching for NHS Read codes corresponding to any of these diagnoses; we used code lists developed for the CALIBER programme.23

ar disease (abdominal aortic aneurism and intermittent claudication) dated at least a year after the start of a patient’s CPRD follow-up. These events were identified within patient records by searching for NHS Read codes corresponding to any of these diagnoses; we used code lists developed for the CALIBER programme.23 Defining initiation We then calculated the proportion of patients who started taking statins for primary and secondary prevention separately for each calendar month throughout the study period. Patients were defined as starting a statin for primary prevention if they had received a first prescription within 28 days of the date of the risk score being recorded. The denominator was all patients eligible to start treatment for primary prevention (as above) who remained alive, under follow-up, and free from cardiovascular disease for the full 28 day period after their risk score. Patients were defined as starting a statin for secondary prevention if they received a first prescription in primary care ≤60 days after their first cardiovascular event; the 60 day grace period was chosen to allow for a period of admission to hospital, based on preliminary analyses (appendix part 2). The denominator for this calculation was all those eligible to start treatment for secondary prevention (as above) who remained alive and under follow-up for the full 60 day grace period.

the 60 day grace period was chosen to allow for a period of admission to hospital, based on preliminary analyses (appendix part 2). The denominator for this calculation was all those eligible to start treatment for secondary prevention (as above) who remained alive and under follow-up for the full 60 day grace period. Statin cessation Study population For each calendar month, we identified all individuals aged >40 and in receipt of a statin prescription that ended within that calendar month. Prescription end dates were calculated based on the date of prescription and quantity of tablets prescribed (appendix part 3). Patients could be included in more than one monthly cohort if they had multiple prescriptions ending during the study period. The study population was then stratified into those taking statins for primary prevention (defined as those with no record of a previous cardiovascular event) and those taking statins for secondary prevention (those with any previous event).

thly cohort if they had multiple prescriptions ending during the study period. The study population was then stratified into those taking statins for primary prevention (defined as those with no record of a previous cardiovascular event) and those taking statins for secondary prevention (those with any previous event). Defining cessation We calculated the proportion of patients who stopped their statins each month: stopping was defined as receiving no further prescription within 28 days of the end date of the previous prescription. This 28 day grace period allowed time for patients with previously overlapping prescriptions to use their excess tablets, based on a preliminary analysis in which we identified all prescriptions from January 2011 to October 2013 and calculated that 90% of prescriptions were followed up with a new prescription within 28 days of the initial prescription ending. Only those remaining alive, under follow-up, and free from cardiovascular disease (for the primary prevention analysis) for the full 28 day grace period were included in the denominator.

and calculated that 90% of prescriptions were followed up with a new prescription within 28 days of the initial prescription ending. Only those remaining alive, under follow-up, and free from cardiovascular disease (for the primary prevention analysis) for the full 28 day grace period were included in the denominator. Period of exposure to high media coverage We defined an exposure period of October 2013 to March 2014, and we compared patterns of statin initiation and cessation before and after this time period. The start date of the exposure was chosen to coincide with the publication of The BMJ papers about statins.8 9 The end date was determined by carrying out a Google trend analytics search for the term “statin side effects” in the UK, which tracks the popularity of this search term over time, and by taking the date of peak searching for this term after October 2013 as the end of the exposure period (appendix part 4). This led to the choice of March 2014, which coincided with a spell of widespread coverage of the debate over statin side effects across most major national media outlets in the UK.

m over time, and by taking the date of peak searching for this term after October 2013 as the end of the exposure period (appendix part 4). This led to the choice of March 2014, which coincided with a spell of widespread coverage of the debate over statin side effects across most major national media outlets in the UK. Statistical analysis We carried out an interrupted time series analysis using a generalised linear model with a binomial error structure, which accounts for the month by month variation in denominators (that is, the number of people eligible to initiate or stop taking a statin each month).22 Seasonal effects were accounted for by adjustment for calendar month,24 and first order lagged residuals were included to account for autocorrelation.25 Standard errors were scaled to account for overdispersion.26 Time was divided into three segments: before, during, and after the exposure period of high media coverage.

easonal effects were accounted for by adjustment for calendar month,24 and first order lagged residuals were included to account for autocorrelation.25 Standard errors were scaled to account for overdispersion.26 Time was divided into three segments: before, during, and after the exposure period of high media coverage. Within this modelling framework, we conducted separate analyses to investigate changes in statin initiation for primary prevention; statin initiation for secondary prevention; statin cessation in primary prevention; and statin cessation in secondary prevention. In each analysis, we looked at whether there was a step change in the log odds of initiating or stopping a statin after the exposure period compared with before, assuming an underlying linear month on month trend throughout the study period. We also investigated whether the underlying trend over time in the log odds of initiating/stopping a statin changed after the exposure period, compared with before. A Wald test was used to compare the trends (log odds ratio per month) before and after exposure. Linear predictions of the log odds and 95% confidence intervals of an event were calculated from the models and converted into probabilities, which we plotted along with a scatter of the raw proportion of patients that initiated/stopped treatment with a statin. We did not directly estimate trends during the exposure period itself because in this period individual level exposure to the coverage would probably have been dynamically changing, and the small number of data points in this period would have led to imprecise and inconclusive estimates.

topped treatment with a statin. We did not directly estimate trends during the exposure period itself because in this period individual level exposure to the coverage would probably have been dynamically changing, and the small number of data points in this period would have led to imprecise and inconclusive estimates. We then investigated effect modification by the following prespecified factors: age group (40-49, 50-59, 60-69, 70-79, ≥80), sex, diabetes (identified with Read codes), and length of previous continuous prescription (up to six months, six months to one year, one to two years, two to four years, longer than four years). When stratifying by length of previous continuous prescription, we restricted the analysis to patients whose most recent continuous prescription began at least a year after their current registration into CPRD, to ensure that the initiation date was accurate and duration could be reliably categorised. By re-fitting the models with the monthly counts stratified by the above factors, and an interaction term added between the potential effect modifier and the parameter representing the post-exposure change in cessation, we examined each potential effect modifier in a separate model and generated P values with likelihood ratio tests. For ordinal covariates, we further explored effect modification by testing for linear trend.

dded between the potential effect modifier and the parameter representing the post-exposure change in cessation, we examined each potential effect modifier in a separate model and generated P values with likelihood ratio tests. For ordinal covariates, we further explored effect modification by testing for linear trend. Post hoc analyses After observing an increase in statin cessation that seemed to be transient in our primary analyses, we then separated post-exposure time into two periods of six months, and tested whether there was a difference between the modelled levels of cessation in each post-exposure period versus the pre-exposure period. We also carried out analyses investigating both a step change and trend change simultaneously after the exposure period, which allowed us to determine the monthly rate at which cessation fell after the initial level increase. To examine changes in the application of risk scoring by GPs over time, we calculated the monthly proportion of patients in the whole of CPRD with any recorded 10 year risk score for cardiovascular disease, as well as the proportions of patients with very high (≥30%), high (20-30%), intermediate (10-20%), or low (<10%) risk scores recorded; these data were then analysed with similar methods as for the main analysis.

roportion of patients in the whole of CPRD with any recorded 10 year risk score for cardiovascular disease, as well as the proportions of patients with very high (≥30%), high (20-30%), intermediate (10-20%), or low (<10%) risk scores recorded; these data were then analysed with similar methods as for the main analysis. Negative control analysis In a preplanned negative control analysis, we replicated the main analyses using drugs prescribed for glaucoma (appendix part 5). Like statins, these drugs are given to those at high risk of disease as a preventive measure, are prescribed with similar frequency, and are typically intended to be continued for life after initiation. The purpose of this analysis was to ensure that any changes in statin use after the exposure period were not explained by broader unrelated underlying trends in prescriptions coincident with our exposure period or by other biases arising from the methods. In a second negative control analysis, we re-ran our main analyses using an alternative exposure period of 12 months earlier as we had no reason to expect any changes in prescribing trends around this time.

derlying trends in prescriptions coincident with our exposure period or by other biases arising from the methods. In a second negative control analysis, we re-ran our main analyses using an alternative exposure period of 12 months earlier as we had no reason to expect any changes in prescribing trends around this time. Public health impact To estimate the potential public health impact of changes in statin cessation, we compared the modelled cessation level in the first six months after the high media coverage exposure period (see “post hoc analyses” above) with the expected cessation level in the same period had there been no changes after exposure (that is, simply projecting the modelled “before” trend line for cessation forward). This estimated the number of patients who might have stopped taking a statin because of the controversy reported in the media, under the assumption of causality. We then estimated the number of excess cardiovascular events among these patients, assuming an average 10 year risk of 20% among those stopping, and assuming that statins would reduce risk by 19%, based on statin efficacy estimates from the Cholesterol Treatment Trialists’ Collaboration.27 We also used historical data from CPRD to estimate and take account of the proportion of patients who would have stopped taking statins or died in the following 10 years regardless of the media controversy. On the basis of results from a published study,28 we assumed that 66% of patients who stopped statins would restart within the following 12 months with no loss of protection. In a second calculation, to obtain an upper bound on the impact, we made the more pessimistic assumption that all those who stopped taking statins did not ever take them again. Full details of these calculations are given in appendix part 6.

atins would restart within the following 12 months with no loss of protection. In a second calculation, to obtain an upper bound on the impact, we made the more pessimistic assumption that all those who stopped taking statins did not ever take them again. Full details of these calculations are given in appendix part 6. All data analyses were carried out in Stata version 14, and all code lists are available at https://clinicalcodes.rss.mhs.man.ac.uk/medcodes/article/46/. Patient involvement No patients were involved in setting the research question or the outcome measures, nor were they involved in developing plans for design or implementation of the study. No patients were asked to advise on interpretation or writing up of results. There are no plans to disseminate the results of the research to study participants or the relevant patient community Results Throughout the study period we identified 88 010 records of a 10 year risk score of ≥20% for cardiovascular disease, 28 593 incident cardiovascular events, 9 286 148 prescriptions of statins for primary prevention, and 5 130 148 prescriptions of statins for secondary prevention. Table 1 provides an overview of study populations. Table 1 Characteristics of study populations from CPRD according initiation or cessation of statins for primary and secondary prevention of cardiovascular disease. Figures are numbers (percentage) unless stated otherwise

ers of events for intracranial abscess, and mixed effects models did not converge, we omitted the random effect for general practice for this outcome. Regression models were not fitted for Lemierre’s syndrome because this condition was rare. The ggplot226 and forestplot27 packages in R were used to present the results. To present the clinical implications of these findings, we calculated the number of events expected in a general practice with 7000 patients (the general practice mean list size for England) during 10 years of follow-up. To estimate the expected number of consultations for RTIs we used the median (95% range) for the RTI consultation rate. To estimate expected numbers of complications and antibiotic prescriptions we used the disease incidence and distribution of antibiotic prescribing proportion for the highest prescribing fourth. We used the relative risk increase for a 10% change in antibiotic prescribing from the Poisson model to estimate the expected change in number of infective complications. Patient involvement No patients were involved in setting the research question or the outcome measures, nor were they involved in developing plans for design or implementation of the study. No patients were asked to advise on interpretation or writing up of results. Results will be disseminated to relevant patient communities through news media.

Results Throughout the study period we identified 88 010 records of a 10 year risk score of ≥20% for cardiovascular disease, 28 593 incident cardiovascular events, 9 286 148 prescriptions of statins for primary prevention, and 5 130 148 prescriptions of statins for secondary prevention. Table 1 provides an overview of study populations. Table 1 Characteristics of study populations from CPRD according initiation or cessation of statins for primary and secondary prevention of cardiovascular disease. Figures are numbers (percentage) unless stated otherwise Study populations Primary initiation Secondary initiation Primary cessation Secondary cessation No of events* 88 010 28 593 9 286 148 5 130 148 No of outcomes† 20 249 17 207 751 243 328 595 No of patients 70 409 28 593 457 073 230 610 Men 48 136 (68.4) 16 512 (57.8) 237 802 (52.0) 137 776 (59.7) Women 22 237 (31.6) 12 081 (42.2) 218 271 (48.0) 92 833 (40.3) Indeterminate 0 (0) 0 (0) 1 (0) 1 (0) Age (years)‡: 40-49 2502 (3.6) 2396 (8.4) 39 806 (8.7) 7534 (3.3) 50-59 12 537 (17.9) 5377 (18.8) 95 485 (20.9) 26 533 (11.5) 60-69 30 491 (43.3) 7152 (25.0) 153 395 (33.6) 57 405 (24.9) 70-79 21 939 (31.2) 6306 (22.1) 116 008 (25.4) 73 013 (31.7) ≥80 2904 (4.1) 27 362 (5.8) 52 379 (11.5) 66 125 (28.7) Median (IQR) 66 (61-72) 69 (58-80) 66 (58-74) 73 (64-81) Diabetes§ 5644 (8.0) 2466 (8.62) 114 910 (31.7) 63 868 (27.7) IQR=interquartile range.

0-69 30 491 (43.3) 7152 (25.0) 153 395 (33.6) 57 405 (24.9) 70-79 21 939 (31.2) 6306 (22.1) 116 008 (25.4) 73 013 (31.7) ≥80 2904 (4.1) 27 362 (5.8) 52 379 (11.5) 66 125 (28.7) Median (IQR) 66 (61-72) 69 (58-80) 66 (58-74) 73 (64-81) Diabetes§ 5644 (8.0) 2466 (8.62) 114 910 (31.7) 63 868 (27.7) IQR=interquartile range. *No of opportunities for patients to either initiate or stop statins, either risk score >20% or incident cardiovascular event in initiation populations, or end of statin prescription in cessation populations. †No of occurrences of initiation in initiation populations and No of occurrences of stopping in cessation populations. ‡Age at first risk score >20% or incident cardiovascular event in initiation populations and age at first prescription in study period in cessation populations. §Diagnosis of diabetes before or within study period. Figure 1 shows the estimated step change in the likelihood of patients initiating and stopping taking a statin for primary and secondary prevention after the exposure period, over and above the underlying time trend. Figures showing the estimated change in the month-on-month trends in statin initiation/cessation are in appendix part 7.

e estimated step change in the likelihood of patients initiating and stopping taking a statin for primary and secondary prevention after the exposure period, over and above the underlying time trend. Figures showing the estimated change in the month-on-month trends in statin initiation/cessation are in appendix part 7. Fig 1 Primary analyses evaluating step change in proportion of patients initiating and stopping statin for primary and secondary prevention of cardiovascular disease after exposure period (October 2013 to March 2014). Model used interrupted time series analysis with generalised linear model with binomial error structure, with break points at beginning and end of exposure period. Models allowed for change in level of proportion of patients initiating/stopping statin. Odds ratios therefore relate to relative change in odds of initiating/stopping statins after exposure period, in comparison with expected change based on pre-exposure predictions. In A and B denominators are patients with opportunity to initiate statin each month within study period, and numerators are patients who did initiate statin after indication. In C and D denominators are patients with statin prescription ending each month within study period, and numerators are patients who did not renew that prescription and hence were defined as stopping. Solid lines and shaded confidence intervals relate to linear predictions of log odds and 95% CI of event, respectively, calculated from model and converted into probabilities. Dotted lines are extrapolation of pre-exposure linear predictions of log odds converted to probabilities, to give hypothetical proportions of post-exposure period under counterfactual scenario of no changes after exposure

og odds and 95% CI of event, respectively, calculated from model and converted into probabilities. Dotted lines are extrapolation of pre-exposure linear predictions of log odds converted to probabilities, to give hypothetical proportions of post-exposure period under counterfactual scenario of no changes after exposure Changes in statin initiation There was no evidence of a stepped change in statin initiation for primary prevention after the exposure period compared with before, adjusted for the underlying trend over time (odds ratio 0.99, 95% confidence interval 0.87 to 1.13; fig 1 A), and no evidence of a stepped change in statin initiation for secondary prevention (1.04, 0.92 to 1.18; fig 1 B). When we looked at month on month trends in initiation over time, we estimated an underlying increase in initiation for primary prevention before the exposure period, which seemed to accelerate after the exposure period (appendix part 7a, P<0.001). We observed no change in the underlying trend in initiation for secondary prevention after the exposure period (appendix part 7b, P=0.41).

e estimated an underlying increase in initiation for primary prevention before the exposure period, which seemed to accelerate after the exposure period (appendix part 7a, P<0.001). We observed no change in the underlying trend in initiation for secondary prevention after the exposure period (appendix part 7b, P=0.41). Changes in statin cessation Patients were more likely to stop taking statins after the exposure period compared with before, after we accounted for the underlying trend over time, for both primary and secondary prevention (odds ratio 1.11 (95% confidence interval 1.05 to 1.18) and 1.12 (1.04 to 1.21), respectively; fig 1 C/D). We found no evidence of any change in the underlying month on month trends in statin cessation (appendix part 7c and 7d, P=0.17 and P=0.16, respectively, for statins used for primary and secondary prevention).

ion (odds ratio 1.11 (95% confidence interval 1.05 to 1.18) and 1.12 (1.04 to 1.21), respectively; fig 1 C/D). We found no evidence of any change in the underlying month on month trends in statin cessation (appendix part 7c and 7d, P=0.17 and P=0.16, respectively, for statins used for primary and secondary prevention). Stratified analysis The increase in statin cessation after the exposure period seemed to vary by both duration of previous statin use and age group (fig 2). The increase in the likelihood of stopping after the exposure period was more pronounced among those who had taken statins for longer than those with shorter previous use (P<0.001 for trend in both primary and secondary prevention analyses). To aid comparability with other studies, we also divided duration of previous continuous prescription into less than and more than a year, and again we found a larger increase in the odds of stopping among those prescribed for longer (for cessation for primary prevention odds ratios were 1.10 (95% confidence interval 1.03 to 1.17) among those with less than one year of continuous prescription and 1.23 (1.15 to 1.32) among those with more than one year of continuous prescription; for cessation for secondary prevention the odds ratios were 1.10 (1.01 to 1.19) and 1.23 (1.13 to 1.33), respectively). The increased likelihood of stopping a statin used for primary and secondary prevention after the exposure period also became more pronounced in older age groups (P<0.001 for trend in both cases). We did not carry out stratified analyses for initiation as there were no measurable effects in the primary analyses.

vely). The increased likelihood of stopping a statin used for primary and secondary prevention after the exposure period also became more pronounced in older age groups (P<0.001 for trend in both cases). We did not carry out stratified analyses for initiation as there were no measurable effects in the primary analyses. Fig 2 Stratified cessation analyses evaluating step change in proportion of patients stopping statin for primary and secondary prevention of cardiovascular disease after exposure period (October 2013 to March 2014). Models used interrupted time series analysis with generalised linear model with binomial error structure, with break points at beginning and end of exposure period. Models allowed for change in level of proportion of patients stopping statin. Odds ratios therefore relate to relative change in odds of stopping statins after exposure period, in comparison with expected change based on pre-exposure predictions

rror structure, with break points at beginning and end of exposure period. Models allowed for change in level of proportion of patients stopping statin. Odds ratios therefore relate to relative change in odds of stopping statins after exposure period, in comparison with expected change based on pre-exposure predictions Post hoc analyses In primary and secondary prevention, the increase in cessation seemed to be restricted to the first six months after exposure, after which cessation fell to a level similar to that expected based on pre-exposure trends for the next six months (fig 3). Consistent with this pattern, when we investigated both a step change and trend change simultaneously in a single model, there was evidence that immediately after the exposure period, patients were more likely to stop statins used for primary prevention (odds ratio 1.19, 95% confidence interval 1.02 to 1.39), followed by a month on month reduction back towards baseline (0.98 (0.97 to 0.99) per month). A similar pattern was seen for cessation of statins used in secondary prevention (1.25 (1.02 to 1.53) for immediate step change after exposure period and 0.98 (0.97 to 0.99) per month for subsequent month on month trend) (appendix part 8).

month on month reduction back towards baseline (0.98 (0.97 to 0.99) per month). A similar pattern was seen for cessation of statins used in secondary prevention (1.25 (1.02 to 1.53) for immediate step change after exposure period and 0.98 (0.97 to 0.99) per month for subsequent month on month trend) (appendix part 8). Fig 3 Post hoc cessation analysis evaluating step change in proportion of patients stopping statin for primary and secondary prevention of cardiovascular disease, with post-exposure period stratified into ≤6 and >6 months. Denominators are patients with statin prescription ending each month within study period, and numerators are patients who did not renew that prescription and hence were defined as stopping. Models used interrupted time series analysis with generalised linear model with binomial error structure, with break points at beginning and end of exposure period. Models allowed for change in level of proportion of patients stopping statin. Odds ratios therefore relate to relative change in odds of stopping statins after for each 6 month section exposure period, in comparison with expected change based on pre-exposure predictions. Solid lines and shaded confidence intervals relate to linear predictions of log odds and 95% CI of event, respectively, calculated from model and converted into probabilities. Dotted lines are extrapolation of pre-exposure linear predictions of log odds converted to probabilities, to give hypothetical proportions of post-exposure period under counterfactual scenario of no changes after exposure

og odds and 95% CI of event, respectively, calculated from model and converted into probabilities. Dotted lines are extrapolation of pre-exposure linear predictions of log odds converted to probabilities, to give hypothetical proportions of post-exposure period under counterfactual scenario of no changes after exposure We also found evidence that patients were less likely to have any recorded risk score in the post-exposure period (odds ratio 0.85 (95% confidence interval 0.78 to 0.93) compared with pre-exposure); a similar pattern was seen for specific categories of risk score (fig 4).

og odds and 95% CI of event, respectively, calculated from model and converted into probabilities. Dotted lines are extrapolation of pre-exposure linear predictions of log odds converted to probabilities, to give hypothetical proportions of post-exposure period under counterfactual scenario of no changes after exposure We also found evidence that patients were less likely to have any recorded risk score in the post-exposure period (odds ratio 0.85 (95% confidence interval 0.78 to 0.93) compared with pre-exposure); a similar pattern was seen for specific categories of risk score (fig 4). Fig 4 Post hoc analyses evaluating step change in proportion of recorded 10 year risk scores for cardiovascular disease in each category after exposure period (October 2013-March 2014), using denominator of total number of patients under follow-up each month in CPRD. Denominators are all patients under follow-up in CPRD each month within study period, and numerators are patients that had recorded 10 year risk score for cardiovascular disease within each category in that month. Models used interrupted time series analysis with generalised linear model with binomial error structure, with break points at beginning and end of exposure period. Models allowed for change in level of proportion of patients with a recorded cardiovascular risk score. Odds ratios therefore relate to relative change in odds of having a recorded risk score after the exposure period, in comparison with expected change based on pre-exposure predictions. Solid lines and shaded confidence intervals relate to linear predictions of log odds and 95% CI of event, respectively, calculated from model and converted into probabilities. Dotted lines are extrapolation of pre-exposure linear predictions of log odds converted to probabilities, to give hypothetical proportions of post-exposure period under counterfactual scenario of no changes after exposure

og odds and 95% CI of event, respectively, calculated from model and converted into probabilities. Dotted lines are extrapolation of pre-exposure linear predictions of log odds converted to probabilities, to give hypothetical proportions of post-exposure period under counterfactual scenario of no changes after exposure Negative control analysis We found no evidence of changes in prescribing of treatment for glaucoma after the exposure period (appendix parts 9 and 10). When we moved the exposure period to 12 months earlier, the “post exposure” change in statin cessation rates disappeared as expected (odds ratio 1.01 (95% confidence interval 0.99 to 1.03) and 1.00 (0.98 to 1.03), respectively, for cessation in primary and secondary prevention). Impact on public health We estimated that across the UK there was an excess of 218 971 patients who stopped taking a statin in the six months after the media coverage. When we applied a previously estimated restart rate of 66% among patients who stopped taking a statin without a statin related event,28 we calculated the number of excess cardiovascular events to be at least 2173 within the subsequent 10 years. Under the most pessimistic assumption that all the patients who stopped taking statins did so indefinitely, the estimated number of excess cardiovascular events rose to 6372.

in without a statin related event,28 we calculated the number of excess cardiovascular events to be at least 2173 within the subsequent 10 years. Under the most pessimistic assumption that all the patients who stopped taking statins did so indefinitely, the estimated number of excess cardiovascular events rose to 6372. Discussion Key findings A period of intense media coverage of statins and their side effects was followed by an increase in cessation of statins prescribed for both primary and secondary prevention of cardiovascular disease in UK primary care. This increase seemed to be temporary, and cessation had returned to expected levels after six months. We also identified that the tendency to stop was higher among patients who had used statins for longer and among older patients. Among those defined as newly eligible to receive statins, we did not observe a change in patients’ likelihood of initiating, over and above the underlying trend. Further investigation, however, showed a marked decrease in the proportion of patients having any cardiovascular disease risk score recorded after the media coverage, and hence a smaller pool of patients whose records met the criteria for initiation of statins for primary prevention.

ver and above the underlying trend. Further investigation, however, showed a marked decrease in the proportion of patients having any cardiovascular disease risk score recorded after the media coverage, and hence a smaller pool of patients whose records met the criteria for initiation of statins for primary prevention. Findings in context of previous research Our study is the first to attempt to quantify the effects of the UK media coverage of statins on prescribing in primary care and, to our knowledge, the first study in any country to look comprehensively at the effect of negative media coverage on rates of statin initiation and cessation, in both primary and secondary prevention of cardiovascular disease. A few recent studies have looked at specific aspects of media impacts on statin use. A study in Denmark reported that patients newly starting to take a statin were less likely to fill a second prescription if there were more negative statin related media stories in the period immediately after initiation.15 This is consistent with our findings that statin cessation rates are affected by negative news stories, and, importantly, our results suggest a similar effect is detectable even among those with longer established use. In Australia, Schaffer and colleagues reported a 28.8% (95% confidence interval 15.4% to 43.7%) increase in the discontinuation of statins in the week of the controversial TV programme Catalyst, which was deemed to be critical of statins.13 It should be noted that this refers to the peak increase in discontinuation at a weekly resolution, so is not comparable with our odds ratio estimates. In addition, the type, duration, and intensity of media coverage differed; nevertheless the results are qualitatively consistent with our findings. Kocas and colleagues reported that in Turkey, as news articles about statins increased from 2011 to 2013, the percentage of days covered by statins decreased (57% (interquartile range 8-83) in 2011, 58% (17-83) in 2012, and 50% (8-83) in 2013) (P=0.01).16 The overarching message is consistent with our study, but they focused on trends in statin adherence over a three year period. In France, Saib and colleagues also reported an increase in the number of patients that intended to stop statin therapy after media controversy,14 but their study was questionnaire based and no prescribing data were used, so, while again consistent with our findings, the results are difficult to compare directly with those in our study.

d colleagues also reported an increase in the number of patients that intended to stop statin therapy after media controversy,14 but their study was questionnaire based and no prescribing data were used, so, while again consistent with our findings, the results are difficult to compare directly with those in our study. Finally, consistent with our findings, a Dutch study by van Hunsel and colleagues used adverse drug reaction reports and found that, after a television programme about the benefits and risks of statins, there was a transient increase in the number of patients reporting reactions, a substantial proportion of whom also reported stopping treatment.29

findings, a Dutch study by van Hunsel and colleagues used adverse drug reaction reports and found that, after a television programme about the benefits and risks of statins, there was a transient increase in the number of patients reporting reactions, a substantial proportion of whom also reported stopping treatment.29 Further studies have examined the impact of various other examples of health related media coverage on patients’ behaviour. A 2002 Cochrane review identified 15 mass media health interventions, and five studies of media coverage outside the context of a planned intervention (including coverage of breast cancer surgery for a public figure, side effects of drugs, and the disclosure of a sporting personality’s HIV status); all but one was associated with a change in health service use in the direction expected.30 Several more recent examples are also noteworthy. A New Zealand study found that media coverage about adverse events after a formula change in the thyroid drug Eltroxin (levothyroxine; GlaxoSmithKline) was followed by an increase in adverse event reports to a national medicines monitoring database, most markedly for the specific symptoms mentioned in television coverage.31 Data from Australia suggested that coverage of the singer and actress Kylie Minogue’s breast cancer was followed by a six week period in which self referral rates for breast cancer screening doubled among eligible women who had never previously presented for screening,32 33 while UK celebrity Jade Goody’s diagnosis of cervical cancer was similarly associated with increased attendance for cervical screening,34 an increased incidence of referrals for colposcopy, and increased diagnoses of high grade cervical neoplasia.35

eligible women who had never previously presented for screening,32 33 while UK celebrity Jade Goody’s diagnosis of cervical cancer was similarly associated with increased attendance for cervical screening,34 an increased incidence of referrals for colposcopy, and increased diagnoses of high grade cervical neoplasia.35 It is difficult to disentangle whether the changes in statin prescribing that we observed were driven by changes in the behaviour of physicians, patients, or both.36 It is noteworthy that we observed increases in cessation rates after the exposure period, but no corresponding decrease in statin initiation. The absence of a decrease in initiation rates for primary prevention, however, needs to be considered in the context of the observed decrease in the number of risk scores recorded after the exposure period, both overall and within risk strata, which indicates that the widespread debate and media coverage on risks and benefits of statins could have changed the general discourse between GPs and patients about management of risk of cardiovascular disease. It is possible that this was driven by a change in GP behaviour or by refusal of patients to engage with risk scoring for cardiovascular disease and discussions about statin use because of concerns about the potential side effects, which is a known worry for patients.18 Conversations between GPs and the most vocally concerned patients might have led to a reluctance to carry out risk scores because of the known worries surrounding the possible side effects of statins, leaving only those more likely to initiate in the denominator. These apparent changes in patterns of recording risk scores might also explain the unexpected acceleration in the already increasing trend in initiation of statins for primary prevention of cardiovascular disease among those with a recorded high risk score (appendix 7a). It is perhaps less surprising that there was no change in patients initiating statins for secondary prevention after the intense media coverage. After a potentially serious cardiovascular event, it is highly likely that a patient will be willing to take drugs proved to reduce their risk of having a recurrent event, regardless of critical media stories and potential side effects.

ients initiating statins for secondary prevention after the intense media coverage. After a potentially serious cardiovascular event, it is highly likely that a patient will be willing to take drugs proved to reduce their risk of having a recurrent event, regardless of critical media stories and potential side effects. We observed an increasing tendency to stop taking statins as the length of previous continuous prescription increased. This might have been driven by immediate concerns about a recent indication for treatment (such as a cardiovascular event or high 10 year risk score) causing patients to be more reluctant to quit their treatment. In comparison, patients whose original indication occurred several years in the past might be comparatively comfortable to stop treatment after witnessing critical stories. Further qualitative research exploring attitudes among patients at different stages of statin use, however, would be needed to confirm this. Another possible explanation for apparent variation in the effect of media coverage is that particular groups of people might pay more attention to, or be more influenced by, health new stories; this might in particular explain why older patients were more likely to stop taking statins after the exposure period.

this. Another possible explanation for apparent variation in the effect of media coverage is that particular groups of people might pay more attention to, or be more influenced by, health new stories; this might in particular explain why older patients were more likely to stop taking statins after the exposure period. Strengths and limitations The CPRD is a large dataset and broadly represents the UK population,21 meaning our findings are generalisable to the wider population, and we were able to detect small effect sizes with great precision. We also believe that our populations will be similar to other developed populations as the indications for statins are broadly the same worldwide. Data were available for only a limited number of months after the exposure period, limiting the timescale over which we could detect effects, but as our analyses suggested that cessation rates were affected for only up to six months after the exposure period, this is unlikely to have been an important limitation.

ents were involved in setting the research question or the outcome measures, nor were they involved in developing plans for design or implementation of the study. No patients were asked to advise on interpretation or writing up of results. Results will be disseminated to relevant patient communities through news media. Results Data were analysed for 610 UK Clinical Practice Research Datalink (CPRD) general practices, with 45 465 201 registered person years of observation from 2005 to 2014. Figure 1 presents data for RTI consultations and associated antibiotic prescribing between 2005 and 2014. The RTI consultation rate per 100 000 continued a long term decline7 during the period, decreasing from 256 to 220 per 100 000 in men and from 351 to 307 per 100 000 in women. The antibiotic prescribing rate for RTIs declined from 128 to 106 per 100 000 in men and from 184 to 155 per 100 000 in women. The proportion of RTI consultations with antibiotics prescribed declined from 53.9% to 50.5% in men and from 54.5% to 51.5% in women. Fig 1 Age standardised consultation rate for self limiting respiratory tract infections (RTIs), antibiotic prescribing rate for RTIs, and proportion of RTI consultations with antibiotics prescribed in 610 general practices contributing to the UK Clinical Practice Research Datalink. Red open circles represent females; blue filled circles represent males. Lines fitted by least squares

wide. Data were available for only a limited number of months after the exposure period, limiting the timescale over which we could detect effects, but as our analyses suggested that cessation rates were affected for only up to six months after the exposure period, this is unlikely to have been an important limitation. Recording of prescriptions in the CPRD is automatic at the point of issue and therefore complete, though one limitation is that we could not be certain that prescriptions were subsequently dispensed at a pharmacy or taken by the patient. The actual proportions of patients initiating and stopping taking statins are likely to be subject to some measurement error as some patients will have initiated or restarted statins after the end of the predefined grace periods that we used. Also, if a patient began to self manage their prescription by taking a lower dose, it is possible that we could have incorrectly classified them as stopping statins because of longer than expected gaps between prescriptions. We believe that such error will be minimal because the grace periods were specifically selected such that in preliminary analyses only a small minority of patients went on to start (or restart) a statin after the selected times. Furthermore, our definitions for initiation and cessation remained the same throughout our study period, and it is unlikely that any measurement error would have changed after the exposure period and affected our main results. We did not capture use of low dose over-the-counter statins, which have been available since 2004, though it seems that uptake has been low,37 and there is no reason to think that the high level of media coverage would have led people to switch to over-the-counter statins.

exposure period and affected our main results. We did not capture use of low dose over-the-counter statins, which have been available since 2004, though it seems that uptake has been low,37 and there is no reason to think that the high level of media coverage would have led people to switch to over-the-counter statins. Interrupted time series cannot confirm a causal link between the media coverage and the observed changes in the likelihood of stopping taking statins. The design avoids confounding by individual level factors such as smoking and obesity that are unlikely to vary over short term timescales, but it is possible that other external factors played a role in the observed changes. We carried out two negative control analyses in an attempt to exclude this, using both a different class of drug and a different time period, and we found no post-exposure changes in either of these analyses, strengthening our main finding. Nevertheless, it is still possible that other changes in the same time period, unrelated to the media controversy and affecting only statin use, could have driven the observed results.

of drug and a different time period, and we found no post-exposure changes in either of these analyses, strengthening our main finding. Nevertheless, it is still possible that other changes in the same time period, unrelated to the media controversy and affecting only statin use, could have driven the observed results. If we assume causality, we estimated that increases in statin cessation due to the period of media coverage of side effects could result in at least 2173 excess cardiovascular events over 10 years, depending on the proportion of “stoppers” who re-started later. Our calculations were based on several assumptions and approximations and clearly could not take account of future changes in statin use and perceptions and other developments in prevention of cardiovascular disease. Varying assumptions also lead to substantial changes in the outcome, meaning these estimations should be interpreted with caution. We also cannot know from our data the extent to which patients were appropriately informed about the risk:benefit balance of statins and whether those who stopped would have been aware and accepting of the consequent increases in risk of cardiovascular disease. Patients can vary widely in the choices that they make about long term preventive drug treatment, and some choose not to take drugs that will extend their life.38 Finally, we did not attempt to take into account any possible benefits of stopping treatment with statins, which might have offset the increase in risk.

ase. Patients can vary widely in the choices that they make about long term preventive drug treatment, and some choose not to take drugs that will extend their life.38 Finally, we did not attempt to take into account any possible benefits of stopping treatment with statins, which might have offset the increase in risk. Conclusion Controversy over the risks and benefits of statins reported in both the medical and popular press was followed by a transient increase in patients stopping treatment prescribed for primary and secondary prevention. Additionally, a marked reduction in the proportion of patients receiving a risk score for cardiovascular disease suggests other important impacts on GP and/or patient behaviour. This research highlights the potential for widely covered health stories in the media to have an effect on real world behaviour related to healthcare and could be used to inform future interactions between clinicians, researchers, the academic press, and the wider media. What is already known on this topic Studies from Denmark, Australia, Turkey, and France have suggested that negative media stories can affect statin cessation and prescribing rates Two controversial articles about statins were published in October 2013 in the UK, with a subsequent high volume of debate in the media about the associated potential risks and benefits

What is already known on this topic Studies from Denmark, Australia, Turkey, and France have suggested that negative media stories can affect statin cessation and prescribing rates Two controversial articles about statins were published in October 2013 in the UK, with a subsequent high volume of debate in the media about the associated potential risks and benefits What this study adds After the media coverage, there were no changes in statin initiation among those with a recorded new indication but an 11% and 12% increase in the likelihood of existing users stopping their treatment, for primary and secondary prevention, respectively Across the UK these effects were estimated to result in over 200 000 patients stopping treatment with a statin in the six months after the media coverage This research provides unique evidence describing the potential for widely covered health stories in the media to affect real world behaviour related to healthcare, with implications for public health, and has the potential to inform future interactions between clinicians, researchers, the academic press, and the wider media Web Extra Extra material supplied by the author Appendix: Supplementary material Click here for additional data file. Contributors: All authors contributed to the design of the study. AM extracted the data, wrote the statistical programmes, and wrote the first draft. All authors contributed to further drafts and approved the final manuscript. AM and EH contributed equally. KB is guarantor.

Appendix: Supplementary material Click here for additional data file. Contributors: All authors contributed to the design of the study. AM extracted the data, wrote the statistical programmes, and wrote the first draft. All authors contributed to further drafts and approved the final manuscript. AM and EH contributed equally. KB is guarantor. Funding: This study was funded by the British Heart Foundation. KB holds a Sir Henry Dale Fellowship jointly funded by the Wellcome Trust and the Royal Society (grant No 107731/Z/15/Z). LS is funded by a Wellcome Trust senior fellowship in clinical science. The British Heart Foundation, the Wellcome Trust, and the Royal Society had no role in the design, analysis, or writing up of this study.

Sir Henry Dale Fellowship jointly funded by the Wellcome Trust and the Royal Society (grant No 107731/Z/15/Z). LS is funded by a Wellcome Trust senior fellowship in clinical science. The British Heart Foundation, the Wellcome Trust, and the Royal Society had no role in the design, analysis, or writing up of this study. Competing interests: All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf and declare: AG received grants from Medical Research Council, during the conduct of the study, and personal fees from University of Umea, unrelated to the submitted work; TVS received personal fees from GSK, Roche, and Sanofi for presenting on methods for pragmatic trials and grant funding from GSK unrelated to the submitted work; BG received grants from the Laura and John Arnold Foundation, Wellcome Trust, and the Health Foundation and receives additional income from speaking, writing, and broadcasting on problems in science and medicine; LS reports grants from Wellcome Trust and British Heart Foundation during the conduct of the study, grants from Wellcome Trust, Medical Research Council, National Institute for Health Research and the European Union outside the submitted work, personal fees from GSK for advisory work unrelated to the submitted work, grant funding from GSK for academic research unrelated to the submitted work, acts as an unpaid steering committee chair for AstraZeneca for a randomised trial unrelated to the submitted work, and is a trustee of the British Heart Foundation; KB received grants from British Heart Foundation, Wellcome Trust, and Royal Society during the conduct of the study, and grant funding from Medical Research Council and National Institute for Health Research unrelated to the submitted work

related to the submitted work, and is a trustee of the British Heart Foundation; KB received grants from British Heart Foundation, Wellcome Trust, and Royal Society during the conduct of the study, and grant funding from Medical Research Council and National Institute for Health Research unrelated to the submitted work Ethical approval: The prespecified study protocol was approved by the Independent Scientific Advisory Committee for MHRA Database Research (ISAC), and the approved protocol including amendments is supplied in appendix part 11. Approval was also received from the London School of Hygiene and Tropical Medicine ethics committee. Data sharing: The data were obtained from the Clinical Practice Research Datalink (CPRD). CPRD is a research service that provides primary care and linked data for public health research. CPRD data governance and our own license to use CPRD data do not allow us to distribute or make available patient data directly to other parties. Researchers can apply for data access at www.cprd.com, and must have their study protocol approved by the Independent Scientific Advisory Committee for MHRA database research (details at www.cprd.com/isac). Transparency: The lead author affirms that the manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained.

Introduction Concern is growing that the widespread and sometimes unnecessary use of antibiotics is leading to the development of antimicrobial drug resistance and potentially to infections caused by resistant organisms that are difficult to treat.1 Reducing the inappropriate use of antibiotics, as well as ensuring that they can be used when needed, represent important components of a strategy to control infectious diseases.2 In the healthcare sector, attention has focused on primary care, including first point-of-contact ambulatory care, where a high proportion of antibiotics are prescribed. About 60% of antibiotics prescribed in primary care are for respiratory tract infections (RTIs).3 RTIs, including common colds, sore throat, cough, acute bronchitis, otitis media, and sinusitis are often self limiting and usually improve without specific treatment.4 Antibiotic treatment of RTIs offers negligible benefit to affected patients5 and is often associated with side effects.6 Guidance in the United Kingdom recommends that either a no antibiotic prescribing strategy or a delayed antibiotic prescribing strategy should be agreed for most patients with RTIs.3 Nevertheless, around 36% of common colds continue to be treated with antibiotics, as do 40% of episodes of sore throat, 70% of otitis media, and 90% of sinusitis.7 Wide variation exists among general practices. About 50% of all general practice consultations for RTIs result in an antibiotic prescription, but some general practices issue prescriptions at a rate of more than 80% and others at less than 20%.8 This contrasts with practice in the Netherlands, where 22.5% of RTI episodes in 2010 were treated with antibiotics.9 A considerable amount of research has been done10 11 12 to develop interventions that might help general practitioners to reduce the rate of antibiotic prescribing for RTIs. This has been translated into policy guidance3 and public campaigns13 to control unnecessary antibiotic use.

2010 were treated with antibiotics.9 A considerable amount of research has been done10 11 12 to develop interventions that might help general practitioners to reduce the rate of antibiotic prescribing for RTIs. This has been translated into policy guidance3 and public campaigns13 to control unnecessary antibiotic use. Clinical concern that reducing antibiotic use might increase the risk of complications following RTIs might be realistic. Evidence from clinical trials suggests that antibiotics may reduce the risk of suppurative complications of RTIs,5 but the more serious complications are generally too rare to evaluate precisely in randomised studies. A cohort study in 162 general practices in the General Practice Research Database from 1991 to 2001 evaluated the effect of antibiotic treatment on the incidence of pneumonia after upper RTI, peritonsillar abscess after sore throat, and mastoiditis after otitis media.14 The results suggested that antibiotic treatment was associated with lower odds of each of these complications, but the overall risk of complications was generally small and the number of patients who would have to be treated to avoid one complication was estimated to be in excess of 4000. However, pneumonia was more common; in people aged 65 or more an estimated one case for every 39 antibiotic prescriptions might be avoided. Infections of the middle ear or sinuses may rarely be complicated by intracranial abscess.15 Lemierre’s syndrome,16 from thrombophlebitis of the internal jugular vein associated with Fusobacterium necrophorum infection, is a rare complication of sore throat,16 but F necrophorum might be frequently detectable in patients with symptoms of sore throat.17 18 The annual number of cases of Lemierre’s syndrome in England was reported to have increased from 19 in 1997 to 34 in 1999, prompting a reminder from the chief medical officer that some symptoms of sore throat may require antibiotic treatment.19 In addition to concerns about complications, medical practitioners may be concerned about the potential consequences of diagnostic misclassification. The initial symptoms of meningitis may sometimes resemble an influenza-like illness.20 Awareness of the possibility of a more serious diagnosis might prompt general practitioners to issue an antibiotic prescription for conditions in which antibiotics are not usually indicated.

sequences of diagnostic misclassification. The initial symptoms of meningitis may sometimes resemble an influenza-like illness.20 Awareness of the possibility of a more serious diagnosis might prompt general practitioners to issue an antibiotic prescription for conditions in which antibiotics are not usually indicated. These observations raise important questions for a policy to reduce antibiotic prescribing for RTIs in primary care: Is there a safe level of antibiotic prescribing for RTIs? What target can general practices safely adopt in reducing the proportion of consultations for RTIs with antibiotics prescribed? Is there a threshold for antibiotic prescribing below which complications may increase? We evaluated the safety of a policy to reduce antibiotic prescribing for RTIs in primary care and the incidence of pneumonia, peritonsillar abscess, mastoiditis, empyema, meningitis, intracranial abscess, and Lemierre’s syndrome. We determined whether these complications were more common in general practices that prescribe fewer antibiotics for self limiting RTIs than higher prescribing practices. We aimed to use this information to quantify the potential clinical and public health impact of changes in antibiotic prescribing practice.

syndrome. We determined whether these complications were more common in general practices that prescribe fewer antibiotics for self limiting RTIs than higher prescribing practices. We aimed to use this information to quantify the potential clinical and public health impact of changes in antibiotic prescribing practice. Methods The data source for the study was the UK Clinical Practice Research Datalink (CPRD).21 This is a large database containing fully anonymised electronic records from about 7% of UK general practices from 1987 to the present. CPRD data are considered representative of the UK population, and the high quality of CPRD data have been confirmed in many studies.21 For the present study, we included data for the 10 year period from 2005 to 2014. During this period the CPRD included data for an open cohort of about 4.5 million registered patients.

CPRD data are considered representative of the UK population, and the high quality of CPRD data have been confirmed in many studies.21 For the present study, we included data for the 10 year period from 2005 to 2014. During this period the CPRD included data for an open cohort of about 4.5 million registered patients. Definition of infective complications of RTIs We evaluated the number of first episodes of infective complications in the entire registered population of CPRD from 2005 to 2014. Such complications were defined using Read medical codes recorded in participants’ electronic health records. The Read code classification represents a terminology used to code primary care electronic health records in the UK.22 Electronic health records include diagnoses recorded at primary care consultations and home visits. In addition, the CPRD referral file includes coded data for hospital referrals and hospital discharges. In analyses we evaluated pneumonia (57 codes), empyema (14 codes), peritonsillar abscess (5 codes), mastoiditis (13 codes), bacterial meningitis (19 codes), and intracranial abscess (14 codes). Codes for “pneumonia” were drawn from section H2 of the Read code classification, which includes codes for “pneumonia and influenza.” Codes were included if they indicated the presence of pneumonia without a viral cause. “Bacterial meningitis” included codes for meningococcal meningitis, meningococcal septicaemia, pneumococcal meningitis, and haemophilus meningitis, as well as unspecified bacterial meningitis. Code lists are available from the authors. Data were extracted for all participants with records of infective complications from 2005 to 2014. We defined incident events as the first record of an event in a participant that was recorded more than 12 months after the start of the participant’s CPRD record. Sex, year, and age group were included as individual level covariates. Nine 10 year age groups were employed, with categories of 0 to 14 years and 85 years or more. We aggregated incident events by year, age group, sex, and general practice. Person time for the registered CPRD population was estimated by year, age group, sex, and general practice to estimate rates of infective complications. Cluster level covariates included CPRD region, with 10 regions in England, as well as Wales, Scotland, and Northern Ireland.

ear, age group, sex, and general practice. Person time for the registered CPRD population was estimated by year, age group, sex, and general practice to estimate rates of infective complications. Cluster level covariates included CPRD region, with 10 regions in England, as well as Wales, Scotland, and Northern Ireland. Deprivation fifth was included, based on general practice level data for indices of multiple deprivation score (IMD 2010) for England, and equivalent scores in Scotland, Wales, and Northern Ireland.

ear, age group, sex, and general practice. Person time for the registered CPRD population was estimated by year, age group, sex, and general practice to estimate rates of infective complications. Cluster level covariates included CPRD region, with 10 regions in England, as well as Wales, Scotland, and Northern Ireland. Deprivation fifth was included, based on general practice level data for indices of multiple deprivation score (IMD 2010) for England, and equivalent scores in Scotland, Wales, and Northern Ireland. Definition of RTI consultation and antibiotic prescribing rates We estimated age standardised measures for RTI consultations and antibiotic prescribing as reported previously.7 23 For each CPRD general practice we estimated the rate of RTI consultations per 1000 registered patients, the antibiotic prescribing rate for RTI per 1000 registered patients, and the proportion (%) of RTI consultations with antibiotics prescribed. These prescribing measures were estimated on a sample of CPRD data because it was not feasible and our licence did not allow us to perform the analysis on the entire CPRD database. Participants were sampled from all acceptable patients included in CPRD. A random sample of 75 currently registered patients was drawn without replacement for each year from 2005 to 2014. This gave a maximum sample of 750 participants, with up to 7500 person years of observation for each practice. We aimed to achieve a total sample of fewer than 0.5 million, and the total sample for analysis was 411 226 participants from 643 general practices. This allowed us to estimate practice specific proportions with a 1% margin of error. For participants in the sample, we estimated person years as denominator from the start of CPRD registration or 1 January 2005, to the end of the participant’s CPRD record or 31 December 2014. We identified self limiting RTIs using medical codes recorded during general practice consultations. These were classified into five groups following the recommendations of the National Institute for Health and Care Excellence3: colds and “upper respiratory tract infections”; sore throat, including pharyngitis and laryngitis; cough and acute bronchitis; otitis media; and rhinosinusitis. Acute bronchitis was included because current recommendations are to avoid antibiotic treatment.3 Consultations for RTIs were identified, and we selected first consultations within a 14 day time window. Data for participants aged 100 or older were excluded.

cough and acute bronchitis; otitis media; and rhinosinusitis. Acute bronchitis was included because current recommendations are to avoid antibiotic treatment.3 Consultations for RTIs were identified, and we selected first consultations within a 14 day time window. Data for participants aged 100 or older were excluded. We identified antibiotic prescriptions issued on the same day as consultations for respiratory problems and then estimated for each general practice the rate of consultations for RTIs per 1000 person years, rate of antibiotic prescribing for RTIs per 1000 person years, and proportion (%) of RTI consultations with antibiotics prescribed. Rates and proportions were standardised for age and sex using the 2013 European standard population. After excluding practices with insufficient data, because of short periods of contributing to CPRD, we estimated rates for 610 CPRD general practices.

atory tract infections (RTIs), antibiotic prescribing rate for RTIs, and proportion of RTI consultations with antibiotics prescribed in 610 general practices contributing to the UK Clinical Practice Research Datalink. Red open circles represent females; blue filled circles represent males. Lines fitted by least squares Figure 2 and table 1 show changes between 2005 and 2014 in rates of outcome measures for males and females registered in CPRD. Over the period there were declining trends in incidence of peritonsillar abscess (1% yearly), mastoiditis (4.6%), and meningitis (5.3%), pneumonia showed an increase of 0.4% yearly, and empyema and intracranial abscess showed no clear change over time. Fig 2 Incidence of infective complications in 610 general practices contributing to the UK Clinical Practice Research Datalink. Red open circles represent females; blue filled circles represent males. Lines fitted by least squares Table 1 Annual percentage relative change in incidence of infective complications from 2005 to 2014 Infective complications Annual % change in relative incidence (95% CI)* P value Pneumonia 0.36 (0.09 to 0.64) 0.008 Peritonsillar abscess −0.99 (−1.45 to −0.53) <0.001 Mastoiditis −4.64 (−5.21 to −4.07) <0.001 Empyema −0.56 (−1.16 to 0.05) 0.07 Bacterial meningitis −5.28 (−4.69 to 5.87) <0.001 Intracranial abscess −1.36 (−6.66 to 3.68) 0.60 *Adjusted for age group, sex, region, deprivation fifth, consultation rate for respiratory tract infections (RTIs), and proportion of RTI consultations with antibiotics prescribed and clustering by general practice.

) 0.07 Bacterial meningitis −5.28 (−4.69 to 5.87) <0.001 Intracranial abscess −1.36 (−6.66 to 3.68) 0.60 *Adjusted for age group, sex, region, deprivation fifth, consultation rate for respiratory tract infections (RTIs), and proportion of RTI consultations with antibiotics prescribed and clustering by general practice. General practices were divided into fourths according to the proportion of RTI consultations with antibiotics prescribed (table 2). General practices in the highest fourth prescribed antibiotics at a median 65% (range 58% to 79%) of RTI consultations, whereas general practices in the lowest fourth prescribed antibiotics at a median 38% (29% to 44%) of RTI consultations. Table 2 shows the age standardised incidence rates for each of the infective complications. The incidence of pneumonia was 157 (95% confidence interval 154 to 159) per 100 000 at low prescribing practices but 119 (117 to 121) per 100 000 at high prescribing practices. The corresponding values for peritonsillar abscess were 15.6 (15.5 to 15.8) per 100 000 and 12.9 (12.8 to 13.0) per 100 000. Mastoiditis, empyema, bacterial meningitis, and intracranial abscess showed lower incidence rates, which did not appear to be associated with antibiotic prescribing category. Fourteen cases of Lemierre’s syndrome occurred, which were evenly distributed between prescribing categories, with an overall incidence rate of 0.31 per million.

yema, bacterial meningitis, and intracranial abscess showed lower incidence rates, which did not appear to be associated with antibiotic prescribing category. Fourteen cases of Lemierre’s syndrome occurred, which were evenly distributed between prescribing categories, with an overall incidence rate of 0.31 per million. Table 2 Distribution of general practices and person years follow-up for registered patients from 2005 to 2014 for 610 general practices contributing to the UK Clinical Practice Research Datalink Variables Fourths of proportion of RTI consultations with antibiotics prescribed High ≥58% 51-57% 44-50% Low <44% No of general practices 152 153 152 153 No of person years from registered patients 10 573 885 12 135 183 12 109 005 10 647 128 Median (95% range) proportion of RTI consultations with antibiotics prescribed 65 (58-79) 54 (51-57) 48 (45-50) 38 (29-44) Infective complications*: Pneumonia 119.2 (117.0 to 121.3) 129.1 (126.9 to 131.2) 156.4 (154.0 to 158.7) 156.6 (154.0 to 159.1) Peritonsillar abscess 12.9 (12.8 to 13.0) 13.2 (13.1 to 13.3) 14.1 (13.9 to 14.2) 15.6 (15.5 to 15.8) Mastoiditis 3.48 (3.37 to 3.60) 3.31 (3.21 to 3.42) 3.32 (3.19 to 3.46) 3.38 (3.25 to 3.51) Empyema 3.64 (3.27 to 4.01) 4.00 (3.63 to 4.37) 3.66 (3.31 to 4.01) 4.00 (3.61 to 4.40) Bacterial meningitis 2.19 (1.90 to 2.47) 2.16 (1.90 to 2.42) 2.24 (1.97 to 2.51) 2.45 (2.15 to 2.75) Intracranial abscess 0.37 (0.25 to 0.48) 0.35 (0.24 to 0.46) 0.55 (0.42 to 0.69) 0.42 (0.29 to 0.55) Lemierre’s syndrome 4 cases 3 cases 2 cases 5 cases RTI=respiratory tract infection.

3.31 to 4.01) 4.00 (3.61 to 4.40) Bacterial meningitis 2.19 (1.90 to 2.47) 2.16 (1.90 to 2.42) 2.24 (1.97 to 2.51) 2.45 (2.15 to 2.75) Intracranial abscess 0.37 (0.25 to 0.48) 0.35 (0.24 to 0.46) 0.55 (0.42 to 0.69) 0.42 (0.29 to 0.55) Lemierre’s syndrome 4 cases 3 cases 2 cases 5 cases RTI=respiratory tract infection. *Values are age and sex standardised incidence rate per 100 000 (95% CI). Figure 3 presents the occurrence of infective complications according to the rate of RTI consultations. General practices in the highest fourth for RTI consultation rate had higher incidence rates for pneumonia and mastoiditis (1.35, 1.14 to 1.61, P=0.001 and 1.67, 1.20 to 2.33, P=0.002, respectively) compared with general practices in the lowest fourth. Peritonsillar abscess, empyema, bacterial meninigitis, and intracranial abscess were not associated with the general practice RTI consultation rate. Fig 3 Association of incidence of infective complications with fourth of consultation rate for self limiting respiratory tract infections (RTIs). Rates are number of incident events per 100 000 person years. Incidence rate ratios (IRRs) were adjusted for sex, age group, region, deprivation fifth, and clustering by general practice

ation of incidence of infective complications with fourth of consultation rate for self limiting respiratory tract infections (RTIs). Rates are number of incident events per 100 000 person years. Incidence rate ratios (IRRs) were adjusted for sex, age group, region, deprivation fifth, and clustering by general practice Figure 4 shows the association between the proportion of RTI consultations with antibiotics prescribed and the incidence of infective complications. The risk of pneumonia and peritonsillar abscess decreased as the antibiotic prescribing proportion for RTI increased, but there was no clear evidence of an association for mastoiditis, empyema, meningitis, and intracranial abscess. For general practices in the highest fourth of antibiotic prescribing, the incidence rate ratio for pneumonia was 0.70 (95% confidence interval 0.59 to 0.82, P<0.001), and for peritonsillar abscess was 0.78 (0.68 to 0.90, P<0.001) compared with the lowest fourth. Fig 4 Association of incidence of infective complications with fourth of antibiotic prescribing proportion. Incidence rate ratios (IRRs) were adjusted for consultation rate for respiratory tract infections, sex, age group, region, deprivation fifth, and clustering by general practice

Figure 4 shows the association between the proportion of RTI consultations with antibiotics prescribed and the incidence of infective complications. The risk of pneumonia and peritonsillar abscess decreased as the antibiotic prescribing proportion for RTI increased, but there was no clear evidence of an association for mastoiditis, empyema, meningitis, and intracranial abscess. For general practices in the highest fourth of antibiotic prescribing, the incidence rate ratio for pneumonia was 0.70 (95% confidence interval 0.59 to 0.82, P<0.001), and for peritonsillar abscess was 0.78 (0.68 to 0.90, P<0.001) compared with the lowest fourth. Fig 4 Association of incidence of infective complications with fourth of antibiotic prescribing proportion. Incidence rate ratios (IRRs) were adjusted for consultation rate for respiratory tract infections, sex, age group, region, deprivation fifth, and clustering by general practice Figure 5 shows the association between the antibiotic prescribing rate and infective complications. Pneumonia showed an association with the antibiotic prescribing rate (incidence rate ratio for highest fourth 0.74, 0.58 to 0.95, P=0.02). Peritonsillar abscess showed a weak association (0.84, 0.68 to 1.03, P=0.09) but the other infective complications did not. The antibiotic prescribing rate is determined by the RTI consultation rate and the proportion of consultations with antibiotics prescribed. It is correlated with both the RTI consultation rate (r=0.82) and the antibiotic prescribing proportion (r=0.66).

68 to 1.03, P=0.09) but the other infective complications did not. The antibiotic prescribing rate is determined by the RTI consultation rate and the proportion of consultations with antibiotics prescribed. It is correlated with both the RTI consultation rate (r=0.82) and the antibiotic prescribing proportion (r=0.66). Fig 5 Association of incidence of infective complications with fourth of antibiotic prescribing rate. Incidence rate ratios (IRRs) were adjusted for consultation rate for respiratory tract infections, sex, age group, region, deprivation fifth, and clustering by general practice

68 to 1.03, P=0.09) but the other infective complications did not. The antibiotic prescribing rate is determined by the RTI consultation rate and the proportion of consultations with antibiotics prescribed. It is correlated with both the RTI consultation rate (r=0.82) and the antibiotic prescribing proportion (r=0.66). Fig 5 Association of incidence of infective complications with fourth of antibiotic prescribing rate. Incidence rate ratios (IRRs) were adjusted for consultation rate for respiratory tract infections, sex, age group, region, deprivation fifth, and clustering by general practice Table 3 shows incident rate ratios estimated after fitting the predictors as continuous variables. These estimates are consistent with linear associations; adding quadratic terms did not improve the goodness of fit. An increasing RTI consultation rate was associated with an increasing incidence of pneumonia and mastoiditis. An increasing antibiotic prescribing proportion was associated with a declining incidence of pneumonia and peritonsillar abscess. Each 10% increase in antibiotic prescribing proportion was associated with a 12.8% (95% confidence interval 7.8% to 17.5%) relative decrease in pneumonia and a 9.9% (5.6% to 14.0%) decrease in peritonsillar abscess. Associations with the antibiotic prescribing rate fitted as a linear predictor were consistent with those for the antibiotic prescribing proportion. In additional analyses, we found that the incidence rate ratio associating the antibiotic prescribing proportion with pneumonia was similar for the population aged less than 65 years and 65 years or older.

rescribing rate fitted as a linear predictor were consistent with those for the antibiotic prescribing proportion. In additional analyses, we found that the incidence rate ratio associating the antibiotic prescribing proportion with pneumonia was similar for the population aged less than 65 years and 65 years or older. Table 3 Associations of consultation and prescribing rates and proportions with infective complications. Incident rate ratios (IRR) are for a 10 unit increment in rate or proportion Infective complications RTI consultation rate (for 10 per 1000 increase) Antibiotic prescribing rate for RTI (for 10 per 1000 increase) Proportion of RTI consultations with antibiotic prescribed (for 10% increase) IRR* (95% CI) P value IRR† (95% CI) P value IRR† (95% CI) P value Pneumonia 1.015 (1.008 to 1.023) <0.001 0.959 (0.941 to 0.976) <0.001 0.87 (0.83 to 0.92) <0.001 Peritonsillar abcess 1.004 (0.998 to 1.010) 0.18 0.968 (0.953 to 0.982) <0.001 0.90 (0.86 to 0.94) <0.001 Mastoiditis 1.020 (1.005 to 1.034) 0.007 1.008 (0.973 to 1.044) 0.67 1.00 (0.90 to 1.12) 0.95 Empyema 1.005 (0.995 to 1.016) 0.35 0.979 (0.953 to 1.005) 0.11 0.93 (0.86 to 1.01) 0.10 Bacterial meninigitis 1.001 (0.987 to 1.016) 0.86 0.986 (0.949 to 1.023) 0.45 0.94 (0.84 to 1.06) 0.30 Intracranial abscess‡ 1.003 (0.984 to 1.022) 0.75 0.986 (0.938 to 1.035) 0.57 0.94 (0.81 to 1.09) 0.40 *Adjusted for sex, age group, region, deprivation fifth, and clustering by general practice.

3 (0.86 to 1.01) 0.10 Bacterial meninigitis 1.001 (0.987 to 1.016) 0.86 0.986 (0.949 to 1.023) 0.45 0.94 (0.84 to 1.06) 0.30 Intracranial abscess‡ 1.003 (0.984 to 1.022) 0.75 0.986 (0.938 to 1.035) 0.57 0.94 (0.81 to 1.09) 0.40 *Adjusted for sex, age group, region, deprivation fifth, and clustering by general practice. †Adjusted for consultation rate for respiratory tract infections, sex, age group, region, deprivation fifth, and clustering by general practice. ‡Adjustment for general practice omitted. A general practice with the mean list size for England of 7000 registered patients is expected to have 20 300 consultations (95% range 11 340 to 30 380) for RTIs over 10 years (table 4). A general practice of this size, with an average RTI consultation rate, might issue 13 195 (11 744 to 16 037) antibiotic prescriptions during this period if it is in the highest prescribing fourth. If the practice reduces the proportion of RTI consultations with antibiotics prescribed by 10% it will issue 2030 (1134 to 3038) fewer antibiotic prescriptions for RTIs. This reduction in antibiotic prescribing is expected to be associated with 1.1 (0.6 to 1.5) more cases of pneumonia each year and 0.9 (0.5 to 1.3) more cases of peritonsillar abscess each decade (table 4). The number of cases of mastoiditis, empyema, bacterial meningitis, intracranial abscess, and Lemierre’s syndrome are not expected to increase. Table 4 Expected number of events over 10 years in a hypothetical high antibiotic prescribing general practice with 7000 patients

A general practice with the mean list size for England of 7000 registered patients is expected to have 20 300 consultations (95% range 11 340 to 30 380) for RTIs over 10 years (table 4). A general practice of this size, with an average RTI consultation rate, might issue 13 195 (11 744 to 16 037) antibiotic prescriptions during this period if it is in the highest prescribing fourth. If the practice reduces the proportion of RTI consultations with antibiotics prescribed by 10% it will issue 2030 (1134 to 3038) fewer antibiotic prescriptions for RTIs. This reduction in antibiotic prescribing is expected to be associated with 1.1 (0.6 to 1.5) more cases of pneumonia each year and 0.9 (0.5 to 1.3) more cases of peritonsillar abscess each decade (table 4). The number of cases of mastoiditis, empyema, bacterial meningitis, intracranial abscess, and Lemierre’s syndrome are not expected to increase. Table 4 Expected number of events over 10 years in a hypothetical high antibiotic prescribing general practice with 7000 patients Measures Median (95% range) over 10 years No expected in general practice with 7000 patients Change after 10% absolute decrease in proportion of RTI consultations with antibiotics prescribed No of RTI consultations 20 300 (11 340 to 30 380) 0 Antibiotic prescriptions for RTI 13 195 (11 744 to 16 037)* −2030 (−1134 to −3038)† No of first episodes: Pneumonia 83 (82 to 85) 11 (6 to 15) Peritonsillar abscess 9 (9 to 9) 0.9 (0.5 to 1.3) Mastoiditis 2 (2 to 3) 0 Empyema 3 (2 to 3) 0 Bacterial meningitis 2 (1 to 2) 0 Intracranial abscess <1 0 RTI=respiratory tract infection.

prescriptions for RTI 13 195 (11 744 to 16 037)* −2030 (−1134 to −3038)† No of first episodes: Pneumonia 83 (82 to 85) 11 (6 to 15) Peritonsillar abscess 9 (9 to 9) 0.9 (0.5 to 1.3) Mastoiditis 2 (2 to 3) 0 Empyema 3 (2 to 3) 0 Bacterial meningitis 2 (1 to 2) 0 Intracranial abscess <1 0 RTI=respiratory tract infection. *Assuming average RTI consultation rate. †Assuming no change in RTI consultation rate.

prescriptions for RTI 13 195 (11 744 to 16 037)* −2030 (−1134 to −3038)† No of first episodes: Pneumonia 83 (82 to 85) 11 (6 to 15) Peritonsillar abscess 9 (9 to 9) 0.9 (0.5 to 1.3) Mastoiditis 2 (2 to 3) 0 Empyema 3 (2 to 3) 0 Bacterial meningitis 2 (1 to 2) 0 Intracranial abscess <1 0 RTI=respiratory tract infection. *Assuming average RTI consultation rate. †Assuming no change in RTI consultation rate. Discussion We used a large dataset of electronic health records to investigate the safety of reducing unnecessary antibiotic prescribing for respiratory tract infections (RTIs) in primary care. The results show that general practices prescribing fewer antibiotics for RTIs may expect to have a slightly higher incidence of pneumonia and peritonsillar abscess than higher prescribing general practices. If a general practice with an average list size of 7000 patients reduced the proportion of RTI consultations with antibiotics prescribed by 10%, it might encounter about one additional case of pneumonia each year and one additional case of peritonsillar abscess each decade. Changes will be proportionately greater for larger reductions in antibiotic prescribing. These estimates represent averages across general practice populations, but complications might be fewer than expected if general practitioners are able effectively to stratify antibiotic prescribing according to level of risk. There was no evidence that diagnoses of mastoiditis, empyema, bacterial meningitis, or intracranial abscess might increase. Lemierre’s syndrome was rare, with about one case per two million person years, but there was no evidence that this was more common at low prescribing practices. This is reassuring in view of recent suggestions that F necrophorum may often be present in patients with sore throat.18 These estimates must be viewed in the context of quantitatively important declining secular trends in incidence for several infective complications of RTI, including peritonsillar abscess, mastoiditis, and meningitis. Bacterial meningitis from pneumococcal, meningococcal, or haemophilus infection has declined after the introduction of vaccination programmes.28 However, the incidence of pneumonia showed a slight increase over time, consistent with previous studies based on hospital admissions.29 30

scess, mastoiditis, and meningitis. Bacterial meningitis from pneumococcal, meningococcal, or haemophilus infection has declined after the introduction of vaccination programmes.28 However, the incidence of pneumonia showed a slight increase over time, consistent with previous studies based on hospital admissions.29 30 Reducing the proportion of RTI consultations with antibiotics prescribed by 10% is expected to be accompanied by some 2000 fewer antibiotic prescriptions for each practice over 10 years. Benefits to individual patients from avoiding antibiotics include reductions in common adverse reactions to antibiotics, such as rashes, vomiting, and diarrhoea, which may affect 10% of patients,6 as well as less common side effects such as anaphylaxis. Benefits to general practices may include a demedicalisation of RTIs followed by a decline in the rate of consultations, since previous observational studies show that higher prescribing general practices receive more consultations for RTIs.31 Trial evidence shows that even one antibiotic prescription increases the likelihood of reconsultation for a new episode of an RTI.32 Most of the complications identified do not require hospital admission and currently respond well to antibiotics, so simply the occurrence of an uncommon complication rate is not in itself a justification for more widespread prescribing of antibiotics for initially uncomplicated presentations. The results did not support a threshold for safe or unsafe prescribing levels. Inspection of forest plots suggested some departure from linearity, but addition of non-linear terms did not improve the goodness of fit of regression models.

e widespread prescribing of antibiotics for initially uncomplicated presentations. The results did not support a threshold for safe or unsafe prescribing levels. Inspection of forest plots suggested some departure from linearity, but addition of non-linear terms did not improve the goodness of fit of regression models. Strengths and weaknesses in relation to other studies Previous studies have consistently shown high levels of unnecessary prescribing of antibiotics for RTIs in primary care.33 While there has been a declining trend in the consultation rate for RTIs,7 the proportion of consultations with antibiotics prescribed has changed little,7 despite the efforts of researchers, clinicians, and policymakers to bring about changes. General practitioners may often be concerned to meet patients’ expectations for antibiotic prescriptions,34 but both patients and prescribers might also have concerns about the safety of non-prescribing strategies.34 One study14 provided evidence that antibiotics reduced the risk of pneumonia, mastoiditis, and peritonsillar abscess but did not quantify the potential population impact of these complications.

prescriptions,34 but both patients and prescribers might also have concerns about the safety of non-prescribing strategies.34 One study14 provided evidence that antibiotics reduced the risk of pneumonia, mastoiditis, and peritonsillar abscess but did not quantify the potential population impact of these complications. The present results represent averages across general practice populations. Diversity among the population of patients at risk of RTIs is considerable. Current management guidelines for RTIs recommend that specific groups of patients should be considered to have positive indications for antibiotic treatment. An immediate antibiotic prescription is recommended if patients have clinical features suggestive of serious illness or complications,35 have comorbidities, or are very young or very old.3 Further research is needed to evaluate whether the present results will be confirmed when subgroups that might be at higher risk, including older adults, are analysed separately. It is possible that general practices with the same overall level of antibiotic prescribing may differ in the appropriateness of their management of patients with defined markers of vulnerability, and this could influence the rate of complications. However, the clinical features of an RTI episode may have only limited predictive value for the future occurrence of complications, and a high proportion of complications might occur in patients who seem to be at low risk.36 A delayed antibiotic prescribing strategy, in which a prescription is issued but only used if symptoms fail to improve, is sometimes recommended as a method for reducing antibiotic utilisation in the management of RTIs.3 37 Delayed antibiotic prescribing may be as effective as immediate use of antibiotics in the prevention of complications of sore throat.4 The development and application of point-of-care testing to guide antibiotic prescribing might have a future role in identifying those who would potentially benefit from antibiotic treatment.38 39

ibiotic prescribing may be as effective as immediate use of antibiotics in the prevention of complications of sore throat.4 The development and application of point-of-care testing to guide antibiotic prescribing might have a future role in identifying those who would potentially benefit from antibiotic treatment.38 39 Strengths and weaknesses of this study This study included more than 600 general practices, with a registered population of more than four million patients and 45 million person years of observation. Consequently, the study provided precise estimates for the more common outcomes evaluated. We acknowledge that there was lower power to evaluate potential changes in less common outcomes. We can conclude that the absolute risks of mastoiditis, empyema, intracranial abscess, and Lemierre’s syndrome remain small, even in practices with low rates of antibiotic prescribing. This study adopted a population perspective, aiming to quantify the outcomes of either high prescribing or low prescribing strategies in the management of RTIs. Consequently, we evaluated changes in infective complications at the level of the general practice population. The research did not deal with variation in prescribing at the level of the individual doctor. The research did not show whether individual patients who experienced complications received antibiotics. Conclusions might differ if individual level analyses showed that complications arise in patients who were treated with antibiotics. We did not evaluate the outcomes of individual patients identified as having complications in this study. Further research is required to evaluate the severity of complications, such as pneumonia, and their outcomes, including mortality. Future studies might also make use of linked hospital episode data, which in recent years have become available for selected CPRD practices in England to evaluate patients in more detail who have been admitted to hospital. The risk of complications associated with different classes of antibiotics also merits study. We acknowledge that there may be other complications, such as a proportion of all cases of septicaemia diagnosed in primary care, which might follow from an RTI.

uate patients in more detail who have been admitted to hospital. The risk of complications associated with different classes of antibiotics also merits study. We acknowledge that there may be other complications, such as a proportion of all cases of septicaemia diagnosed in primary care, which might follow from an RTI. We acknowledge several sources of misclassification: we used a sample of the UK Clinical Practice Research Datalink (CPRD) to estimate consultation and prescribing rates; there is variation among general practices in the use of diagnostic categories40; general practice populations may vary in their use of out-of-hours and emergency services, whose generally higher antibiotic prescribing may not be captured in CPRD; and some general practices may use delayed antibiotic prescribing strategies,4 but these were not distinguished in the analysis of prescriptions issued. Use of near patient testing might possibly have contributed to better diagnosis during the period. These forms of misclassification generally tend to diminish estimated associations but might cause bias if effects are differentially distributed across prescribing categories. Diagnostic coding may have a subjective element41 and bias might arise if low prescribing practices are more likely than high prescribing practices to code pneumonia to justify the prescription of an antibiotic. The research utilised non-randomised data, and we adjusted for age, sex, region, deprivation category, and general practice, but it is possible unmeasured confounders might have biased the reported associations. We also caution that, in the analysis of large datasets “significant” results must be judged in relation to their clinical importance. Antibiotic prescribing in the UK is high compared with some international comparators, and we cannot be sure that the associations reported here would also hold at very low antibiotic prescribing levels.

at, in the analysis of large datasets “significant” results must be judged in relation to their clinical importance. Antibiotic prescribing in the UK is high compared with some international comparators, and we cannot be sure that the associations reported here would also hold at very low antibiotic prescribing levels. Meaning of the study: possible explanations and implications for clinicians and policymakers This study provides evidence that general practices prescribing antibiotics less often at consultations for RTIs may experience a slight increase in the incidence of pneumonia and peritonsillar abscess, both of which would be expected to respond to treatment while bacterial pathogens remain sensitive to antibiotics. No increase is likely in mastoiditis, empyema, meningitis, intracranial abscess, or Lemierre’s syndrome. Even a large reduction in antibiotic prescribing was predicted to be associated with only a small increase in numbers of cases observed over a 10 year period, and this would be expected to reduce the risks of antibiotic resistance, the side effects of antibiotics, and the medicalisation of largely self limiting illnesses. The safety outcomes of no antibiotic prescribing strategies for RTIs are an important aspect for communication to patients and the public in the context of wider communication strategies to support antimicrobial stewardship.42

, the side effects of antibiotics, and the medicalisation of largely self limiting illnesses. The safety outcomes of no antibiotic prescribing strategies for RTIs are an important aspect for communication to patients and the public in the context of wider communication strategies to support antimicrobial stewardship.42 Unanswered questions and future research Further research is needed to quantify associations based on individual patient characteristics and consultation patterns in primary care, particularly in children and older adults, as we used age standardised prescribing measures. However, associations might vary in different age groups. We also recommend that future randomised studies should be sufficiently large to evaluate safety outcomes of strategies to reduce antibiotic prescribing. What is already known on this topic Widespread unnecessary utilisation of antibiotics is leading to an increase in antimicrobial drug resistance Many respiratory tract infections (RTIs) are largely self limiting, but antibiotics continue to be prescribed for about 50% of consultations for RTIs in primary care RTIs are infrequently associated with complications, including pneumonia, peritonsillar abscess, mastoiditis, meningitis, intracranial abscess, and Lemierre’s syndrome, but it is not known whether these are more common in general practices that prescribe antibiotics less often What this study adds General practices prescribing antibiotics less often for RTIs had slightly higher rates of pneumonia and peritonsillar abscess than higher prescribing practices

RTIs are infrequently associated with complications, including pneumonia, peritonsillar abscess, mastoiditis, meningitis, intracranial abscess, and Lemierre’s syndrome, but it is not known whether these are more common in general practices that prescribe antibiotics less often What this study adds General practices prescribing antibiotics less often for RTIs had slightly higher rates of pneumonia and peritonsillar abscess than higher prescribing practices There was no evidence that mastoiditis, empyema, meningitis, intracranial abscess, or Lemierre’s syndrome were more frequent at low prescribing practices Even a substantial reduction in antibiotic prescribing was predicted to be associated with only a small increase in numbers of cases observed, and this would be expected to reduce the risks of antibiotic resistance, the side effects of antibiotics, and the medicalisation of largely self limiting illnesses Contributors: MCG, MVM, PL, ADH, RF, DJ, and MA contributed to the idea and design of the study. MCG analysed the data. JC checked and replicated the analysis. ATP advised on statistical analysis. MCG drafted the paper. All authors contributed to and approved the final draft. MCG is guarantor.

Even a substantial reduction in antibiotic prescribing was predicted to be associated with only a small increase in numbers of cases observed, and this would be expected to reduce the risks of antibiotic resistance, the side effects of antibiotics, and the medicalisation of largely self limiting illnesses Contributors: MCG, MVM, PL, ADH, RF, DJ, and MA contributed to the idea and design of the study. MCG analysed the data. JC checked and replicated the analysis. ATP advised on statistical analysis. MCG drafted the paper. All authors contributed to and approved the final draft. MCG is guarantor. Funding: The research was supported by the UK National Institute for Health Research Health Technology Assessment programme initiative on antimicrobial drug resistance. The funders did not engage in the design and conduct of the study; collection, management, analysis, and interpretation of the data; or preparation, review, and approval of the manuscript. MCG and ATP were also supported by the National Institute of Health Research (NIHR) Biomedical Research Centre at Guy’s and St Thomas’ Hospitals. The views expressed are those of the author(s) and not necessarily those of the National Health Service, the NIHR, or the Department of Health. This study is based in part on data from the Clinical Practice Research Datalink obtained under licence from the UK Medicines and Healthcare products Regulatory Agency. However, the interpretation and conclusions contained in this report are those of the authors alone.

th Service, the NIHR, or the Department of Health. This study is based in part on data from the Clinical Practice Research Datalink obtained under licence from the UK Medicines and Healthcare products Regulatory Agency. However, the interpretation and conclusions contained in this report are those of the authors alone. Competing interests: All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf and declare: no support from any organisation for the submitted work; no financial relationships with any organisations that might have an interest in the submitted work in the previous three years; no other relationships or activities that could appear to have influenced the submitted work. Ethical approval: The protocol for the study was approved by the Clinical Practice Research Datalink (CPRD) independent scientific advisory committee, reference 14_130A2. The CPRD has broad National Research Ethics Service Committee ethics approval for observational research studies. The research was conducted to inform a study funded by the NIHR Health Technology Assessment programme. The outcomes were selected from a wider range of safety outcomes proposed for the study because they were identified as being potentially associated with respiratory tract infections. Data sharing: CPRD data were analysed under licence and are not available for sharing.

Ethical approval: The protocol for the study was approved by the Clinical Practice Research Datalink (CPRD) independent scientific advisory committee, reference 14_130A2. The CPRD has broad National Research Ethics Service Committee ethics approval for observational research studies. The research was conducted to inform a study funded by the NIHR Health Technology Assessment programme. The outcomes were selected from a wider range of safety outcomes proposed for the study because they were identified as being potentially associated with respiratory tract infections. Data sharing: CPRD data were analysed under licence and are not available for sharing. Transparency: The guarantor (MCG) affirms that the manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained.

Introduction Cardiovascular disease and heart failure are major causes of morbidity and mortality in people with type 2 diabetes.1 2 Once heart failure is present in people with diabetes, mortality is increased 10-fold and five year survival is only 12.5%, a prognosis worse than for metastatic breast cancer.2 Several diabetes drugs have been associated with an unexpected increase in risk of heart failure during both clinical trials3 and post-marketing surveillance raising concerns about the overall risks and benefits for people with diabetes.4 5

r survival is only 12.5%, a prognosis worse than for metastatic breast cancer.2 Several diabetes drugs have been associated with an unexpected increase in risk of heart failure during both clinical trials3 and post-marketing surveillance raising concerns about the overall risks and benefits for people with diabetes.4 5 Following its launch in 2000,6 rosiglitazone, the first drug in the “insulin sensitising” thiazolidinedione class, was associated with an increased rate of heart failure.5 This resulted in its withdrawal from Europe, India, New Zealand, and South Africa in 2010-11. Rosiglitazone is, however, still prescribed in the United States—a controversial decision7 informed by the open label “non-inferiority” rosiglitazone evaluated for cardiovascular outcomes in oral agent combination therapy for type 2 diabetes (RECORD) trial funded by the manufacturers of rosiglitazone (Avandia; GlaxoSmithKline).4 The RECORD study assessed cardiovascular outcomes in 2220 people prescribed rosiglitazone in combination with either metformin or sulphonylureas compared with 2227 prescribed both metformin and sulphonylureas between 2001 and 2008. Although the numbers of events were low, the trial reported an increased risk of heart failure with rosiglitazone and was unable to rule out an increased risk of myocardial infarction.4 The design, results, and interpretation of this trial have been heavily criticised.8

tformin and sulphonylureas between 2001 and 2008. Although the numbers of events were low, the trial reported an increased risk of heart failure with rosiglitazone and was unable to rule out an increased risk of myocardial infarction.4 The design, results, and interpretation of this trial have been heavily criticised.8 Pioglitazone is another thiazolidinedione that decreases blood glucose levels. The placebo controlled PROactive trial of pioglitazone9 was also controversial in its design and interpretation, largely because of the choice of composite endpoints.10 Although the trial failed to clearly show improved cardiovascular outcomes for patients, it reported increased hospital admissions for heart failure as an adverse effect,9 as did a subsequent meta-analysis of 8554 patients prescribed pioglitazone.11 Meanwhile, a Canadian cohort study comparing pioglitazone with rosiglitazone between 2002 and 2008 reported a lower risk of heart failure and death in pioglitazone users.12 Similarly, a US cohort study reported a lower risk of stroke, heart failure, and all cause mortality among patients prescribed pioglitazone compared with rosiglitazone.13 Pioglitazone continues to be prescribed in the United Kingdom and the US although it has been withdrawn elsewhere owing to concerns about an increased risk of bladder cancer.14

y reported a lower risk of stroke, heart failure, and all cause mortality among patients prescribed pioglitazone compared with rosiglitazone.13 Pioglitazone continues to be prescribed in the United Kingdom and the US although it has been withdrawn elsewhere owing to concerns about an increased risk of bladder cancer.14 Dipeptidyl peptidase-4 (DPP-4) inhibitors, also known as gliptins, are a relatively new class of diabetes drug that are included in international guidelines15 as second line agents after metformin, although data on long term clinical benefits and safety are inconclusive.16 In a placebo controlled clinical trial, saxagliptin was associated with an unexpected 27% increase in admissions to hospital from heart failure.3 In the Examination of Cardiovascular Outcomes with Alogliptin versus Standard of Care (EXAMINE) trial ,17 alogliptin did not significantly increase overall risk of hospital admissions for heart failure compared with placebo. Similarly, in the Trial Evaluating Cardiovascular Outcomes with Sitagliptin (TECOS) trial, sitagliptin was not associated with an increased risk of heart failure compared with placebo.18 Observational studies have also found inconsistent associations. Sitagliptin was associated with an increased risk of heart failure in a cohort study of 8288 Taiwanese patients using sitagliptin over 1.5 years.19 Conversely, a US cohort study of 8032 sitagliptin users showed no excess risk of hospital admission or death compared with users of other glucose lowering agents.20 Although heart failure was included within the composite endpoint, it was not evaluated separately. A meta-analysis of 25 trials of 7726 patients receiving sitagliptin or a comparator agent between 12 weeks and two years was undertaken by the manufacturers of sitagliptin (Januvia; MSD).21 The composite study endpoints included major adverse cardiovascular events (defined as ischaemic events or cardiovascular deaths), but the study did not specifically evaluate risk of heart failure. In summary, the findings from clinical trials and observational studies are inconsistent, which may reflect differences in study design, study duration, individual drugs, or outcome measures.

vents (defined as ischaemic events or cardiovascular deaths), but the study did not specifically evaluate risk of heart failure. In summary, the findings from clinical trials and observational studies are inconsistent, which may reflect differences in study design, study duration, individual drugs, or outcome measures. Uncertainty remains over the longer term comparative risks among patients prescribed different diabetes drugs, particularly gliptins and glitazones alone and in combination with other diabetes drugs.22 23 Regulatory agencies have responded to this uncertainty by requiring evidence that new diabetes drugs are not associated with harmful increases in cardiovascular events rather than the more stringent requirement that the drugs result in evidence of clinical benefit.24 25

nd in combination with other diabetes drugs.22 23 Regulatory agencies have responded to this uncertainty by requiring evidence that new diabetes drugs are not associated with harmful increases in cardiovascular events rather than the more stringent requirement that the drugs result in evidence of clinical benefit.24 25 Concerns have also been raised about the safety of sulphonylureas, an older class of oral diabetes drug, as these have been linked with increased adverse cardiovascular events in some5 but not all studies.26 The lifelong nature of diabetes, the noticeable increase in its incidence and prevalence, and prescribing recommendations in guidelines,15 mean that the number of people prescribed diabetes drugs is likely to increase. Given the impracticability and ethical difficulties of head-to-head trials comparing different agents, the risks of clinical outcomes need to be quantified in large representative populations of people prescribed these drugs over longer periods. This information, which is available from large longitudinal observational databases, can complement information from meta-analyses of clinical trials that, while valuable, are prone to publication bias and lack sufficient detail, duration of follow-up, or the power to make relevant comparisons for unintended effects.21 23

information, which is available from large longitudinal observational databases, can complement information from meta-analyses of clinical trials that, while valuable, are prone to publication bias and lack sufficient detail, duration of follow-up, or the power to make relevant comparisons for unintended effects.21 23 We therefore carried out a cohort study using a large UK primary care database with linked general practitioner, mortality, and hospital admissions data to investigate the associations between different classes of diabetes drugs and the risks of heart failure, cardiovascular disease, and all cause mortality for people with type 2 diabetes. We were particularly interested in the risks associated with the newer agents, including glitazones and gliptins. In a companion paper we reported on a similar analysis examining the risks of microvascular complications (severe kidney disease, blindness, lower limb amputation), hyperglycaemia, and hypoglycaemia between different classes of diabetes drugs in people with type 2 diabetes.27

agents, including glitazones and gliptins. In a companion paper we reported on a similar analysis examining the risks of microvascular complications (severe kidney disease, blindness, lower limb amputation), hyperglycaemia, and hypoglycaemia between different classes of diabetes drugs in people with type 2 diabetes.27 Methods We did a population based open cohort study of people in England aged 25-84 years with a diagnosis of type 2 diabetes. We used a large population of primary care patients derived from version 40 of the QResearch database (www.qresearch.org). QResearch is a continually updated patient level pseudonymised database with event level data extending back to 1989. QResearch currently includes clinical and demographic data from over 1243 general practices in England and two practices in Scotland, covering a population of over 24 million patients, and collected in the course of routine healthcare by general practitioners and associated staff. The primary care data include demographic information, diagnoses, prescriptions, referrals, laboratory test results, and clinical values. Diagnoses are recorded using the Read code classification.28 QResearch has been used for a wide range of clinical research, including the assessment of unintended effects of commonly prescribed medicines.29 30 31 32 33 34 The primary care data are linked at individual patient level to hospital episode statistics and mortality records from the Office for National Statistics. Hospital episode statistics provides details of all National Health Service inpatient admissions since 1997, including primary and secondary causes coded using the ICD-10 (international classification of diseases, 10th revision) classifications and OPCS-4 (Office of Population Censuses and Surveys, fourth revision) codes for operations and interventions. ONS provides details of all deaths in England with primary and underlying causes, also coded using the ICD-10 classification. Patient records are linked using a project specific pseudonymised NHS number, which is valid and complete for 99.8% of primary care patients, 99.9% of ONS mortality records, and 98% of hospital admissions records.1

of all deaths in England with primary and underlying causes, also coded using the ICD-10 classification. Patient records are linked using a project specific pseudonymised NHS number, which is valid and complete for 99.8% of primary care patients, 99.9% of ONS mortality records, and 98% of hospital admissions records.1 Inclusion and exclusion criteria The study included all QResearch practices in England that had been using the Egton Medical Information Systems (EMIS) computer system for at least a year. We initially identified an open cohort of people aged 25-84 years registered with eligible practices between 1 April 2007 and 31 January 2015. This time interval was chosen because both pioglitazone and gliptins were available in the UK during the full study period. We then selected people with diabetes if they had a Read code for diabetes or more than one prescription for a diabetes drug.

ered with eligible practices between 1 April 2007 and 31 January 2015. This time interval was chosen because both pioglitazone and gliptins were available in the UK during the full study period. We then selected people with diabetes if they had a Read code for diabetes or more than one prescription for a diabetes drug. We excluded people as having type 1 diabetes if they had received a diagnosis aged less than 35 and had been prescribed insulin.35 We also excluded patients without a postcode related deprivation score. For each patient we determined an entry date to the cohort, which was the latest of the following: date of diagnosis of diabetes, 25th birthday, date of registration with the practice plus one year, date on which the practice computer system was installed plus one year, and the beginning of the study period. To reduce bias we used an incident user design for people prescribed glitazones, gliptins (our main drugs of interest), or insulin.36 As in other studies,20 we defined incident users as people without a prescription for these drugs in the 12 months before the study entry date, and we excluded people who had received any of these drugs in the previous 12 months. We included prevalent users of metformin or sulphonylureas in the study cohort; if we had excluded them the numbers of new users of glitazones and gliptins—our main exposures of interest—would have been substantially reduced because these drugs are usually prescribed after monotherapy with metformin or sulphonylureas. People with an existing diagnosis of an outcome of interest at the study entry date were also excluded from the analysis of that outcome. Patients were censored at the earliest date of the first recorded diagnosis of the outcome of interest, death, deregistration with the practice, last upload of computerised data, or the end date of the study (31 January 2015).

ome of interest at the study entry date were also excluded from the analysis of that outcome. Patients were censored at the earliest date of the first recorded diagnosis of the outcome of interest, death, deregistration with the practice, last upload of computerised data, or the end date of the study (31 January 2015). Outcomes Our primary outcomes were incident heart failure, cardiovascular disease, and all cause mortality, recorded in either the patient’s primary care record, linked hospital record, or mortality record. Definition of outcomes We used Read codes to identify recorded diagnoses of heart failure from the primary care records (G58%, G5yy9, G5yyA, 662f, 662g, 662h, and 662i). To identify incident cases of heart failure from hospital and mortality records, we used ICD-10 clinical codes (I110, I130, I42, and I50). We used the earliest recorded date of heart failure on any of the three data sources as the index date for the diagnosis of heart failure.

G5yy9, G5yyA, 662f, 662g, 662h, and 662i). To identify incident cases of heart failure from hospital and mortality records, we used ICD-10 clinical codes (I110, I130, I42, and I50). We used the earliest recorded date of heart failure on any of the three data sources as the index date for the diagnosis of heart failure. Our definition of cardiovascular disease included coronary heart disease (angina and myocardial infarction), stroke, or transient ischaemic attacks but not peripheral vascular disease. The supplementary file lists the Read codes used for case identification on the primary care record. The ICD-10 codes used for case identification on the ONS death certificate or hospital admission records were: angina pectoris (I20), acute myocardial infarction (I22), complications after acute myocardial infarction (I23), other acute ischaemic heart disease (I24), chronic ischaemic heart disease (I25), and ischaemic stroke (I63, I64) or transient ischaemic attack (G45). We used the earliest recorded date of cardiovascular disease on any of the three data sources as the index date for the diagnosis of cardiovascular disease. All cause mortality was defined by the status of death recorded in the general practice systems linked to the date and cause of death as recorded on the ONS mortality record.

Our definition of cardiovascular disease included coronary heart disease (angina and myocardial infarction), stroke, or transient ischaemic attacks but not peripheral vascular disease. The supplementary file lists the Read codes used for case identification on the primary care record. The ICD-10 codes used for case identification on the ONS death certificate or hospital admission records were: angina pectoris (I20), acute myocardial infarction (I22), complications after acute myocardial infarction (I23), other acute ischaemic heart disease (I24), chronic ischaemic heart disease (I25), and ischaemic stroke (I63, I64) or transient ischaemic attack (G45). We used the earliest recorded date of cardiovascular disease on any of the three data sources as the index date for the diagnosis of cardiovascular disease. All cause mortality was defined by the status of death recorded in the general practice systems linked to the date and cause of death as recorded on the ONS mortality record. Exposure data Our primary exposures of interest were new use of gliptins and new use of glitazones during the study period. For each participant we extracted details of all individual prescriptions for all types of diabetes drug, including the prescription date and the type of drug. We partitioned the follow-up time into treatment periods, where each period corresponded to treatment with a particular type or combination of diabetes drug, or could be a period of no treatment with any diabetes drugs. If the patient changed to a different type of treatment or to a different combination of treatments, we classified that as a separate treatment period. For example, if a patient was prescribed metformin alone on entry to the cohort for 12 months and then was prescribed both glitazones and metformin for a further 24 months and then had a treatment free period for six months until they were censored, they would have three treatment periods (metformin only for 12 months, metformin and glitazones for 24 months, and no treatment for six months).

ohort for 12 months and then was prescribed both glitazones and metformin for a further 24 months and then had a treatment free period for six months until they were censored, they would have three treatment periods (metformin only for 12 months, metformin and glitazones for 24 months, and no treatment for six months). We determined the duration of each treatment period by calculating the number of days between the earliest issue date and the latest issue date plus 90 days for the type of treatment prescribed. If another treatment was added before the initial treatment was stopped then we calculated the duration of the treatment period on the initial treatment alone to be the number of days between the earliest issue date for the initial treatment and the earliest issue date for the next treatment. We added 90 days to the last prescription date as an estimate of the date on which the patient stopped treatment (the “stop date”), we made this assumption to allow for events that occur during a withdrawal period to be attributed to the drug rather than being counted as unexposed time. For the analysis, we used six binary exposure variables for each treatment period to indicate treatment with any of the diabetes drugs, grouped into six drug classes—glitazones (including rosiglitazone and pioglitazone), gliptins, metformin, sulphonylureas, insulin, and other oral diabetes drugs (including α-glucosidase inhibitors, sodium-glucose cotransporter 2 inhibitors, glinides, guar). This accounted for patients receiving different combinations of these drugs during a treatment period. To further assess associations for different specific treatment combinations (such as dual treatment with metformin and glitazones) we also categorised treatments during each treatment period into one categorical variable with 21 mutually exclusive treatment categories, including a no current treatment group and 20 categories for monotherapy, dual treatment, and triple combinations of drugs.

ns (such as dual treatment with metformin and glitazones) we also categorised treatments during each treatment period into one categorical variable with 21 mutually exclusive treatment categories, including a no current treatment group and 20 categories for monotherapy, dual treatment, and triple combinations of drugs. Confounding variables We considered confounding variables that were likely to be associated with the risk of the complications from diabetes20 37 38 39 40 or with the likelihood of receiving treatment with different diabetes drugs. These included age at study entry, sex, number of years since diabetes was diagnosed (<1 year and 1-3, 4-6, 7-10, and ≥ 11 years),40 calendar year, smoking status (non-smoker; former smoker; light smoker, 1-9 cigarettes/day; moderate smoker, 10-19 cigarettes/day; heavy smoker, ≥20 cigarettes/day; not recorded), ethnic group (white/not recorded, Indian, Pakistani, Bangladeshi, other Asian, black African, black Caribbean, Chinese, other, including mixed),37 Townsend deprivation score, previous diabetes complications (severe kidney failure,40 ≥1 episodes of hyperglycaemia, ≥1 episodes of hypoglycaemia, lower limb amputation, blindness), comorbidities (cardiovascular disease40 (other than when cardiovascular disease was the outcome of interest), heart failure (other than when heart failure was the outcome of interest), peripheral vascular disease, valvular heart disease, chronic kidney disease, atrial fibrillation,37 hypertension,37 rheumatoid arthritis37)), prescription drugs (statins, aspirin, anticoagulants, thiazides, angiotensin converting enzyme inhibitors, angiotensin receptor blockers, calcium channel blockers), and clinical values (body mass index kg/m,2 40 cholesterol to high density lipoprotein cholesterol ratio,37 systolic blood pressure (mm Hg),37 serum creatinine level, glycated haemoglobin A1c (mmol/mol).40 41 We evaluated confounders at the start of each treatment period for comorbidities, previous complications, other prescribed drugs, smoking status, and clinical values. For comorbidities and previous complications, we identified whether patients had a diagnosis recorded before the relevant treatment period. For prescribed drugs, we defined patients as treated at the start of the relevant period of diabetes drug treatment if they had at least two prescriptions for the other type of drug, including one in the 28 days before the treatment period and one after the start date.

sis recorded before the relevant treatment period. For prescribed drugs, we defined patients as treated at the start of the relevant period of diabetes drug treatment if they had at least two prescriptions for the other type of drug, including one in the 28 days before the treatment period and one after the start date. For smoking status and continuous variables (systolic blood pressure, body mass index, creatinine level, cholesterol to high density lipoprotein cholesterol ratio, and haemoglobin A1c), we used the most recent recorded value immediately before the relevant treatment period.

sis recorded before the relevant treatment period. For prescribed drugs, we defined patients as treated at the start of the relevant period of diabetes drug treatment if they had at least two prescriptions for the other type of drug, including one in the 28 days before the treatment period and one after the start date. For smoking status and continuous variables (systolic blood pressure, body mass index, creatinine level, cholesterol to high density lipoprotein cholesterol ratio, and haemoglobin A1c), we used the most recent recorded value immediately before the relevant treatment period. Statistical analysis Using Cox proportional hazards models we assessed the associations between the six different classes of diabetes drugs and risk of each of our three outcomes, adjusting for potential confounding variables. Rather than using a competing risks analysis we used the Cox model as it is considered more appropriate for analyses of causes such as in this study, whereas competing risks analyses tend to be more useful for prediction modelling or estimating absolute risks.42 43 44 To account for patients starting and stopping different diabetes treatments and changing between treatments, we included use of different diabetes drugs as time varying exposures. In the analysis, we calculated unadjusted and adjusted hazard ratios for the six different classes of diabetes drug (each as a binary variable indicating use or no use), with adjustment for the confounding variables and the other classes of diabetes drugs. We also calculated unadjusted and adjusted hazard ratios for the mutually exclusive treatment combinations comparing each treatment category with no current treatment. To determine whether there were significant differences between classes or individual drugs, we carried out Wald’s tests. We tested for interactions between the six different drug classes and age, sex, haemoglobin A1c, and body mass index. We used multiple imputation with chained equations to replace missing values for continuous values and smoking status and used these values in our main analyses.45 46 47 We did this for each of the study outcomes and in the imputation model included the censoring indicator for the outcome, the log of survival time, all the confounding variables, and the diabetes drug treatment variables. Before imputation we log transformed body mass index, haemoglobin A1c, creatinine level, and cholesterol to high density lipoprotein cholesterol ratio, as they had skewed distributions. We carried out five imputations, and combined the results using Rubin’s rules.

g variables, and the diabetes drug treatment variables. Before imputation we log transformed body mass index, haemoglobin A1c, creatinine level, and cholesterol to high density lipoprotein cholesterol ratio, as they had skewed distributions. We carried out five imputations, and combined the results using Rubin’s rules. To evaluate the robustness of our results and assess the impact of confounding variables we added the confounding variables to our model in blocks and compared the adjusted hazard ratios. Box 1 lists the models we assessed. Box 1: Types of models used in study Model A: diabetes drug classes adjusted for age, sex, ethnicity, deprivation, calendar year, duration of diabetes, plus other diabetes drugs Model B: model A plus comorbidities (hypertension, cardiovascular disease, heart failure, atrial fibrillation, chronic kidney disease, rheumatoid arthritis, valvular heart disease, peripheral vascular disease) plus previous complications (hypoglycaemia, hyperglycaemia, amputation, severe kidney failure, blindness) plus use of other drugs (statins, aspirin, anticoagulants, thiazides, angiotensin converting enzyme inhibitors or angiotensin receptor blockers, calcium channel blockers) Model C (primary analysis model): model B plus clinical values (body mass index, cholesterol to high density lipoprotein cholesterol ratio, systolic blood pressure, serum creatinine level, haemoglobin A1c) Model D: model C plus interaction terms Model E: treatment combinations categorical variable plus confounders in model C (models compared with no treatment and with metformin monotherapy)

Model C (primary analysis model): model B plus clinical values (body mass index, cholesterol to high density lipoprotein cholesterol ratio, systolic blood pressure, serum creatinine level, haemoglobin A1c) Model D: model C plus interaction terms Model E: treatment combinations categorical variable plus confounders in model C (models compared with no treatment and with metformin monotherapy) Model F: model C with prevalent users of sulphonylureas excluded Model G: model E with prevalent users of sulphonylureas excluded We also carried out a sensitivity analysis where we excluded prevalent users of sulphonylureas from the study cohort so that the hazard ratios for sulphonylureas are based on incident users, and we fitted the models F and G (see box 1). We used all the available data in the database to maximise the power and generalisability of the results. P values less than 0.01 (two tailed) were considered as significant and hazard ratios of 1.10 or more or 0.90 or less as clinically important. STATA (version 13.1) was used for all analyses. Patient involvement Patients were not involved in setting the research question, in the outcome measures, in the design, or in the implementation of the study. Patient representatives from the QResearch advisory board have written the information for patients on the QResearch website about the use of the database for research. They have also advised on dissemination, including the use of lay summaries describing the research and its results.

n the implementation of the study. Patient representatives from the QResearch advisory board have written the information for patients on the QResearch website about the use of the database for research. They have also advised on dissemination, including the use of lay summaries describing the research and its results. Results Overall, 1243 practices contributing to QResearch in England met the inclusion criteria. A cohort of 601 405 patients aged 25-84 years with diabetes was identified (fig 1). We sequentially excluded 31 224 people with type 1 diabetes (5.1%), 748 (0.1%) without a Townsend deprivation score, and 99 745 prescribed glitazones, gliptins, or insulin in the 12 months before the study entry date, leaving 469 688 patients with type 2 diabetes in the study cohort. Figure 1 also shows the numbers of patients with each outcome at baseline who were excluded from the analysis of that outcome, as well as the numbers of incident outcomes observed during follow-up. Fig 1 Flow of people through study Baseline characteristics In total 274 324 (58.4%) of the patients in the study cohort received prescriptions for one or more diabetes drugs during follow-up: 21 308 (4.5%) for glitazones, 32 533 (6.9%) for gliptins, 256 024 (54.5%) for metformin, 134 570 (28.7%) for sulphonylureas, 19 791 (4.2%) for insulin, and 12 062 (2.6%) for other oral diabetes drugs.

324 (58.4%) of the patients in the study cohort received prescriptions for one or more diabetes drugs during follow-up: 21 308 (4.5%) for glitazones, 32 533 (6.9%) for gliptins, 256 024 (54.5%) for metformin, 134 570 (28.7%) for sulphonylureas, 19 791 (4.2%) for insulin, and 12 062 (2.6%) for other oral diabetes drugs. Table 1 shows the characteristics of patients who started each of the six classes of diabetes drugs during follow-up based on the last recorded value before the drug was first prescribed (or at study entry for patients already prescribed sulphonylureas, metformin, or other diabetes drugs at baseline). The groups were similar for most characteristics except for higher levels of comorbidities other than hypertension in patients prescribed insulin, and lower levels of prescriptions for statins and aspirin in patients prescribed metformin compared with the other drugs. Table 1 Characteristics of patients with type 2 diabetes when starting drugs or at study entry for prevalent users. Values are numbers (percentages) unless stated otherwise

Table 1 shows the characteristics of patients who started each of the six classes of diabetes drugs during follow-up based on the last recorded value before the drug was first prescribed (or at study entry for patients already prescribed sulphonylureas, metformin, or other diabetes drugs at baseline). The groups were similar for most characteristics except for higher levels of comorbidities other than hypertension in patients prescribed insulin, and lower levels of prescriptions for statins and aspirin in patients prescribed metformin compared with the other drugs. Table 1 Characteristics of patients with type 2 diabetes when starting drugs or at study entry for prevalent users. Values are numbers (percentages) unless stated otherwise Characteristics Glitazones Gliptins Metformin Sulphonylureas Insulin Other diabetes drugs Total No of patients exposed 21 308 32 533 256 024 134 570 19 791 12 062 Median No of years exposed 4.5 5.7 4.8 4.9 5.9 4.9 Mean (SD) age at study entry 63.0 (11.9) 63.3 (12.1) 64.6 (13.1) 66.2 (12.9) 64.5 (12.7) 60.0 (11.9) Mean (SD) Townsend score 0.4 (3.5) 0.5 (3.5) 0.6 (3.6) 0.6 (3.6) 0.5 (3.6) 0.8 (3.6) Male 12 658 (59.4) 18 871 (58.0) 146 690 (57.3) 79 284 (58.9) 11 499 (58.1) 6509 (54.0) Ethnicity: Ethnicity recorded 19 130 (89.8) 29 396 (90.4) 228 962 (89.4) 119 507 (88.8) 17 264 (87.2) 10 947 (90.8) White or not recorded 17 112 (80.3) 26 104 (80.2) 204 915 (80.0) 107 537 (79.9) 17 001 (85.9) 10 135 (84.0) Indian 997 (4.7) 1662 (5.1) 11 732 (4.6) 5978 (4.4) 476 (2.4) 420 (3.5) Pakistani 811 (3.8) 1132 (3.5) 7425 (2.9) 3972 (3.0) 389 (2.0) 290 (2.4) Bangladeshi 586 (2.8) 713 (2.2) 7282 (2.8) 3980 (3.0) 370 (1.9) 374 (3.1) Other Asian 476 (2.2) 720 (2.2) 5873 (2.3) 2947 (2.2) 234 (1.2) 164 (1.4) Caribbean 473 (2.2) 795 (2.4) 6376 (2.5) 3700 (2.7) 549 (2.8) 278 (2.3) Black African 392 (1.8) 676 (2.1) 5715 (2.2) 2977 (2.2) 350 (1.8) 161 (1.3) Chinese 84 (0.4) 95 (0.3) 983 (0.4) 513 (0.4) 36 (0.2) 34 (0.3) Other 377 (1.8) 636 (2.0) 5723 (2.2) 2966 (2.2) 386 (2.0) 206 (1.7) Smoking status: Smoking status recorded 21 215 (99.6) 32 399 (99.6) 255 186 (99.7) 134 080 (99.6) 19 569 (98.9) 12 003 (99.5) Non-smoker 11 374 (53.4) 17 116 (52.6) 132 634 (51.8) 69 849 (51.9) 9393 (47.5) 6126 (50.8) Former smoker 6252 (29.3) 9725 (29.9) 78 935 (30.8) 41 438 (30.8) 6142 (31.0) 3726 (30.9) Light smoker 2170 (10.2) 3358 (10.3) 25 678 (10.0) 13 846 (10.3) 2413 (12.2) 1252 (10.4) Moderate smoker 730 (3.4) 1121 (3.4) 9395 (3.7) 4661 (3.5) 832 (4.2) 441 (3.7) Heavy smoker 689 (3.2) 1079 (3.3) 8544 (3.3) 4286 (3.2) 789 (4.0) 458 (3.8) Comorbidities: Cardiovascular disease 2962 (13.9) 5325 (16.4) 48 066 (18.8) 28 895 (21.5) 4596 (23.2) 1992 (16.5) Heart failure 302 (1.4) 737 (2.3) 6943 (2.7) 5069 (3.8) 960 (4.9) 374 (3.1) Peripheral vascular disease 1008 (4.7) 1576 (4.8) 12458 (4.9) 8467 (6.3) 1519 (7.7) 544 (4.5) Valvular heart disease 379 (1.8) 914 (2.8) 7378 (2.9) 4606 (3.4) 765 (3.9) 292

(13.9) 5325 (16.4) 48 066 (18.8) 28 895 (21.5) 4596 (23.2) 1992 (16.5) Heart failure 302 (1.4) 737 (2.3) 6943 (2.7) 5069 (3.8) 960 (4.9) 374 (3.1) Peripheral vascular disease 1008 (4.7) 1576 (4.8) 12458 (4.9) 8467 (6.3) 1519 (7.7) 544 (4.5) Valvular heart disease 379 (1.8) 914 (2.8) 7378 (2.9) 4606 (3.4) 765 (3.9) 292 (2.4) Hypertension 12 520 (58.8) 19 293 (59.3) 150 219 (58.7) 80 776 (60.0) 11 117 (56.2) 7310 (60.6) Atrial fibrillation 929 (4.4) 1980 (6.1) 17327 (6.8) 10574 (7.9) 1890 (9.5) 657 (5.4) Chronic kidney disease 388 (1.8) 593 (1.8) 3067 (1.2) 4183 (3.1) 1165 (5.9) 224 (1.9) Rheumatoid arthritis 719 (3.4) 1237 (3.8) 9718 (3.8) 5382 (4.0) 842 (4.3) 460 (3.8) Previous complications: Severe kidney disease 54 (0.3) 74 (0.2) 509 (0.2) 825 (0.6) 213 (1.1) 36 (0.3) Blindness 260 (1.2) 383 (1.2) 3715 (1.5) 2404 (1.8) 360 (1.8) 170 (1.4) Amputation 85 (0.4) 125 (0.4) 1239 (0.5) 894 (0.7) 161 (0.8) 65 (0.5) ≥1 previous episode of hypoglycaemia 288 (1.4) 286 (0.9) 2247 (0.9) 1946 (1.4) 337 (1.7) 215 (1.8) ≥1 previous episode of hyperglycaemia 7921 (37.2) 10 054 (30.9) 68 839 (26.9) 46 341 (34.4) 7279 (36.8) 3914 (32.4) Other drugs: anticoagulant 642 (3.0) 1419 (4.4) 9409 (3.7) 5989 (4.5) 1344 (6.8) 540 (4.5) Thiazides 3444 (16.2) 4346 (13.4) 31 291 (12.2) 16 972 (12.6) 2386 (12.1) 1844 (15.3) ACE inhibitors 9318 (43.7) 12 939 (39.8) 83 847 (32.7) 48 960 (36.4) 7750 (39.2) 5362 (44.5) Angiotension receptor blockers 3399 (16.0) 4895 (15.0) 28 629 (11.2) 16 976 (12.6) 2633 (13.3) 2088 (17.3) Calcium channel blockers 5613 (26.3) 8105 (24.9) 55 674 (21.7) 32 141 (23.9) 5034 (25.4) 3328 (27.6) Statins 15 512 (72.8) 21 383 (65.7) 137 574 (53.7) 77 865 (57.9) 12 640 (63.9) 8451 (70.1) Aspirin 7890 (37.0) 9684 (29.8) 68 013 (26.6) 41 647 (30.9) 7057 (35.7) 4096 (34.0) ACE=angiotensin converting enzyme.

88 (17.3) Calcium channel blockers 5613 (26.3) 8105 (24.9) 55 674 (21.7) 32 141 (23.9) 5034 (25.4) 3328 (27.6) Statins 15 512 (72.8) 21 383 (65.7) 137 574 (53.7) 77 865 (57.9) 12 640 (63.9) 8451 (70.1) Aspirin 7890 (37.0) 9684 (29.8) 68 013 (26.6) 41 647 (30.9) 7057 (35.7) 4096 (34.0) ACE=angiotensin converting enzyme. Values represent those recorded before starting drugs or at study entry for prevalent users. Treatment groups not mutually exclusive. Table 2 shows levels of recording and mean values for haemoglobin A1c, body mass index, cholesterol to high density lipoprotein cholesterol ratio, systolic blood pressure, and serum creatinine level before starting treatment. The highest levels of recording were for haemoglobin A1c, which were in excess of 97% for all six drug groups. Lowest levels of recording were for cholesterol to high density lipoprotein cholesterol ratios, which were more than 84% for all drug groups. Overall, at least 82% of patients had complete data for each clinical value across all drug groups. Apart from higher mean levels of haemoglobin A1c in patients before use of insulin or the group of other diabetes drugs, and higher levels of creatinine among those prescribed sulphonylureas or insulin, the mean values were similar across the six groups. Supplementary table 1 shows mean values before starting the 20 different treatment combinations. The mean values for haemoglobin A1c tended to be higher for patients starting triple treatment (as high values tend to trigger changes in treatment).