Browse the corpus

, 5, 6, 7 The choice of fourth-line drug treatment for resistant hypertension has been entirely empirical, reflecting an absence of data from prospective randomised controlled trials comparing different drug treatment options. The underlying pathophysiological basis for resistant hypertension is also poorly understood. One hypothesis is that resistant hypertension is predominantly caused by sodium retention, due in part to the reduced doses of diuretics prescribed in recent years; if so, drugs with a diuretic action would be the most effective additional treatment.8, 9 An alternative hypothesis is that resistant hypertension is a heterogeneous state, with average responses in study cohorts masking substantial individual patient differences. In the latter case, treatment could be stratified by use of biomarkers of sodium/volume status, particularly plasma renin level, to which sodium status is inversely related.10

Myanmar (Burma) is undergoing a complex political and economic transformation, from a long civil war and military regime to a peace process and democratisation. Since 2011, the Myanmar Ministry of Health has started to rehabilitate the fragile health system, setting the goal of achieving universal health coverage by 2030.1 To achieve this target, Myanmar will have to face substantial challenges; arguably one of the most important difficulties is how to allocate limited health-care resources equitably and effectively. Attention to the most vulnerable people would substantially improve national health outcomes. Myanmar's life expectancy at birth of 66·8 years and infant mortality of 62·0 years are the worst in southeast Asia. Persistent inequalities exist in health outcomes in Myanmar's seven states and seven regions.2 Residents of mountainous peripheral states suffer from remoteness, civil conflicts, and low socioeconomic development.3 For example, infant mortality is 94·2 per 1000 births and under-five mortality is 149·1 per 1000 births near the eastern border of Myanmar, where malaria and tuberculosis prevalence are also high.4 The gap in life expectancy at birth between the areas with the highest and lowest values within Myanmar is more than 11 years, almost as large as that between Myanmar and the USA.

under-five mortality is 149·1 per 1000 births near the eastern border of Myanmar, where malaria and tuberculosis prevalence are also high.4 The gap in life expectancy at birth between the areas with the highest and lowest values within Myanmar is more than 11 years, almost as large as that between Myanmar and the USA. Such disparities are not surprising when reliance on out-of-pocket financing is among the highest globally (81% of Myanmar's total health expenditures) because of no reliable health insurance system and the tight fiscal space for health. As an important step towards universal health coverage, the government increased health-care expenditures by 8·7 times from 2011 to 2015. However, resource allocation does not seem to be closely aligned with the goal of reducing health disparities.5 Conventional budget allocation, tied to population and infrastructure (figure3), gives disproportionately more resources to regions with better health, and fewer resources to several states with high health needs (as measured by infant mortality in the figure).

rticoid receptor blockade, reflecting overlap between primary aldosteronism and resistant hypertension, will require comparison of spironolactone with other types of diuretic. Meanwhile, truly resistant hypertension can now be considered rare and redefined as blood pressure not controlled by A + C + D + spironolactone. For drug treatment of resistant hypertension, international guidelines specifically refer to fourth-line therapy for patients, whose blood pressure is not controlled by treatment with three drugs, typically A + C + D, where “A” is an angiotensin-converting-enzyme (ACE) inhibitor or an angiotensin II receptor blocker (ARB), “C” is a calcium channel blocker (CCB), and “D” is a thiazide or thiazide-like diuretic.4, 5, 6, 7 The choice of fourth-line drug treatment for resistant hypertension has been entirely empirical, reflecting an absence of data from prospective randomised controlled trials comparing different drug treatment options. The underlying pathophysiological basis for resistant hypertension is also poorly understood.

ely aligned with the goal of reducing health disparities.5 Conventional budget allocation, tied to population and infrastructure (figure3), gives disproportionately more resources to regions with better health, and fewer resources to several states with high health needs (as measured by infant mortality in the figure). Crafting policies to mitigate rather than exacerbate health disparities needs professional and innovative leadership, which is one reason why Myanmar's physicians are protesting militarisation of the Ministry of Health through the Black Ribbon Movement. Development of a workable health policy to distribute resources equitably is not only crucial to improve the health status of the population—raising the average by prioritising the needs of the most vulnerable people—but also to build trust, to support peace and reconciliation, and to control the spread of drug-resistant strains of tuberculosis and malaria. These diseases fester in the border areas of Myanmar, and could develop into important regional public health concerns. We declare no competing interests. Figure Health budgets for Myanmar's regions and states are not proportionate to health needs The lines are smooth curves fitted to the data by use of local regression and dots are roughly proportional to population size. Data taken from the Department of Population, Ministry of Immigration and Population.3

Introduction Resistant hypertension, defined as suboptimal blood pressure control despite treatment with at least three blood pressure-lowering drugs, is associated with a poor prognosis. This is caused by organ damage from prolonged exposure to suboptimal blood pressure control, and to the association with diabetes, chronic kidney disease, and obesity.1, 2 The prevalence of resistant hypertension is estimated to be at least 10% of treated hypertensive patients, which would equate to a potential prevalence of about 100 million people globally.1, 3 There has been a growing perception that controlling blood pressure in resistant hypertension is beyond the reach of existing drug therapies, leading to the emergence of device-based therapies such as renal denervation and baroreceptor stimulation. Research in context Evidence before this study

Introduction Resistant hypertension, defined as suboptimal blood pressure control despite treatment with at least three blood pressure-lowering drugs, is associated with a poor prognosis. This is caused by organ damage from prolonged exposure to suboptimal blood pressure control, and to the association with diabetes, chronic kidney disease, and obesity.1, 2 The prevalence of resistant hypertension is estimated to be at least 10% of treated hypertensive patients, which would equate to a potential prevalence of about 100 million people globally.1, 3 There has been a growing perception that controlling blood pressure in resistant hypertension is beyond the reach of existing drug therapies, leading to the emergence of device-based therapies such as renal denervation and baroreceptor stimulation. Research in context Evidence before this study International guidelines have converged on a definition of resistant hypertension as blood pressure that is not controlled to target despite treatment with three recommended blood pressure-lowering drugs at maximum tolerated doses—namely, an angiotensin-converting-enzyme (ACE) inhibitor or an angiotensin II receptor blocker (ARB; “A”), plus a calcium channel blocker (CCB; “C”), plus a thiazide-like diuretic (“D”)—ie, A + C + D. We searched MEDLINE, Embase, and the Cochrane CENTRAL register, for English language publications published up to July, 2015, for randomised controlled trials, and open and observational studies, of the drug treatment of resistant hypertension. This review was cross-referenced to the National Institute for Health and Care Excellence (NICE) hypertension clinical guideline review of drug treatment of resistant hypertension (CG 127) from 2011 and a recently published meta-analysis of randomised controlled trials and non-randomised studies of drug treatment of resistant hypertension with aldosterone antagonists. The meta-analysis suggested that spironolactone can be an effective blood pressure-lowering treatment for patients with resistant hypertension, however, the NICE review concluded that the quality of existing evidence was low. Three randomised controlled trials, of 135 patients (combined), had shown spironolactone to be superior to the placebo in reducing seated clinic blood pressure, when added to existing treatment for resistant hypertension. But there had been no previous randomised controlled trials comparing spironolactone with other blood pressure-lowering drugs to determine if it is the most effective treatment for resistant hypertension.

the placebo in reducing seated clinic blood pressure, when added to existing treatment for resistant hypertension. But there had been no previous randomised controlled trials comparing spironolactone with other blood pressure-lowering drugs to determine if it is the most effective treatment for resistant hypertension. Added value of this study PATHWAY-2 is, to our knowledge, the first randomised controlled trial to compare different blood pressure-lowering treatments in rigorously assessed patients with resistant hypertension, and the first comparison of mineralocorticoid receptor blockade with alternative recommended classes that block the sympathetic nervous system (α blockers and β blockers). The size, crossover design, and hierarchical primary endpoints of PATHWAY-2 enabled demonstration at high significance (p<0·0001) that spironolactone 25–50 mg/day is by far the most effective drug added to A + C + D, for the treatment of resistant hypertension; blood pressure was controlled (home systolic blood pressure <135 mm Hg) in 60% of patients. The role of sodium retention in causing resistant hypertension was strongly suggested by a low baseline plasma renin, despite treatment with three drugs which usually elevate renin, and by a significant inverse correlation between renin and blood pressure reduction by spironolactone. The individual crossover data show that spironolactone is the most effective add-on drug, by a large margin, in the overwhelming majority of patients confirmed as adherent, but resistant, to treatment with A + C + D. Implications of all the available evidence

PATHWAY-2 is, to our knowledge, the first randomised controlled trial to compare different blood pressure-lowering treatments in rigorously assessed patients with resistant hypertension, and the first comparison of mineralocorticoid receptor blockade with alternative recommended classes that block the sympathetic nervous system (α blockers and β blockers). The size, crossover design, and hierarchical primary endpoints of PATHWAY-2 enabled demonstration at high significance (p<0·0001) that spironolactone 25–50 mg/day is by far the most effective drug added to A + C + D, for the treatment of resistant hypertension; blood pressure was controlled (home systolic blood pressure <135 mm Hg) in 60% of patients. The role of sodium retention in causing resistant hypertension was strongly suggested by a low baseline plasma renin, despite treatment with three drugs which usually elevate renin, and by a significant inverse correlation between renin and blood pressure reduction by spironolactone. The individual crossover data show that spironolactone is the most effective add-on drug, by a large margin, in the overwhelming majority of patients confirmed as adherent, but resistant, to treatment with A + C + D. Implications of all the available evidence The unequivocal superiority of spironolactone, together with supportive efficacy and safety data from longer term observational studies, should influence treatment guidelines globally. Whether the superiority is specific for mineralocorticoid receptor blockade, reflecting overlap between primary aldosteronism and resistant hypertension, will require comparison of spironolactone with other types of diuretic. Meanwhile, truly resistant hypertension can now be considered rare and redefined as blood pressure not controlled by A + C + D + spironolactone.

hat resistant hypertension is a heterogeneous state, with average responses in study cohorts masking substantial individual patient differences. In the latter case, treatment could be stratified by use of biomarkers of sodium/volume status, particularly plasma renin level, to which sodium status is inversely related.10 We selected spironolactone as the drug with diuretic action (through blocking the mineralocorticoid receptor) because of observational and limited randomised controlled trial data suggesting good blood pressure-lowering efficacy in resistant hypertension, recently summarised in a meta-analysis.11 However, spironolactone has not been compared with alternative drugs recommended for resistant hypertension. It has therefore been unknown whether spironolactone is the most effective treatment, and if so, whether this applies to a subset of patients or the majority. We aimed to compare spironolactone with alternative fourth-line treatments targeting different pathogenetic mechanisms: the α1-adrenoceptor blocker, doxazosin, acting to reduce peripheral resistance, and the β1-adrenoceptor blocker, bisoprolol, which inhibits the release of renin, and reduces cardiac output. Our primary aim was to determine, for the first time, whether spironolactone is overall the most effective add-on drug treatment for resistant hypertension. The second aim was to determine whether plasma renin levels predict the most effective treatment for individual patients, and whether spironolactone would be most effective in patients with a low plasma renin as a marker of sodium retention. We therefore designed a randomised crossover trial so that each patient's best drug and its predictors could be discovered.

a renin levels predict the most effective treatment for individual patients, and whether spironolactone would be most effective in patients with a low plasma renin as a marker of sodium retention. We therefore designed a randomised crossover trial so that each patient's best drug and its predictors could be discovered. Methods Study design and participants In this 12-month double-blind, placebo-controlled, crossover phase 4 trial, patients were enrolled from 12 secondary care and 2 primary care sites in the UK. The protocol has been published.12 The trial enrolled patients aged 18–79 years with seated clinic systolic blood pressure 140 mm Hg or greater (or ≥135 mm Hg for patients with diabetes) and home systolic blood pressure (18 readings over 4 days) 130 mm Hg or greater, despite treatment for at least 3 months with maximally tolerated doses of three drugs. These had to be an ACE inhibitor or an ARB; “A”), a CCB (“C”), and diuretic (“D”). A full list of eligibility and exclusion criteria is provided in the appendix (pp 3,4). Special emphasis was given to assessment of adherence to the patient's baseline medication before randomisation by measurement of home systolic blood pressure 6 h after directly observed therapy, by returned tablet counts, and by measurement of serum ACE activity. All patients gave informed written consent. The protocol was approved by Cambridge South Ethics Committee. There was no data monitoring board.

Methods Study design and participants In this 12-month double-blind, placebo-controlled, crossover phase 4 trial, patients were enrolled from 12 secondary care and 2 primary care sites in the UK. The protocol has been published.12 The trial enrolled patients aged 18–79 years with seated clinic systolic blood pressure 140 mm Hg or greater (or ≥135 mm Hg for patients with diabetes) and home systolic blood pressure (18 readings over 4 days) 130 mm Hg or greater, despite treatment for at least 3 months with maximally tolerated doses of three drugs. These had to be an ACE inhibitor or an ARB; “A”), a CCB (“C”), and diuretic (“D”). A full list of eligibility and exclusion criteria is provided in the appendix (pp 3,4). Special emphasis was given to assessment of adherence to the patient's baseline medication before randomisation by measurement of home systolic blood pressure 6 h after directly observed therapy, by returned tablet counts, and by measurement of serum ACE activity. All patients gave informed written consent. The protocol was approved by Cambridge South Ethics Committee. There was no data monitoring board. Randomisation and masking The full trial protocol is summarised in the appendix (p 15). After a month's single-blind placebo run-in, patients rotated through four cycles of once daily oral treatment with: (1) spironolactone 25–50 mg, (2) doxazosin modified release 4–8 mg, (3) bisoprolol 5–10 mg, and (4) placebo. The complete set of permutations of the sequence order for the four treatments in the crossover design were randomly ordered within blocks using computer generated pseudo random numbers. Study sites received the allocation for a particular participant by accessing a web-based randomisation system within the Robertson Centre for Biostatistics, Glasgow, Scotland, UK. The study drugs were masked by re-encapsulation at the Royal Free Hospital Pharmacy. Investigators and patients were masked to the identity of drugs, and to their sequence allocation.

rticular participant by accessing a web-based randomisation system within the Robertson Centre for Biostatistics, Glasgow, Scotland, UK. The study drugs were masked by re-encapsulation at the Royal Free Hospital Pharmacy. Investigators and patients were masked to the identity of drugs, and to their sequence allocation. Procedures The treatment cycles were initiated for 6 weeks at the lower dose, followed by forced titration to twice this dose for a further 6 weeks (total of 12 weeks). Patients unable to tolerate a drug in a cycle were allowed to move to the next drug in sequence. There was no washout period between the four treatment cycles (three active, one placebo). The entire study, including placebo run-in, lasted 1 year. After this, patients were invited to participate in a further 12-week cycle of open-label amiloride 10 mg, titrated to 20 mg after 6 weeks. After initial screening and enrolment, there were nine subsequent visits on blinded medication: one after the placebo run-in, the remaining eight after the 6-weekly periods on each of the two doses of the three active drugs and placebo.

Procedures The treatment cycles were initiated for 6 weeks at the lower dose, followed by forced titration to twice this dose for a further 6 weeks (total of 12 weeks). Patients unable to tolerate a drug in a cycle were allowed to move to the next drug in sequence. There was no washout period between the four treatment cycles (three active, one placebo). The entire study, including placebo run-in, lasted 1 year. After this, patients were invited to participate in a further 12-week cycle of open-label amiloride 10 mg, titrated to 20 mg after 6 weeks. After initial screening and enrolment, there were nine subsequent visits on blinded medication: one after the placebo run-in, the remaining eight after the 6-weekly periods on each of the two doses of the three active drugs and placebo. The primary endpoint measurement was average home systolic blood pressure, recorded in the morning and the evening in triplicate, on 4 consecutive days before study visits. For analysis, a maximum of the last 18 recordings for each measurement period—ie, days 2–4—if all completed, were used. A minimum of six blood pressure recordings per measurement period was required for a valid measurement of the home systolic blood pressure average. The home systolic blood pressure average for the primary endpoint included all of the aforementioned measurements throughout each treatment cycle—ie, at week 6 and week 12). For seated clinic blood pressure, the mean of the last two measurements was recorded as the clinic blood pressure. The home and clinic blood pressures were measured for every patient using the approved, automated blood pressure monitor (WatchBP Home, Microlife, Clearwater, FL, USA), which was allocated to the patient for their sole use for the duration of the trial. Patients were instructed by the specialist nurses in self-measurement of blood pressure and technique was visibly re-enforced at each visit, when the research nurse measured patients' blood pressure using the same monitor.

r, FL, USA), which was allocated to the patient for their sole use for the duration of the trial. Patients were instructed by the specialist nurses in self-measurement of blood pressure and technique was visibly re-enforced at each visit, when the research nurse measured patients' blood pressure using the same monitor. Plasma renin was measured at baseline (following run-in on background “A + C + D” and placebo) with a Diasorin Liaison automated chemiluminescent immunoassay for direct renin mass.13 Serum electrolytes were measured at every visit. Outcomes The primary objective was to test the hypothesis that spironolactone is the most effective add-on treatment for patients with resistant hypertension. The primary analysis used an average of home systolic blood pressure recorded throughout the treatment cycle. We prespecified hierarchical primary endpoints; (1) the difference in the home systolic blood pressure between spironolactone and placebo, followed if significant by (2) the difference in home systolic blood pressure between spironolactone and the average of the other two active drugs, (doxazosin and bisoprolol), followed if significant by (3) the difference in home systolic blood pressure between spironolactone and each of the other two active drugs.

bo, followed if significant by (2) the difference in home systolic blood pressure between spironolactone and the average of the other two active drugs, (doxazosin and bisoprolol), followed if significant by (3) the difference in home systolic blood pressure between spironolactone and each of the other two active drugs. The secondary objectives included evaluation of; (1) clinic blood pressure responses to randomised treatments; (2) blood pressure control rates—ie, home systolic blood pressure less than 135 mm Hg; (3) whether plasma renin concentrations and other baseline characteristics could help personalise treatment by predicting the best drug treatment; and (4) adverse event rates during each treatment cycle. To test the hypothesis that plasma renin (measured on a background of three drugs—ie, A + C + D), will predict the most effective fourth-line drug, we examined the relationship between plasma renin and the reduction of home systolic blood pressure with each drug, adjusted for the placebo response. We also identified the best treatment for each patient—ie, the one on which they achieved the lowest blood pressure, and estimated for each drug the relationship between baseline renin and the likelihood that it would provide the best response. Adverse events were recorded at each visit.

usted for the placebo response. We also identified the best treatment for each patient—ie, the one on which they achieved the lowest blood pressure, and estimated for each drug the relationship between baseline renin and the likelihood that it would provide the best response. Adverse events were recorded at each visit. Statistical analysis The sample size was estimated to be 294 patients, based on detecting a difference of 3 mm Hg (SD 12) in home systolic blood pressure between each of the experimental drugs and the placebo treatment, with 90% power using a single sample t test at the 0·003 significance level (this was chosen in order that the 0·01 level could be adjusted for three planned comparisons). However, the hierarchical analysis subsequently adopted negated the need to adjust p values. We tested hypotheses with the mixed effect models to analyse continuous variables, with unstructured covariances for repeated measures within a patient. We included prespecified baseline covariates (sex, age, height, weight, smoking history, and the baseline value of the outcome being analysed) in the models. Least squares means for each treatment estimated from these models are presented. Blood pressure control and response rates were analysed with logistic regression models, which also included the baseline covariates. Comparisons of adverse event rates between treatments were done with χ2 tests and Fisher's exact p values are given. The probability that a drug would provide the best response as a function of baseline renin, was estimated from multinomial logistic regression.

ssion models, which also included the baseline covariates. Comparisons of adverse event rates between treatments were done with χ2 tests and Fisher's exact p values are given. The probability that a drug would provide the best response as a function of baseline renin, was estimated from multinomial logistic regression. Intention-to-treat analyses excluded only those participants with no primary outcome data at any follow-up visit (21 patients). Other participants with missing data were included, and we assumed that data was missing at random (ie, its absence was unrelated to the unobserved value). Analyses were done with SAS (Cary, USA) version 9.3. The trial is registered with EudraCT number 2008-007149-30, and ClinicalTrials.gov number, NCT02369081. Role of the funding source The funders had no role in the data collection, data analysis, or data interpretation, or the writing of the report. The investigators and all authors had sole discretion in the data analysis and interpretation, writing of the report, and the decision to submit for publication. The corresponding author had full access to all of the data and the final responsibility to submit for publication.

erpretation, or the writing of the report. The investigators and all authors had sole discretion in the data analysis and interpretation, writing of the report, and the decision to submit for publication. The corresponding author had full access to all of the data and the final responsibility to submit for publication. Results Between May 15, 2009, and July 8, 2014, we screened 436 patients and randomised 335, of which 21 had no follow-up for any drug and were excluded from the intention-to-treat analysis, which comprised 314 patients who had any follow-up data. 285 patients received spironolactone, 282 doxazosin, 285 bisoprolol, and 274 placebo; 230 patients completed all treatment cycles (figure 1; appendix p 15). Last patient visit was on June 5, 2015. Table 1 shows the baseline of characteristics of the randomised patients. The average reduction in home systolic blood pressure throughout the treatment cycle with spironolactone was superior to each of: placebo (–8·70 mm Hg [95% CI −9·72 to −7·69]; p<0·0001); the mean of the other two active treatments (doxazosin and bisoprolol, −4·26 [–5·13 to 3·38]; p<0·0001); and each of the other individual treatments; doxazosin (–4·03 [–5·04 to 3·02]; p<0·0001) and bisoprolol (–4·48 [–5·50 to −3·46]; p<0·0001; figure 2; Table 2, Table 3, Table 4).

Hg [95% CI −9·72 to −7·69]; p<0·0001); the mean of the other two active treatments (doxazosin and bisoprolol, −4·26 [–5·13 to 3·38]; p<0·0001); and each of the other individual treatments; doxazosin (–4·03 [–5·04 to 3·02]; p<0·0001) and bisoprolol (–4·48 [–5·50 to −3·46]; p<0·0001; figure 2; Table 2, Table 3, Table 4). The differences in favour of spironolactone were greater when restricted to home systolic blood pressure values measured on the higher doses at the end of each treatment cycle (table 3). Spironolactone showed the largest difference between high and low doses (table 4); this was true irrespective of which treatment was assigned in the previous cycle. In further, prespecified sensitivity analyses, similar differences in favour of spironolactone were seen in 230 patients who received all four study drugs and 216 patients on three “A + C + D” background medications (appendix pp 5–8). The steep dose response for spironolactone, and superiority over other treatments, were seen also in a parallel group analysis of the first treatment cycle (appendix p 17). The results for seated clinic systolic blood pressure largely mirror those seen with home systolic blood pressure except that there was a large placebo effect on clinic blood pressure that was not seen with home blood pressure measurement (appendix p 9). Full details of all blood pressure data (home and clinic), including diastolic pressures and heart rate are shown in appendix (p 10).

those seen with home systolic blood pressure except that there was a large placebo effect on clinic blood pressure that was not seen with home blood pressure measurement (appendix p 9). Full details of all blood pressure data (home and clinic), including diastolic pressures and heart rate are shown in appendix (p 10). Overall 219 (68·9% [95% CI 63·6–73·8]) of 314 patients achieved target home systolic blood pressure of less than 135 mm Hg. The comparison of control rates is shown in the appendix (p 11). 58% of patients had their blood pressure controlled with spironolactone, which was superior to rates for other treatments. Most patients who were controlled by doxazosin or bisoprolol had a still greater fall in blood pressure on spironolactone, which was consequently the most effective treatment in almost 60% of patients. This was at least three times the proportion in whom doxazosin or bisoprolol were the most effective.

ther treatments. Most patients who were controlled by doxazosin or bisoprolol had a still greater fall in blood pressure on spironolactone, which was consequently the most effective treatment in almost 60% of patients. This was at least three times the proportion in whom doxazosin or bisoprolol were the most effective. The proportion of patients in whom spironolactone was their best drug for blood pressure lowering was most evident on the planned analyses of prediction by plasma renin of blood pressure response to each drug and the likelihood that different drugs would be best at different points in the plasma renin distribution. Figure 3 shows the relation between plasma renin (measured at baseline whilst patients were receiving their usual medication—ie, A + C + D) and the blood pressure-lowering response to each active treatment corrected for the placebo effect. There was a clear inverse relation between the home systolic blood pressure fall with spironolactone and plasma renin, not seen with bisoprolol or doxazosin. Moreover, the blood pressure response to spironolactone was superior to bisoprolol and doxazosin across most of the plasma renin distribution (figure 3). Only in a small minority of patients, with very high plasma renin levels, did the mean home systolic blood pressure response to doxazosin or bisoprolol overlap that to spironolactone. Analysis of each patient's best drug clearly showed that, although spironolactone was the best blood pressure-lowering treatment throughout almost the entire renin distribution, the likelihood of being superior, and the magnitude of this superiority, was several-fold higher for spironolactone than the other drugs at the lower end of the distribution (figure 3; appendix pp 18,19).

ough spironolactone was the best blood pressure-lowering treatment throughout almost the entire renin distribution, the likelihood of being superior, and the magnitude of this superiority, was several-fold higher for spironolactone than the other drugs at the lower end of the distribution (figure 3; appendix pp 18,19). All active treatments were well tolerated with similar low rates of adverse events and withdrawals due to adverse events (table 5; appendix p 12 shows the numbers and rates of the commonest adverse events, and appendix p 13 shows all serious adverse events). Notably, discontinuations due to renal impairment, hyperkalaemia, and gynaecomastia were not increased with spironolactone relative to other treatments and placebo. Serum sodium was reduced with spironolactone (–1·91 mmol/L, −1·35%), but not the other treatments, whereas some increase in potassium levels was observed with both bisoprolol and spironolactone (appendix p 14). Serum creatinine levels increased and estimated glomerular filtration rate (eGFR) decreased (appendix p 14). Stacking of the distribution of serum sodium, potassium, and eGFR at the end of the treatment cycle for each drug shows that all drugs lowering blood pressure in this population cause some fall in eGFR, and that the frequency of abnormal numbers for each parameter is low (figure 4). None of these was clinically serious or led to withdrawals from the trial. Only six (2%) of 285 patients exposed to spironolactone developed a serum potassium on a single occasion greater than 6·0 mmol/L, with a maximum of 6·5 mmol/L.

GFR, and that the frequency of abnormal numbers for each parameter is low (figure 4). None of these was clinically serious or led to withdrawals from the trial. Only six (2%) of 285 patients exposed to spironolactone developed a serum potassium on a single occasion greater than 6·0 mmol/L, with a maximum of 6·5 mmol/L. Discussion PATHWAY-2 is the first randomised controlled trial to compare spironolactone with other blood pressure-lowering drug treatments in a well-characterised population of patients with resistant hypertension. The study shows that spironolactone was by far the most effective blood pressure-lowering treatment for patients with resistant hypertension. This was true in terms of the magnitude of the blood pressure response, the proportion of patients achieving a stringent measure of blood pressure control (home systolic blood pressure <135 mm Hg), and the proportion in whom it was more effective than either of the non-diuretic alternative drugs. These findings suggest that the predominant underlying pathophysiological cause of resistant hypertension is sodium retention, despite existing baseline diuretic therapy. This conclusion is supported by our finding that the response to spironolactone had a clear inverse relation with plasma renin, was especially effective at lower plasma renin levels, and yet the most effective drug throughout the range of plasma renin.

on is sodium retention, despite existing baseline diuretic therapy. This conclusion is supported by our finding that the response to spironolactone had a clear inverse relation with plasma renin, was especially effective at lower plasma renin levels, and yet the most effective drug throughout the range of plasma renin. Bisoprolol and doxazosin were more effective than placebo at reducing blood pressure as so-called add-on therapy for resistant hypertension, but significantly less effective than spironolactone. Thus, this study has for the first time, established a clear hierarchy for drug treatment of resistant hypertension in which spironolactone is the most effective add-on therapy (ie, fourth-line drug in addition to A + C + D) for most patients. Bisoprolol or doxazosin are less effective alternatives for those intolerant of spironolactone.

s for the first time, established a clear hierarchy for drug treatment of resistant hypertension in which spironolactone is the most effective add-on therapy (ie, fourth-line drug in addition to A + C + D) for most patients. Bisoprolol or doxazosin are less effective alternatives for those intolerant of spironolactone. PATHWAY-2 is the first study to use home blood pressure averages rather than clinic blood pressure to assess the primary outcome of blood pressure response in patients with resistant hypertension. This is important because home blood pressure measurement reduces the placebo effect, as was evident in our study, and eliminated patients whose blood pressure could have been spuriously elevated at baseline due to so-called white coat hypertension. Indeed, we noted a large placebo effect on clinic systolic blood pressure readings, exceeding 10 mm Hg (table 3), which points to important confounding in interpreting data from previous studies that have relied on clinic blood pressure readings alone to assess the efficacy of drug and non-drug-based interventions in these patients. The magnitude of home systolic blood pressure reduction with spironolactone versus placebo (–8·7 mm Hg) in the present study is consistent with data from two smaller randomised controlled trials versus placebo that have used ambulatory blood pressure monitoring (secondary analysis) to measure changes in mean daytime pressures (–7·31 mm Hg).11, 14, 15 The reduction in seated clinic systolic blood pressure with spironolactone (–20·7 mm Hg) was also similar to the reduction in clinic blood pressure from meta-analysis of previous small randomised contolled trials versus placebo (–24·3 mm Hg) and single arm studies with no comparator (–22·7 mm Hg).11

1 mm Hg).11, 14, 15 The reduction in seated clinic systolic blood pressure with spironolactone (–20·7 mm Hg) was also similar to the reduction in clinic blood pressure from meta-analysis of previous small randomised contolled trials versus placebo (–24·3 mm Hg) and single arm studies with no comparator (–22·7 mm Hg).11 Spironolactone substantially increased the likelihood of achieving blood pressure control relative to bisoprolol or doxasosin, with almost 60% achieving blood pressure control within 3 months of starting treatment. This challenges the concept that resistant hypertension cannot be treated adequately with existing drug therapies, a concept that might have contributed to the growth of non-drug-based therapies such as renal denervation. Indeed, only 15 of 285 patients assessed on spironolactone failed to achieve a home systolic blood pressure lower than 150 mm Hg (equivalent to a clinic systolic blood pressure of roughly 160 mm Hg), the usual eligibility criterion for denervation. Furthermore, it is clear from our data that spironolactone, unlike bisoprolol or doxazosin, exhibited a significant dose response with regard to the magnitude of blood pressure lowering. A previous crossover comparison of spironolactone with even higher doses (ie, 50–100 mg) in patients without resistant hypertension, also showed a dose response,16 suggesting that the highest dose of spironolactone used in our study (ie, 50 mg), might not be at the top of the dose range, and hence the potential for even better control rates with higher doses.

olactone with even higher doses (ie, 50–100 mg) in patients without resistant hypertension, also showed a dose response,16 suggesting that the highest dose of spironolactone used in our study (ie, 50 mg), might not be at the top of the dose range, and hence the potential for even better control rates with higher doses. The superior response to spironolactone, compared with the other drugs, particularly in patients at the lower end of the distribution of plasma renin, supports the hypothesis that the predominant cause of resistant hypertension is sodium retention. The fact that spironolactone was the most effective drug across a wide range of baseline plasma renin values does not negate the hypothesis, because one would expect plasma renin levels to be elevated in patients receiving treatment with A + C + D, all of which usually increase plasma renin levels. Indeed, because plasma renin is substantially affected by background antihypertensive treatment, the interpretation and recognition of so-called low renin status has been uncertain in such populations. The median renin in PATHWAY-2, 34 mU/L, is roughly three times higher than that in the PATHWAY-1 study, of 600 patients with untreated hypertension (unpublished). However, 34 mU/L is a lower median than expected if the sole influence on baseline renin was drug treatment.10 It is probable that plasma renin in resistant hypertension is relatively suppressed by sodium retention, even though absolute values appear normal or high.

0 patients with untreated hypertension (unpublished). However, 34 mU/L is a lower median than expected if the sole influence on baseline renin was drug treatment.10 It is probable that plasma renin in resistant hypertension is relatively suppressed by sodium retention, even though absolute values appear normal or high. One contributor to the sodium retention could be under-dosing of background diuretic treatment. An alternative or additional possibility is that some patients with resistant hypertension have undetected aldosterone producing adenomas (APAs). Recent studies have shown that specific somatic mutations in the adrenal gland can result in micro-APAs, which are difficult to detect by conventional imaging.17 Additionally, some high-renin patients will have an element of secondary aldosteronism, explaining why spironolactone retains some efficacy at the upper end of the renin distribution. Determination of whether spironolactone is particularly effective treatment for resistant hypertension because it antagonises the effects of aldosterone will require a head-to-head comparison of spironolactone with an increase in dose of the background diuretic.

some efficacy at the upper end of the renin distribution. Determination of whether spironolactone is particularly effective treatment for resistant hypertension because it antagonises the effects of aldosterone will require a head-to-head comparison of spironolactone with an increase in dose of the background diuretic. As well as being the most effective treatment for resistant hypertension, spironolactone was well tolerated. The doses of spironolactone used in the present study are low compared with the 200–400 mg daily used in other clinical circumstances. The 25–50 mg daily doses in PATHWAY-2 are consistent with previous studies of resistant hypertension in which 25 mg was the most common daily dose.11 These studies have shown that the main biochemical effects associated with spironolactone treatment are a reduction in serum sodium and an increase in potassium. We noted a magnitude of change (–1·19 mmol/L in sodium and 0·45 mmol/L in potassium) very similar to that reported in summary data from previous studies.11 Despite almost 14% of our patients having type 2 diabetes, only six patients receiving spironolactone developed potassium levels in excess of 6·0 mmol/L, which was detected by our routine monitoring and had no clinical consequence. We also recorded reductions in eGFR with all active blood pressure-lowering treatments that most likely reflect a reduction in renal perfusion pressure with blood pressure lowering. Thus, although spironolactone is both very effective and safe in resistant hypertension, it is important to monitor electrolytes (especially potassium) and renal function during the weeks after initiation of treatment, after dose escalation and periodically thereafter. A recent large longitudinal population study of the use of spironolactone showed no evidence of any increased incidence of admission to hospital or outpatient hyperkalaemia.18 Another recognised adverse effect of spironolactone treatment relates to its anti-androgen effects and the development of gynaecomastia, which has been reported to occur in roughly 6% of men.11, 19 We did not observe any cases in the present study, but this most likely reflects the relatively short duration of our study (3 months exposure).

erse effect of spironolactone treatment relates to its anti-androgen effects and the development of gynaecomastia, which has been reported to occur in roughly 6% of men.11, 19 We did not observe any cases in the present study, but this most likely reflects the relatively short duration of our study (3 months exposure). The PATHWAY-2 study has some limitations. 3 months treatment exposure is relatively short. Nevertheless, observational studies of longer spironolactone treatment duration for resistant hypertension suggest that the magnitude of the initial blood pressure response is durable and that among adverse events, only gynaecomastia is exposure dependent.11, 19 The results of open-label treatment with amiloride (10–20 mg daily) during the run-out phase of our study will be reported later and will help to determine whether amiloride is an effective alternative to spironolactone. Our study excluded patients with an eGFR less than 45 mL/min, as have previous studies, thus there are no data on the safety profile of spironolactone in patients with resistant hypertension and an eGFR less than 45 mL/min. Our study included predominantly white Caucasian patients, thus it is unclear whether the results are applicable to other ethnic groups, however, spironolactone has been shown to be just as effective in a small observational study that included black American patients.20 The absence of washout periods, inherent in a study design already 1 year in length, might be considered a concern. However, we were confident from our previous crossover studies16 that carry-over would not be a problem, and are supported by the sensitivity analyses, including the absence of change in blood pressure during the placebo cycle, and similarity of the primary outcome result to the retrospective parallel group analysis of cycle 1 alone. At worst, we could have under estimated the superiority of spironolactone. Finally, the study does not include data for morbidity and mortality outcomes but blood pressure lowering is a powerful surrogate for clinical benefit, especially in this high-risk group of patients.

ive parallel group analysis of cycle 1 alone. At worst, we could have under estimated the superiority of spironolactone. Finally, the study does not include data for morbidity and mortality outcomes but blood pressure lowering is a powerful surrogate for clinical benefit, especially in this high-risk group of patients. Our study also has a number of strengths. The patients in this study were particularly well characterised as resistant hypertension, with the use of home blood pressure monitoring to exclude so-called white coat hypertension, standardisation of background medication (A + C + D), directly observed therapy to exclude patients non-compliant with background medication, measurement of serum ACE to allow retrospective confirmation of the expected difference in distribution between patients receiving ACE inhibitor or ARB as one of their background drugs, and oversight by national specialist hypertension centres to exclude secondary hypertension. In the late stages of the study, we incorporated a new assay for monitoring all commonly administered antihypertensive drugs in patients' urine,21 and will report results from this substudy that strongly support a high adherence rate among our patients. Another strength was the design of the study, which incorporated a random cross over design, allowing each patient's best drug to be determined, and predictive testing of this to be assessed. Mechanistic haemodynamic substudies, which could add to the predictive value of renin and help us to understand the pathophysiology of resistant hypertension, were included at most sites and will be reported separately.12 Finally, this was the first randomised controlled trial directly comparing different active drug treatments in resistant hypertension and it produced an unequivocal result.

ue of renin and help us to understand the pathophysiology of resistant hypertension, were included at most sites and will be reported separately.12 Finally, this was the first randomised controlled trial directly comparing different active drug treatments in resistant hypertension and it produced an unequivocal result. The results of PATHWAY-2 have broad international relevance because of convergent guideline recommendations, which recommend A + C + D as the preferred three-drug combination at step 3.22 Our finding, that spironolactone was clearly the most effective treatment for resistant hypertension, should influence future treatment guidelines and clinical practice globally. The finding could indeed stimulate an early redefinition of resistant hypertension to include a trial of spironolactone before the label is applied. A longer-term question is whether the antecedent to resistant hypertension is under treatment or wrong treatment, with the resistance to conventional drugs marking a subpopulation in whom spironolactone should be used at an earlier stage. Supplementary Material Supplementary appendix

The results of PATHWAY-2 have broad international relevance because of convergent guideline recommendations, which recommend A + C + D as the preferred three-drug combination at step 3.22 Our finding, that spironolactone was clearly the most effective treatment for resistant hypertension, should influence future treatment guidelines and clinical practice globally. The finding could indeed stimulate an early redefinition of resistant hypertension to include a trial of spironolactone before the label is applied. A longer-term question is whether the antecedent to resistant hypertension is under treatment or wrong treatment, with the resistance to conventional drugs marking a subpopulation in whom spironolactone should be used at an earlier stage. Supplementary Material Supplementary appendix Acknowledgments The study was funded by a special project grant from the British Heart Foundation (number SP/08/002). Further funding was provided by the National Institute for Health Research (NIHR) Comprehensive Local Research Networks. BW, PS, MJC, and MJB are NIHR Senior Investigators, and are supported by, respectively, the NIHR UCL/UCL Hospitals Biomedical Research Centre, the Biomedical Research Centre award to Imperial College Healthcare NHS Trust, the NIHR Cardiovascular Biomedical research Unit at St Bartholomew's Hospital, London, and the NIHR Biomedical Research Centre award to Cambridge University Hospitals NHS Trust. Blinded medication was packed by Alan Wong and colleagues at the Royal Free Hospital pharmacy.

award to Imperial College Healthcare NHS Trust, the NIHR Cardiovascular Biomedical research Unit at St Bartholomew's Hospital, London, and the NIHR Biomedical Research Centre award to Cambridge University Hospitals NHS Trust. Blinded medication was packed by Alan Wong and colleagues at the Royal Free Hospital pharmacy. Contributors BW, MJB, TMM, MJC, GM, PS, DJW, and IF designed the trial and were coapplicants to the British Heart Foundation for funding of the PATHWAY programme, led by MJB. All applicants constituted the steering committee, joined by JKC, IM, and SP during the trial. MJB and JS drafted the protocol aided by TMM and BW. JS was Trial Coordinator. IF and SM advised on statistical analysis. MJB, BW, and TMM constituted the trial executive committee. BW drafted the manuscript, with MJB, TMM, and SM. All authors approved the final draft.

ittee, joined by JKC, IM, and SP during the trial. MJB and JS drafted the protocol aided by TMM and BW. JS was Trial Coordinator. IF and SM advised on statistical analysis. MJB, BW, and TMM constituted the trial executive committee. BW drafted the manuscript, with MJB, TMM, and SM. All authors approved the final draft. Declaration of interests BW has received honoraria for lectures on hypertension from Novartis, Boehringer Ingelheim, Servier, Daiichi Sankyo, and Pfizer. MJC declares no competing interests, but has received honoraria from Medtronic and Quantum Genomics during the life of this trial. JKC is Vice-President of the Artery Society, with no competing interest to declare. GM has received honoraria from Novartis. PS, JS, and SP declare no competing interests. DJW has received funding for membership of Independent Data Monitoring Committees for Abbvie in relation to clinical trials in diabetic nephropathy. DJW is President-elect of the British Pharmacological Society and a Board Member of MHRA. IF declares no competing interests. IS declares no competing interests. TMM is chief investigator on two large investigator initiated, industry funded but university sponsored cardiovascular outcome studies (funded by Pfizer and Menarini/IPSEN/Teijin pharmaceuticals). His research unit also does industry funded studies by Novartis and Amgen, but none of these studies focus on blood pressure. He has provided consultancy or received honoraria for speaking from Novartis, Takeda, Daiichi Sankyo, Shire, and Astellus. TMM is the current president of the British Hypertension Society. IM is a member of the Executive Committee of the British Hypertension Society. MJB has received honoraria from Novartis.

pressure. He has provided consultancy or received honoraria for speaking from Novartis, Takeda, Daiichi Sankyo, Shire, and Astellus. TMM is the current president of the British Hypertension Society. IM is a member of the Executive Committee of the British Hypertension Society. MJB has received honoraria from Novartis. Figure 1 Trial profile *Randomised but instructed not to take any study drug after the result of directly observed therapy. Participants with any follow-up were included in the intent-to-treat analysis and the full analysis dataset consisted of all available data for these participants. Per-protocol analyses included participants who completed all follow-up visits without major deviation from the protocol. ITT=intention to treat. Figure 2 Home systolic and diastolic blood pressures comparing spironolactone with each of the other cycles The top and bottom of each column represents the unadjusted home systolic and diastolic blood pressures, respectively, averaged across the mid-cycle (low-dose) and end-of-cycle (high-dose) visits (6 weeks and 12 weeks) in which patients received the drug. Error bars represent 95% CI. Comparisons are as described under methods for the primary endpoint. Figure 3 Blood pressure response versus renin

The top and bottom of each column represents the unadjusted home systolic and diastolic blood pressures, respectively, averaged across the mid-cycle (low-dose) and end-of-cycle (high-dose) visits (6 weeks and 12 weeks) in which patients received the drug. Error bars represent 95% CI. Comparisons are as described under methods for the primary endpoint. Figure 3 Blood pressure response versus renin Regression (90% CI) of placebo corrected change in home systolic blood pressure versus renin for spironolactone (r2=0·037, p=0·003), doxazosin (r2=0·007, p=0·183), and bisoprolol (r2=0·0004, p=0·750). Blood pressures were averaged across the mid-cycle and end-of-cycle visits (6 and 12 weeks) for every treatment cycle. The distribution curve is fitted to the baseline renins observed in the study. The vertical dashed line shows that the blood pressure fall on bisoprolol numerically exceeds that on spironolactone only in the top 3% of the renin distribution. A more detailed histogram for plasma renin is shown in the appendix (p 20). Figure 4 Distribution of potassium (A), sodium (B), and estimated glomerular filtration rate (eGFR; C) on each drug Values on the x axis are the measurement at the end of each 12-week cycle, and the y axis represents the number of patients with values in each bin on the x axis. Table 1 Baseline characteristics of the patients randomised into the PATHWAY-2 study (n=335)

Figure 4 Distribution of potassium (A), sodium (B), and estimated glomerular filtration rate (eGFR; C) on each drug Values on the x axis are the measurement at the end of each 12-week cycle, and the y axis represents the number of patients with values in each bin on the x axis. Table 1 Baseline characteristics of the patients randomised into the PATHWAY-2 study (n=335) Mean (SD) or N (%) Age (years) 61·4 (9·6) Sex Male 230 (69%) Female 105 (31%) Weight (kg) 93·5 (18·1) Smoker 26 (7·8%) Home Systolic blood pressure (mm Hg) 147·6 (13·2) Diastolic blood pressure (mm Hg) 84·2 (10·9) Heart rate (beats per min) 73·3 (9·9) Clinic Systolic blood pressure (mm Hg) 157·0 (14·3) Diastolic blood pressure (mm Hg) 90·0 (1·5) Heart rate (beats per min) 77·2 (12·2) 24 h urine (mmol/24 h) Sodium 137·1 (71·8) Potassium 70·5 (29·5) Blood electrolytes (mmol/L) Sodium 139·6 (3·0) Potassium 4·1 (0·5) eGFR (mL/min) 91·1 (26·8) Diabetic 46 (14%) eGFR=estimated glomerular filtration rate. Table 2 Home systolic blood pressure averaged across both visits for each cycle

Mean (SD) or N (%) Age (years) 61·4 (9·6) Sex Male 230 (69%) Female 105 (31%) Weight (kg) 93·5 (18·1) Smoker 26 (7·8%) Home Systolic blood pressure (mm Hg) 147·6 (13·2) Diastolic blood pressure (mm Hg) 84·2 (10·9) Heart rate (beats per min) 73·3 (9·9) Clinic Systolic blood pressure (mm Hg) 157·0 (14·3) Diastolic blood pressure (mm Hg) 90·0 (1·5) Heart rate (beats per min) 77·2 (12·2) 24 h urine (mmol/24 h) Sodium 137·1 (71·8) Potassium 70·5 (29·5) Blood electrolytes (mmol/L) Sodium 139·6 (3·0) Potassium 4·1 (0·5) eGFR (mL/min) 91·1 (26·8) Diabetic 46 (14%) eGFR=estimated glomerular filtration rate. Table 2 Home systolic blood pressure averaged across both visits for each cycle Blood pressure (mm Hg) Change from baseline (mm Hg) Mean Spironolactone 134·9 (134·0 to 135·9) −12·8 (−13·8 to −11·8) Doxazosin 139·0 (138·0 to 140·0) −8·7 (−9·7 to −7·7) Bisoprolol 139·4 (138·4 to 140·4) −8·3 (−9·3 to −7·3) Placebo 143·6 (142·6 to 144·6) −4·1 (−5·1 to −3·1) Mean differences Spironolactone vs placebo 8·70 (−9·72 to −7·69) p<0·0001 Spironolactone vs mean bisoprolol and doxazosin −4·26 (−5·13 to −3·38) p<0·0001 Spironolactone vs doxazosin −4·03 (−5·04 to −3·02) p<0·0001 Spironolactone vs bisoprolol −4·48 (−5·50 to −3·46) p<0·0001 Data are mean (95% CI). Home systolic blood pressure throughout the treatment cycle for each drug (includes data from mid-cycle at week 6 and the final visit at week 12). Least squares means from mixed effects models adjusted for baseline covariates. Hierarchical primary endpoints each tested only if the preceding tests were significant.

CI). Home systolic blood pressure throughout the treatment cycle for each drug (includes data from mid-cycle at week 6 and the final visit at week 12). Least squares means from mixed effects models adjusted for baseline covariates. Hierarchical primary endpoints each tested only if the preceding tests were significant. Table 3 Home systolic blood pressure at final visit of each cycle Blood pressure (mm Hg) Change from baseline (mm Hg) Mean Spironolactone 133·5 (132·3 to 134·8) −14·4 (−15·6 to −13·1) Doxazosin 138·8 (137·6 to 140·1) −9·1 (−10·3 to −7·8) Bisoprolol 139·5 (138·2 to 140·8) −8·4 (−9·7 to −7·1) Placebo 143·7 (142·5 to 145·0) −4·2 (−5·4 to −2·9) Mean differences Spironolactone vs placebo −10·2 (−11·7 to −8·74) p<0·0001 Spironolactone vs mean bisoprolol and doxazosin −5·64 (−6·91 to −4·36) p<0·0001 Spironactone vs doxazosin −5·30 (−6·77 to −3·83) p<0·0001 Spironolactone vs bisoprolol −5·98 (−7·45 to −4·51) p<0·0001 Data are mean (95% CI). Sensitivity analysis using only the mean home systolic blood pressure at the final visit of each cycle (week 12). Table 4 Home systolic blood pressure dose response (higher vs lower dose) Blood pressure (mm Hg) p value Spironolactone −3·86 (−5·28 to −2·45) <0·0001 Doxazosin −0·88 (−2·32 to 0·56) 0·23 Bisoprolol −1·49 (−2·94 to −0·04) 0·04 Placebo −0·68 (−2·10 to 0·75) 0·35 Difference in mean home systolic blood pressure after treatment with the lower (week 6) and higher doses (week 12) of each treatment. Table 5 Adverse events and withdrawals

Blood pressure (mm Hg) p value Spironolactone −3·86 (−5·28 to −2·45) <0·0001 Doxazosin −0·88 (−2·32 to 0·56) 0·23 Bisoprolol −1·49 (−2·94 to −0·04) 0·04 Placebo −0·68 (−2·10 to 0·75) 0·35 Difference in mean home systolic blood pressure after treatment with the lower (week 6) and higher doses (week 12) of each treatment. Table 5 Adverse events and withdrawals Spironolactone Doxazosin Bisoprolol Placebo p value* Serious adverse events 7 (2%) 5 (2%) 8 (3%) 5 (2%) 0·82 Any adverse event 58 (19%) 67 (23%) 68 (23%) 42 (15%) 0·036 Withdrawals for adverse events 4 (1%) 9 (3%) 4 (1%) 3 (1%) 0·28 Data are n (%). * p values for Fisher's exact test. The most common adverse events in at least 5% of patients on any treatment are shown in appendix p 12.

Introduction Use of ultrasonography to identify small-for-gestational-age (SGA) infants is widespread in contemporary obstetric practice.1, 2 In the USA, UK, and many other countries women are not routinely scanned in late pregnancy, but are selected for third trimester ultrasonography on the basis of pre-pregnancy risk factors, development of obstetric complications, and serial measurement of symphyseal-fundal height.2, 3 This approach identifies a third of SGA infants or fewer,4, 5, 6 and unidentified SGA is a common finding in perinatal deaths.7, 8 However, a meta-analysis9 of nine trials assessing universal late pregnancy ultrasonography, including about 27 000 women, showed no beneficial effect, which led to the recommendation that it should not be offered routinely in the third trimester.2, 3

6 and unidentified SGA is a common finding in perinatal deaths.7, 8 However, a meta-analysis9 of nine trials assessing universal late pregnancy ultrasonography, including about 27 000 women, showed no beneficial effect, which led to the recommendation that it should not be offered routinely in the third trimester.2, 3 Assessment of a screening programme could yield a negative result for three major reasons. First, the screening test could perform poorly—ie, have poor diagnostic effectiveness. Second, screening might not be coupled with use of an effective intervention—ie, the screening programme would not be clinically effective. Third, both the screening test and intervention could be effective, but the studies analysed might be methodologically flawed—eg, they might be underpowered.10 A screening study can be designed only if the diagnostic effectiveness of the screening test has been well characterised. The National Institute for Health and Care Excellence (NICE) in the UK did a thorough systematic review of the evidence about the diagnostic effectiveness of universal screening for SGA using ultrasonography for their 2008 Antenatal Care guideline.3 They concluded that, “the methods by which [SGA] can be identified antenatally are poorly developed or not tested by rigorous methodology”. Furthermore, SGA is frequently used as a proxy for fetal growth restriction (FGR). However, in reality many SGA infants are physiologically small. Very little information is available about the ability of universal ultrasonography to identify those SGA fetuses that are at increased risk of morbidity (panel).

hodology”. Furthermore, SGA is frequently used as a proxy for fetal growth restriction (FGR). However, in reality many SGA infants are physiologically small. Very little information is available about the ability of universal ultrasonography to identify those SGA fetuses that are at increased risk of morbidity (panel). The aims of this study were to compare the diagnostic effectiveness of universal ultrasound as a screening test for SGA compared with selective ultrasound and to establish which, if any, of a series of previously described ultrasonic markers of FGR identified those SGA fetuses at an increased risk of an adverse outcome.

The aims of this study were to compare the diagnostic effectiveness of universal ultrasound as a screening test for SGA compared with selective ultrasound and to establish which, if any, of a series of previously described ultrasonic markers of FGR identified those SGA fetuses at an increased risk of an adverse outcome. Methods Study design and participants In this prospective cohort Pregnancy Outcome Prediction (POP) study, nulliparous women attending for their dating ultrasound scan at the Rosie Hospital (Cambridge, UK) between Jan 14, 2008, and July 31, 2012, with a viable pregnancy were eligible to participate. The protocol has been published elsewhere.11 The only clinical exclusion criterion was multiple pregnancy. Women were selected for clinically indicated ultrasound scans in the third trimester as per routine clinical care, and the results of these scans were reported (selective ultrasonography). All women in the cohort also had research ultrasound scans in which both the women and the clinicians caring for them were masked to the results (universal ultrasonography). After delivery, the results of the research scans were unmasked and their associations with birthweight less than the 10th percentile and neonatal morbidity were assessed. This study was designed to generate level 1 evidence of diagnostic effectiveness, as defined by the latest NICE Guideline at the time.12 Reporting of this study conforms to the STROBE (The Strengthening the Reporting of Observational Studies in Epidemiology) statement.

h percentile and neonatal morbidity were assessed. This study was designed to generate level 1 evidence of diagnostic effectiveness, as defined by the latest NICE Guideline at the time.12 Reporting of this study conforms to the STROBE (The Strengthening the Reporting of Observational Studies in Epidemiology) statement. Ethical approval for the study was obtained from the Cambridgeshire 2 Research Ethics Committee (reference 07/H0308/163) and approval to study data routinely gathered from non-participants was obtained from the South Central (Berkshire) Research Ethics Committee (reference 12/SC/0344). Participants provided written informed consent.

In conclusion, we showed that universal third trimester ultrasound tripled the detection of SGA infants and could identify FGR fetuses that were at increased risk of neonatal morbidity. The guideline1 from the Royal College of Obstetricians and Gynaecologists (RCOG) lists a series of evidence-based recommendations for the management of suspected FGR—including fetal monitoring, timing of induction of labour, and how to undertake delivery.1 We believe that a programme of screening that includes universal ultrasonography and intervention following a care bundle based on the latest RCOG guideline1 has the potential to reduce the number of adverse perinatal outcomes caused by FGR. This online publication has been corrected. The corrected version first appeared at thelancet.com on November 19, 2015 Supplementary Material Supplementary appendix Acknowledgments This work was supported by the National Institute for Health Research (NIHR) Cambridge Comprehensive Biomedical Research Centre and the Stillbirth and Neonatal Death Society. DP was supported by a Medical Research Council (MRC) Clinical Training Fellowship. IRW is supported by a MRC Unit Programme (number U105260558). GE Healthcare (Fairfield, CT, USA) donated two Voluson i ultrasound systems for this study. This study was also supported by the NIHR Cambridge Clinical Research Facility, where all visits at about 20, 28, and 36 weeks took place. No direct or indirectly supporting bodies for the project were involved in any aspect of preparation of this paper for publication. We thank the Perinatal Institute for providing a bulk calculator for customised percentiles of estimated fetal weight. We thank all the women who participated in the study, and all the staff in the Rosie Hospital (Cambridge, UK) and NIHR Cambridge Clinical Research Facility who provided direct or indirect assistance for the study.

al for the study was obtained from the Cambridgeshire 2 Research Ethics Committee (reference 07/H0308/163) and approval to study data routinely gathered from non-participants was obtained from the South Central (Berkshire) Research Ethics Committee (reference 12/SC/0344). Participants provided written informed consent. Procedures Women who agreed to participate were given follow-up appointments at about 20, 28, and 36 weeks' gestation in the National Institute for Health Research Cambridge Clinical Research Facility (Cambridge, UK). All research scans after the dating scan were done with a Voluson i system (GE Healthcare, Fairfield CT, USA) by one of a team of six sonographers, all of whom received standard training. All ultrasound examinations followed the same protocols as those used in the clinical service.13, 14 At the 20 week research appointment, participants were given a novel questionnaire we created to obtain details about their medical history and demographic characteristics.11 The 20 week scan had both routine (review of fetal anatomy and biometric measurements) and research (uterine and umbilical artery Doppler flow velocimetry) elements. Women were informed about routine elements (any concerns about the fetal anatomy and of the fetal measurements at the 20 week scan), but women and clinicians were masked to the research elements (results of the uterine and umbilical Dopplers). At the 28 and 36 week research appointments, umbilical and uterine artery Doppler flow velocimetry were repeated, and ultrasonographic measurement of fetal biparietal diameter, head circumference, abdominal circumference, and femur length were also done using standard techniques. An estimated fetal weight (EFW) percentile was calculated by use of the Hadlock equations and reference standard.15, 16 Uteroplacental Dopplers, biometry, and EFW results from the research ultrasound scans at 28 and 36 weeks were not reported to the participant or the clinician. However, both were informed about incidental findings, specifically previously undiagnosed placenta praevia, severe oligohydramnios (amniotic fluid index <5), a previously undiagnosed fetal abnormality, or non-cephalic presentation at the time of the 36 week scan.

were not reported to the participant or the clinician. However, both were informed about incidental findings, specifically previously undiagnosed placenta praevia, severe oligohydramnios (amniotic fluid index <5), a previously undiagnosed fetal abnormality, or non-cephalic presentation at the time of the 36 week scan. Gestational age was defined on the basis of ultrasonographic estimation at the time of the first scan, as recommended.3 Distributions of all measurements in the research scans were similar to previously reported reference cohorts (appendix). Summary statistics for reproducibility and reliability of research scans (assessed by two sonographers, scanning the same woman twice at the same appointment, each masked to the results of the other's scan) are tabulated for 45 women at 20 weeks' gestation and 44 women at 36 weeks' gestation (appendix). Coefficients of variation were less than 5% for fetal biometry and EFW, and between 5% and 10% for uteroplacental Doppler at both timepoints.

n twice at the same appointment, each masked to the results of the other's scan) are tabulated for 45 women at 20 weeks' gestation and 44 women at 36 weeks' gestation (appendix). Coefficients of variation were less than 5% for fetal biometry and EFW, and between 5% and 10% for uteroplacental Doppler at both timepoints. Women were selected for additional, clinically indicated scans in the third trimester of pregnancy as per routine clinical care, using local and national guidelines (eg, the NICE Guidelines on low risk women,3 women with diabetes,17 and women with hypertensive disorders18). Women were also screened with serial measurement of the symphyseal-fundal height. All women carried their maternity notes, which included a chart of the normal range of measurements for fetuses in relation to gestational age. Referral for an ultrasound scan was made by the midwife or doctor providing clinical care. Results of all clinically indicated scans were reported and paper copies were filed in both the participant's hand-held notes and hospital case records.

the normal range of measurements for fetuses in relation to gestational age. Referral for an ultrasound scan was made by the midwife or doctor providing clinical care. Results of all clinically indicated scans were reported and paper copies were filed in both the participant's hand-held notes and hospital case records. Screening status in relation to EFW was classified on the basis of the last scan before birth (which could be the 28 week scan or the 36 week scan for universal ultrasonography, depending on the gestational age at delivery). Screen positive was defined as an EFW less than the 10th percentile, using an externally derived reference range15, 16 (for both selective and universal ultrasonography). Screen negative was defined as an EFW of the 10th percentile or more (both selective and universal ultrasonography), or if no clinically indicated scan had been done at gestational age of 26 weeks or later (only selective ultrasonography).

erived reference range15, 16 (for both selective and universal ultrasonography). Screen negative was defined as an EFW of the 10th percentile or more (both selective and universal ultrasonography), or if no clinically indicated scan had been done at gestational age of 26 weeks or later (only selective ultrasonography). Outcomes Inclusion criteria for analysis were that women attended research scans booked before delivery and had a live birth at the Rosie Hospital. We excluded women who delivered before their 28 week scan appointment from the analysis. The results of clinically indicated scans and the outcome of the pregnancy were ascertained by individual review of all paper-case records by research midwives, and by linkage of the research data to the hospitals' electronic databases of ultrasonography (Astraia; Munich, Germany), delivery (Protos; iSoft, Banbury, UK), biochemical tests (Meditech; Westwood, MA, USA), and neonatal intensive care (Badgernet, Clevermed, Edinburgh, UK). The gold standard for SGA was birthweight of less than the 10th percentile for sex and gestational age, calculated from a UK reference.19 We also studied severe SGA (birthweight <3rd percentile) as a secondary outcome.

sts (Meditech; Westwood, MA, USA), and neonatal intensive care (Badgernet, Clevermed, Edinburgh, UK). The gold standard for SGA was birthweight of less than the 10th percentile for sex and gestational age, calculated from a UK reference.19 We also studied severe SGA (birthweight <3rd percentile) as a secondary outcome. We defined neonatal morbidity as one or more of the following criteria: a 5 min Apgar score of less than 7, delivery with metabolic acidosis (defined as a cord blood pH <7·1 and base deficit >10 mmol/L), or admission to the neonatal unit at term (defined as admission <48 h after birth at ≥37 weeks' gestational age and discharge ≥48 h after admission). We defined severe adverse perinatal outcome as stillbirth or term livebirth associated with neonatal death, hypoxic ischaemic encephalopathy, use of inotropes, need for mechanical ventilation, or severe metabolic acidosis (defined as a cord blood pH <7·0 and base deficit >12 mmol/L: the criteria by which international guidelines define fetal metabolic acidosis, which can be regarded as cause for cerebral palsy during childhood20). This group included seven stillbirths, which had been excluded from the main study cohort.

tabolic acidosis (defined as a cord blood pH <7·0 and base deficit >12 mmol/L: the criteria by which international guidelines define fetal metabolic acidosis, which can be regarded as cause for cerebral palsy during childhood20). This group included seven stillbirths, which had been excluded from the main study cohort. We calculated customised percentiles of EFW on the basis of reported methods,21 but used coefficients from the latest model of Gestation-Related Optimal Weight (GROW; version 6.7.3_13 [UK]). We compared associations between population-based EFW and customised EFW of less than the 10th percentile and neonatal morbidity. We analysed other indicators of growth restriction through comparison of the association between an EFW of less than the 10th percentile and neonatal morbidity, in the presence or absence of the given factor. We quantified Doppler flow velocimetry with the pulsatility index,14 and uterine artery pulsatility index as the mean pulsatility index of the left and right uterine arteries, classified by the measurement at the 20 week scan.14 We classified umbilical artery pulsatility index, head circumference-to-abdominal circumference ratio, and abdominal circumference-to-femur length ratio by the last measurement taken before birth. We quantified all measurements as gestational age adjusted Z scores, to account for variation in the exact timing of ultrasound scans (appendix). We quantified growth velocity as the difference in abdominal circumference Z score, comparing the last scan before birth and the scan at 20 weeks. For all five of these indices, we generated deciles by use of the distribution in the study cohort. We defined as abnormal the highest deciles of head circumference-to-abdominal circumference ratio, uterine Doppler, and umbilical Doppler in addition to the lowest deciles of abdominal circumference-to-femur length ratio and abdominal circumference growth velocity. We did not investigate other growth indices to reduce the possibility of chance findings due to repeated hypothesis tests. Our study did not include Doppler assessment of blood flow in fetal vessels (eg, ductus venosus or middle cerebral artery).

erence-to-femur length ratio and abdominal circumference growth velocity. We did not investigate other growth indices to reduce the possibility of chance findings due to repeated hypothesis tests. Our study did not include Doppler assessment of blood flow in fetal vessels (eg, ductus venosus or middle cerebral artery). Statistical analysis We used an open-ended recruitment approach, to provide increased power for the less common adverse outcomes with greater cohort size. In our protocol paper11 we identified a sample size of 4000 women as providing reasonably precise estimates of sensitivity for outcomes affecting 3% of the population.

Statistical analysis We used an open-ended recruitment approach, to provide increased power for the less common adverse outcomes with greater cohort size. In our protocol paper11 we identified a sample size of 4000 women as providing reasonably precise estimates of sensitivity for outcomes affecting 3% of the population. We compared continuous variables with a two-sample Wilcoxon rank-sum test and categorical variables with the Pearson χ2 test, with a trend test if appropriate, or Fisher's exact test if numbers were small. We compared sensitivity, specificity, false positive rate, and false negative rate using McNemar's test; positive and negative predictive values using weighted generalised score tests;22 and likelihood ratios using regression model-based tests.23 We did a series of post-hoc sensitivity analyses in which we included women who had defaulted from their research scans (at 28 or 36 weeks), and excluded women who had their research scan results shown to them for any reason, and in which we combined the results of universal and selective ultrasonography. We tested interactions between EFW and ultrasonic markers of FGR in their associations with neonatal morbidity by the Mantel-Haenszel test. We defined significance as p<0·05 (two-sided). We did not make formal adjustments for multiple comparisons and did not adjust for maternal baseline characteristics. Analyses were done with Stata software (version 13.1) and R software (version 3.0.2).

We compared continuous variables with a two-sample Wilcoxon rank-sum test and categorical variables with the Pearson χ2 test, with a trend test if appropriate, or Fisher's exact test if numbers were small. We compared sensitivity, specificity, false positive rate, and false negative rate using McNemar's test; positive and negative predictive values using weighted generalised score tests;22 and likelihood ratios using regression model-based tests.23 We did a series of post-hoc sensitivity analyses in which we included women who had defaulted from their research scans (at 28 or 36 weeks), and excluded women who had their research scan results shown to them for any reason, and in which we combined the results of universal and selective ultrasonography. We tested interactions between EFW and ultrasonic markers of FGR in their associations with neonatal morbidity by the Mantel-Haenszel test. We defined significance as p<0·05 (two-sided). We did not make formal adjustments for multiple comparisons and did not adjust for maternal baseline characteristics. Analyses were done with Stata software (version 13.1) and R software (version 3.0.2). Role of the funding source The funders of this study had no role in study design, data analysis, data interpretation, or writing of the report. The corresponding author and a coauthor (US) had full access to data used in the study. The corresponding author had final responsibility for the decision to submit the paper for publication.

ding source The funders of this study had no role in study design, data analysis, data interpretation, or writing of the report. The corresponding author and a coauthor (US) had full access to data used in the study. The corresponding author had final responsibility for the decision to submit the paper for publication. Results Between Jan 14, 2008, and July 31, 2012, we identified 8028 eligible women of whom 4512 (56%) provided written informed consent and were enrolled, whereas 3516 (44%) women declined to participate or were not approached by a study recruiter. Although recruited and non-recruited women were broadly comparable with each other, those recruited were slightly older in age, more often of white ethnic origin, less likely to smoke, more likely to have a caesarean delivery, and had infants of slightly heavier birthweights (appendix). 3977 (88%) of 4512 women recruited were included in the data analysis (figure 1).

roadly comparable with each other, those recruited were slightly older in age, more often of white ethnic origin, less likely to smoke, more likely to have a caesarean delivery, and had infants of slightly heavier birthweights (appendix). 3977 (88%) of 4512 women recruited were included in the data analysis (figure 1). 1666 (42%) women had a clinically indicated scan including biometry at gestation of 26 weeks or more, and 2311 (58%) women did not (table 1). Women having clinically indicated scans were more likely to be at extremes of maternal age (<20 years and ≥40 years) than those women not having clinically indicated scans and were more likely to have discontinued education early in life (<19 years), a body-mass index greater than 30 kg/m2, had previous miscarriages, and to have pre-existing diabetes or to develop gestational diabetes. Average birthweight of the infants born to this group of women was lower, and they had a greater proportion of SGA infants, preterm births, induced labours, and caesarean deliveries than women who did not have clinically indicated scans.

iscarriages, and to have pre-existing diabetes or to develop gestational diabetes. Average birthweight of the infants born to this group of women was lower, and they had a greater proportion of SGA infants, preterm births, induced labours, and caesarean deliveries than women who did not have clinically indicated scans. 352 (9%) infants had a birthweight of less than the 10th percentile. The last clinically indicated scan before birth recorded an EFW of less than the 10th percentile in 138 (8%) of 1666 women, with 69 of these women going on to have babies with birthweight less than the 10th percentile, yielding a sensitivity of 20% (69 of 352) infants. The last research ultrasound scan before birth recorded an EFW of less than the 10th percentile in 562 women; 199 of these women had babies with a birthweight less than the 10th percentile, yielding a sensitivity of 57% (199 of 352). Table 2, Table 3 compare universal and selective ultrasonography as a two-by-two table and as showing screening summary statistics, respectively. All analyses were repeated with the outcome of severe SGA. Areas under the receiver operating characteristic curve for universal ultrasonography were 0·87 (95% CI 0·85–0·88) for SGA and 0·91 (0·89–0·94) for severe SGA (appendix). Diagnostic effectiveness of the 28 and 36 week research scans are presented separately, and also described for each in relation to the interval between the scan and the delivery date (appendix). Sensitivity analyses generated very similar results to the main analysis (appendix).

1 (0·89–0·94) for severe SGA (appendix). Diagnostic effectiveness of the 28 and 36 week research scans are presented separately, and also described for each in relation to the interval between the scan and the delivery date (appendix). Sensitivity analyses generated very similar results to the main analysis (appendix). The relative risk of any neonatal morbidity associated with EFW of less than the 10th percentile was 1·6 (95% CI 1·2–2·1, p=0·001; table 4). Definition of EFW with customised percentiles did not result in a stronger association. The association between an EFW lower than the 10th percentile and the risk of neonatal morbidity was then assessed in relation to five previously reported indices of fetal growth restriction (figure 2). Only the measurement of abdominal circumference growth velocity was associated with strong evidence for an interaction (p=0·005). Screen positive fetuses with normal growth velocity were not at increased risk of neonatal morbidity, whereas an EFW of less than the 10th percentile was associated with an increase of 3·9 times (95% CI 1·9–8·1) of neonatal morbidity in infants in the lowest decile of abdominal circumference growth velocity. 172 (4·3%) fetuses had the combination of an ultrasonic diagnosis of SGA plus the lowest decile of abdominal circumference growth velocity from universal ultrasonography. This combination was associated with a relative risk of any morbidity of 2·5 (95% CI 1·7–3·6) and relative risk of delivering an SGA infant with neonatal morbidity of 17·6 (9·2–34·0; table 4). Similar associations were reported when the analysis was repeated for severe adverse perinatal outcome, with a relative risk of 2·9 (95% CI 1·0–8·3, p=0·058) for any severe adverse outcome and 39·8 (95% CI 3·6–436·6, p=0·007) for severe adverse outcome in an SGA infant. We repeated all analyses of abdominal circumference growth velocity using abdominal circumference growth charts generated by the Fetal Growth Longitudinal Study component of the INTERGROWTH-21st Project,25 an international consortium that established fetal growth standards using methods recommended by WHO. All associations were very similar when the standards from the INTERGROWTH-21st Project were used (appendix).

growth charts generated by the Fetal Growth Longitudinal Study component of the INTERGROWTH-21st Project,25 an international consortium that established fetal growth standards using methods recommended by WHO. All associations were very similar when the standards from the INTERGROWTH-21st Project were used (appendix). The combination of ultrasonic diagnosis of SGA infants plus lowest decile of abdominal circumference growth velocity (defined by the INTERGROWTH-21st Project standards) was associated with a relative risk of 2·5 (95% CI 1·7–3·5, p<0·0001) for any morbidity, 17·6 (9·4–33·0, p<0·0001) for delivering an SGA infant with neonatal morbidity, 2·5 (0·9–7·0, p=0·09) for severe adverse perinatal outcome, and 33·4 (3·0–366·6, p=0·009) for delivering an SGA infant with severe adverse perinatal outcome). Finally, no indicator of FGR was associated with adverse outcome if the EFW was above the 10th percentile (appendix).

ring an SGA infant with neonatal morbidity, 2·5 (0·9–7·0, p=0·09) for severe adverse perinatal outcome, and 33·4 (3·0–366·6, p=0·009) for delivering an SGA infant with severe adverse perinatal outcome). Finally, no indicator of FGR was associated with adverse outcome if the EFW was above the 10th percentile (appendix). Discussion Fetal growth restriction is associated with many adverse outcomes including stillbirth,26 neonatal death,27 hypoxic ischaemic encephalopathy,28 cerebral palsy,29 special educational needs,30 and many diseases in adult life.31 The present standard of care in the USA,2 UK,3 and many other countries is that women are selected for third trimester ultrasonographic fetal biometry on the basis of specific indications. From our study of a population of nulliparous women of mixed risk with a singleton pregnancy, we showed that selective use of ultrasonography identified one in five infants with a birthweight of less than the 10th percentile, which is similar to reports from other centres.4, 5, 6 Additionally, a policy of screening with universal ultrasonographic estimation of fetal weight at 28 and 36 weeks' gestational age roughly tripled the sensitivity of detection of SGA infants. However, the specificity was higher for selective ultrasonography (98%) than universal ultrasonography (90%). After the absolute numbers of true and false positives were calculated, our findings showed that for every additional SGA infant correctly identified by universal ultrasonography, about two additional results were false positives.

city was higher for selective ultrasonography (98%) than universal ultrasonography (90%). After the absolute numbers of true and false positives were calculated, our findings showed that for every additional SGA infant correctly identified by universal ultrasonography, about two additional results were false positives. On the basis of these results, implementation of universal ultrasonographic screening would likely increase the detection of SGA infants. However, it would also substantially increase the number of false positive results. The net effect on clinical outcomes would depend on the balance between any benefits that arise from identification of true positives versus any harm caused by false positives. However, even correct identification of SGA infants has the potential to cause unnecessary intervention. The population of SGA infants is well recognised to consist of both those that are healthy but small in size and those with restricted growth. We postulated that effective markers of growth restriction would identify the small fetuses who were at increased risk of neonatal morbidity and assessed five previously described ultrasonographic markers of FGR (figure 2). The only measurement that had strong evidence for an interaction was the fetal abdominal circumference growth velocity. In all other cases, SGA infants were still at increased risk of morbidity if the indicator of FGR was normal. By contrast, an EFW of less than the 10th percentile was not associated with neonatal morbidity if the infant's abdominal circumference growth velocity was normal, but was associated with about a four times increased risk of neonatal morbidity if the abdominal circumference growth velocity was in the lowest decile. Furthermore, combination of an EFW of less than the 10th percentile plus an abdominal circumference growth velocity in the lowest decile was associated with about an 18 times increased risk of mothers delivering an SGA infant with neonatal morbidity, and about a 40 times increased risk of mothers delivering an SGA infant with a severe adverse perinatal outcome. We repeated the analyses using the 2014 international reference standard to quantify abdominal circumference growth velocity25 and the results were largely identical.

g an SGA infant with neonatal morbidity, and about a 40 times increased risk of mothers delivering an SGA infant with a severe adverse perinatal outcome. We repeated the analyses using the 2014 international reference standard to quantify abdominal circumference growth velocity25 and the results were largely identical. Our results showed that customisation of the EFW did not increase the strength of association between SGA and neonatal morbidity. The process of customisation attempts to relate the estimated size of a fetus to its genetic potential, using the maternal characteristics. We interpreted the findings of this study to suggest that the size of a fetus at the 20 week scan might be a better proxy of its genetic growth potential than the maternal characteristics. Our finding that abdominal circumference growth velocity was better than either uterine or umbilical Doppler to distinguish between SGA infants at low risk and high risk is consistent with the view that poor growth could be an endpoint of several pathological changes. Hence, assessment of growth velocity might be a more appropriate marker of adverse outcome as additional specific tests only provide information about a subset of FGR caused by a specific pathophysiological pathway.

risk is consistent with the view that poor growth could be an endpoint of several pathological changes. Hence, assessment of growth velocity might be a more appropriate marker of adverse outcome as additional specific tests only provide information about a subset of FGR caused by a specific pathophysiological pathway. Our study has strengths and weaknesses. One of the strengths is that clinicians were blinded to the results of research ultrasonographic assessments of fetal biometry and uteroplacental Doppler. The justification for concealment of the results of the research biometry and uteroplacental Doppler was taken from the NICE recommendation that these scans should not be offered routinely.3 The rationale for concealment of the biometry was that if the results had been known, they might have biased subsequent assessment of symphyseal-fundal height. A limitation of our study is that it was confined to nulliparous women. The rationale for selection of this group was that nulliparous women have higher rates of SGA than multiparous women and, by definition, no information is available about any history of previous SGA births in this group, which is one of the strongest predictors of SGA in a pregnancy. However, further studies are needed to establish whether universal ultrasonography is also effective in multiparous women. The major single cause of non-anomalous perinatal deaths at term is antepartum stillbirth, and about 30% of these stillbirths are associated with poor fetal growth.32 Although stillbirth was included in our composite of severe adverse outcome, our study was underpowered to investigate stillbirth directly. However, we speculate that the same ultrasonic features associated with neonatal morbidity and the composite of severe adverse perinatal outcome are likely to be associated with the risk of stillbirth.

as included in our composite of severe adverse outcome, our study was underpowered to investigate stillbirth directly. However, we speculate that the same ultrasonic features associated with neonatal morbidity and the composite of severe adverse perinatal outcome are likely to be associated with the risk of stillbirth. In conclusion, we showed that universal third trimester ultrasound tripled the detection of SGA infants and could identify FGR fetuses that were at increased risk of neonatal morbidity. The guideline1 from the Royal College of Obstetricians and Gynaecologists (RCOG) lists a series of evidence-based recommendations for the management of suspected FGR—including fetal monitoring, timing of induction of labour, and how to undertake delivery.1 We believe that a programme of screening that includes universal ultrasonography and intervention following a care bundle based on the latest RCOG guideline1 has the potential to reduce the number of adverse perinatal outcomes caused by FGR. This online publication has been corrected. The corrected version first appeared at thelancet.com on November 19, 2015 Supplementary Material Supplementary appendix

Perinatal Institute for providing a bulk calculator for customised percentiles of estimated fetal weight. We thank all the women who participated in the study, and all the staff in the Rosie Hospital (Cambridge, UK) and NIHR Cambridge Clinical Research Facility who provided direct or indirect assistance for the study. Contributors GCSS created the study concept and design. AD and DP acquired data. US, IRW, and GCSS did the data analysis and interpretation. US and GCSS drafted the manuscript. All authors contributed to the critical revision of the manuscript for important intellectual content and approved the final version to be published. US and GCSS had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Declaration of interests DP reports a grant from the Medical Research Council (MRC), during the study. GCSS reports grants from the National Institute for Health Research (NIHR), Stillbirth and Neonatal Death Society, and MRC, and reports other from GE and NIHR Cambridge Clinical Research Facility, during the study; reports personal fees and other from GlaxoSmithKline and Roche, and reports other from Chiesi, outside of the study. US, IRW, and AD declare no competing interests. Figure 1 Study profile Figure 2 Stratified analyses of the risk of the neonatal composite adverse outcome associated with diagnosis of small-for-gestational-age infants

Declaration of interests DP reports a grant from the Medical Research Council (MRC), during the study. GCSS reports grants from the National Institute for Health Research (NIHR), Stillbirth and Neonatal Death Society, and MRC, and reports other from GE and NIHR Cambridge Clinical Research Facility, during the study; reports personal fees and other from GlaxoSmithKline and Roche, and reports other from Chiesi, outside of the study. US, IRW, and AD declare no competing interests. Figure 1 Study profile Figure 2 Stratified analyses of the risk of the neonatal composite adverse outcome associated with diagnosis of small-for-gestational-age infants Diagnosis of infants by universal ultrasonography in relation to indicators of fetal growth restriction. The five previously described indices of fetal growth restriction were classified as the extreme decile associated with fetal growth restriction (highest or lowest, as appropriate) compared with the other nine deciles in the cohort. Points are relative risks of neonatal morbidity associated with an ultrasonic diagnosis of a small-for-gestational-age infant at the last scan before birth. p values are a Mantel-Haenszel test calculation of interaction (ie, testing the hypothesis that the association between diagnosis of a small-for-gestation-age infant and neonatal morbidity varies in the two strata). Interactions tested using logistic regression showed almost identical p values. AC=abdominal circumference. FL=femur length. HC=head circumference. Table 1 Characteristics of the study cohort

Diagnosis of infants by universal ultrasonography in relation to indicators of fetal growth restriction. The five previously described indices of fetal growth restriction were classified as the extreme decile associated with fetal growth restriction (highest or lowest, as appropriate) compared with the other nine deciles in the cohort. Points are relative risks of neonatal morbidity associated with an ultrasonic diagnosis of a small-for-gestational-age infant at the last scan before birth. p values are a Mantel-Haenszel test calculation of interaction (ie, testing the hypothesis that the association between diagnosis of a small-for-gestation-age infant and neonatal morbidity varies in the two strata). Interactions tested using logistic regression showed almost identical p values. AC=abdominal circumference. FL=femur length. HC=head circumference. Table 1 Characteristics of the study cohort No clinically indicated scan ≥26 weeks (n=2311) One or more clinically indicated scan ≥26 weeks (n=1666) p value Overall baseline characteristics (N=3977) Maternal characteristics Age (years) <0·0001 <20 66 (3%) 73 (4%) .. 139 (4%) 20–24·9 311 (13%) 209 (13%) .. 520 (13%) 25–29·9 757 (33%) 468 (28%) .. 1225 (31%) 30–34·9 887 (38%) 598 (36%) .. 1485 (37%) 35–39·9 274 (12%) 260 (16%) .. 534 (13%) ≥40 16 (1%) 58 (3%) .. 74 (2%) Age stopped FTE (years) 0·01 <19 728 (32%) 593 (36%) .. 1321 (33%) 19–22 828 (36%) 555 (33%) .. 1383 (35%) ≥23 697 (30%) 463 (28%) .. 1160 (29%) Missing 58 (3%) 55 (3%) NA 113 (3%) Deprivation quartile 0·18 1 (lowest) 574 (25%) 400 (24%) .. 974 (24%) 2 550 (24%) 393 (24%) .. 943 (24%) 3 561 (24%) 399 (24%) .. 960 (24%) 4 (highest) 523 (23%) 416 (25%) .. 939 (24%) Missing 103 (4%) 58 (3%) NA 161 (4%) Postcode area 0·17 CB1–5 709 (31%) 514 (31%) .. 1223 (31%) CB21–25 528 (23%) 374 (22%) .. 902 (23%) CB6–11 549 (24%) 446 (27%) .. 995 (25%) Outside Cambridgeshire 459 (20%) 302 (18%) .. 761 (19%) Missing 66 (3%) 30 (2%) NA 96 (2%) Ethnic origin White 2151 (93%) 1545 (93%) 0·65 3696 (93%) Missing 40 (2%) 29 (2%) NA 69 (2%) Married 1576 (68%) 1151 (69%) 0·55 2727 (69%) Smoker 106 (5%) 79 (5%) 0·82 185 (5%) Alcohol consumption Any 117 (5%) 66 (4%) 0·10 183 (5%) Missing 1 (<1%) 0 (0%) NA 1 (<1%) BMI (kg/m2) <0·0001 <25 1416 (61%) 909 (55%) .. 2325 (58%) 25–29·9 667 (29%) 450 (27%) .. 1117 (28%) 30–34·9 209 (9%) 168 (10%) .. 377 (9%) 35–39·9 18 (1%) 92 (6%) .. 110 (3%) ≥40 1 (<1%) 46 (3%) .. 47 (1%) Missing 0 (0%) 1 (<1%) NA 1 (<1%) ≥1 previous miscarriage 207 (9%) 199 (12%) 0·002 406 (10%) Diabetes <0·0001 Type 1 or type 2 0 (0%) 14 (1%) .. 14 (<1%) Gestational 2 (<1%) 160 (10%) .. 162 (4%) Missing 3 (<1%) 2 (<1%) NA 5 (<1%) Birth outcomes Birthweight (g) 3480 (3175–3770) 3345 (3010–3685) <0·0001 3420 (3105–3740) SGA (<10th) 178 (8%) 174 (10%) 0·003 352 (9%) Severe SGA (<3rd) 34 (1%) 53 (3%) 0·0003 87 (2%) Gestational age (weeks) <0·0001 Preterm (26–32) 15 (1%) 15 (1%) .. 30 (1%) Preterm (33–36) 53 (2%) 80 (5%) .. 133 (3%) Term (≥37) 2243 (97%) 1571 (94%) ..

hweight (g) 3480 (3175–3770) 3345 (3010–3685) <0·0001 3420 (3105–3740) SGA (<10th) 178 (8%) 174 (10%) 0·003 352 (9%) Severe SGA (<3rd) 34 (1%) 53 (3%) 0·0003 87 (2%) Gestational age (weeks) <0·0001 Preterm (26–32) 15 (1%) 15 (1%) .. 30 (1%) Preterm (33–36) 53 (2%) 80 (5%) .. 133 (3%) Term (≥37) 2243 (97%) 1571 (94%) .. 3814 (96%) Induction of labour Yes 629 (27%) 629 (38%) <0·0001 1258 (32%) Missing 4 (<1%) 2 (<1%) NA 6 (<1%) Mode of delivery <0·0001 Spontaneous vaginal 1218 (53%) 706 (42%) .. 1924 (48%) Assisted vaginal 596 (26%) 353 (21%) .. 949 (24%) Intrapartum caesarean 415 (18%) 283 (17%) .. 698 (18%) Prelabour caesarean 74 (3%) 317 (19%) .. 391 (10%) Missing 8 (<1%) 7 (<1%) NA 15 (<1%) Data are n (%) or median (IQR). p values are for difference between groups calculated using the two-sample Wilcoxon rank-sum (Mann-Whitney) test for continuous variables and the Pearson χ2 test for categorical variables, with trend tests if appropriate. The “missing” category was not included in statistical tests. For characteristics that have no “missing” category, data were 100% complete. Maternal age was defined as age at recruitment to study. All other maternal characteristics were defined by self-report at the 20 weeks questionnaire, from examination of the clinical case record, or linkage to the hospital's electronic databases. Socioeconomic status was quantified by use of the Index of Multiple Deprivation24 2007, which is based on census data from the area of the mother's postcode. FTE=full-time education. CB1–5=central Cambridge city. CB21–25=peripheral Cambridge city. CB6–11=Cambridgeshire, outside city. NA=not applicable. BMI=body-mass index. SGA=small for gestational age.

by use of the Index of Multiple Deprivation24 2007, which is based on census data from the area of the mother's postcode. FTE=full-time education. CB1–5=central Cambridge city. CB21–25=peripheral Cambridge city. CB6–11=Cambridgeshire, outside city. NA=not applicable. BMI=body-mass index. SGA=small for gestational age. Table 2 Screening effectiveness for selective and universal ultrasonographic screening for infants who are small and severely small for gestational age

obesity (male 1·70, female 1·37; appendix p 22) suggest that men have almost double the proportional excess mortality of women— but, as age-specific death rates are typically more than 50% higher in men, the absolute excess death rate associated with grade 1 obesity is about three times as great in men (appendix p 45). Because the prevalence of obesity differs by region, for all-cause mortality there was wide variation across regions in the approximate population-attributable fraction due to overweight and obesity. These findings suggest that if the overweight and obese population had WHO-defined normal levels of BMI, the proportion of premature deaths that could be avoided would be about one in five in North America, one in six in Australia and New Zealand, one in seven in Europe, and one in 20 in east Asia, assuming that the associations of overweight and obesity with mortality in our primary analyses largely reflect causal effects. Moreover, BMI is increasing in many populations, so the pattern of high mortality from adiposity in North America might become typical elsewhere.31 At the opposite extreme, there was a substantially higher mortality not only among those in WHO's underweight category, but also in those with BMI 18·5 kg/m2 to <20 kg/m2, suggesting that in excessively lean adult populations underweight remains a cause for concern. We have no information about whether the BMI in underweight individuals was always low.

by use of the Index of Multiple Deprivation24 2007, which is based on census data from the area of the mother's postcode. FTE=full-time education. CB1–5=central Cambridge city. CB21–25=peripheral Cambridge city. CB6–11=Cambridgeshire, outside city. NA=not applicable. BMI=body-mass index. SGA=small for gestational age. Table 2 Screening effectiveness for selective and universal ultrasonographic screening for infants who are small and severely small for gestational age SGA Severe SGA Yes No Total Yes No Total Selective ultrasonography EFW <10th 69 69 138 28 110 138 EFW ≥10th or no scan 283 3556 3839 59 3780 3839 Total 352 3625 3977 87 3890 3977 Universal ultrasonography EFW <10th 199 363 562 67 495 562 EFW ≥10th 153 3262 3415 20 3395 3415 Total 352 3625 3977 87 3890 3977 Selective ultrasonography shows the results of clinically indicated scans. Of the 1666 women selected for ultrasonography at 26 weeks or later, 1388 (83%) had one or two scans, 245 (15%) had three or four scans, and 33 (2%) had five or more scans. If a woman did not have a clinically indicated scan after the routine anomaly scan she was defined as screen negative by selective ultrasonography. Universal ultrasonography shows the results of the last research scan done before birth (either the 28 week scan or the 36 week scan, depending on the gestational age at delivery). Median time interval (IQR) between the last selective scan and birth was 3·1 weeks (1·6–5·6 weeks), and between the last universal scan and birth was 4·1 weeks (3·1–5·0 weeks). SGA=small for gestational age (birthweight <10th percentile; severe SGA birthweight <3rd percentile). EFW=estimated fetal weight (from the last scan before birth).

(IQR) between the last selective scan and birth was 3·1 weeks (1·6–5·6 weeks), and between the last universal scan and birth was 4·1 weeks (3·1–5·0 weeks). SGA=small for gestational age (birthweight <10th percentile; severe SGA birthweight <3rd percentile). EFW=estimated fetal weight (from the last scan before birth). Table 3 Diagnostic effectiveness of selective versus universal ultrasonographic screening for infants who are small and severely small for gestational age

(IQR) between the last selective scan and birth was 3·1 weeks (1·6–5·6 weeks), and between the last universal scan and birth was 4·1 weeks (3·1–5·0 weeks). SGA=small for gestational age (birthweight <10th percentile; severe SGA birthweight <3rd percentile). EFW=estimated fetal weight (from the last scan before birth). Table 3 Diagnostic effectiveness of selective versus universal ultrasonographic screening for infants who are small and severely small for gestational age SGA Severe SGA Selective Universal Selective Universal Sensitivity (%) 20% (15–24) 57% (51–62) 32% (22–42) 77% (68–86) Specificity (%) 98% (98–99) 90% (89–91) 97% (97–98) 87% (86–88) Positive predictive value (%) 50% (42–58) 35% (31–39) 20% (14–27) 12% (9–15) Negative predictive value (%) 93% (92–93) 96% (95–96) 98% (98–99) 99% (99–100) False positive rate* (%) 2% (1–2) 10% (9–11) 3% (2–3) 13% (12–14) False negative rate† (%) 80% (76–85) 43% (38–49) 68% (58–78) 23% (14–32) Positive likelihood ratio 10·3 (7·5–14·1) 5·6 (4·9–6·5)‡ 11·4 (8·0–16·3) 6·1 (5·3–7·0) Negative likelihood ratio 0·8 (0·8–0·9) 0·5 (0·4–0·5)‡ 0·7 (0·6–0·8) 0·3 (0·2–0·4) Relative sensitivity 1·0 (reference) 2·9 (2·4–3·5) 1·0 (reference) 2·4 (1·8–3·2) 95% CIs are given in brackets. All values were calculated with estimated fetal weight <10th percentile as screen positive. Statistical comparison by McNemar, weighted generalised score tests, or regression model-based tests as appropriate. All comparisons of selective vs universal had p<0·0001 for both outcomes, except for SGA positive likelihood ratio (p=0·0001), severe SGA positive predictive value (p=0·0002), and positive likelihood ratio (p=0·0003). SGA=small for gestational age (birthweight <10th percentile; severe SGA birthweight <3rd percentile).

All comparisons of selective vs universal had p<0·0001 for both outcomes, except for SGA positive likelihood ratio (p=0·0001), severe SGA positive predictive value (p=0·0002), and positive likelihood ratio (p=0·0003). SGA=small for gestational age (birthweight <10th percentile; severe SGA birthweight <3rd percentile). * Defined as proportion of screen positives among non-SGA infants. † Defined as proportion of screen negatives among SGA infants. ‡ Sample calculation: positive likelihood ratio= (199 ÷ 363)/(352 ÷ 3625)=5·6; negative likelihood ratio= (153 ÷ 3262)/(352 ÷ 3625)=0·5. Table 4 Association between perinatal outcomes of estimated fetal weight less than the 10th percentile and abdominal circumference growth velocity

† Defined as proportion of screen negatives among SGA infants. ‡ Sample calculation: positive likelihood ratio= (199 ÷ 363)/(352 ÷ 3625)=5·6; negative likelihood ratio= (153 ÷ 3262)/(352 ÷ 3625)=0·5. Table 4 Association between perinatal outcomes of estimated fetal weight less than the 10th percentile and abdominal circumference growth velocity Any neonatal morbidity (n=275) Metabolic acidosis (n=42) 5 min Apgar <7 (n=36) Neonatal unit admission (n=229) SGA plus any neonatal morbidity (n=49) Severe adverse perinatal outcome (n=33) SGA plus severe adverse perinatal outcome (n=5) RR (95% CI) p value RR (95% CI) p value RR (95% CI) p value RR (95% CI) p value RR (95% CI) p value RR (95% CI) p value RR (95% CI) p value EFW <10th (population) 1·6 (1·2–2·1) 0·001 1·4 (0·7–3·1) 0·37 2·3 (1.1–4·8) 0·03 1·6 (1·2–2·1) 0·006 10·5 (5·9–18·6) <0·0001 1·4 (0·6–3·3) 0·45 24·3 (2·7–217·1) 0·002 EFW <10th (customised*) 1·7 (1·3–2·3) 0·001 1·5 (0·6–3·4) 0·44 1·7 (0·7–4.2) 0·26 1·6 (1·2–2·3) 0·01 9·8 (5·7–17·1) <0·0001 1·9 (0·8–4·7) 0·14 34·8 (3·9–310·4) 0·0005 EFW <10th plus normal ACGV 1·3 (0·9–1·8) 0·23 0·3 (0·0–1·9) 0·25 1·4 (0.5–3·9) 0·54 1·4 (0·9–2·0) 0·13 7·3 (3·7–14·4) <0·0001 0·7 (0·2–2·7) 0·76 17·6 (1·6–193·5) 0·03 EFW <10th plus lowest decile ACGV 2·5 (1·7–3·6) <0·0001 4·1 (1·8–9·1) 0·003 4·6 (1·9–11·0) 0·004 2·1 (1·3–3·2) 0·003 17·6 (9·2–34·0) <0·0001 2·9 (1·0–8·3) 0·06 39·8 (3·6–436·6) 0·007 All estimated fetal weights (EFWs) are based on population-based percentiles, unless stated otherwise. All relative risks (RRs) are referent to infants with an EFW of ≥10th percentile by population-based standards, except for the RRs for customised EFW <10th percentile, which are referent to infants with an EFW of the ≥10th percentile by customised standards. Appendix has n/N for every cell. Small for gestational age (SGA) is defined as birthweight of <10th percentile by population standards. Abdominal circumference growth velocity (ACGV) is based on the change in the gestational age adjusted Z score, comparing the result at the 20 week scan with the last scan before birth. Neonatal morbidity is a composite outcome—ie, one or more of these three outcomes: metabolic acidosis (defined as pH <7·1 and base deficit >10 mmol/L), 5 min Apgar score less than 7, and neonatal unit admission (defined as admission to the neonatal intensive care unit, the high dependency unit, or the special care baby unit).

l morbidity is a composite outcome—ie, one or more of these three outcomes: metabolic acidosis (defined as pH <7·1 and base deficit >10 mmol/L), 5 min Apgar score less than 7, and neonatal unit admission (defined as admission to the neonatal intensive care unit, the high dependency unit, or the special care baby unit). Severe adverse perinatal outcome is a composite outcome—ie, one or more of the following outcomes specified: stillbirth (not due to congenital anomaly), neonatal death at term (not due to congenital anomaly), hypoxic ischaemic encephalopathy at term, use of inotropes at term, mechanical ventilation at term, severe metabolic acidosis at term (defined as pH <7·0 and base deficit >12 mmol/L). p values (all two sided) are from Fisher's exact test. * Customised percentiles of EFW were calculated with the Gestation-Related Optimal Weight Customised Weight Centile Calculator (version 6.7 [UK]). Panel Research in context Systematic review

Severe adverse perinatal outcome is a composite outcome—ie, one or more of the following outcomes specified: stillbirth (not due to congenital anomaly), neonatal death at term (not due to congenital anomaly), hypoxic ischaemic encephalopathy at term, use of inotropes at term, mechanical ventilation at term, severe metabolic acidosis at term (defined as pH <7·0 and base deficit >12 mmol/L). p values (all two sided) are from Fisher's exact test. * Customised percentiles of EFW were calculated with the Gestation-Related Optimal Weight Customised Weight Centile Calculator (version 6.7 [UK]). Panel Research in context Systematic review Previously reported large-scale systematic reviews have addressed the ability of universal ultrasonography to result in improved clinical outcome. Reported Cochrane reviews have examined both conventional ultrasound and use of Doppler flow velocimetry of the uteroplacental circulation. Neither of these methods has been shown to improve outcome. The National Institute for Health and Care Excellence (NICE) did a highly detailed review of the methods of screening women at low risk for fetal wellbeing for the 2008 Guideline on Antenatal Care. This guideline included both results of the Cochrane reviews of clinical effectiveness of universal ultrasound and results of the new meta-analyses of diagnostic effectiveness of screening for small-for-gestational-age (SGA) infants. The guideline concluded that high quality evidence on the diagnostic effectiveness of routine ultrasonography was scarce, and recommended several aspects of screening for SGA infants as research priorities. It might seem paradoxical that universal ultrasonography has not been shown to be beneficial in randomised controlled trials, but NICE recommended further research into its diagnostic effectiveness. However, this absence of benefit might be attributed to the impossibility of designing an interventional study of screening without knowledge of the diagnostic effectiveness of the screening test in the given population. Moreover, many observational studies reported that SGA infants are at increased risk of complications, methods used to screen for SGA have low sensitivity and specificity, and undiagnosed SGA is a common finding in perinatal deaths.

nowledge of the diagnostic effectiveness of the screening test in the given population. Moreover, many observational studies reported that SGA infants are at increased risk of complications, methods used to screen for SGA have low sensitivity and specificity, and undiagnosed SGA is a common finding in perinatal deaths. Interpretation

nowledge of the diagnostic effectiveness of the screening test in the given population. Moreover, many observational studies reported that SGA infants are at increased risk of complications, methods used to screen for SGA have low sensitivity and specificity, and undiagnosed SGA is a common finding in perinatal deaths. Interpretation Our study shows that universal use of ultrasound roughly tripled the detection of SGA infants. Moreover, by use of a combination of assessment of the estimated fetal weight of a baby with a previously reported measurement of fetal growth restriction (abdominal circumference growth velocity), the population of small fetuses could be divided into about 30% that were growth restricted and at increased risk of neonatal morbidity, and about 70% that did not have growth restriction and were at the same risk of neonatal morbidity as those babies not identified as small by universal ultrasound. The statistical associations (for both main effects and interactions) reported were sufficiently strong that these findings were unlikely to be chance or data driven. These results suggest that the screening of unselected nulliparous women and identification of growth-restricted fetuses at increased risk of an adverse outcome is possible. Additionally, the same combination of findings associated with the risk of neonatal morbidity could plausibly be predictive of other adverse outcomes associated with fetal growth restriction, such as antepartum stillbirth. As a result, a programme to screen with universal ultrasonography and intervention using a care bundle based on the latest Royal College of Obstetricians and Gynaecologists guideline might reduce the number of adverse perinatal outcomes caused by fetal growth restriction.

Introduction For national disease control strategies, and for individual decisions about smoking, what matters is not just current but also future tobacco-attributed mortality rates. China now consumes over a third of the world's cigarettes,1 but the increase is too recent for the full effect of current cigarette consumption levels on the hazard per continuing smoker to have emerged.2, 3 There has been a large intergenerational increase in cigarette smoking by young men, first in urban and then in rural areas. In contrast, there has been a large intergenerational decrease in smoking among women, at least until women born in the 1970s.4, 5, 6, 7 So, assessment of future tobacco hazards must allow for urban versus rural and, particularly, male versus female differences in lifelong smoking patterns. Research in context Systematic review

Introduction For national disease control strategies, and for individual decisions about smoking, what matters is not just current but also future tobacco-attributed mortality rates. China now consumes over a third of the world's cigarettes,1 but the increase is too recent for the full effect of current cigarette consumption levels on the hazard per continuing smoker to have emerged.2, 3 There has been a large intergenerational increase in cigarette smoking by young men, first in urban and then in rural areas. In contrast, there has been a large intergenerational decrease in smoking among women, at least until women born in the 1970s.4, 5, 6, 7 So, assessment of future tobacco hazards must allow for urban versus rural and, particularly, male versus female differences in lifelong smoking patterns. Research in context Systematic review Existing reviews of smoking and death in China include mainly studies established during the 1980s or 1990s.23, 24 We searched PubMed for further such studies using the terms “smoking” AND “mortality” AND “China”, but found no large studies reported in English. The previous studies mainly involved people born before 1950 who, unlike those born more recently, had not smoked cigarettes persistently since early adulthood, or had smoked forms of tobacco (eg, pipes) that carry a lower risk than manufactured cigarettes. These previous studies could not assess directly the growing risks of tobacco in the present century. For, the hazards among smokers depend importantly not only on recent smoking patterns, but also on patterns in early adult life,2, 13, 14 and consumption of substantial numbers of cigarettes from early adulthood used to be uncommon, particularly in rural areas. Future hazards can be assessed only by studies of large numbers who have smoked cigarettes ever since early adulthood. Our two nationally representative studies can help to assess how the epidemic of smoking in mainland China has developed, and how it will evolve.

thood used to be uncommon, particularly in rural areas. Future hazards can be assessed only by studies of large numbers who have smoked cigarettes ever since early adulthood. Our two nationally representative studies can help to assess how the epidemic of smoking in mainland China has developed, and how it will evolve. Interpretation There are contrasting male and female trends. Among women, tobacco-attributed mortality is currently about 5%, 3%, 1%, and <1% of all mortality in those born in the 1930s, 1940s, 1950s, and since 1960. Female tobacco-attributed mortality is likely to fall as the pre-1950 generation, in which some women smoke, is replaced by the next generation, in which far fewer do so. Among men, the situation is opposite; the proportion of male deaths from smoking has been increasing, first in urban and then in rural areas, and already by 2010 tobacco accounted for a quarter of urban male deaths at ages 40–79 years. Among urban male smokers who started to smoke cigarettes (rather than other forms of tobacco) before age 20 years, which is the uptake pattern that is now typical among young men throughout China, overall mortality is already twice as great as among otherwise similar men who never smoked. As the epidemic matures, first in urban and then in rural areas, and the population grows, Chinese tobacco deaths will rise from about 1 million in 2010, to 2 million in 2030, and 3 million in 2050, unless there is widespread cessation.

ortality is already twice as great as among otherwise similar men who never smoked. As the epidemic matures, first in urban and then in rural areas, and the population grows, Chinese tobacco deaths will rise from about 1 million in 2010, to 2 million in 2030, and 3 million in 2050, unless there is widespread cessation. Assessment of current and future tobacco hazards must also allow for other changes in mortality. Communicable disease mortality is decreasing steeply and non-communicable disease mortality is also decreasing, albeit more slowly.8, 9 This health transition is continuing, so all-cause mortality is falling in children and in adults, despite the net effects of changes in smoking and in adiposity. At 1970 Chinese rates, half would die before age 70 years (including 10% before age 5 plus another 10% before age 50). At 2010 rates, however, only a quarter would die before age 70 years (including only 1% before age 5 plus another 4% before age 50).9 These improvements are likely to continue (except perhaps in male smokers), due to better treatment and reductions in many other causes of disease.

age 5 plus another 10% before age 50). At 2010 rates, however, only a quarter would die before age 70 years (including only 1% before age 5 plus another 4% before age 50).9 These improvements are likely to continue (except perhaps in male smokers), due to better treatment and reductions in many other causes of disease. The Chinese population has stabilised below age 50, but at ages 60–79 years it will double from 2010 to 2030, then stabilise.10 The rising over-50 population and falling age-specific death rates affect projections of tobacco-attributed mortality, and of the eventual risks for those smokers who, as has become usual for men in urban and then rural China, started substantial cigarette usage before age 20 years. The hazards currently seen in urban men who started cigarette smoking before age 20 years are particularly important, as they foreshadow what nationwide smoker versus non-smoker mortality rate ratios (RRs) will eventually become.2, 3

n urban and then rural China, started substantial cigarette usage before age 20 years. The hazards currently seen in urban men who started cigarette smoking before age 20 years are particularly important, as they foreshadow what nationwide smoker versus non-smoker mortality rate ratios (RRs) will eventually become.2, 3 Successive nationally representative epidemiological studies, conducted some time apart, can help to monitor and predict trends in tobacco-attributed mortality.11, 12, 13, 14, 15 We report two large studies in China, one monitoring deaths from 1991–99 (mid-year 1995),16 and the second monitoring deaths 15 years later, from 2006–14 (mid-year 2010).17, 18 Changing smoker versus non-smoker RRs between the two studies show how the epidemic is evolving. The baseline survey of the second study (which was more detailed than that of the first) yields data on intergenerational changes in smoking patterns among urban and rural men and women, and prospective follow-up of mortality yields the RRs associated with each pattern, including the particularly informative hazards among urban men who have smoked cigarettes since before age 20 years.

an that of the first) yields data on intergenerational changes in smoking patterns among urban and rural men and women, and prospective follow-up of mortality yields the RRs associated with each pattern, including the particularly informative hazards among urban men who have smoked cigarettes since before age 20 years. Methods Study populations After a pilot phase in 1990–91, the first nationwide prospective study (Chinese Prospective Smoking Study [CPSS]) had its main recruitment phase from April 9, to Dec 31, 1991, with follow-up to Dec 31, 1999 (mid-year 1995). The second (China Kadoorie Biobank [CKB] study) recruited from June 25, 2004, to July 15, 2008, with follow-up to Jan 1, 2014 (mid-year 2010). The designs and methods of both studies have been described previously.16, 17, 18

a substantially higher mortality not only among those in WHO's underweight category, but also in those with BMI 18·5 kg/m2 to <20 kg/m2, suggesting that in excessively lean adult populations underweight remains a cause for concern. We have no information about whether the BMI in underweight individuals was always low. Our primary analyses used three main approaches to help avoid bias. First, we restricted analysis to never-smokers to avoid residual confounding by smoking as far as possible because merely adjusting for smoking habits would be unlikely to eliminate important residual biases due to the effect on BMI of different intensities of smoking.13 Second, we sought to exclude people known to have specific pre-existing chronic diseases (although full information about this variable was often unavailable). Finally, we omitted the initial 5 years of follow-up from the analysis because diseases at baseline that might cause death over the next 5 years could result in reverse causation (where lower BMI at recruitment is the result, rather than the cause, of the underlying pathology).14, 15, 16

ecruitment phase from April 9, to Dec 31, 1991, with follow-up to Dec 31, 1999 (mid-year 1995). The second (China Kadoorie Biobank [CKB] study) recruited from June 25, 2004, to July 15, 2008, with follow-up to Jan 1, 2014 (mid-year 2010). The designs and methods of both studies have been described previously.16, 17, 18 In the first, 225  721 men were recruited from 45 study areas, chosen at random from 145 nationally representative Disease Surveillance Points (DSPs). A typical DSP monitors cause-specific deaths in 50 000–100 000 residents in five to ten nearby residential units (groups of rural villages or urban street committees). In each study area (22 rural, 23 urban), two or three such units were randomly selected and all men aged 40 years or older were identified through local residential records and invited to take part; about four-fifths participated (including 219 893 at ages 40–79 years). In local study clinics, trained health workers measured blood pressure, height, weight, and peak expiratory flow rate and administered a standardised questionnaire on demographics, education, occupation, smoking, drinking, diet, and self-reported medical history. Survey records were handwritten on paper, with subsequent data entry double-punched.

ned health workers measured blood pressure, height, weight, and peak expiratory flow rate and administered a standardised questionnaire on demographics, education, occupation, smoking, drinking, diet, and self-reported medical history. Survey records were handwritten on paper, with subsequent data entry double-punched. In the second study, 210 222 men and 302 669 women were recruited in 2004–08 from ten diverse locations across China, four urban and six rural (or semi-rural in the case of Suzhou). These study areas chosen from the DSPs to span a range of socioeconomic levels, risk factor patterns, and disease patterns. The set of all DSPs is nationally representative, and after excluding a few with organisational difficulties, the choice of ten study sites from the remaining DSPs was made centrally, knowing local characteristics and cause-specific death rates. The exact choice of areas was decided carefully, aiming successfully to retain geographic and social balance and balance across deciles of mortality from major diseases, so the set of ten study sites should still be approximately nationally representative. All 1 801 167 registered residents of age 35–74 years were identified through local records and invited to survey clinics, and 500 223 (28%) participated; another 12 668 just outside this age range also participated. As a substantial minority of registered residents would actually have been living elsewhere, we estimate that about a third of the invitees actually living in the study areas participated. Trained health workers took blood for long-term storage; measured height, weight, waist, hips, bioimpedance, blood pressure, heart rate, and lung function; and administered laptop-based questionnaires on tobacco, alcohol, diet, indoor air pollution, physical activity, education, socio-demographic status, medical history, and female reproductive history.17

long-term storage; measured height, weight, waist, hips, bioimpedance, blood pressure, heart rate, and lung function; and administered laptop-based questionnaires on tobacco, alcohol, diet, indoor air pollution, physical activity, education, socio-demographic status, medical history, and female reproductive history.17 Assessment of smoking In both studies, questions about smoking included frequency, type, amount (current and past), age first began, age stopped, and main reason for cessation (already ill or stopped by choice), with additional information in the second study on inhalation and exhaled CO (MicroCO meter, Carefusion, San Diego, CA, USA).19 Regular smokers used one or more cigarettes (or ≥1 g tobacco) daily for at least 6 months. Of smokers who had stopped (≥6 months), about half had done so because they were ill, and (to avoid bias) were still counted with smokers in the main analyses; the remainder, who are described as having stopped by choice, helped assess the effects of cessation.

cigarettes (or ≥1 g tobacco) daily for at least 6 months. Of smokers who had stopped (≥6 months), about half had done so because they were ill, and (to avoid bias) were still counted with smokers in the main analyses; the remainder, who are described as having stopped by choice, helped assess the effects of cessation. Mortality follow-up Cause-specific mortality in both studies was monitored through DSP death registries,20 and checked annually against local residential records, with active confirmation of survival through street committees or village administrators.18 In the second study, deaths were also monitored through the new nationwide health insurance system (yielding few additional cases). Causes were checked against any available medical records. For the few deaths (about 5% in both studies) without recent medical attention, standardised procedures determined probable causes from symptoms or signs described by informants (usually family). Deaths were coded by trained staff, blind to baseline data, using International Classification of Diseases (ICD)-9 in the first study and ICD-10 in the second (appendix p 3). At ages 40–79 years, each study had about 25 000 deaths.

ermined probable causes from symptoms or signs described by informants (usually family). Deaths were coded by trained staff, blind to baseline data, using International Classification of Diseases (ICD)-9 in the first study and ICD-10 in the second (appendix p 3). At ages 40–79 years, each study had about 25 000 deaths. Statistical analysis Cox regression yielded multi-covariate-adjusted smoker versus never-smoker RRs at ages 40–79 years. Pack-years were not used, as ten cigarettes per day for 40 years may have effects very different from those of 20 cigarettes per day for 20 years. Analyses were stratified for location (first study 45 areas, second study ten areas) and 5-year age-at-risk groups, and adjusted for education (tertiary, secondary, primary, or none completed) and alcohol consumption (never, occasional, or ever-regular). The RR for each smoking category is accompanied by a CI derived only from the variance of the log risk in that one category. Hence, each RR, including that for the reference group, is associated with a group-specific CI that can be thought of as reflecting the amount of data only in that one category.21 The 95% group-specific CI for RR is (RR/T, RR×T), where T=exp (1·96√v) and v is the variance of the log risk. If the reference group with RR=1 and another group with RR=R have, respectively, group-specific CIs (a, b) and (x, y), then the CI for R that allows for the variation in both of the groups is (√ [xy/k], √ [xyk]), where log (k) is given by √(log2 [y/x] + log2 [b/a]); since k>y/x, this CI is wider than (x, y).

The RR for each smoking category is accompanied by a CI derived only from the variance of the log risk in that one category. Hence, each RR, including that for the reference group, is associated with a group-specific CI that can be thought of as reflecting the amount of data only in that one category.21 The 95% group-specific CI for RR is (RR/T, RR×T), where T=exp (1·96√v) and v is the variance of the log risk. If the reference group with RR=1 and another group with RR=R have, respectively, group-specific CIs (a, b) and (x, y), then the CI for R that allows for the variation in both of the groups is (√ [xy/k], √ [xyk]), where log (k) is given by √(log2 [y/x] + log2 [b/a]); since k>y/x, this CI is wider than (x, y). If RR is causal, the fraction of all deaths in the population that is attributed to smoking (ie, the population-attributed fraction; PAF) is P(RR-1) divided by RR, P being the prevalence of smoking among those dying of the relevant cause.22 Setting P=1 yields (RR–1) divided by RR, the fraction of the mortality among smokers that is attributed to smoking. Analyses used SAS version 9.3. Role of the funding sources The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. ZC, RP, and LL had full access to all the data in the study and had final responsibility for the decision to submit for publication.

If RR is causal, the fraction of all deaths in the population that is attributed to smoking (ie, the population-attributed fraction; PAF) is P(RR-1) divided by RR, P being the prevalence of smoking among those dying of the relevant cause.22 Setting P=1 yields (RR–1) divided by RR, the fraction of the mortality among smokers that is attributed to smoking. Analyses used SAS version 9.3. Role of the funding sources The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. ZC, RP, and LL had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Women were included only in the second study. Of 302 669 women enrolled in 2004–08, only 3·2% were ever regular smokers, including 0·5% who stopped by choice. Smoking uptake decreased steeply across successive generations of women. Hence, there were much lower prevalences among those born in more recent decades (figure 1). Taking all study areas together, the prevalences of smoking among women born in the 1930s, 1940s, 1950s, and since the 1960s were 10%, 5%, 2%, and 1% (see figure 1 legend). In most locations, few women smoked, although in urban Harbin (northeast China) and rural Sichuan (southwest China), appreciable numbers of older women did so. Both in these two locations and elsewhere, the prevalence was about tenfold lower in women born in the 1960s (age about 40 years in 2006) than in women born in the 1930s (age about 70 years in 2006). On average, female smokers had started at age 27 years and currently smoked about 10 g/day (ten cigarette equivalents/day, though not necessarily as cigarettes).

, the prevalence was about tenfold lower in women born in the 1960s (age about 40 years in 2006) than in women born in the 1930s (age about 70 years in 2006). On average, female smokers had started at age 27 years and currently smoked about 10 g/day (ten cigarette equivalents/day, though not necessarily as cigarettes). During 2·0 million woman-years of follow-up (mid-year 2010), 9934 women died at ages 40–79 years. Female mortality rates were significantly related to smoking for all causes (RR 1·51, 1·40–1·63) and for lung cancer, ischaemic heart disease, and chronic obstructive pulmonary disease (COPD) (appendix p 4). The proportional increase in overall mortality among smokers was greater in urban (RR 1·72, 1·52–1·96) than in rural (RR 1·40, 1·27–1·54) women, but numbers were too small for subdivision by smoking patterns. If these associations are causal, then about a third (0·51/1·51) of all deaths of female smokers were due to tobacco. As, however, few women smoked, the fraction of female mortality attributed to smoking was only 3% for deaths at ages 40–79 years (and 1·1%, 2·7%, and 5·2% for deaths at ages 40–59, 60–69, and 70–79 years). This dependence on age can also be expressed in terms of decade of birth: smoking caused about 5%, 3%, 1%, and less than 1% of all deaths among Chinese women born in the 1930s, 1940s, 1950s, and since 1960. The prevalences of smoking in these four birth cohorts were 10%, 5%, 2%, and 1%, respectively.

and 70–79 years). This dependence on age can also be expressed in terms of decade of birth: smoking caused about 5%, 3%, 1%, and less than 1% of all deaths among Chinese women born in the 1930s, 1940s, 1950s, and since 1960. The prevalences of smoking in these four birth cohorts were 10%, 5%, 2%, and 1%, respectively. Among men, the smoking patterns were very different. The first of these two prospective studies began about 15 years before the second (1991 vs 2006), so those in it had, on average, an earlier year of birth (1938 vs 1954) and were less likely to have smoked only cigarettes when they started smoking, or last smoked (appendix p 5). The prevalences of ever-smoking among men in the two studies were similar, and did not depend strongly on age. In the second study, 68% were smokers (including those who had stopped because they were ill) and another 7% were ex-smokers (described as stopping by choice). Figure 2, from the baseline survey for the second study, describes smoking patterns by birth year. Male smokers born in 1970 had started around age 20 years and used only cigarettes, but male smokers born in the 1930s (an earlier generation) had started around age 25 years, smoked other tobacco types, and (not shown) were slightly less likely to inhale deeply. The proportion of smokers who had used cigarettes throughout adulthood was higher in urban than in rural men.

ears and used only cigarettes, but male smokers born in the 1930s (an earlier generation) had started around age 25 years, smoked other tobacco types, and (not shown) were slightly less likely to inhale deeply. The proportion of smokers who had used cigarettes throughout adulthood was higher in urban than in rural men. Male deaths at ages 40–79 years numbered 25 548 in the first study and 14 241 in the second. In both, smokers' mortality rates were significantly elevated (first study: RR 1·17, 1·14–1·21; second study: RR 1·33, 1·28–1·39; table 1). So, although the studies were only 15 years apart (mid-years of follow-up 1995 and 2010), the proportional excess mortality among smokers (RR-1) had approximately doubled. Within each study, the RRs for all-cause mortality were not as extreme in rural as in urban men (first study: rural RR 1·13, 1·09–1·17 and urban RR 1·32, 1·24–1·41; second study: rural RR 1·22, 1·16–1·29 and urban RR 1·65, 1·53–1·79). Apart from the rural versus urban difference, the RRs in the second study differed little across study sites (appendix p 6). If the associations of smoking with death are largely causal, then the proportion of all male deaths at ages 40–79 years attributed to smoking rose between 1995 and 2000 from about 11% in the first study (PAF 9% rural, 17% urban) to about 18% in the second study (PAF 14% rural, 26% urban). If the PAF rose from 11% in 1995 to 18% in 2010, it will be at least 20% by the mid-2010s. Hence, smoking will cause about 20% of all male deaths at ages 40–79 years during the 2010s (ie, 2010–19).

0 from about 11% in the first study (PAF 9% rural, 17% urban) to about 18% in the second study (PAF 14% rural, 26% urban). If the PAF rose from 11% in 1995 to 18% in 2010, it will be at least 20% by the mid-2010s. Hence, smoking will cause about 20% of all male deaths at ages 40–79 years during the 2010s (ie, 2010–19). Similarly, the proportional excess mortality among smokers (ie, RR-1) increased substantially between 1995 and 2010 for lung cancer (first study: RR 1·95, 1·68–2·26; second study: RR 2·58, 2·17–3·05), all vascular disease (first study: RR 1·12, 1·07–1·18; second study: RR 1·24, 1·16–1·33), and COPD (first study: RR 1·28, 1·20–1·36; second study: RR 1·62, 1·38–1·90). Figure 3 shows, for both studies, the urban and rural RRs for overall mortality by age started smoking regularly. In each case, men who had started before age 20 years (mean age 17 years) were at substantially greater risk than were those who had started at ages 20–24 years (mean age 21 years). Differences in the reported ages at which men had started smoking regularly did not, however, suffice to explain the more extreme RRs in the second study, as even within each category of age started, the RR was substantially more extreme in the second study (table 1).

ad started at ages 20–24 years (mean age 21 years). Differences in the reported ages at which men had started smoking regularly did not, however, suffice to explain the more extreme RRs in the second study, as even within each category of age started, the RR was substantially more extreme in the second study (table 1). Among urban male smokers in the second study who had started before age 20 years (as is now typical throughout urban and rural China), most had always used manufactured cigarettes, and their all-cause mortality rate was double that of non-smokers (RR 1·98, 1·79–2·19). This suggests that about half (0·98/1·98) of their deaths were due to tobacco. This proportion could well rise still further. The few urban men who had started before age 15 years were at even greater risk (RR 2·64, 2·19–3·19). The remaining analyses are given separately for urban and rural men in the two studies. The discussion particularly emphasises findings in the second study for urban smokers, among whom the RR was 1·65 (1·53–1·79) for all causes, 2·98 (2·28–3·89) for lung cancer, 1·63 (1·39–1·90) for ischaemic stroke or ischaemic heart disease, and 4·61 (2·75–7·73) for COPD; the RRs were even more extreme for men who had started before age 20 years (table 1). Smokers also had elevated mortality from some other diseases (appendix p 4), including stomach and oesophagus cancer, which are still major causes of death in some parts of China. In rural men, the RRs were generally less extreme, especially for vascular disease and COPD.

who had started before age 20 years (table 1). Smokers also had elevated mortality from some other diseases (appendix p 4), including stomach and oesophagus cancer, which are still major causes of death in some parts of China. In rural men, the RRs were generally less extreme, especially for vascular disease and COPD. For the same disease groupings, table 2 shows urban and rural RRs in both studies by amount last smoked (<15, 15–24 and ≥25 g/day, adding together cigarettes per day and g per day of tobacco). It is only the RRs (not the absolute mortality rates) that are compared between areas and studies, as in table 1. For smokers in the second study, among urban men there were highly significant dose-response relationships for mortality from all causes, lung cancer, ischaemic stroke and ischaemic heart disease (each trend p<0·0001), and COPD (trend p=0·0021), but in rural men, the dose-response relationships among smokers (ignoring the non-smokers) were significant only for lung cancer, a disease with little time between symptom onset and death. The lack of a dose-response relationship for COPD (a major cause of death in rural areas) may reflect biases from reverse causality, whereby COPD symptoms could reduce the amount smoked. For all four disease groupings, tests for trend that included the never-smokers as having zero dose were highly significant in urban and rural areas in both studies (16 trend tests each p<0·0001; table 2).

h in rural areas) may reflect biases from reverse causality, whereby COPD symptoms could reduce the amount smoked. For all four disease groupings, tests for trend that included the never-smokers as having zero dose were highly significant in urban and rural areas in both studies (16 trend tests each p<0·0001; table 2). The proportion of male smokers who had stopped by choice rose appreciably over the 15 years from 1991 to 2006, from only 3% (4306/160 971) in the first baseline survey to 9% (14 080/156 313) in the second, and stopping by choice avoided nearly all the excess risk that would have been seen at the mortality rates of continuing smokers (figure 4). Among ex-smokers who had stopped by choice, the RR for all-cause mortality was 1·02 (0·95–1·10), and it attenuated with quitting duration. For men who had stopped by choice less than 5 years, 5–14 years, and 15 or more years before baseline, the RRs were 1·21 (1·07–1·37), 1·00 (0·90–1·11), and 0·98 (0·87–1·11) (trend p=0·01). These findings were not materially changed by inclusion of the first-study results (appendix p 7). Numbers who stopped by choice and died from specific conditions were too small for separate analysis. Among men who had stopped because they were already ill, the protective effects of quitting cannot be assessed straightforwardly. For, even if cessation is substantially protective, the illness that had made them stop could still cause them to be at misleadingly elevated risk (figure 4).

ariable was often unavailable). Finally, we omitted the initial 5 years of follow-up from the analysis because diseases at baseline that might cause death over the next 5 years could result in reverse causation (where lower BMI at recruitment is the result, rather than the cause, of the underlying pathology).14, 15, 16 Our findings are consistent with other (albeit less precise) studies that have used effective methods to reduce potential bias in evaluations of a causal relationship between excess BMI and mortality, such as Mendelian randomisation analyses,32, 33 other instrumental variable analyses,34 and a meta-analysis of randomised trials.35 Our findings are also broadly consistent with the stricter analyses done in a 2015 study36 of 12 million Korean adults and with a 2016 review that attempted to limit the effects of reverse causality.37

oo small for separate analysis. Among men who had stopped because they were already ill, the protective effects of quitting cannot be assessed straightforwardly. For, even if cessation is substantially protective, the illness that had made them stop could still cause them to be at misleadingly elevated risk (figure 4). The smoker versus non-smoker RRs for lung cancer in these studies are much less extreme than in recent US prospective studies (appendix p 8–10), but the absolute excess lung cancer mortality in Chinese smokers is still substantial, as Chinese people have much higher non-smoker lung cancer rates than Americans do, especially in later middle age.3 Within China, never-smoker lung cancer rates fell slightly between the two prospective studies, whereas the smokers' lung cancer rates rose, resulting in greater RRs in the second study. Discussion Among Chinese men, tobacco-attributed mortality has grown considerably since the 1990s, and during the 2010s, smoking will cause about 20% of all male deaths at ages 40–79 years, up from only about 10% in the early 1990s. Moreover, the mortality rate ratio of 2 already seen among urban male smokers who started before age 20 years (the uptake pattern now typical in both urban and rural China) suggests that about half of these men's deaths were caused by smoking. This mortality rate ratio of 2 is still increasing, foreshadowing substantially greater future hazards for Chinese men (panel). Also, although this was not assessed in the present report, tobacco causes many non-fatal disease episodes, and much disability.

s that about half of these men's deaths were caused by smoking. This mortality rate ratio of 2 is still increasing, foreshadowing substantially greater future hazards for Chinese men (panel). Also, although this was not assessed in the present report, tobacco causes many non-fatal disease episodes, and much disability. Among Chinese women, however, tobacco-attributed mortality is much smaller, and is currently still falling. Fewer women than men are smokers, and there was a large intergenerational decrease in the female prevalence of smoking, from about 10% among women born in the 1930s down to only about 1% among women born around 1970, which has been confirmed in an independent nationwide survey.6 Hence, although the present study suggests that the hazard per smoker is comparable for men and women, as did previous studies,13, 14, 23 tobacco-attributed mortality in the entire female population is low, and will fall further by 2030 as the present generation of women in which few smoke is succeeded by a generation in which even fewer smoke. (A recent WHO report mistakenly suggested, using methods3 inappropriate for China, that male and female tobacco-attributed proportions of deaths are similar;24 this is impossible, given the 20-fold difference in smoking prevalence.) But, this favourable trend may cease (with women born around 1980 and now in their thirties being the least exposed generation), as tobacco use by adolescent females has recently begun to increase in some parts of China.7 This underlines the danger that, as in many Western countries, social changes may well lead young Chinese women to start smoking. It will, however, take decades, probably until after 2050, for an increase in female smoking to cause a substantial increase in mortality, and over the next 20 years, while tobacco-attributed deaths increase among Chinese men, they should decrease among Chinese women, perhaps from about 3% of all female deaths now to less than 1% by the 2030s.2

ades, probably until after 2050, for an increase in female smoking to cause a substantial increase in mortality, and over the next 20 years, while tobacco-attributed deaths increase among Chinese men, they should decrease among Chinese women, perhaps from about 3% of all female deaths now to less than 1% by the 2030s.2 A few large studies have provided reasonably robust evidence about the hazards of smoking in specific Chinese male populations, but most of the people who died in them were born before 1950,16, 25, 26, 27, 28, 29 so the smokers were at limited risk, as in the USA 60 years ago.11, 14 Some studies29, 30 have attempted to assess the extent of the Chinese epidemic in the early 2000s, but they involved cohorts established decades ago of men born before 1950, or used data from atypical regions (eg, Shanghai) where manufactured cigarettes have long been available.31

isk, as in the USA 60 years ago.11, 14 Some studies29, 30 have attempted to assess the extent of the Chinese epidemic in the early 2000s, but they involved cohorts established decades ago of men born before 1950, or used data from atypical regions (eg, Shanghai) where manufactured cigarettes have long been available.31 In China, cigarette consumption became widespread earlier in urban than in rural areas, mainly because of limited rural availability (and, until recent decades, affordability) of cigarettes.32 Hence, the hazard associated with a given current smoking pattern is more extreme in urban than in rural areas. However, this urban versus rural difference is likely to diminish, or even be reversed, over the next few decades, because rural men born after the 1960s not only tended to start at the same age as urban men and to smoke only cigarettes, but also had a somewhat higher smoking prevalence (figure 2). As earlier generations of urban and rural men get replaced by generations who have smoked cigarettes persistently since early adulthood, tobacco-attributable risks in middle age may soon reach those seen in many Western populations, as has almost happened in the subgroup of urban men who started smoking before age 20 years in the present study (and in Hong Kong, where cigarette use peaked about 20 years earlier than in mainland China33).

arly adulthood, tobacco-attributable risks in middle age may soon reach those seen in many Western populations, as has almost happened in the subgroup of urban men who started smoking before age 20 years in the present study (and in Hong Kong, where cigarette use peaked about 20 years earlier than in mainland China33). For the chief diseases by which tobacco causes death, there are large quantitative differences between China and elsewhere, between urban and rural China, and between past and future decades.2 In many Western populations, tobacco used to cause far more deaths from vascular than from respiratory disease,11, 12, 13, 14, 15 whereas in China the opposite is true, especially in rural areas. Although the RRs for COPD, lung cancer, and stroke may at present be smaller than in many Western populations, the absolute risks associated with smoking are not, as Chinese non-smoker death rates are high. Indeed, 1980s lung cancer mortality among Chinese never-smokers was more than three times that in US never-smokers, perhaps due partly to indoor air pollution from heating and cooking.2 While US never-smoker lung cancer rates have remained roughly constant over the past half-century,3 those in Chinese never-smokers seem to be decreasing while those in smokers are increasing, causing increasing smoker versus non-smoker lung cancer mortality rate ratios.

o indoor air pollution from heating and cooking.2 While US never-smoker lung cancer rates have remained roughly constant over the past half-century,3 those in Chinese never-smokers seem to be decreasing while those in smokers are increasing, causing increasing smoker versus non-smoker lung cancer mortality rate ratios. Absolute mortality rates are likely to be lower in prospective studies than in the general population. Hence, to estimate absolute numbers of tobacco-attributed deaths in China in 2010, the smoking-attributed fractions of all deaths in our second prospective study have been applied to independent estimates of male and female cause-specific numbers of deaths in mainland China34 at ages 35–69, 70–79, and 80 years or older (appendix p 4). This shows that there were about 1 million smoking-attributed deaths in 2010 (840 000 male, 130 000 female; table 3), mainly from diseases already known to be affected by smoking (lung cancer, ischaemic heart disease, ischaemic stroke, COPD, and other neoplastic, vascular, and respiratory conditions).3, 11 Counterbalancing the increasing RRs for all-cause mortality, age-specific under-70 mortality rates in China are decreasing due to many social, occupational and health-care changes, falling by about 15% during 2000–10,9 so the absolute death rate from smoking is not increasing as fast as would be suggested just by the increasing RRs.

ng the increasing RRs for all-cause mortality, age-specific under-70 mortality rates in China are decreasing due to many social, occupational and health-care changes, falling by about 15% during 2000–10,9 so the absolute death rate from smoking is not increasing as fast as would be suggested just by the increasing RRs. About two-thirds of young Chinese men become cigarette smokers in early adult life. Unless they stop, the present study suggests that at least half of them will eventually be killed by their habit, and future studies may well show that a somewhat greater proportion will be killed by it. The first generation of men to experience the full hazards will probably be those born during the 1970s or 1980s, who reached adulthood when nationwide cigarette consumption was high. Conversely, this may well be the least exposed female generation.

may well show that a somewhat greater proportion will be killed by it. The first generation of men to experience the full hazards will probably be those born during the 1970s or 1980s, who reached adulthood when nationwide cigarette consumption was high. Conversely, this may well be the least exposed female generation. China's 2030 sustainable development goals include reducing non-communicable disease mortality by a third, and monitoring the changes. If current smoking patterns persist, then as the smoker versus non-smoker RRs increases, mortality from other causes decreases, and the over-60 population doubles, Chinese tobacco deaths are likely to rise from 1 million in 2010 to about 2 million in 2030. Nowadays, in China more than 6 million young men a year begin smoking. If most persist, and (as in the USA14 and UK12, 13) smokers eventually have more than double the non-smoker mortality rates, there will in 2050 be about 3 million Chinese tobacco deaths, when those born in 1970 reach age 80 years. Although continuation of our second prospective study will monitor how the epidemic develops over the next decade or two, large new prospective studies of people born after 1970 will be needed to continue monitoring it thereafter. Fortunately, China's nationally representative household surveys regularly record smoking and the reasons for smoking cessation, and electronic linkage of this information with routine mortality records should allow reliable monitoring of the evolution of the epidemic for many decades to come.

monitoring it thereafter. Fortunately, China's nationally representative household surveys regularly record smoking and the reasons for smoking cessation, and electronic linkage of this information with routine mortality records should allow reliable monitoring of the evolution of the epidemic for many decades to come. Avoiding uptake of smoking by young people will greatly reduce tobacco deaths in the second half of the century. Moreover, stopping before the onset of life-threatening illness is remarkably protective, and an increasing proportion of smokers have stopped by choice (9% in 2006 vs only 3% in 1991). With effective measures to accelerate cessation, the growing epidemic of premature death from tobacco can be halted and then reversed, as in other countries. Widespread smoking cessation offers China one of the most effective, and cost-effective,35 strategies to avoid disability and premature death over the next few decades. Supplementary Material Supplementary appendix

Avoiding uptake of smoking by young people will greatly reduce tobacco deaths in the second half of the century. Moreover, stopping before the onset of life-threatening illness is remarkably protective, and an increasing proportion of smokers have stopped by choice (9% in 2006 vs only 3% in 1991). With effective measures to accelerate cessation, the growing epidemic of premature death from tobacco can be halted and then reversed, as in other countries. Widespread smoking cessation offers China one of the most effective, and cost-effective,35 strategies to avoid disability and premature death over the next few decades. Supplementary Material Supplementary appendix Acknowledgments The chief acknowledgment is to the participants, the project staff, the China National Centre for Disease Control and Prevention, and its regional offices. The two prospective studies already span 25 years, throughout which Cancer Research UK, the UK Medical Research Council, and the British Heart Foundation core-funded the Oxford CTSU. The first study's baseline survey was funded by the World Bank and Canadian International Development Research Centre, with long-term continuation funded by the CTSU and organised by Gonghuang Yang; fuller acknowledgments have been published.16 The second study's baseline survey was funded by the Kadoorie Charitable Foundation, with blood storage through the core-funded CTSU Wolfson Laboratories (director Mike Hill) and Fu Wai China National Cardiovascular Centre, with long-term continuation funded by the CTSU, Wellcome Trust (088158/Z/09/Z, 104085/Z/14/Z), MRC (Newton Fund MCPC13049), Chinese Ministry of Science and Technology (2011BAI09B01), and National Natural Science Foundation of China (81390541); the Chinese National Health Insurance scheme provides linkage to all hospital admission data.

ntinuation funded by the CTSU, Wellcome Trust (088158/Z/09/Z, 104085/Z/14/Z), MRC (Newton Fund MCPC13049), Chinese Ministry of Science and Technology (2011BAI09B01), and National Natural Science Foundation of China (81390541); the Chinese National Health Insurance scheme provides linkage to all hospital admission data. China Kadoorie Biobank collaborative group International Steering Committee—Junshi Chen (co-chair), Zhengming Chen (co-principal investigator), Rory Collins, Liming Li (co-principal investigator), and Richard Peto (co-chair). International Co-ordinating Centre, Oxford—Daniel Avery, Derrick Bennett, Yumei Chang, Yiping Chen, Zhengming Chen, Robert Clarke, Huaidong Du, Xuejuan Fan, Haiyan Gao, Simon Gilbert, Michael Holmes, Andri Iona, Rene Kerosi, Ling Kong, Om Kurmi, Garry Lancaster, Sarah Lewington, John McDonnell, Winnie Mei, Iona Millwood, Qunhua Nie, Jayakrishnan Radhakrishnan, Paul Ryder, Sam Sansome, Dan Schmidt, Paul Sherliker, Margaret Smith, Rajani Sohoni, Robin Walters, Jenny Wang, Lin Wang, Alex Williams, Ling Yang, and Xiaoming Yang. National Co-ordinating Centre, Beijing—Zheng Bian, Ge Chen, Lei Guo, Yu Guo, Bingyang Han, Can Hou, Peng Liu, Jun Lv, Pei Pei, Shuzhen Qu, Yunlong Tan, Canqing Yu, and Huiyan Zhou. Qingdao Regional Co-ordinating Centre—Zengchang Pang, Shaojie Wang, Yun Zhang, and Kui Zhang (Qingdao CDC); Silu Liu and Wei Hou (Licang CDC). Heilongjiang Regional Co-ordinating Centre—Zhonghou Zhao, Shumei Liu, and Zhigang Pang (Provincial CDC); Weijia Feng, Shuling Wu, Liqiu Yang, Huili Han, Hui He, and Bo Yu (Nangang CDC). Hainan Regional Co-ordinating Centre—Xianhai Pan, Shanqing Wang, and Hongmei Wang (Provincial CDC); Xinhua Hao, Chunxing Chen, Shuxiong Lin, and Xiangyang Zheng (Meilan CDC); Jiangsu Regional Co-ordinating Centre—Xiaoshu Hu, Minghao Zhou, Ming Wu, and Ran Tao (Provincial CDC); Yeyuan Wang, Yihe Hu, Liangcai Ma, Renxian Zhou, Guanqun Xu, and Yan Lu (Suzhou CDC). Guangxi Regional Co-ordinating Centre—Baiqing Dong, Naying Chen, and Ying Huang (Provincial CDC); Mingqiang Li, Jinhuai Meng, Zhigao Gan, Jiujiu Xu, Yun Liu, and Jingxin Qing (Liuzhou CDC). Sichuan Regional Co-ordinating Centre—Xianping Wu, Yali Gao, and Ningmei Zhang (Provincial CDC); Guojin Luo, Xiangsan Que, Xiaofang Chen (Pengzhou CDC). Gansu Regional Co-ordinating Centre—Pengfei Ge, Jian He, and Xiaolan Ren (Provincial CDC); Hui Zhang, Enke Mao, Guanzhong Li, Zhongxiao Li, Jun He, Yulong Lei, and Xiaoping Wang (Maiji CDC).

inating Centre—Xianping Wu, Yali Gao, and Ningmei Zhang (Provincial CDC); Guojin Luo, Xiangsan Que, Xiaofang Chen (Pengzhou CDC). Gansu Regional Co-ordinating Centre—Pengfei Ge, Jian He, and Xiaolan Ren (Provincial CDC); Hui Zhang, Enke Mao, Guanzhong Li, Zhongxiao Li, Jun He, Yulong Lei, and Xiaoping Wang (Maiji CDC). Henan Regional Co-ordinating Centre—Guohua Liu, Baoyu Zhu, Gang Zhou, and Shixian Feng (Provincial CDC); Yulian Gao, Tianyou He, Li Jiang, Jianhua Qin, and Huarong Sun (Huixian CDC). Zhejiang Regional Co-ordinating Centre—Liqun Liu, Min Yu, Yaping Chen, and Ruying Hu (Provincial CDC); Zhixiang Hu, Jianjin Hu, Yijian Qian, Zhiying Wu, Chunmei Wang, and Lingli Chen (Tongxiang CDC). Hunan Regional Co-ordinating Centre—Wen Liu, Guangchun Li, and Huilin Liu (Provincial CDC); Xiangquan Long, Xin Xu, Youping Xiong, Zhongwen Tan, Xuqiu Xie, Yunfang Peng, and Weifang Jia (Liuyang CDC). Contributors ZC, RP, and LL had full access to the data. All authors were involved in study design, conduct, long-term follow-up, analysis of data, interpretation, or writing the report. Declaration of interests We declare no competing interests. Figure 1 Chinese female smoking uptake rate by year of birth and locality

Henan Regional Co-ordinating Centre—Guohua Liu, Baoyu Zhu, Gang Zhou, and Shixian Feng (Provincial CDC); Yulian Gao, Tianyou He, Li Jiang, Jianhua Qin, and Huarong Sun (Huixian CDC). Zhejiang Regional Co-ordinating Centre—Liqun Liu, Min Yu, Yaping Chen, and Ruying Hu (Provincial CDC); Zhixiang Hu, Jianjin Hu, Yijian Qian, Zhiying Wu, Chunmei Wang, and Lingli Chen (Tongxiang CDC). Hunan Regional Co-ordinating Centre—Wen Liu, Guangchun Li, and Huilin Liu (Provincial CDC); Xiangquan Long, Xin Xu, Youping Xiong, Zhongwen Tan, Xuqiu Xie, Yunfang Peng, and Weifang Jia (Liuyang CDC). Contributors ZC, RP, and LL had full access to the data. All authors were involved in study design, conduct, long-term follow-up, analysis of data, interpretation, or writing the report. Declaration of interests We declare no competing interests. Figure 1 Chinese female smoking uptake rate by year of birth and locality 300 000 women seen in ten study areas in about 2006, with birth years grouped as: before 1935, 1935–44, 1945–54, 1955–64, and 1965 or later. The two areas where many older women smoked are in Harbin (urban northeast China) and Sichuan (rural southwest China). Taking all ten areas together, the prevalences of ever-smoking among women born in the 1930s, 1940s, 1950s, 1960s, and 1970s were, respectively, 10%, 5%, 2%, 1%, and <1% (3097/30 943, 3265/62 246, 2339/97 344, 926/94 772, and 142/17 161). Figure 2 Urban and rural Chinese male smoking patterns, by year of birth—prevalence, consumption, age started, and tobacco type smoked initially

300 000 women seen in ten study areas in about 2006, with birth years grouped as: before 1935, 1935–44, 1945–54, 1955–64, and 1965 or later. The two areas where many older women smoked are in Harbin (urban northeast China) and Sichuan (rural southwest China). Taking all ten areas together, the prevalences of ever-smoking among women born in the 1930s, 1940s, 1950s, 1960s, and 1970s were, respectively, 10%, 5%, 2%, 1%, and <1% (3097/30 943, 3265/62 246, 2339/97 344, 926/94 772, and 142/17 161). Figure 2 Urban and rural Chinese male smoking patterns, by year of birth—prevalence, consumption, age started, and tobacco type smoked initially 210 000 men seen in about 2006 (at the 2004–08 baseline survey for the second prospective study). Prevalence of smoking (A); amount smoked per day when last smoked (B); Mean age started smoking regularly (C); and percentage of all smokers who used cigarettes when first started (D). To avoid reverse causality biasing the apparent effects of smoking and of cessation, in panel (A) and in the main analyses, the few men who had stopped smoking because they were ill are combined with the continuing smokers, leaving the ex-smokers who had stopped by choice. The overall proportions of men who had stopped because they were ill were 2·16%, 2·47%, 2·19%, 1·08% and 0·12% for those born during the 1930s, 1940s, 1950s, 1960s, and 1970s respectively. Figure 3 All-cause smoker versus never-smoker mortality rate ratio (RR) among urban and rural Chinese men in two prospective studies, by time period (about 1995 or about 2010) and by age started smoking regularly

210 000 men seen in about 2006 (at the 2004–08 baseline survey for the second prospective study). Prevalence of smoking (A); amount smoked per day when last smoked (B); Mean age started smoking regularly (C); and percentage of all smokers who used cigarettes when first started (D). To avoid reverse causality biasing the apparent effects of smoking and of cessation, in panel (A) and in the main analyses, the few men who had stopped smoking because they were ill are combined with the continuing smokers, leaving the ex-smokers who had stopped by choice. The overall proportions of men who had stopped because they were ill were 2·16%, 2·47%, 2·19%, 1·08% and 0·12% for those born during the 1930s, 1940s, 1950s, 1960s, and 1970s respectively. Figure 3 All-cause smoker versus never-smoker mortality rate ratio (RR) among urban and rural Chinese men in two prospective studies, by time period (about 1995 or about 2010) and by age started smoking regularly Each study followed about 200 000 men. Each group-specific CI (including that for never-smokers, given by the width of the shaded strip) reflects the variance of the log risk in that one group, so comparisons of RRs use variances from more than one group. Figure 4 Ex-smoker versus never-smoker all-cause mortality rate ratio (RR), by years stopped smoking and reason stopped, for men in the second study Each group-specific CI (including that for never-smokers, given by the width of the shaded strip) reflects the variance of the log risk in that 1 group, so comparisons use variances from more than one group.

Figure 4 Ex-smoker versus never-smoker all-cause mortality rate ratio (RR), by years stopped smoking and reason stopped, for men in the second study Each group-specific CI (including that for never-smokers, given by the width of the shaded strip) reflects the variance of the log risk in that 1 group, so comparisons use variances from more than one group. Table 1 Age started smoking regularly versus cause-specific mortality rate ratio (RR) among urban and rural men in about 2010 (second study) and about 1995 (first study)

Each group-specific CI (including that for never-smokers, given by the width of the shaded strip) reflects the variance of the log risk in that 1 group, so comparisons use variances from more than one group. Table 1 Age started smoking regularly versus cause-specific mortality rate ratio (RR) among urban and rural men in about 2010 (second study) and about 1995 (first study) Breath CO, ppm All causes Lung cancer Ischaemic stroke or ischaemic heart disease COPD Number of deaths RR* (95% group-specific CI) Number of deaths RR* (95% group-specific CI) Number of deaths RR* (95% group-specific CI) Number of deaths RR* (95% group-specific CI) Urban men, about 2010 (2006–14) Age started smoking (mean), years Age <20 (16·9) 16·5 843 1·98 (1·85–2·13) 127 3·78 (3·15–4·54) 200 2·03 (1·75–2·35) 43 9·09 (6·60–12·50) Age 20–24 (21·0) 15·2 797 1·61 (1·50–1·73) 127 3·17 (2·67–3·77) 182 1·50 (1·30–1·74) 24 3·89 (2·62–5·79) Age ≥25 (30·4) 12·6 754 1·44 (1·34–1·55) 96 2·23 (1·82–2·73) 196 1·49 (1·29–1·72) 24 2·89 (1·93–4·32) All smokers 14·9 2394 1·65 (1·59–1·73) 350 2·98 (2·66–3·33) 578 1·63 (1·49–1·77) 91 4·61 (3·71–5·71) Non-smokers 4·5 924 1·00 (0·94–1·07) 69 1·00 (0·79–1·27) 242 1·00 (0·88–1·14) 18 1·00 (0·62–1·60) Trend p value† <0·0001 0·0002 0·0032 <0·0001 Rural men, about 2010 (2006–14) Age started smoking (mean), years <20 (16·8) 15·1 2710 1·39 (1·34–1·45) 271 2·91 (2·58–3·29) 334 1·37 (1·23–1·53) 315 1·88 (1·68–2·10) 20–24 (21·2) 13·6 2788 1·26 (1·21–1·31) 275 2·45 (2·17–2·75) 397 1·35 (1·23–1·49) 261 1·41 (1·24–1·59) ≥25 (31·3) 11·8 2434 1·05 (1·01–1·10) 177 1·63 (1·41–1·90) 347 1·04 (0·93–1·16) 233 1·07 (0·94–1·22) All smokers 13·6 7932 1·22 (1·20–1·25) 723 2·30 (2·13–2·48) 1078 1·24 (1·17–1·32) 809 1·41 (1·31–1·51) Non-smokers 6·0 2031 1·00 (0·96–1·05) 92 1·00 (0·81–1·23) 316 1·00 (0·89–1·12) 172 1·00 (0·86–1·16) Trend p value† <0·0001 <0·0001 0·0005 <0·0001 Urban men, about 1995 (1991–99) Age started smoking (mean), years <20 (16·7) .. 1158 1·60 (1·51–1·70) 142 3·20 (2·69–3·80) 178 1·58 (1·36–1·84) 194 1·82 (1·57–2·21) 20–24 (21·2) .. 1358 1·37 (1·30–1·44) 130 2·31 (1·95–2·75) 208 1·37 (1·19–1·57) 209 1·43 (1·25–1·64) ≥25 (30·6) .. 1136 1·11 (1·04–1·17) 97 1·73 (1·42–2·12) 178 1·17 (1·01–1·36) 186 1·16 (1·00–1·34) All smokers .. 3652 1·32 (1·28–1·37) 369 2·32 (2·08–2·59) 564 1·35 (1·23–1·47) 589 1·42 (1·30–1·55) Non-smokers .. 1381 1·00 (0·95–1·06) 82 1·00 (0·80–1·25) 240 1·00 (0·87–1·14) 197 1·00 (0·87–1·15) Trend p value† <0·0001 <0·0001 0·0052 <0·0001 Rural men, about 1995 (1991–99) Age started smoking (mean), years <20 (16·8) ..

smokers .. 3652 1·32 (1·28–1·37) 369 2·32 (2·08–2·59) 564 1·35 (1·23–1·47) 589 1·42 (1·30–1·55) Non-smokers .. 1381 1·00 (0·95–1·06) 82 1·00 (0·80–1·25) 240 1·00 (0·87–1·14) 197 1·00 (0·87–1·15) Trend p value† <0·0001 <0·0001 0·0052 <0·0001 Rural men, about 1995 (1991–99) Age started smoking (mean), years <20 (16·8) .. 5107 1·20 (1·16–1·23) 236 1·98 (1·73–2·26) 435 1·12 (1·02–1·23) 1373 1·40 (1·33–1·48) 20–24 (20·9) .. 6215 1·14 (1·11–1·17) 280 1·82 (1·62–2·05) 605 1·17 (1·08–1·27) 1573 1·28 (1·22–1·34) ≥25 (29·6) .. 3901 1·04 (1·01–1·07) 139 1·37 (1·16–1·62) 385 1·08 (0·98–1·18) 986 1·06 (0·99–1·13) All smokers .. 15 223 1·13 (1·11–1·15) 655 1·76 (1·62–1·91) 1425 1·13 (1·07–1·20) 3932 1·25 (1·21–1·30) Non-smokers .. 4855 1·00 (0·97–1·03) 150 1·00 (0·85–1·18) 499 1·00 (0·91–1·09) 1114 1·00 (0·94–1·06) Trend p value† <0·0001 <0·0001 0·0620 <0·0001 Group-specific CI for the non-smoker RR of 1·00 reflects the variance of the log risk in non-smokers; for each of the other RRs the CI is correspondingly wider than the group-specific CI. COPD=chronic obstructive pulmonary disease. * RRs (smokers vs non-smokers) were adjusted for 5-year age group and region, alcohol, and education; additional adjustment for occupation made no material difference to the RRs. † Trend test in smokers, ignoring non-smokers; if trend tests in this table had included non-smokers, each would have yielded p<0·0001. Table 2 Amount last smoked versus cause-specific mortality rate ratio (RR) among urban and rural men in about 2010 (second study) and about 1995 (first study)

* RRs (smokers vs non-smokers) were adjusted for 5-year age group and region, alcohol, and education; additional adjustment for occupation made no material difference to the RRs. † Trend test in smokers, ignoring non-smokers; if trend tests in this table had included non-smokers, each would have yielded p<0·0001. Table 2 Amount last smoked versus cause-specific mortality rate ratio (RR) among urban and rural men in about 2010 (second study) and about 1995 (first study) Breath CO, ppm All causes Lung cancer Ischaemic stroke or ischaemic heart disease COPD Number of deaths RR* (95% group-specific CI) Number of deaths RR* (95% group-specific CI) Number of deaths RR* (95% group-specific CI) Number of deaths RR* (95% group-specific CI) Urban men, about 2010 (2006–14) Daily amount smoked (mean) <15 (8·3) 12·0 890 1·48 (1·39–1·59) 115 2·28 (1·90–2·74) 209 1·36 (1·19–1·56) 26 2·94 (1·99–4·35) 15–24 (19·2) 16·2 1091 1·70 (1·60–1·80) 167 3·28 (2·82–3·83) 261 1·73 (1·53–1·96) 45 5·40 (4·04–7·23) ≥25 (34·9) 17·6 413 1·93 (1·75–2·12) 68 4·12 (3·24–5·24) 108 2·24 (1·85–2·71) 20 7·26 (4·66–11·32) All smokers 14·9 2394 1·65 (1·59–1·73) 350 2·98 (2·66–3·33) 578 1·63 (1·49–1·77) 91 4·61 (3·71–5·71) Non-smokers 4·5 924 1·00 (0·94–1·07) 69 1·00 (0·79–1·27) 242 1·00 (0·88–1·14) 18 1·00 (0·62–1·60) Trend p value† <0·0001 <0·0001 <0·0001 0·0021 Rural men, about 2010 (2006–14) Daily amount smoked (mean) <15 (7·7) 12·1 3298 1·25 (1·20–1·29) 202 1·81 (1·57–2·09) 537 1·27 (1·16–1·39) 347 1·52 (1·36–1·70) 15–24 (19·3) 14·5 3113 1·17 (1·13–1·22) 333 2·38 (2·14–2·65) 380 1·14 (1·03–1·27) 306 1·32 (1·18–1·47) ≥25 (35·3) 14·0 1521 1·27 (1·20–1·34) 188 3·20 (2·75–3·72) 161 1·37 (1·16–1·61) 156 1·34 (1·13–1·59) All smokers 13·6 7932 1·22 (1·20–1·25) 723 2·30 (2·13–2·48) 1078 1·24 (1·17–1·32) 809 1·41 (1·31–1·51) Non-smokers 6·0 2031 1·00 (0·96–1·05) 92 1·00 (0·81–1·23) 316 1·00 (0·89–1·12) 172 1·00 (0·86–1·16) Trend p value† 0·7875 <0·0001 0·9951 0·1126 Urban men, about 1995 (1991–99) Daily amount smoked (mean) <15 (8·2) .. 1332 1·23 (1·16–1·30) 104 1·77 (1·46–2·15) 202 1·22 (1·06–1·40) 223 1·28 (1·12–1·46) 15–24 (19·2) .. 1660 1·35 (1·29–1·42) 190 2·62 (2·27–3·03) 261 1·40 (1·24–1·59) 256 1·45 (1·28–1·64) ≥25 (36·1) .. 660 1·51 (1·39–1·63) 75 2·80 (2·21–3·53) 101 1·55 (1·27–1·89) 110 1·78 (1·46–2·15) All smokers .. 3652 1·32 (1·28–1·37) 369 2·32 (2·08–2·59) 564 1·35 (1·23–1·47) 589 1·42 (1·30–1·55) Non-smokers .. 1381 1·00 (0·95–1·06) 82 1·00 (0·80–1·25) 240 1·00 (0·87–1·14) 197 1·00 (0·87–1·15) Trend p value† <0·0001 0·0015 0·0389 0·0071 Rural men, about 1995 (1991–99) Daily amount smoked (mean) <15 (8·9) ..

2·15) All smokers .. 3652 1·32 (1·28–1·37) 369 2·32 (2·08–2·59) 564 1·35 (1·23–1·47) 589 1·42 (1·30–1·55) Non-smokers .. 1381 1·00 (0·95–1·06) 82 1·00 (0·80–1·25) 240 1·00 (0·87–1·14) 197 1·00 (0·87–1·15) Trend p value† <0·0001 0·0015 0·0389 0·0071 Rural men, about 1995 (1991–99) Daily amount smoked (mean) <15 (8·9) .. 3959 1·16 (1·09–1·16) 132 1·37 (1·15–1·63) 434 1·10 (1·00–1·21) 1005 1·25 (1·17–1·33) 15–24 (18·8) .. 6886 1·13 (1·10–1·15) 327 1·90 (1·70–2·12) 659 1·18 (1·09–1·28) 1800 1·23 (1·18–1·29) ≥25 (36·5) .. 4378 1·13 (1·10–1·17) 196 1·84 (1·58–2·13) 332 1·08 (0·97–1·21) 1127 1·29 (1·22–1·38) All smokers .. 15 223 1·13 (1·11–1·15) 655 1·76 (1·62–1·91) 1425 1·13 (1·07–1·20) 3932 1·25 (1·21–1·30) Non-smokers .. 4855 1·00 (0·97–1·03) 150 1·00 (0·85–1·18) 499 1·00 (0·91–1·09) 1114 1·00 (0·94–1·06) Trend p value† 0·7163 0·0233 0·8946 0·4014 Daily amount smoked adds together cigarettes plus g of other tobacco. Group-specific CI for the non-smoker RR of 1·00 reflects the variance of the log risk in non-smokers. COPD=chronic obstructive pulmonary disease. * RRs (smokers vs non-smokers) were adjusted for 5-year age group and region, alcohol, and education; additional adjustment for occupation made no material difference to the RRs. † Trend test in smokers, ignoring non-smokers; if trend tests in this table had included non-smokers, each would have yielded p<0·0001. Table 3 Deaths attributed to tobacco in China, 2010

* RRs (smokers vs non-smokers) were adjusted for 5-year age group and region, alcohol, and education; additional adjustment for occupation made no material difference to the RRs. † Trend test in smokers, ignoring non-smokers; if trend tests in this table had included non-smokers, each would have yielded p<0·0001. Table 3 Deaths attributed to tobacco in China, 2010 Male Female Both 30–69 years 375/2030 20/968 395/2998 70–79 years 250/1370 45/907 295/2277 ≥80 years* 215/1178 65/1269 280/2447 All ages 840/4578 130/3144 970/7722 Data are number of deaths caused by tobacco (thousands)/total number of deaths (thousands). Below age 30 years, the total number of deaths (thousands) was 398 in men and 200 in women. * Estimated indirectly by applying mortality rate ratios at age 70–79 years.

Introduction The worldwide prevalence of overweight and obesity is high and is increasing.1, 2 WHO estimates that more than 1·3 billion adults worldwide are overweight, defined by WHO as a body-mass index (BMI) of 25–<30 kg/m2, and a further 600 million are obese (BMI ≥30 kg/m2).3 Appropriate analyses of large-scale prospective studies with prolonged follow-up generally indicate that both overweight and obesity are associated with increased mortality, as is underweight (defined conservatively by WHO as BMI <18·5 kg/m2). However, it is not known how such associations vary across major global regions, an uncertainty relevant to international strategies for overweight and obesity.4 Most previous analyses have focused on people living in one particular country or continent,5, 6, 7, 8, 9, 10, 11, 12 even though associations with overweight and underweight might differ from one population to another. Estimation of the relationships between BMI and mortality in various populations can help to assess the adverse physiological effects of excessive adiposity (and the adverse physiological effects of various determinants of low BMI). However, reliable estimates of the causal relevance of BMI to mortality need to limit the effects of reverse causality, because chronic disease and smoking can themselves affect BMI. To help achieve more valid estimates, prospective studies of BMI and mortality should, when possible, exclude: smokers, participants who already have some chronic disease at recruitment that could affect BMI, and those dying within 5 years of recruitment.13, 14, 15, 16

nic disease and smoking can themselves affect BMI. To help achieve more valid estimates, prospective studies of BMI and mortality should, when possible, exclude: smokers, participants who already have some chronic disease at recruitment that could affect BMI, and those dying within 5 years of recruitment.13, 14, 15, 16 The Global BMI Mortality Collaboration was established to provide a standardised comparison of associations of BMI with mortality across different populations. It includes individual-participant data for 10·6 million adults in 239 prospective cohort studies in 32 countries, mainly located in Asia, Australia and New Zealand, Europe, or North America, about 4 million of whom were never-smokers without reported chronic diseases (mainly cardiovascular disease, cancer, or chronic respiratory disease) at recruitment and who were still being followed up 5 years afterwards. Research in context Evidence before this study

The Global BMI Mortality Collaboration was established to provide a standardised comparison of associations of BMI with mortality across different populations. It includes individual-participant data for 10·6 million adults in 239 prospective cohort studies in 32 countries, mainly located in Asia, Australia and New Zealand, Europe, or North America, about 4 million of whom were never-smokers without reported chronic diseases (mainly cardiovascular disease, cancer, or chronic respiratory disease) at recruitment and who were still being followed up 5 years afterwards. Research in context Evidence before this study A previous study has claimed that relative to normal weight (defined by WHO as a body-mass index [BMI] of 18·5–<25·0 kg/m2), overweight (BMI 25·0–<30·0 kg/m2) and grade 1 obesity (30·0–<35·0 kg/m2) are not associated with higher all-cause mortality. However, reliable estimates of the causal relevance of BMI to mortality should limit the effects of reverse causality, because chronic disease and smoking can themselves affect BMI. To help achieve such estimates, we established the Global BMI Mortality Collaboration, which involved analysis of individual-participant data from about 10·6 million adults in 239 prospective studies in 32 countries in Asia, Australia and New Zealand, Europe, or North America, about 4 million of whom were never-smokers without chronic disease at recruitment who were still being followed up at least 5 years afterwards. Added value of this study

A previous study has claimed that relative to normal weight (defined by WHO as a body-mass index [BMI] of 18·5–<25·0 kg/m2), overweight (BMI 25·0–<30·0 kg/m2) and grade 1 obesity (30·0–<35·0 kg/m2) are not associated with higher all-cause mortality. However, reliable estimates of the causal relevance of BMI to mortality should limit the effects of reverse causality, because chronic disease and smoking can themselves affect BMI. To help achieve such estimates, we established the Global BMI Mortality Collaboration, which involved analysis of individual-participant data from about 10·6 million adults in 239 prospective studies in 32 countries in Asia, Australia and New Zealand, Europe, or North America, about 4 million of whom were never-smokers without chronic disease at recruitment who were still being followed up at least 5 years afterwards. Added value of this study The Global BMI Mortality Collaboration has combined several features to help guide international public health policy. First, it involved a detailed and standardised comparison of the associations of BMI with mortality across prospective studies in four continents. Second, this analysis has been comprehensive, entailing data from 97% of eligible participants in relevant prospective cohort studies. Third, the study further subdivided the WHO's normal BMI range, which is excessively wide. Finally, the study's approach should have reduced the potentially distorting effects of smoking and ill health on BMI because the primary analyses were of never-smokers without previous disease who survived at least 5 years.

ies. Third, the study further subdivided the WHO's normal BMI range, which is excessively wide. Finally, the study's approach should have reduced the potentially distorting effects of smoking and ill health on BMI because the primary analyses were of never-smokers without previous disease who survived at least 5 years. Implications of all the available evidence This analysis has shown that both overweight and obesity (all grades) were associated with increased all-cause mortality. In the BMI range above 25 kg/m2 (the upper limit of the WHO's normal range), the relationship of BMI to mortality was strong and positive in every global region we studied (except perhaps south Asia, where numbers of deaths were small), lending support to strategies to combat the entire spectrum of excess adiposity worldwide. Our results challenge recent suggestions that overweight and moderate obesity are not associated with higher mortality, bypassing speculation about hypothetical protective metabolic effects of increased body fat in apparently healthy individuals.

o strategies to combat the entire spectrum of excess adiposity worldwide. Our results challenge recent suggestions that overweight and moderate obesity are not associated with higher mortality, bypassing speculation about hypothetical protective metabolic effects of increased body fat in apparently healthy individuals. Methods Search strategy and selection criteria In 2013, over 500 investigators (appendix pp 49, 50) from over 300 institutions in 32 countries agreed an analysis plan for combining individual-participant data from contributing studies. This prespecified analysis plan is provided in the appendix (pp 51–53). The goal was to produce reliable estimates of potentially causal associations of overweight and obesity with mortality using data from studies in several regions. The prespecified analysis methods were designed to maximise the internal validity by reducing the scope for bias. This Article follows PRISMA for Individual Patient Data reporting guidelines (appendix pp 54, 55).17

causal associations of overweight and obesity with mortality using data from studies in several regions. The prespecified analysis methods were designed to maximise the internal validity by reducing the scope for bias. This Article follows PRISMA for Individual Patient Data reporting guidelines (appendix pp 54, 55).17 We sought data from large prospective studies (≥100 000 participants at baseline) or large multicohort consortia (total ≥100 000 participants at baseline). We identified studies published from January, 1970, to January, 2015, through systematic literature searches and discussion with investigators (appendix pp 56, 57). Electronic searches were done with MEDLINE, Embase, and Scopus, and with the terms ‘“body-mass index”, “mortality” or “death”, “cohort” or “prospective”, and combinations of the words “risk”, “relative”, “ratio”, “hazard”, or “rate”. Prospective cohort studies or consortia thereof were eligible if they: (1) had information about weight, height, age, and sex; (2) did not select participants on the basis of having any previous chronic disease; (3) recorded overall or cause-specific deaths; and (4) had accrued 5 years or more of median follow-up. We identified only two eligible studies that were unable to contribute (appendix p 37).18, 19 Details of the included studies are provided in the appendix (pp 3–14). The contributing studies classified deaths according to the primary cause (or, in its absence, the underlying cause), on the basis of coding from the International Classification of Diseases, revisions 8–10, to at least three digits (appendix p 15), or according to study-specific classification systems. Ascertainment of outcomes was generally based on death certificates, supplemented in some studies by additional data.

erlying cause), on the basis of coding from the International Classification of Diseases, revisions 8–10, to at least three digits (appendix p 15), or according to study-specific classification systems. Ascertainment of outcomes was generally based on death certificates, supplemented in some studies by additional data. The appendix (p 37) describes the inclusion and exclusion criteria. We excluded participants with a BMI of less than 15 kg/m2 or 60 kg/m2 or more, or baseline age younger than 20 years or older than 90 years. To limit residual confounding by smoking and bias due to effects of pre-existing disease on baseline BMI (ie, reverse causality), the primary analysis was restricted to never-smokers without specific known chronic diseases at baseline (eg, cardiovascular disease, cancer, or respiratory diseases), and omitted the first 5 years of follow-up.

founding by smoking and bias due to effects of pre-existing disease on baseline BMI (ie, reverse causality), the primary analysis was restricted to never-smokers without specific known chronic diseases at baseline (eg, cardiovascular disease, cancer, or respiratory diseases), and omitted the first 5 years of follow-up. Statistical analysis Associations of all-cause mortality with BMI depend not only on the associations of specific causes of death with BMI in different regions (which might differ quantitatively), but also on how relatively common each specific cause of death is in the particular region (which can differ substantially by region and over time). Hence, the association of all-cause mortality with BMI might differ in regions with different underlying mortality patterns. Therefore, the prespecified primary analysis was stratified by five major geographical regions, three with extensive data (east Asia, Europe, and North America) and two with more limited data (Australia and New Zealand, and south Asia). Data from some or all regions are shown separately, in the main text or in the appendix.

lian randomisation analyses,32, 33 other instrumental variable analyses,34 and a meta-analysis of randomised trials.35 Our findings are also broadly consistent with the stricter analyses done in a 2015 study36 of 12 million Korean adults and with a 2016 review that attempted to limit the effects of reverse causality.37 The most important limitation is that our only measure of adiposity was BMI, so we could not directly address aspects of body composition such as visceral fat or fat distribution,38, 39 nor could we consider modification of HRs by metabolic factors.40 Such factors might have different effects in different populations because, at the same BMI, people of Asian ancestry might have higher amounts of body fat and greater risk of developing metabolic diseases than people of European ancestry.41 Moreover, south Asia, Africa, and Latin America were either unrepresented or poorly represented, and large studies in those areas might yield different findings. The study-specific results were in general not adjusted for ethnicity or for socioeconomic status. We did not adjust for regression dilution because previous surveys have reported high levels of concordance in replicate BMI measures taken from the same adults some years apart.42

fore, the prespecified primary analysis was stratified by five major geographical regions, three with extensive data (east Asia, Europe, and North America) and two with more limited data (Australia and New Zealand, and south Asia). Data from some or all regions are shown separately, in the main text or in the appendix. Each study (or consortium of studies) analysed individual-participant data according to a common analytical plan with SAS version 9.3 (SAS Institute, Cary, NC, USA) or Stata version 12 (StataCorp, College Station, TX, USA) provided by the coordinating centres. These separate results were then meta-analysed at Cambridge University, UK. To facilitate standardised comparisons with other meta-analyses, we calculated hazard ratios (HRs) for mortality in the six WHO-defined baseline BMI categories: underweight (15·0–<18·5 kg/m2), normal (18·5–<25·0 kg/m2; the reference category for analyses of these six BMI groups), overweight (25–<30·0 kg/m2), and obesity grade 1 (30·0–<35·0 kg/m2), grade 2 (35·0–<40·0 kg/m2), and grade 3 (40·0–<60·0 kg/m2).20 Because, however, most people are of normal weight or overweight, these two categories were subdivided, yielding nine groups (15·0–<18·5 kg/m2; 18·5–<20·0 kg/m2; 20·0–<22·5 kg/m2; 22·5–<25·0 kg/m2, the reference category for analyses of nine BMI groups; 25·0–<27·5 kg/m2; 27·5–<30·0 kg/m2; 30·0–<35·0 kg/m2; 35·0–<40·0 kg/m2; and 40·0–<60·0 kg/m2). The BMI group with the largest number of participants was chosen as the reference group.

groups (15·0–<18·5 kg/m2; 18·5–<20·0 kg/m2; 20·0–<22·5 kg/m2; 22·5–<25·0 kg/m2, the reference category for analyses of nine BMI groups; 25·0–<27·5 kg/m2; 27·5–<30·0 kg/m2; 30·0–<35·0 kg/m2; 35·0–<40·0 kg/m2; and 40·0–<60·0 kg/m2). The BMI group with the largest number of participants was chosen as the reference group. Study-specific log HRs in specific BMI categories were pooled by inverse-variance-weighted random-effects meta-analyses (an extension of the DerSimonian and Laird procedure) and plotted against the mean BMI value within each category. Sensitivity analyses used other statistical methods (eg, fixed-effect models). To enable comparisons across BMI groups irrespective of the choice of a reference group, a floating variance estimate (reflecting independent variability within each group, including the reference group) was attributed to each category using Plummer's method and used to calculate group-specific confidence intervals.21

To enable comparisons across BMI groups irrespective of the choice of a reference group, a floating variance estimate (reflecting independent variability within each group, including the reference group) was attributed to each category using Plummer's method and used to calculate group-specific confidence intervals.21 To estimate the BMI levels at which mortality risk was lowest (ie, the nadir), weighted linear regression yielded the best-fitting second-degree fractional polynomial model relating pooled log HRs to pooled mean BMI levels (weighted by the inverse of the floating variance of the log HR), and the minimum of this polynomial was the nadir. We assessed all-cause mortality and its main components, coronary heart disease, stroke, other cardiovascular disease, cancer, and respiratory disease (appendix p 15). HRs were calculated separately within each study with Cox regression models stratified for baseline age and sex (appendix pp 51–53), with participants contributing from the baseline survey in crude analyses or from year 5 in the primary analyses. HRs in sex-specific and baseline-age-specific groups (and, when appropriate, by trial groups) were combined across studies.22 To avoid over-fitting of statistical models, studies with ten or fewer deaths from a particular cause were excluded from meta-analyses of that cause.23, 24

r from year 5 in the primary analyses. HRs in sex-specific and baseline-age-specific groups (and, when appropriate, by trial groups) were combined across studies.22 To avoid over-fitting of statistical models, studies with ten or fewer deaths from a particular cause were excluded from meta-analyses of that cause.23, 24 Because the associations of BMI with mortality were approximately log-linear above a BMI of 25 kg/m2, we calculated HRs per 5 kg/m2 higher BMI increase by inverse-variance-weighted regression of the pooled log HRs on mean BMI values in each category.17 For all-cause mortality, we estimated population-attributable fractions for underweight, overweight, and obesity by combining the proportional excess mortality (X0, X1, and X2, where X=HR-1) in these BMI categories with the corresponding prevalences (P0, P1, and P2, taken from Global Burden of Disease25 region-specific prevalences). The population-attributable fractions for overweight and obesity are then P1X1/k and P2X2/k, where k=1 + P0X0 + P1X1 + P2X2. Between-study heterogeneity was quantified by the I2 statistic.26 We used two-sided p values and 95% CIs. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. SK, PG, EDA, and JD had access to all the data, and, together with SNB and FBH, were responsible for the decision to submit for publication.

Because the associations of BMI with mortality were approximately log-linear above a BMI of 25 kg/m2, we calculated HRs per 5 kg/m2 higher BMI increase by inverse-variance-weighted regression of the pooled log HRs on mean BMI values in each category.17 For all-cause mortality, we estimated population-attributable fractions for underweight, overweight, and obesity by combining the proportional excess mortality (X0, X1, and X2, where X=HR-1) in these BMI categories with the corresponding prevalences (P0, P1, and P2, taken from Global Burden of Disease25 region-specific prevalences). The population-attributable fractions for overweight and obesity are then P1X1/k and P2X2/k, where k=1 + P0X0 + P1X1 + P2X2. Between-study heterogeneity was quantified by the I2 statistic.26 We used two-sided p values and 95% CIs. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. SK, PG, EDA, and JD had access to all the data, and, together with SNB and FBH, were responsible for the decision to submit for publication. Results Of 10 625 411 participants from 239 studies (median follow-up 13·7 years, IQR 11·4–14·7), 3 951 455 people in 189 studies were never-smokers without specific chronic diseases at recruitment who survived 5 years, of whom 385 879 died. To limit bias, the prespecified primary analyses involved this restricted population. To avoid merging importantly different risks, many of these primary analyses further subdivided the WHO-defined normal and overweight BMI categories, yielding nine BMI groups rather than six.

ho survived 5 years, of whom 385 879 died. To limit bias, the prespecified primary analyses involved this restricted population. To avoid merging importantly different risks, many of these primary analyses further subdivided the WHO-defined normal and overweight BMI categories, yielding nine BMI groups rather than six. Table 1 shows the substantial relevance of successively stricter exclusions, going from crude analyses of about 10·6 million to prespecified analyses of about 4 million adults. With BMI in only six groups, the whole range from 18·5 kg/m2 to less than 25 kg/m2 is the reference group, and HRs were: underweight 1·47 (95% CI 1·39–1·55), overweight 1·11 (1·10–1·11), grade 1 obesity 1·44 (1·41–1·47), grade 2 obesity 1·92 (1·86–1·98), grade 3 obesity 2·71 (2·55–2·86), and any obesity 1·64 (1·61–1·67; appendix pp 16, 17, 25). With normal and overweight groups more finely subdivided, however, BMI 22·5 kg/m2 to less than 25·0 kg/m2 becomes the reference group, and with this more precise reference group, the HRs for grade 1, 2, and 3 obesity increased slightly (table 1, 2). Mortality was lowest in the BMI range from 20·0 kg/m2 to less than 25·0 kg/m2, and was significantly increased just below this BMI range and in the overweight range just above it (table 2).

eference group, and with this more precise reference group, the HRs for grade 1, 2, and 3 obesity increased slightly (table 1, 2). Mortality was lowest in the BMI range from 20·0 kg/m2 to less than 25·0 kg/m2, and was significantly increased just below this BMI range and in the overweight range just above it (table 2). In these prespecified analyses of almost 4 million adults, the HRs for overweight and for obesity grade 1 were broadly similar across different geographical regions (Europe, North America, east Asia, and Australia and New Zealand; numbers of deaths in south Asia were too small to be reliable), but the HRs for underweight and grade 3 obesity appeared somewhat higher in Europe than in east Asia (figure 1, table 3, appendix pp 16–21). Combining all regions, the HRs for overweight and obesity were higher at younger ages than older ages, and in men than women (Figure 2, Figure 3); this finding held in each major geographical region (appendix pp 22–24 38–40). In each region, BMI was non-linearly associated with all-cause mortality, with nadir at BMI 20·0 kg/m2 to less than 25·0 kg/m2 and excess mortality in underweight, overweight, and at BMI 18·5 kg/m2 to less than 20·0 kg/m2, at the lower end of the WHO-defined normal range. The nadir depended on age, and was BMI=22 kg/m2 for baseline age 35–49 years, BMI=23 kg/m2 for baseline age 50–69 years, and BMI=24 kg/m2 for baseline age 70–89 years.

kg/m2 and excess mortality in underweight, overweight, and at BMI 18·5 kg/m2 to less than 20·0 kg/m2, at the lower end of the WHO-defined normal range. The nadir depended on age, and was BMI=22 kg/m2 for baseline age 35–49 years, BMI=23 kg/m2 for baseline age 50–69 years, and BMI=24 kg/m2 for baseline age 70–89 years. Population-attributable fractions for all-cause mortality due to overweight or obesity were 19% in North America, 16% in Australia and New Zealand, and 14% in Europe, but only 5% in east Asia (appendix p 25). For BMI 25 kg/m2 or more, the association of BMI with all-cause mortality was approximately log-linear, and of similar strength in each region (except perhaps south Asia, where numbers of deaths were small), with HR per 5 kg/m2 units higher BMI 1·31 (95% CI 1·29–1·33) overall, 1·39 (1·34–1·44) in east Asia, 1·39 (1·34–1·43) in Europe, 1·29 (1·26–1·32) in North America, and 1·31 (1·27–1·35) in Australia and New Zealand. The HR decreased with age from 1·52 (1·47–1·56) for ages 35–49 years at baseline to 1·21 (1·17–1·25) for ages 70–89 years at baseline (trend p<0·0001; appendix p 26). The HR was 1·51 (1·46–1·56) for men versus 1·30 (1·26–1·33) for women (heterogeneity p<0·0001; figure 3). Hence, a given increase in BMI is associated with a far greater absolute mortality increase in men than in women (appendix p 45). As there were far more women than men, particularly among obese people, the HR among all participants was similar to the HR just among women.

3) for women (heterogeneity p<0·0001; figure 3). Hence, a given increase in BMI is associated with a far greater absolute mortality increase in men than in women (appendix p 45). As there were far more women than men, particularly among obese people, the HR among all participants was similar to the HR just among women. For each major cause of death, BMI was non-linearly associated with mortality in each major region we studied (appendix pp 27–29, 41, 42). Above 25 kg/m2, BMI was strongly positively related to coronary heart disease, stroke, and respiratory disease mortality, and moderately positively related to cancer mortality (figure 4); these findings were broadly similar in Europe, North America, and east Asia (appendix pp 28, 29). Within WHO's wide normal BMI range (18·5–<25·0 kg/m2) the main geographical difference was that, in east Asia, mortality from coronary heart disease had its nadir at 18·5–<20·0 kg/m2, lower than in other regions (appendix p 28). In all regions, underweight was associated with substantially higher respiratory disease mortality and somewhat higher mortality from coronary heart disease, stroke, and cancer (figure 4). HRs comparing underweight versus normal-weight cardiovascular disease mortality were more extreme in Europe than elsewhere (appendix pp 28, 29).

s, underweight was associated with substantially higher respiratory disease mortality and somewhat higher mortality from coronary heart disease, stroke, and cancer (figure 4). HRs comparing underweight versus normal-weight cardiovascular disease mortality were more extreme in Europe than elsewhere (appendix pp 28, 29). Compared with the strict primary analyses described above, crude analyses that ignored smoking and any effects of prior disease at baseline, and did not exclude the first 5 years of follow-up, yielded different (presumably substantially biased) results, with exaggerated HRs for underweight, inverted HRs for overweight, and less than half of the excess risk for grade 1 obesity suggested by the strict primary analyses (Table 1, Table 2, appendix p 43). In sensitivity analyses (appendix pp 30–36, 46–48), HRs were little changed in analyses that used fixed effect models or restricted follow-up to years 5–15; considered age at risk rather than age at baseline; adjusted additionally for race or excluded participants with diabetes at baseline; used only studies that included both sexes; used only studies with baseline data for heart disease, stroke, and cancer; or subdivided studies by mean baseline BMI or median recruitment year (HRs were somewhat higher in studies starting before 1990 than those after 1990, but meta-regression of HRs on year of recruitment was not significant). HRs did not vary substantially between larger and smaller studies, between studies with measured and self-reported BMI, or between occupational and other studies.

year (HRs were somewhat higher in studies starting before 1990 than those after 1990, but meta-regression of HRs on year of recruitment was not significant). HRs did not vary substantially between larger and smaller studies, between studies with measured and self-reported BMI, or between occupational and other studies. Discussion Associations between BMI and mortality can help to estimate the public health impact of excess adiposity only if the estimated relationships are not substantially distorted by the effects of smoking or ill health on BMI. Hence, our primary analyses were of never-smokers without previous disease who survived at least 5 years. Both overweight and obesity were associated with increased all-cause mortality. In the BMI range above 25 kg/m2 (ie, above the upper limit of the WHO's normal range) the relationship to mortality was steep in every global region we studied, except perhaps south Asia where numbers of deaths were small.27

least 5 years. Both overweight and obesity were associated with increased all-cause mortality. In the BMI range above 25 kg/m2 (ie, above the upper limit of the WHO's normal range) the relationship to mortality was steep in every global region we studied, except perhaps south Asia where numbers of deaths were small.27 Our primary analyses challenge previous suggestions that overweight (25–<30 kg/m2) and grade 1 obesity (30–<35 kg/m2) are not associated with higher mortality,28 bypassing speculation about hypothetical protective metabolic effects of increased body fat in apparently healthy individuals.29 In particular, the findings here contrast with those of a 2013 review that claimed that, relative to normal weight, grade 1 obesity was not associated with excess all-cause mortality and that overweight was associated with lower all-cause mortality.28 That review could not, however, control for the biases controlled for in our analysis. Indeed, the results of the current analysis (eg, table 1, table 2, and appendix pp 16, 17) show how the limited ability of that literature-based review to control for bias could have accounted for its misleading findings. Our study was able to reproduce such findings when conducting crude analyses with inadequate control of reverse causality, but not when we conducted appropriately strict analyses.

p 16, 17) show how the limited ability of that literature-based review to control for bias could have accounted for its misleading findings. Our study was able to reproduce such findings when conducting crude analyses with inadequate control of reverse causality, but not when we conducted appropriately strict analyses. Despite broadly similar overall findings across different continents, we found some differences. HRs per 5 kg/m2 higher BMI above 25 kg/m2 appeared to be somewhat greater in Europe than in North America. In each major region we studied, HRs were substantially higher at younger than at older ages, although the absolute excess mortality was higher in older people. HRs were substantially higher in men than in women, consistent with previous observations that, at equivalent BMI levels, men have greater insulin resistance, ectopic (eg, liver) fat levels, and type 2 diabetes prevalence.30 Our primary analyses of never-smokers included, however, far more women than men, particularly at higher BMI levels. Hence, our HRs for obesity (and, above 25·0 kg/m2, the excess HR per 5 kg/m2 increase in BMI) mainly describe effects in women, despite the substantially larger HRs in men. Our HRs for grade 1 obesity (male 1·70, female 1·37; appendix p 22) suggest that men have almost double the proportional excess mortality of women— but, as age-specific death rates are typically more than 50% higher in men, the absolute excess death rate associated with grade 1 obesity is about three times as great in men (appendix p 45).

se areas might yield different findings. The study-specific results were in general not adjusted for ethnicity or for socioeconomic status. We did not adjust for regression dilution because previous surveys have reported high levels of concordance in replicate BMI measures taken from the same adults some years apart.42 There are, however, particular strengths. Compared with single-country studies, we enhanced generalisability by combining findings from 239 studies across four continents. We had access to data for about 97% of the participants in the studies eligible for this analysis (giving large numbers and negligible bias from unavailability of particular studies), we used a prespecified analysis plan, we analysed individual-participant data to avoid the potentially important limitations of literature-based reviews,43 and we analysed clinically relevant subpopulations reliably, exploiting the considerable statistical power of the study. We avoided potential over-adjustment by not adjusting for variables (eg, diabetes status and physical activity) that could mediate associations between BMI and mortality.44 Finally, our results were robust to a variety of sensitivity analyses. We conclude that wherever overweight and obesity are common their associations with higher all-cause mortality are broadly similar in different populations, supporting strategies to combat the entire spectrum of excessive adiposity worldwide. Correspondence to: Prof John Danesh, Department of Public Health and Primary Care, University of Cambridge, Cambridge CB1 8RN, England, UK gbmc@phpc.cam.ac.uk

We conclude that wherever overweight and obesity are common their associations with higher all-cause mortality are broadly similar in different populations, supporting strategies to combat the entire spectrum of excessive adiposity worldwide. Correspondence to: Prof John Danesh, Department of Public Health and Primary Care, University of Cambridge, Cambridge CB1 8RN, England, UK gbmc@phpc.cam.ac.uk Supplementary Material Supplementary appendix Acknowledgments This paper is dedicated to the memory of Gary Whitlock, who contributed much to developing the collaboration. Global BMI Mortality Collaboration provides links to websites of the component studies (or consortia), many of which describe their funding. The coordinating centre at the University of Cambridge was funded by the UK Medical Research Council (G0800270), British Heart Foundation (SP/09/002), British Heart Foundation Cambridge Cardiovascular Centre of Excellence, and National Institute for Health Research Cambridge Biomedical Research Centre. The work of the coordinating centre at the Harvard TH Chan School of Public Health was funded by grants P01 CA87969, UM1 CA176726, UM1 CA167552, DK58845, P30 DK046200, and U54 CA155626 from the National Institutes of Health. This research has been conducted using the UK Biobank resource.

dge Biomedical Research Centre. The work of the coordinating centre at the Harvard TH Chan School of Public Health was funded by grants P01 CA87969, UM1 CA176726, UM1 CA167552, DK58845, P30 DK046200, and U54 CA155626 from the National Institutes of Health. This research has been conducted using the UK Biobank resource. Contributors All of the authors contributed to data collection, and the design, analysis, interpretation, and re-drafting of this paper. EDA, SNB, DW, SK, BJC, RH, SL, MW, JD, and FBH drafted the study protocol and analysis plan. SK, PG, and DW conducted the combined statistical analysis. EDA, SNB, RP, JD, and FBH drafted the manuscript.

the authors contributed to data collection, and the design, analysis, interpretation, and re-drafting of this paper. EDA, SNB, DW, SK, BJC, RH, SL, MW, JD, and FBH drafted the study protocol and analysis plan. SK, PG, and DW conducted the combined statistical analysis. EDA, SNB, RP, JD, and FBH drafted the manuscript. The Global BMI Mortality Collaboration Writing Committee (*equal contribution)—Emanuele Di Angelantonio (University of Cambridge, Cambridge, UK)*; Shilpa N Bhupathiraju (Harvard TH Chan School of Public Health, Harvard University, Boston, MA, USA)*; David Wormser (University of Cambridge, Cambridge, UK)*; Pei Gao (University of Cambridge, Cambridge, UK and Peking University, Beijing, China)*; Stephen Kaptoge (University of Cambridge, Cambridge, UK)*; Amy Berrington de Gonzalez (National Cancer Institute, Bethesda, MD, USA)*; Benjamin J Cairns (University of Oxford, Oxford, UK)*; Rachel Huxley (Curtin University, Perth, Australia)*; Chandra L Jackson (Harvard Medical School, Harvard University, Boston, MA, USA)*; Grace Joshy (Australian National University, Canberra, Australia)*; Sarah Lewington (University of Oxford, Oxford, UK)*; JoAnn E Manson (Harvard TH Chan School of Public Health and Harvard Medical School, Harvard University, Boston, MA, USA)*; Neil Murphy (Imperial College London, London, UK)*; Alpa V Patel (American Cancer Society, Atlanta, GA, USA)*; Jonathan M Samet (University of Southern California, Los Angeles, CA, USA)*; Mark Woodward (University of Oxford, Oxford, UK; University of Sydney, NSW, Australia; and Johns Hopkins University, Baltimore, MD, USA)*; Wei Zheng (Vanderbilt University Medical Center, Nashville, TN, USA)*; Maigen Zhou (Chinese Center for Disease Control and Prevention, Beijing, China)*; Narinder Bansal (University of Cambridge, Cambridge, UK); Aurelio Barricarte (Navarre Public Health Institute and Consortium for Biomedical Research in Epidemiology and Public Health, Pamplona, Spain); Brian Carter (American Cancer Society, Atlanta, GA, USA); James R Cerhan (Mayo Clinic, Rochester, MN, USA), Rory Collins (University of Oxford, Oxford, UK); George Davey Smith (University of Bristol, Bristol, UK); Xianghua Fang (Capital Medical University, Beijing, China); Oscar H Franco (University Medical Center Rotterdam, Rotterdam, Netherlands); Jane Green (University of Oxford, Oxford, UK); Jim Halsey (University of Oxford, Oxford, UK); Janet S Hildebrand (American Cancer Society, Atlanta, GA, USA); Keum Ji Jung (Yonsei University, Seoul, Ko

ital Medical University, Beijing, China); Oscar H Franco (University Medical Center Rotterdam, Rotterdam, Netherlands); Jane Green (University of Oxford, Oxford, UK); Jim Halsey (University of Oxford, Oxford, UK); Janet S Hildebrand (American Cancer Society, Atlanta, GA, USA); Keum Ji Jung (Yonsei University, Seoul, Ko rea); Rosemary J Korda (Australian National University, Canberra, Australia); Dale F McLerran (Fred Hutchinson Cancer Research Center, Seattle, WA, USA); Steven C Moore (National Cancer Institute, Bethesda, MD, USA); Linda M O'Keeffe (University of Cambridge, Cambridge, UK); Ellie Paige (University of Cambridge, Cambridge, UK); Anna Ramond (University of Cambridge, Cambridge, UK); Gillian K Reeves (University of Oxford, Oxford, UK); Betsy Rolland (National Cancer Institute, Bethesda, MD, USA); Carlotta Sacerdote (University of Turin, Center for Cancer Prevention, Turin, Italy); Naveed Sattar (University of Glasgow, Glasgow, UK); Eleni Sofianopoulou (University of Cambridge, Cambridge, UK); June Stevens (University of North Carolina, Chapel Hill, NC, USA); Michael Thun (American Cancer Society, Atlanta, GA, USA); Hirotsugu Ueshima (Shiga University of Medical Science, Shiga, Japan); Ling Yang (University of Oxford, Oxford, UK); Young Duk Yun (Health Insurance Policy Research Institute, Seoul, South Korea); Peter Willeit (University of Cambridge, Cambridge, UK and Medical University Innsbruck, Innsbruck, Austria); Emily Banks (Australian National University, Canberra, ACT, Australia)*; Valerie Beral (University of Oxford, Oxford, UK)*; Zhengming Chen (University of Oxford, Oxford, UK)*; Susan M Gapstur (American Cancer Society, Atlanta, GA, USA)*; Marc J Gunter (International Agency for Research on Cancer, Lyon, France)*; Patricia Hartge (National Cancer Institute, Bethesda, MD, USA)*; Sun Ha Jee (Yonsei University, Seoul, Korea)*; Tai-Hing Lam (University of Hong Kong, Hong Kong, China)*; Richard Peto (University of Oxford, Oxford, UK)*; John D Potter (Massey University, Wellington, New Zealand)*; Walter C Willett (Harvard TH Chan School of Public Health and Harvard Medical School, Harvard University, Boston, MA, USA)*; Simon G Thompson (University of Cambridge, Cambridge, UK)*; John Danesh (University of Cambridge, Cambridge, UK)*; Frank B Hu (Harvard TH Chan School of Public Health and Harvard Medical School, Harvard University, Boston, MA, USA)*.

ol of Public Health and Harvard Medical School, Harvard University, Boston, MA, USA)*; Simon G Thompson (University of Cambridge, Cambridge, UK)*; John Danesh (University of Cambridge, Cambridge, UK)*; Frank B Hu (Harvard TH Chan School of Public Health and Harvard Medical School, Harvard University, Boston, MA, USA)*. Declaration of interests EDA received research funding from UK Medical Research Council, British Heart Foundation, National Institute of Health Research, NHS Blood and Transplant, European Commission Framework Programme during the conduct of the study; and personal fees from Elsevier (France). Since January, 2014, DW has been a full-time employee of F. Hoffmann-La Roche and received personal fees and holding shares in F. Hoffmann-La Roche. PG received grants from Recruitment Program for Young Professionals in China and British Heart Foundation. BJC, JG, GKR, and VB received research funding from Cancer Research UK and Medical Research Council. BJC received funding from the British Heart Foundation Centre of Research Excellence, Oxford. MW received personal fees from Novartis and Amgen. OHF received research funding from Nestle and Metagenics. RJK received grants from National Health and Medical Research Council. NS received personal fees from AstraZeneca, Boehringer Ingelheim, and Janssen. EB received grants from National Health and Medical Research Council of Australia and National Heart Foundation of Australia. SGT received grants from UK Medical Research Council and British Heart Foundation. JD has received research funding from the British Heart Foundation, NIHR Cambridge Comprehensive Biomedical Research Centre, BUPA Foundation, diaDexus, European Research Council, European Union, Evelyn Trust, Fogarty International Centre, GlaxoSmithKline, Merck, National Heart, Lung and Blood Institute, National Institute for Health Research, National Institute of Neurological Disorders and Stroke, NHS Blood and Transplant, Novartis, Pfizer, UK Medical Research Council, and Wellcome Trust. All other members of the writing committee declare no competing interests.

Kline, Merck, National Heart, Lung and Blood Institute, National Institute for Health Research, National Institute of Neurological Disorders and Stroke, NHS Blood and Transplant, Novartis, Pfizer, UK Medical Research Council, and Wellcome Trust. All other members of the writing committee declare no competing interests. Figure 1 Association of body-mass index with all-cause mortality, by geographical region Boxes are plotted against the mean BMI in each group. The HR per 5 kg/m2 higher body-mass index (BMI) and its 95% CI are calculated only for BMI more than 25·0 kg/m2. Analyses restricted to never-smokers without pre-existing chronic disease, excluding the first 5 years of follow-up. The reference category is shown with the arrow and is 22·5–<25·0 kg/m2. CIs are from floating variance estimates (reflecting independent variability within each category, including reference). Areas of squares are proportional to the information content (ie, inverse of the floating variance). HR=hazard ratio. Figure 2 Association of body-mass index with all-cause mortality, by baseline age group

Boxes are plotted against the mean BMI in each group. The HR per 5 kg/m2 higher body-mass index (BMI) and its 95% CI are calculated only for BMI more than 25·0 kg/m2. Analyses restricted to never-smokers without pre-existing chronic disease, excluding the first 5 years of follow-up. The reference category is shown with the arrow and is 22·5–<25·0 kg/m2. CIs are from floating variance estimates (reflecting independent variability within each category, including reference). Areas of squares are proportional to the information content (ie, inverse of the floating variance). HR=hazard ratio. Figure 2 Association of body-mass index with all-cause mortality, by baseline age group The HR per 5 kg/m2 higher body-mass index (BMI) and its 95% CI are calculated only for BMI more than 25·0 kg/m2. Analyses restricted to never-smokers without pre-existing chronic disease, and excluding the first 5 years of follow-up, and include data from all geographical regions. The reference category is shown with the arrow and is 22·5–<25·0 kg/m2. CIs are from floating variance estimates (reflecting independent variability within each category, including the reference category). Areas of squares are proportional to the information content. Analyses by baseline age and the three main geographical regions are in the appendix (p 38). HR=hazard ratio. Figure 3 Association of body-mass index with all-cause mortality, by sex

The HR per 5 kg/m2 higher body-mass index (BMI) and its 95% CI are calculated only for BMI more than 25·0 kg/m2. Analyses restricted to never-smokers without pre-existing chronic disease, and excluding the first 5 years of follow-up, and include data from all geographical regions. The reference category is shown with the arrow and is 22·5–<25·0 kg/m2. CIs are from floating variance estimates (reflecting independent variability within each category, including the reference category). Areas of squares are proportional to the information content. Analyses by baseline age and the three main geographical regions are in the appendix (p 38). HR=hazard ratio. Figure 3 Association of body-mass index with all-cause mortality, by sex The HR per 5 kg/m2 higher body-mass index (BMI) and its 95% CI are calculated only for BMI more than 25·0 kg/m2. Analyses restricted to never-smokers without pre-existing chronic disease, excluding the first 5 years of follow-up, and include data from all geographical regions. The reference category is shown with the arrow and is 22·5–<25·0 kg/m2. CIs are from floating variance estimates (reflecting independent variability within each category, including reference). Areas of squares are proportional to the information content. Analyses by sex and the three main geographical regions (east Asia, Europe, and North America) are in the appendix (p 39). HR=hazard ratio. Figure 4 Association of body-mass index with mortality, by major underlying cause

The HR per 5 kg/m2 higher body-mass index (BMI) and its 95% CI are calculated only for BMI more than 25·0 kg/m2. Analyses restricted to never-smokers without pre-existing chronic disease, excluding the first 5 years of follow-up, and include data from all geographical regions. The reference category is shown with the arrow and is 22·5–<25·0 kg/m2. CIs are from floating variance estimates (reflecting independent variability within each category, including reference). Areas of squares are proportional to the information content. Analyses by sex and the three main geographical regions (east Asia, Europe, and North America) are in the appendix (p 39). HR=hazard ratio. Figure 4 Association of body-mass index with mortality, by major underlying cause The HR per 5 kg/m2 higher body-mass index (BMI) and its 95% CI are calculated only for BMI more than 25·0 kg/m2. Analyses restricted to never-smokers without pre-existing chronic disease, excluding the first 5 years of follow-up, and include data from all geographical regions. The reference category is shown with the arrow and is 22·5–<25·0 kg/m2. CIs are from floating variance estimates (reflecting independent variability within each category, including reference). Areas of squares are proportional to the information content. Analyses of cause-specific mortality by three geographical regions (east Asia, Europe, and North America) are in the appendix (pp 41, 42). Table 1 Effects of successively stricter precautions against bias on analyses of six WHO defined groups of BMI versus all-cause mortality

The HR per 5 kg/m2 higher body-mass index (BMI) and its 95% CI are calculated only for BMI more than 25·0 kg/m2. Analyses restricted to never-smokers without pre-existing chronic disease, excluding the first 5 years of follow-up, and include data from all geographical regions. The reference category is shown with the arrow and is 22·5–<25·0 kg/m2. CIs are from floating variance estimates (reflecting independent variability within each category, including reference). Areas of squares are proportional to the information content. Analyses of cause-specific mortality by three geographical regions (east Asia, Europe, and North America) are in the appendix (pp 41, 42). Table 1 Effects of successively stricter precautions against bias on analyses of six WHO defined groups of BMI versus all-cause mortality Underweight (15·0– <18·5 kg/m2) Normal weight (18·5– <25·0 kg/m2) Overweight (25·0– <30·0 kg/m2) Obesity grade 1 (30·0– <35·0 kg/m2) Obesity grade 2 (35·0– <40·0 kg/m2) Obesity grade 3 (40·0– <60·0 kg/m2) Crude analysis with no exclusions* Participants/deaths 292 003/68 455 5 586 892/810 838 3 467 617/526 098 946 257/144 871 237 223/36 113 92 458/15 399 HR (95% CI) 1·82 (1·74–1·91) 1·00 (0·98–1·02) 0·95 (0·94–0·97) 1·17 (1·16–1·18) 1·49 (1·47–1·51) 1·95 (1·90–2·01) Participants without known disease at baseline† Participants/deaths 255 000/52 789 4 922 817/631 488 2 916 978/388 781 756 075/102 315 183 689/24 556 696 88/10 321 HR (95% CI) 1·81 (1·72–1·91) 1·00 (0·98–1·02) 0·95 (0·95–0·96) 1·18 (1·16–1·20) 1·52 (1·48–1·55) 2·05 (1·98–2·13) Participants without known chronic disease at baseline, adjusted for smoking status‡ Participants/deaths 245 080/51 170 4 751 019/618 881 2 826 687/381 617 733 108/100 113 178 130/23 945 67 593/10 002 HR (95% CI) 1·70 (1·61–1·80) 1·00 (0·98–1·02) 0·99 (0·98–1·00) 1·25 (1·23–1·27) 1·63 (1·59–1·66) 2·24 (2·15–2·33) Participants without known chronic disease at baseline, adjusted for smoking status, and excluding the first 5 years of follow-up§ Participants/deaths 208 044/33 817 4 234 052/496 310 2 513 128/312 450 641 237/80 037 152 741/18 737 56 232/7 659 HR (95% CI) 1·60 (1·51–1·70) 1·00 (0·98–1·02) 1·03 (1·01–1·04) 1·31 (1·29–1·33) 1·70 (1·67–1·74) 2·36 (2·27–2·45) The primary prespecified analysis: never-smokers without known chronic disease at baseline—excluding the first 5 years of follow-up¶ Participants/deaths 114 091/12 726 2 145 550/192 523; 1 250 103/130 293; 330 840/37 318 80 827/9 179 30 044/3 840 HR (95% CI) 1·47 (1·39–1·55) 1·00 (0·98–1·02) 1·11 (1·10, 1·11) 1·44 (1·41–1·47) 1·92 (1·86–1·98) 2·71 (2·55–2·86) CIs were calculated with floating variance estimates (reflecting independent variability within each group, including the reference group). Reference group is normal weight (18·5–<25·0 kg/m2). All analyses are adjusted for age and sex. Baseline BMI categories were defined by WHO. BMI=body-mass index. HR=hazard ratio.

) CIs were calculated with floating variance estimates (reflecting independent variability within each group, including the reference group). Reference group is normal weight (18·5–<25·0 kg/m2). All analyses are adjusted for age and sex. Baseline BMI categories were defined by WHO. BMI=body-mass index. HR=hazard ratio. * 237 studies; 10 622 450 participants; 1 601 774 deaths. † 236 studies; 9 104 247 participants; 1 210 250 deaths. ‡ 234 studies; 8 801 617 participants; 1 185 728 deaths. § 213 studies; 7 805 434 participants; 949 010 deaths. ¶ 189 studies; 3 951 455 participants; 385 879 deaths. Table 2 Nine groups of BMI versus all-cause mortality, with use of the primary prespecified analysis

* 237 studies; 10 622 450 participants; 1 601 774 deaths. † 236 studies; 9 104 247 participants; 1 210 250 deaths. ‡ 234 studies; 8 801 617 participants; 1 185 728 deaths. § 213 studies; 7 805 434 participants; 949 010 deaths. ¶ 189 studies; 3 951 455 participants; 385 879 deaths. Table 2 Nine groups of BMI versus all-cause mortality, with use of the primary prespecified analysis 15·0–<18·5 kg/m2 18·5–<20·0 kg/m2 20·0– <22·5 kg/m2 22·5–<25·0 kg/m2 25·0–<27·5 kg/m2 27·5–<30·0 kg/m2 30·0–<35·0 kg/m2 35·0–<40·0 kg/m2 40·0–<60·0 kg/m2 Participants/deaths 114 091/12 726 230 749/20 989 838 907/72 701 1075 894/98 833 821 303/84 952 428 800/45 341 330 840/37 318 80 827/9 179 30 044/3 840 HR (95% CI) 1·51 (1·43–1·59) 1·13 (1·09–1·17) 1·00 (0.98–1.02) 1·00 (0·99–1·01) 1·07 (1·07–1·08) 1·20 (1·18–1·22) 1·45 (1·41–1·48) 1·94 (1·87–2·01) 2·76 (2·60–2·92) 189 studies; 3 951 455 participants; 385 879 deaths. The primary prespecified analysis in never-smokers without known chronic disease at baseline, excluding the first 5 years of follow-up (with normal weight and overweight categories further subdivided into: 18·5–<20·0 kg/m2, 20·0–<22·5 kg/m2, 22·5–<25·0 kg/m2, 25·0–<27·5 kg/m2, and 27·5–<30·0 kg/m2). CIs were calculated using floating variance estimates (reflecting independent variability within each group, including the reference group). Reference group is 22·5–<25·0 kg/m2. All analyses are adjusted for age and sex. Baseline BMI categories were defined by WHO. BMI=body-mass index. HR=hazard ratio.

·5–<30·0 kg/m2). CIs were calculated using floating variance estimates (reflecting independent variability within each group, including the reference group). Reference group is 22·5–<25·0 kg/m2. All analyses are adjusted for age and sex. Baseline BMI categories were defined by WHO. BMI=body-mass index. HR=hazard ratio. Table 3 Nine BMI groups versus all-cause mortality in never-smokers, excluding chronic disease at baseline and 5 years of follow-up in geographical regions with more than 1 million participants

·5–<30·0 kg/m2). CIs were calculated using floating variance estimates (reflecting independent variability within each group, including the reference group). Reference group is 22·5–<25·0 kg/m2. All analyses are adjusted for age and sex. Baseline BMI categories were defined by WHO. BMI=body-mass index. HR=hazard ratio. Table 3 Nine BMI groups versus all-cause mortality in never-smokers, excluding chronic disease at baseline and 5 years of follow-up in geographical regions with more than 1 million participants 15·0–<18·5 kg/m2 18·5–<20·0 kg/m2 20·0–<22·5 kg/m2 22·5–<25·0 kg/m2 25·0–<27·5 kg/m2 27·5–<30·0 kg/m2 30·0–<35·0 kg/m2 35·0–<40·0 kg/m2 40·0–<60·0 kg/m2 Europe* Participants/deaths 13 398/675 42 584/1508 199 369/7449 306 566/13278 249 929/12 850 153 147/8935 127 536/8386 32 749/2424 10 322/972 HR (95% CI) 1·79 (1·63–1·97) 1·25 (1·14–1·38) 1·02 (0·97–1·07) 1·00 (0·97–1·03) 1·07 (1·06–1·09) 1·21 (1·18–1·25) 1·52 (1·45–1·58) 1·99 (1·87–2·12) 3·04 (2·84–3·27) North America† Participants/deaths 22 028/3846 67 114/8597 274 883/36 200 359 022/54 995 317 721/53 464 168 183/28 471 149 807/25 348 39 379/6299 16 950/2702 HR (95% CI) 1·51 (1·34–1·70) 1·09 (1·02–1·16) 1·01 (0·96–1·06) 1·00 (0·97–1·03) 1·06 (1·04–1·07) 1·17 (1·12–1·22) 1·39 (1·30–1·49) 1·93 (1·74–2·13) 2·58 (2·26–2·93) East Asia‡ Participants/deaths 46 979/7178 94 409/10 206 301 242/27 537 336 758/28 755 194 857/17 070 72 133/6950 25 658/2753 1941/231 408/104 HR (95% CI) 1·36 (1·25–1·49) 1·11 (1·04–1·18) 0·99 (0·97–1·02) 1·00 (0·97–1·03) 1·07 (1·04–1·11) 1·28 (1·21–1·35) 1·54 (1·42–1·67) 2·01 (1·59–2·54) 2·38 (1·33–4·24) p value for heterogeneityl¶ 0·0045 0·28 0·42 .. 0·89 0·46 0·20 0·48 <0·0001 Normal weight and overweight are subdivided, and the reference category is BMI 22·5 kg/m2 to less than 25·0 kg/m2. Numbers of studies, participants, and deaths are shown after exclusions from these prespecified principal analyses. CIs were calculated using floating variance estimates (reflecting independent variability within each group, including the reference group). Results from studies in south Asia and Australia and New Zealand are in figure 1, with details in the appendix (p 20).

after exclusions from these prespecified principal analyses. CIs were calculated using floating variance estimates (reflecting independent variability within each group, including the reference group). Results from studies in south Asia and Australia and New Zealand are in figure 1, with details in the appendix (p 20). * 89 studies; 1 135 600 participants; 56 477 deaths. † 40 studies; 1 415 087 participants; 219 922 deaths. ‡ 46 studies; 1 074 385 participants; 100 784 deaths. ¶ p value for heterogenity is for all three regions.

Introduction Clinical depression is a common and debilitating mental health disorder, being the second largest cause of global disability.1 Globally, the effect of depression on aggregate economic output is predicted to be US$5·36 trillion between 2011 and 2030.2 Reduction of these substantial costs is a key objective for low-income, middle-income, and high-income countries alike. Antidepressant medication and cognitive behavioural therapy (CBT) have the most clinical evidence. However, although antidepressant medications are cheap, their use is limited by side-effects, poor patient adherence, and discontinuation relapse risk. CBT is as effective as are antidepressants3 and provides long-term protection against relapse, but it is complex and its effectiveness is dependent on the skills of psychological therapists, who are expensive to train and employ. For low-income and middle-income countries in particular, the need for an extensive professional infrastructure of such therapists limits access to CBT. Research in context Evidence before this study

Introduction Clinical depression is a common and debilitating mental health disorder, being the second largest cause of global disability.1 Globally, the effect of depression on aggregate economic output is predicted to be US$5·36 trillion between 2011 and 2030.2 Reduction of these substantial costs is a key objective for low-income, middle-income, and high-income countries alike. Antidepressant medication and cognitive behavioural therapy (CBT) have the most clinical evidence. However, although antidepressant medications are cheap, their use is limited by side-effects, poor patient adherence, and discontinuation relapse risk. CBT is as effective as are antidepressants3 and provides long-term protection against relapse, but it is complex and its effectiveness is dependent on the skills of psychological therapists, who are expensive to train and employ. For low-income and middle-income countries in particular, the need for an extensive professional infrastructure of such therapists limits access to CBT. Research in context Evidence before this study Authors of published systematic reviews, including a Cochrane review, have commented on the limitations of existing evidence for the effectiveness of behavioural activation (BA) for depression compared with cognitive behavioural therapy (CBT) and the scarcity of cost-effectiveness data, with the existing evidence insufficiently robust to establish comparability. Authors of the Cochrane review called for studies that improve the quality of evidence. Our pretrial evidence took published review findings from the UK National Institute for Health and Care Excellence (NICE), who reported no difference in treatment outcome between BA and CBT immediately after treatment (Hedges' g 0·139 [95% CI −0·400 to 0·122]; p=0·296) and subsequent follow-up (0·135 [−0·456 to 0·186]; p=0·409). The authors of NICE's review regarded the existing international evidence as insufficient to recommend BA for first-line treatment in clinical guidelines for depression.

A and CBT immediately after treatment (Hedges' g 0·139 [95% CI −0·400 to 0·122]; p=0·296) and subsequent follow-up (0·135 [−0·456 to 0·186]; p=0·409). The authors of NICE's review regarded the existing international evidence as insufficient to recommend BA for first-line treatment in clinical guidelines for depression. Added value of this study This trial addresses these research recommendations and is, to our knowledge, the only high-quality, fully powered non-inferiority and cost-effectiveness study addressing both the effects and costs of BA compared with CBT for depression. When we combine the data from our study with data from other international studies in the meta-analysis done by NICE, our data reduce the 95% CIs around the effect size for depression symptoms immediately after treatment (Hedges' g 0·054 [95% CI −0·214 to 0·107]; p=0·514) and at follow-up (0·059 [−0·234 to 0·115]; p=0·503) and unequivocally show both non-inferiority of BA compared with CBT and that BA is more cost-effective than is CBT against commonly applied decision maker willingness-to-pay thresholds. Implications of all the available evidence Junior mental health workers with no professional training in psychological therapies can deliver behavioural activation, a simple psychological treatment, with no lesser effect than CBT has and at less cost. Effective psychological therapy for depression can be delivered without the need for costly and highly trained professionals.

health workers with no professional training in psychological therapies can deliver behavioural activation, a simple psychological treatment, with no lesser effect than CBT has and at less cost. Effective psychological therapy for depression can be delivered without the need for costly and highly trained professionals. Globally, health services require effective, easily implemented, and cost-effective psychological treatments for depression that can be delivered by less specialist health workers than are needed at present to close a treatment gap that can be as much as 80–90% in some low-income countries.4 One potential alternative, behavioural activation (BA), is a simple psychological treatment for depression. It might be easy and quick to train junior mental health workers (MHWs) in BA who have no professional training in psychological therapies.5 However, this method is only appropriate if BA delivered in this way is as effective as and more cost-effective than is CBT.

), is a simple psychological treatment for depression. It might be easy and quick to train junior mental health workers (MHWs) in BA who have no professional training in psychological therapies.5 However, this method is only appropriate if BA delivered in this way is as effective as and more cost-effective than is CBT. Although BA compares favourably with CBT in systematic reviews,6, 7 the existing evidence is insufficiently robust to establish comparability.8 Authors of a Cochrane review7 called for more quality studies than have been done so far and the UK National Institute for Health and Care Excellence (NICE) regarded the international evidence as insufficient to recommend BA for first-line treatment in clinical guidelines,8 instead recommending a large non-inferiority study: “to establish whether behavioural activation is an effective alternative to CBT”.8 Given these recommendations, we hypothesised that BA is non-inferior to CBT for depression treatment response in adults with depression and that BA is cost-effective compared with CBT.

instead recommending a large non-inferiority study: “to establish whether behavioural activation is an effective alternative to CBT”.8 Given these recommendations, we hypothesised that BA is non-inferior to CBT for depression treatment response in adults with depression and that BA is cost-effective compared with CBT. Methods Study design and participants In this randomised, controlled, open-label, non-inferiority trial (the Cost and Outcome of Behavioural Activation versus Cognitive Behaviour Therapy for Depression [COBRA] trial), we recruited participants from primary care and psychological therapy services in Devon, Durham, and Leeds (UK). Eligible participants were adults aged 18 years or older who met diagnostic criteria for major depressive disorder assessed by researchers using a standard clinical interview (Structured Clinical Interview for the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition [SCID]9). We excluded people at interview who were receiving psychological therapy, were alcohol or drug dependent, were acutely suicidal or had attempted suicide in the previous 2 months, or were cognitively impaired, or who had bipolar disorder or psychosis or psychotic symptoms.

l of Mental Disorders, Fourth Edition [SCID]9). We excluded people at interview who were receiving psychological therapy, were alcohol or drug dependent, were acutely suicidal or had attempted suicide in the previous 2 months, or were cognitively impaired, or who had bipolar disorder or psychosis or psychotic symptoms. We recruited participants by searching the electronic case records of general practices and psychological therapy services for patients with depression, identifying potential participants from depression classification codes. Practices or services contacted patients to seek permission for researcher contact. The research team interviewed those that responded, provided detailed information on the study, took informed written consent, and assessed people for eligibility. The UK South West Research Ethics Committee gave national approval for the study (NRES/07/H1208/60). The protocol has been published previously.10

contact. The research team interviewed those that responded, provided detailed information on the study, took informed written consent, and assessed people for eligibility. The UK South West Research Ethics Committee gave national approval for the study (NRES/07/H1208/60). The protocol has been published previously.10 Randomisation and masking After eligibility was established, consent agreed, and baseline data collected, we randomly allocated participants (1:1) to BA or CBT using computer-generated allocation, stratified by depression severity according to the Patient Health Questionnaire 9 (PHQ-9)11 (<19 vs ≥19), antidepressant use (taking antidepressants or not), and recruitment site (Devon, Durham, or Leeds). A computer-based system allocated the first 20 participants to each group on a truly random basis. For subsequent participants, allocation was minimised to maximise the likelihood of balance in stratification variables across the two study groups. The registered Peninsula Clinical Trials Unit (Plymouth University, Plymouth, UK) allocated participants remotely after baseline data entry to ensure allocation concealment. Treatment was given open label, but outcome assessors were masked to participants' allocations. Concealment was ensured by use of an externally administered password-protected trial website with retention of a stochastic element to the minimisation algorithm. We recorded instances when outcome assessors were unmasked during interviews if participants informed them of their allocation.

o participants' allocations. Concealment was ensured by use of an externally administered password-protected trial website with retention of a stochastic element to the minimisation algorithm. We recorded instances when outcome assessors were unmasked during interviews if participants informed them of their allocation. Procedures We developed our clinical protocols in line with published treatment protocols,12, 13 including those from our own trials,14, 15 advice from international collaborators, and NICE recommendations8 for duration and frequency of BA and CBT. Full-time National Health Service (NHS) MHWs and therapists worked half of their working week for COBRA (with the other half worked as normal) and followed written manuals to deliver a maximum of 20 sessions over 16 weeks, with the option of four additional booster sessions if the patients wanted them.8 Treatment included core and supplementary techniques appropriate to the BA or CBT protocol to be used as clinically indicated; for example, behavioural or cognitive strategies for management of anxiety. All core components of both treatments were delivered by session eight, which we considered to represent a minimally sufficient dose of therapy (appendix). Sessions were face to face, lasting for 60 min. BA and CBT experts on the trial team trained MHWs and therapists for 5 days in either BA or CBT. MHWs and therapists were assessed for competence at the end of training with use of standardised quality criteria instruments consistent with the relevant treatment: either the Quality of Behavioral Activation Scale (Dimidjian S, University of Colorado, personal communication) or the Revised Cognitive Therapy Scale for CBT.16 Further training was given if competency was not demonstrated. MHWs and therapists received 60 min of clinical supervision fortnightly from NHS psychological therapists clinically experienced in BA or CBT, overseen by trial team experts.

olorado, personal communication) or the Revised Cognitive Therapy Scale for CBT.16 Further training was given if competency was not demonstrated. MHWs and therapists received 60 min of clinical supervision fortnightly from NHS psychological therapists clinically experienced in BA or CBT, overseen by trial team experts. Junior MHWs—graduates trained to deliver guided self-help interventions, but with neither professional mental health qualifications nor formal training in psychological therapies—delivered an individually tailored programme re-engaging participants with positive environmental stimuli and developing depression management strategies. Participants were encouraged to increase their contact with individually specified positive situations and reduce their avoidance of other situations. Specific BA techniques included identification of depressed behaviours, analysis of the triggers and consequences of depressed behaviours, monitoring of activities, development of alternative goal-orientated behaviours, scheduling of activities, and development of alternative behavioural responses to rumination.

tions. Specific BA techniques included identification of depressed behaviours, analysis of the triggers and consequences of depressed behaviours, monitoring of activities, development of alternative goal-orientated behaviours, scheduling of activities, and development of alternative behavioural responses to rumination. Professional or equivalently qualified psychotherapists, accredited as CBT therapists with the British Association of Behavioural and Cognitive Psychotherapy, with a postgraduate diploma in CBT, delivered a personalised treatment programme based on an assessment of how participants' beliefs lead to emotional distress and ineffectual coping. Participants used cognitive and behavioural exercises to specifically test the accuracy of those beliefs by identifying and modifying negative thoughts and beliefs that give rise to them. Specific techniques included participants monitoring moods and activities, planning of exercises to test negative beliefs, and thought records to identify and examine the accuracy of negative automatic thoughts and underlying beliefs. We did follow-up assessments 6 months, 12 months, and 18 months after randomisation.

Specific techniques included participants monitoring moods and activities, planning of exercises to test negative beliefs, and thought records to identify and examine the accuracy of negative automatic thoughts and underlying beliefs. We did follow-up assessments 6 months, 12 months, and 18 months after randomisation. We assessed the quality of and adherence to treatment using audiotapes and written records of therapy sessions. Independent experts in both treatments rated a random (with use of a computer-generated random number sequence) sample of tapes, stratified by therapist, therapy session, and intervention, for competence using the Revised Cognitive Therapy Scale16 for CBT (range 0–72) and the Quality of Behavioral Activation Scale (Dimidjian S, University of Colorado, personal communication) for BA (range 0–96). All therapists recorded the specific therapeutic techniques that they had used for each session on a checklist.

or competence using the Revised Cognitive Therapy Scale16 for CBT (range 0–72) and the Quality of Behavioral Activation Scale (Dimidjian S, University of Colorado, personal communication) for BA (range 0–96). All therapists recorded the specific therapeutic techniques that they had used for each session on a checklist. Outcomes The primary outcome was self-reported depression severity (PHQ-9 score11) at 12 months. Secondary outcomes were PHQ-9 score at 6 months and 18 months and Diagnostic and Statistical Manual of Mental Disorders IV major depressive and anxiety disorder status and number of depression-free days between follow-ups (SCID),9 anxiety (Generalized Anxiety Disorder 7),17 and health-related quality of life (36-Item Short Form Survey)18 at 6 months, 12 months, and 18 months. For adverse events, we recorded deaths from whatever cause and all self-harm and suicide attempts. The independent Data Management Committee reviewed all adverse events and made relevant trial conduct recommendations.

nd health-related quality of life (36-Item Short Form Survey)18 at 6 months, 12 months, and 18 months. For adverse events, we recorded deaths from whatever cause and all self-harm and suicide attempts. The independent Data Management Committee reviewed all adverse events and made relevant trial conduct recommendations. Statistical analysis Previous research has suggested that non-inferiority margins should be half of the mean controlled effect size from historical trials.19 Accordingly, we estimated the non-inferiority margin for the primary outcome using meta-analysis data from trials of BA14 for which BA was superior to controls by a mean of 0·7 SD units (95% CI 0·39–1) or 3·8 PHQ-9 score units (2·1–5·4). Therefore, our non-inferiority margin was 1·9 PHQ-9 points (ie, 0·5 × 3·8). We inflated our sample size by 20% for participant follow-up attrition. We planned to recruit 220 participants per arm to detect a between-group non-inferiority margin of 1·9 PHQ-9 points with a one-sided 2·5% α. Furthermore, although findings from a trial20 of CBT have shown little effect of outcome clustering by therapists, the presence of a small therapist clustering effect (ie, an intracluster correlation coefficient of 0·01) would still provide the same power.

inferiority margin of 1·9 PHQ-9 points with a one-sided 2·5% α. Furthermore, although findings from a trial20 of CBT have shown little effect of outcome clustering by therapists, the presence of a small therapist clustering effect (ie, an intracluster correlation coefficient of 0·01) would still provide the same power. We did all analyses using a statistical analysis plan prepared in the first 6 months of the trial, agreed with the Trial Management Group, Trial Steering Committee, and Data Management Committee. We assessed between-group equivalence of baseline characteristics and outcomes descriptively and did a descriptive analysis of baseline characteristics by recruitment method.

lan prepared in the first 6 months of the trial, agreed with the Trial Management Group, Trial Steering Committee, and Data Management Committee. We assessed between-group equivalence of baseline characteristics and outcomes descriptively and did a descriptive analysis of baseline characteristics by recruitment method. We compared observed primary and secondary outcomes between groups 12 months after randomisation using linear regression models adjusted for baseline outcome values and stratification variables. We did modified intention-to-treat (mITT) and per-protocol (PP) analyses, as security of inference depends on both PP and intention-to-treat analyses showing non-inferiority.21 PP analysis provides some protection for any theoretical increase in the risk of type I error (erroneously concluding non-inferiority). Our mITT population comprises all patients according to and included in random allocation with complete data. We defined the PP population as participants meeting the mITT definition and receiving at least eight treatment sessions (representing a minimally sufficient dose of therapy). We analysed safety in the mITT population. We did sensitivity analyses for our primary outcome and for different definitions of PP (eight, 12, 16, and 20 treatment sessions) to check security of inference of non-inferiority.

eiving at least eight treatment sessions (representing a minimally sufficient dose of therapy). We analysed safety in the mITT population. We did sensitivity analyses for our primary outcome and for different definitions of PP (eight, 12, 16, and 20 treatment sessions) to check security of inference of non-inferiority. We accepted non-inferiority of BA to CBT (in a 0·025 level test) if the lower bound of the two-sided 95% CI (equivalent to the upper bound of one-sided 97·5% CI) was within the non-inferiority margin of −1·9 PHQ-9 points. We checked for non-equivalence of the primary outcome at all follow-up points using the same approach.

ccepted non-inferiority of BA to CBT (in a 0·025 level test) if the lower bound of the two-sided 95% CI (equivalent to the upper bound of one-sided 97·5% CI) was within the non-inferiority margin of −1·9 PHQ-9 points. We checked for non-equivalence of the primary outcome at all follow-up points using the same approach. We did secondary analyses to compare groups at follow-up across 6 months, 12 months, and 18 months using hierarchical linear regression. To ease clinical interpretation, we calculated proportions of recovery (participants with PHQ-9 scores of ≤9) and response (50% reduction from baseline PHQ-9 scores). We ran sensitivity analyses to assess the likely effect of missing data using multiple imputation models. We did imputation by treatment group using chained equations to create 20 complete datasets under the assumption that data were missing at random.22 Imputation models included covariates as defined for the primary analysis model and auxiliary variables that were predictive of outcomes. After analysis, we combined the effect estimates from the imputed datasets using Rubin's rule.23 For economic analyses, we took the UK NHS and personal social services perspective consistent with the NICE reference case,24 also examining a wide societal perspective, adding productivity losses due to time off work in a sensitivity analysis. We collected participants' use of BA and CBT from clinical records, with additional resource information (eg, training, supervision, and other non-face-to-face activities) from therapists and trainers. We used the Adult Service Use Schedule to measure other health and social care services used, including psychotropic medications. We measured productivity losses using the absenteeism and presenteeism questions from the Health and Work Performance Questionnaire.25 We calculated effectiveness in terms of quality-adjusted life-years (QALYs) using the EuroQol-5D-3L measure of health-related quality of life.26 We assigned health states from the EuroQol-5D-3L measure a utility score using responses from a representative sample of adults in the UK.27 We calculated QALYs as the area under the curve defined by the utility values at baseline and each follow-up, assuming that utility score changes over time followed a linear path.

We assigned health states from the EuroQol-5D-3L measure a utility score using responses from a representative sample of adults in the UK.27 We calculated QALYs as the area under the curve defined by the utility values at baseline and each follow-up, assuming that utility score changes over time followed a linear path. We compared the costs and cost-effectiveness of treatments at 18 months to capture the economic effect of events like relapse with unit costs from the 2013–14 financial year.28, 29 We discounted costs and QALYs in year 2 at 3·5%.24 We used complete case analysis with missing data explored in a sensitivity analysis using multiple imputation with chained equations. We calculated the cost of each treatment using a microcosting (bottom-up) approach.30 We based MHW costs on NHS Agenda for Change salary band five (salary range £21 909–28 462; US$31 662–41 130; €27 726–35 993) for BA and band seven (£31 383–41 373; US$45 350–59 786; €39 738–52 388) for CBT therapists and included employer National Insurance and pension contributions plus capital, administrative, and managerial costs. We calculated cost per h using standard working time assumptions,31 weighted to account for time spent on non-patient-facing activities. We applied nationally applicable unit costs for other health and social care services.

National Insurance and pension contributions plus capital, administrative, and managerial costs. We calculated cost per h using standard working time assumptions,31 weighted to account for time spent on non-patient-facing activities. We applied nationally applicable unit costs for other health and social care services. We assessed cost-effectiveness in terms of QALYs using the net benefit approach.32 We analysed differences in mean cost per participant at 18 months using parametric t tests, with the validity of results confirmed using bias-corrected, non-parametric bootstrapping.33 We calculated incremental cost-effectiveness ratios and constructed cost-effectiveness planes using 1000 bootstrapped resamples from regression models of total cost and outcome by treatment group. We used these bootstrapped replications to calculate the probability that each of the treatments is the optimal choice for different values a decision maker is willing to pay for a unit improvement in outcome, representing uncertainty around the cost and effectiveness estimates, with cost-effectiveness acceptability curves illustrating the probability that BA is cost-effective compared with CBT, dependent on willingness to pay per QALY.34 We controlled for stratification variables and baseline values of the variables of interest, truncating data to exclude influential outliers—ie, cases with total costs in the 99th percentile that make a significant difference to the results. We did all analyses using Stata version 14.1. This trial is registered with the ISCRTN registry, number ISRCTN27473954.

s and baseline values of the variables of interest, truncating data to exclude influential outliers—ie, cases with total costs in the 99th percentile that make a significant difference to the results. We did all analyses using Stata version 14.1. This trial is registered with the ISCRTN registry, number ISRCTN27473954. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. DAR, RST, FCW, SB, SR, and BB had full access to all the data in the study and DAR had final responsibility for the decision to submit for publication. Results Between Sept 26, 2012, and April 3, 2014, we recruited 440 participants, randomly allocating 221 (50%) to the BA group and 219 (50%) to the CBT group (figure 1). 175 (79%) participants were assessable for the primary outcome in the mITT population in the BA group compared with 189 (86%) in the CBT group, whereas 135 (61%) were assessable in the PP population in the BA group compared with 151 (69%) in the CBT group. We noted no evidence of a difference in patient characteristics between recruitment methods (appendix). Patient-level and trial-level characteristics at baseline were well balanced between groups (table 1). PHQ-9 score at baseline was negatively skewed, with a high proportion of participants scoring towards the upper end of the distribution (data not shown), but scores were similar between groups (table 2).

s (appendix). Patient-level and trial-level characteristics at baseline were well balanced between groups (table 1). PHQ-9 score at baseline was negatively skewed, with a high proportion of participants scoring towards the upper end of the distribution (data not shown), but scores were similar between groups (table 2). Ten MHWs provided BA (median 22 participants each [IQR 19–25]) and 12 therapists provided CBT (21 [13–23]). MHWs had a mean of 18 months mental health experience (SD 11) and CBT therapists had a mean of 22 months post-CBT qualification (24). We removed one CBT therapist from the trial in the early stages who did not meet acceptable competency. Participants received a mean of 11·5 BA sessions (7·8) or 12·5 CBT sessions (7·8). 305 participants (69%) completed the PP number of at least eight sessions (BA 147 [67%] patients, mean 16·1 sessions [SD 5·3]; CBT 158 [72%] patients, mean 16·4 sessions [5·4]); participants completing less than eight sessions (135 [31%]; BA 74 [33%], CBT 61 [28%]) completed a mean of 2·5 BA sessions (SD 1·9) or 2·6 CBT sessions (2·1). MHWs and therapists met acceptable competency standards: mean Quality of Behavioral Activation Scale BA competence was 55 (7·5) and mean Revised Cognitive Therapy Scale for CBT competence was 37·9 (10·9).

ions (135 [31%]; BA 74 [33%], CBT 61 [28%]) completed a mean of 2·5 BA sessions (SD 1·9) or 2·6 CBT sessions (2·1). MHWs and therapists met acceptable competency standards: mean Quality of Behavioral Activation Scale BA competence was 55 (7·5) and mean Revised Cognitive Therapy Scale for CBT competence was 37·9 (10·9). We found no evidence of inferiority of PHQ-9 score at 12 months in either the mITT (CBT 8·4 PHQ-9 points [SD 7·5]; BA 8·4 PHQ-9 points [7·0]; mean difference 0·1 PHQ-9 points [95% CI −1·3 to 1·5]; p=0·89) or PP (CBT 7·9 PHQ-9 points [7·3]; BA 7·8 [6·5]; mean difference 0·0 [–1·5 to 1·6]; p=0·99) populations (table 2). The non-inferiority of BA to CBT was accepted for both the mITT and PP populations as the lower bound of the 95% CI (one-sided 97·5% CI) of the between-group mean difference lies within the non-inferiority margin of −1·9 PHQ-9 points (appendix). Although we initially planned to include therapist as a random-effects variable, given the low levels of observed clustering, we parsimoniously fitted our models without therapist as a variable. We checked for no inference difference with and without inclusion of a random-effects therapist term. We ruled out superiority of CBT to BA as the lower bound of the 95% CI included zero for the mITT and PP populations. The inference of non-inferiority was robust to sensitivity analysis across different PP definitions. We found no evidence of a significant between-group treatment interaction across the mITT or PP populations with the primary outcome at 12 months as stratified by depression severity, antidepressants use, and recruitment site (appendix).

on-inferiority was robust to sensitivity analysis across different PP definitions. We found no evidence of a significant between-group treatment interaction across the mITT or PP populations with the primary outcome at 12 months as stratified by depression severity, antidepressants use, and recruitment site (appendix). We found that BA was not different from CBT in anxiety (Generalized Anxiety Disorder 7), depression status, and depression-free days and anxiety diagnoses (SCID) for either the mITT or PP populations using observed or imputed data at 12 months (table 2). Because of substantial missing 36-Item Short Form Survey data at baseline, we analysed these data adjusted for stratification variables only. We found no difference in numbers of participants with at least one anxiety diagnosis: BA 43 (28%) of 153; CBT 43 (27%) of 161 (mITT population; χ2 0·08; p=0·78). Between 61% and 70% of mITT and PP participants in both groups met criteria for recovery from depression or response to treatment at 12 months, with no differences in the proportions of patients in each group who recovered or responded (table 3). Using observed data for all outcomes, we found no evidence of a difference between the CBT and BA groups over the period of the trial, as indicated by a non-significant time-by-treatment effect interaction, for both the mITT and PP populations (appendix). We found a small, negligible clustering of primary and secondary outcome scores at follow-up across therapists overall and within BA and CBT groups (intracluster correlation coefficient ≤0·04).

rial, as indicated by a non-significant time-by-treatment effect interaction, for both the mITT and PP populations (appendix). We found a small, negligible clustering of primary and secondary outcome scores at follow-up across therapists overall and within BA and CBT groups (intracluster correlation coefficient ≤0·04). Two (1%) non-trial-related deaths (one [1%] multidrug toxicity in the BA group and one [1%] cancer in the CBT group) and 15 depression-related, but not treatment-related, serious adverse events (three in the BA group and 12 in the CBT group) occurred in three [2%] participants in the BA group (two [1%] patients who overdosed and one [1%] who self-harmed) and eight (4%) participants in the CBT group (seven [4%] who overdosed and one [1%] who self-harmed). Of the 440 participants recruited, 76 (17%) had missing primary outcome data at 12 month follow-up. The proportion of missing PHQ-9 data was higher in the BA than in the CBT group (46 [21%] vs 30 [14%]; odds ratio 1·6 [95% CI 1·0–2·7]; p=0·05). Imputation of data for primary and secondary outcomes at 12 months showed that in accordance with the observed data analysis, no difference existed between groups (Table 2, Table 3), supporting our conclusion of non-inferiority. The odds of missing PHQ-9 data were higher for patients with increased baseline severity of depression (PHQ ≥19, odds ratio 1·6 [95% CI 1·0–2·6]; p=0·05) and increasing age (in years) was associated with lower odds of missing PHQ-9 data (odds ratio 0·97 [0·96–0·99]; p=0·01). We found no evidence of an association between missingness and any other baseline characteristic (data not shown). Outcome assessors reported having been unmasked for 16 (4%) participants (five [2%] in the BA group and 11 [5%] in the CBT group; due to participants informing assessors of their treatment allocation).

1). We found no evidence of an association between missingness and any other baseline characteristic (data not shown). Outcome assessors reported having been unmasked for 16 (4%) participants (five [2%] in the BA group and 11 [5%] in the CBT group; due to participants informing assessors of their treatment allocation). For economic analyses, at 18 months, full service use data was available for 159 (90%) of 176 participants in the BA group and 168 (93%) of 180 participants in the CBT group. We found a significant difference in mean intervention costs between the two groups, but no differences in other categories of cost or in total cost (table 4). Mean health state utility scores according to EuroQoL-5D-3L were slightly higher in the BA group than in the CBT group across the entire follow-up period, with resultant QALYs also higher for BA, but the QALY difference was not significant. Costs were lower and QALY outcomes better in the BA group than in the CBT group, generating an incremental cost-effectiveness ratio of –£6865. The scatterplot of bootstrapped cost and effectiveness pairs for BA versus CBT illustrates dominance of BA over CBT, with the point estimate and two-thirds of scatter points falling in the southeast quadrant of the cost-effectiveness plane, where BA replications are cheaper and more effective than are CBT ones (figure 2). The cost-effectiveness acceptability curve (appendix) showing the probability of BA being cost-effective compared with CBT does not fall below 75% and is closer to 80% at NICE-preferred willingness24 to pay £20 000–30 000 per QALY.

lane, where BA replications are cheaper and more effective than are CBT ones (figure 2). The cost-effectiveness acceptability curve (appendix) showing the probability of BA being cost-effective compared with CBT does not fall below 75% and is closer to 80% at NICE-preferred willingness24 to pay £20 000–30 000 per QALY. In all sensitivity analyses, including complementary therapies and productivity losses, as well as analyses taking narrow intervention and mental health service perspectives, BA was significantly less costly than was CBT, so BA continues to have a higher probability of being cost-effective than does CBT at the NICE threshold (appendix).24 Imputation of missing data increased the difference in total cost (BA £1841·67; CBT £2282·40; difference –£440·73 [95% CI −1007·71 to 126·26]; p=0·13), but reduced the difference in QALYs (BA 1·22; CBT 1·19; difference 0·03 [–0·06 to 0·11]; p=0·55), increasing the incremental cost-effectiveness ratio to –£16 951. The cost-effectiveness acceptability curve for the missing data analysis again supported the likelihood that BA is cost-effective compared with CBT (appendix).

t reduced the difference in QALYs (BA 1·22; CBT 1·19; difference 0·03 [–0·06 to 0·11]; p=0·55), increasing the incremental cost-effectiveness ratio to –£16 951. The cost-effectiveness acceptability curve for the missing data analysis again supported the likelihood that BA is cost-effective compared with CBT (appendix). Discussion We found that BA for depression is not inferior to CBT in terms of reduction of depression symptoms and is more cost-effective than is CBT against commonly applied decision maker willingness to pay thresholds. We observed our results using both mITT and PP analyses, using a conservative non-inferiority margin. Our economic analyses were driven by the lower costs of the MHWs who delivered BA compared with the more experienced psychological therapists who routinely deliver CBT. Our study results therefore substantiate the hypothesis that BA is as effective as is CBT and that its simplicity renders BA suitable for delivery by junior MHWs with no professional training in psychological therapies.5

delivered BA compared with the more experienced psychological therapists who routinely deliver CBT. Our study results therefore substantiate the hypothesis that BA is as effective as is CBT and that its simplicity renders BA suitable for delivery by junior MHWs with no professional training in psychological therapies.5 This trial is the largest trial of BA to date and is one of the largest trials of psychological treatments for depression. We followed up participants for 18 months and our economic analysis is one of few in this field. Therapists and MHWs working in three different routine UK care settings delivered treatment, providing evidence of potential generalisability. We assessed therapy quality using independent raters and ensured that treatment in both arms was delivered to the standard recommended guidelines. Our levels of attrition and outcome loss to follow-up were low at 12 months and 18 months, similar to other trials in this area, but are still a limitation. Although participants in the per-protocol population attended similar numbers of sessions to those in other CBT trials,15 35% of participants chose to not even attend a minimal number of sessions, a problem well known to routine psychological therapy services. This pragmatic trial done in routine environments means that we were unable to quantify or control for the contribution of antidepressant medicines to outcomes. However, most participants who were taking medication had been doing so for a considerable time before entering the trial, making it unlikely that our results were driven by pharmacological treatment. Given the nature of the intervention and comparator, we could not mask patients or the mental health workers or therapists who were delivering the interventions to treatment allocation, but we used self-reported outcome measures and robust outcome assessor-masking procedures to reduce researcher unmasking to less than 5%. Missing data for the primary outcome measure was substantial. However, our between-group inferences were robust to data imputation.

ivering the interventions to treatment allocation, but we used self-reported outcome measures and robust outcome assessor-masking procedures to reduce researcher unmasking to less than 5%. Missing data for the primary outcome measure was substantial. However, our between-group inferences were robust to data imputation. Our findings could have substantial implications for the scalability of psychological treatment for depression internationally4 given the greater availability and ease with which a BA workforce could be trained than could a CBT workforce. For many years, CBT has been the foremost psychological therapy recommended by therapists, researchers, and policy makers. Our results challenge this dominance. Although more work needs to be done than has been done so far to find ways to effectively treat the 20–23% of patients whose depression was unchanged by BA or CBT, our findings suggest that BA should be a front-line treatment for depression, with substantial potential to improve reach and access to psychological therapy globally.

re work needs to be done than has been done so far to find ways to effectively treat the 20–23% of patients whose depression was unchanged by BA or CBT, our findings suggest that BA should be a front-line treatment for depression, with substantial potential to improve reach and access to psychological therapy globally. Our results in both groups compare favourably with a meta-analysis3 of the effects of CBT that estimate proportions of patients with remissions of around 50%. Our cost-effectiveness analyses show the high probability that BA is cost-effective and affordable compared with CBT at standard willingness to pay thresholds. Our most striking finding is that BA leads to similar clinical outcomes for patients with depression, but at a financial saving to clinical providers of 21% compared with the costs of provision of CBT, with no compensatory use of other health-care services by patients.

ith CBT at standard willingness to pay thresholds. Our most striking finding is that BA leads to similar clinical outcomes for patients with depression, but at a financial saving to clinical providers of 21% compared with the costs of provision of CBT, with no compensatory use of other health-care services by patients. Driving these savings is the fact that BA can be delivered by inexperienced MHWs with no professional training in psychological therapies, with no lesser effect than that of more highly trained and experienced psychological therapists giving patients CBT. Although many obstacles exist to successful dissemination in addition to training of MHWs, our findings suggest that health services globally could reduce the need for costly professional training and infrastructure, reduce waiting times, and increase access to psychological therapies.4 Our findings have substantial implications given the increasing global pressure for cost containment across health systems in high-income countries and the need to develop accessible, scalable interventions in low-income and middle-income countries. Such countries might choose to investigate the training and employment of junior workers over expensive groups of psychological professionals. Our results, therefore, offer hope to many societies, cultures, and communities worldwide, rich and poor, struggling with the effect of depression on the health of their people and economies. Supplementary Material Supplementary appendix

Driving these savings is the fact that BA can be delivered by inexperienced MHWs with no professional training in psychological therapies, with no lesser effect than that of more highly trained and experienced psychological therapists giving patients CBT. Although many obstacles exist to successful dissemination in addition to training of MHWs, our findings suggest that health services globally could reduce the need for costly professional training and infrastructure, reduce waiting times, and increase access to psychological therapies.4 Our findings have substantial implications given the increasing global pressure for cost containment across health systems in high-income countries and the need to develop accessible, scalable interventions in low-income and middle-income countries. Such countries might choose to investigate the training and employment of junior workers over expensive groups of psychological professionals. Our results, therefore, offer hope to many societies, cultures, and communities worldwide, rich and poor, struggling with the effect of depression on the health of their people and economies. Supplementary Material Supplementary appendix Acknowledgments This report is independent research funded by the UK National Institute for Health Research (NIHR) Health Technology Assessment Programme. The views expressed in this publication are those of the authors and not necessarily of the NIHR or UK Department of Health. We would like to thank all participants, National Health Service services, mental health workers, therapists, and general practitioners involved in the study and acknowledge the vital contributions of study researchers and administrators in Devon, Durham, and Leeds, the Peninsula Clinical Trials Unit, and the NIHR Clinical Research Network.

all participants, National Health Service services, mental health workers, therapists, and general practitioners involved in the study and acknowledge the vital contributions of study researchers and administrators in Devon, Durham, and Leeds, the Peninsula Clinical Trials Unit, and the NIHR Clinical Research Network. Contributors DAR, DE, DM, RST, SB, PAF, SG, WK, HO'M, ERW, and KAW designed the study and were responsible for its conduct. SR, EF, and KF were responsible for study and data collection management. RST, SB, FCW, and BB did data analysis. SDH and NR provided expert advice on clinical and patient and public involvement. All authors contributed to writing and editing of the manuscript. Declaration of interests All authors report grants from the National Institute for Health Research (NIHR) during the course of the study. DAR reports grants from the European Science Foundation. DAR and RST have received funding support from NIHR Collaborations for Leadership in Applied Health Research and Care and report NIHR panel memberships. WK reports fees from book royalties. Figure 1 Trial profile (A) 6 month, (B) 12 month, and (C) 18 month follow-up. BA=behavioural activation. CBT=cognitive behavioural therapy. *Includes four participants who were initially allocated in error and subsequently excluded. Figure 2 Bootstrapped mean differences in costs and effects of BA compared with CBT BA=behavioural activation. CBT=cognitive behavioural therapy. NE=northeast. NW=northwest. SE=southeast. SW=southwest. QALY=quality-adjusted life-year. Table 1 Baseline characteristics

(A) 6 month, (B) 12 month, and (C) 18 month follow-up. BA=behavioural activation. CBT=cognitive behavioural therapy. *Includes four participants who were initially allocated in error and subsequently excluded. Figure 2 Bootstrapped mean differences in costs and effects of BA compared with CBT BA=behavioural activation. CBT=cognitive behavioural therapy. NE=northeast. NW=northwest. SE=southeast. SW=southwest. QALY=quality-adjusted life-year. Table 1 Baseline characteristics BA (n=221) CBT (n= 219) All (n=440) Trial characteristics Method of recruitment Primary care 192 (87%) 190 (87%) 382 (87%) IAPT 29 (13%) 29 (13%) 58 (13%) Patient characteristics Age (years) 43.9 (14·1) 43·0 (14·1) 43·5 (14·1) Sex Male 79 (36%) 71 (32%) 150 (34%) Female 142 (64%) 148 (68%) 290 (66%) Number of episodes of depression (including current) Mean 7·0 (15·0) 6·3 (13·8) 6·7 (14·4) Median 3·0 (1–5) 2·0 (1–5) 3·0 (1–5) Age of onset of first depression episode (years) 27·2 (15·0) 26·3 (13·5) 26·7 (14·2) Duration of antidepressant treatment (weeks)* Mean; n 215 (817); 160 116 (480); 169 164 (666); 329 Median; n 21 (10–78); 160 18 (7–52); 169 19 (8–71); 329 At least one comorbid anxiety disorder 131 (59%) 141 (64%) 272 (62%) Marital status Single 68 (31%) 59 (27%) 127 (29%) Cohabiting (not married) 29 (13%) 25 (11%) 54 (12%) Civil partnership 1 (<1%) 1 (<1%) 2 (<1%) Married 84 (38%) 92 (42%) 176 (40%) Divorced or separated 39 (18%) 42 (19%) 81 (18%) Number of children 0 74 (33%) 72 (33%) 146 (33%) 1 35 (16%) 31 (14%) 66 (15%) 2 67 (30%) 69 (32%) 136 (31%) 3 31 (14%) 27 (12%) 58 (13%) ≥4 14 (6%) 20 (9%) 34 (8%) Level of education No qualifications 25 (11%) 30 (14%) 55 (13%) GCSEs or O Levels 36 (16%) 43 (20%) 79 (18%) AS or A Levels 28 (13%) 22 (10%) 50 (11%) NVQ or other vocational qualification 54 (24%) 71 (32%) 125 (28%) Undergraduate degree 44 (20%) 35 (16%) 79 (18%) Postgraduate degree 28 (13%) 14 (6%) 42 (10%) Doctoral degree 2 (1%) 1 (<1%) 3 (1%) Professional degree (eg, MD) 4 (2%) 3 (1%) 7 (2%) Ethnicity White British 204 (92%) 197 (90%) 402 (91%) Other 17 (8%) 22 (10%) 38 (9%) Stratification or minimisation variables PHQ-9 category <19 118 (53%) 118 (54%) 236 (54%) ≥19 103 (47%) 101 (46%) 204 (46%) Antidepressant use Yes 172 (78%) 173 (79%) 345 (78%) No 49 (22%) 46 (21%) 95 (22%) Site Devon 74 (33%) 73 (33%) 147 (33%) Durham 79 (36%) 78 (36%) 157 (36%) Leeds 68 (31%) 68 (31%) 136 (31%) Data are n (%), mean (SD), or median (IQR), unless otherwise indicated. IAPT=Improving Access to Psychological Therapies. GCSE=General Certificate of Secondary Education. O Level=Ordinary Level. AS Level=Advanced Subsidiary Level. A Level=Advanced Level.

ham 79 (36%) 78 (36%) 157 (36%) Leeds 68 (31%) 68 (31%) 136 (31%) Data are n (%), mean (SD), or median (IQR), unless otherwise indicated. IAPT=Improving Access to Psychological Therapies. GCSE=General Certificate of Secondary Education. O Level=Ordinary Level. AS Level=Advanced Subsidiary Level. A Level=Advanced Level. NVQ=National Vocational Qualification. MD=Doctor of Medicine. PHQ-9=Patient Health Questionnaire 9. * 16 participants who reported that they were using antidepressant medication at baseline did not report duration of use (12 in the BA group and four in the CBT group). Table 2 Primary and secondary outcomes at 12 months

ham 79 (36%) 78 (36%) 157 (36%) Leeds 68 (31%) 68 (31%) 136 (31%) Data are n (%), mean (SD), or median (IQR), unless otherwise indicated. IAPT=Improving Access to Psychological Therapies. GCSE=General Certificate of Secondary Education. O Level=Ordinary Level. AS Level=Advanced Subsidiary Level. A Level=Advanced Level. NVQ=National Vocational Qualification. MD=Doctor of Medicine. PHQ-9=Patient Health Questionnaire 9. * 16 participants who reported that they were using antidepressant medication at baseline did not report duration of use (12 in the BA group and four in the CBT group). Table 2 Primary and secondary outcomes at 12 months CBT BA Observed data only Observed and imputed data Between-group difference p value Between-group difference p value Primary outcome PHQ-9 Baseline 17·4 (4·8); 219 17·7 (4·8); 221 .. .. mITT 8·4 (7·5); 189 8·4 (7·0); 175 0·1 (−1·3 to 1·5)* 0·89 0·2 (−1·1 to 1·7)* 0·80 PP 7·9 (7·3); 151 7·8 (6·5); 135 0·0 (−1·5 to 1·6)* 0·99 0·0 (−1·6 to 1·6)* 0·99 Secondary outcomes GAD-7 Baseline 12·6 (5·1); 219 12·7 (5·1); 221 .. .. mITT 6·3 (6·0); 176 6·4 (5·9); 161 −0·1 (−1·0 to 1·3)* 0·82 0·0 (−1·3 to 1·4)* 0·96 PP 6·0 (5·8); 146 5·9 (5·5); 129 0·01 (−1·3 to 1·2)* 0·95 −0·4 (−1·7 to 1·0)* 0·60 SCID number of depression-free days Baseline .. .. .. .. mITT 129 (58); 160 120 (56); 150 9 (−3 to 23)* 0·13 7 (−7 to 20)* 0·27 PP 132 (55); 138 119 (55); 125 13 (0 to 26)* 0·06 8 (−4 to 21)* 0·21 SF-36v2 PCS Baseline 50·1 (13·1); 65 51·4 (11·9); 69 .. .. mITT 48·1 (12·2); 168 49·9 (11·6); 150 1·6 (−1·0 to 4·2)† 0·22 1·4 (−1·1 to 4·0)† 0·27 PP 48·0 (12·2); 144 49·9 (12·0); 125 1·6 (−1·3 to 4·4)† 0·28 1·3 (−1·5 to 4·1)† 0·36 SF-36v2 MCS Baseline 23·2 (9·4); 65 22·5 (7·8); 69 .. .. mITT 41·7 (14·1); 168 41·6 (14·0); 150 0·0 (−3·0 to 3·0)† 0·99 0·0 (−2·9 to 2·8)† 0·97 PP 42·9 (13·6); 144 42·3 (13·3); 125 −0·5 (−3·7 to 2·7)† 0·77 −0·6 (−3·8 to 2·7)† 0·73 Data are mean (SD); n or mean (95% CI). CBT=cognitive behavioural therapy. BA=behavioural activation. PHQ-9=Patient Health Questionnaire 9. mITT=modified intention to treat. PP=per protocol. GAD-7=Generalized Anxiety Disorder 7. SCID=Structured Clinical Interview for the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition. SF-36v2=36-Item Short Form Survey version 2. PCS=physical component summary. MCS=mental component summary.

onnaire 9. mITT=modified intention to treat. PP=per protocol. GAD-7=Generalized Anxiety Disorder 7. SCID=Structured Clinical Interview for the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition. SF-36v2=36-Item Short Form Survey version 2. PCS=physical component summary. MCS=mental component summary. * Models adjusted for baseline outcome score and stratification variables (symptom severity [PHQ-9 score of <19 vs ≥19], site [Devon, Durham, or Leeds], and antidepressant use [use or not]). † Models adjusted for stratification variables, but not baseline outcome score because of substantial missing data. Table 3 Depression status, recovery, and response at 12 months CBT BA Observed data only Observed and imputed data Odds ratio p value Odds ratio p value SCID depression Baseline 219/219 (100%) 221/221 (100%) mITT 37/163 (23%) 31/154 (20%) 0·9 (0·5–1·6) 0·71 0·9 (0·5–1·6) 0·70 PP 30/141 (21%) 24/128 (19%) 0·9 (0·5–1·7) 0·80 0·9 (0·5–1·7) 0·75 Depression recovery* mITT 124/189 (66%) 115/175 (66%) 1·0 (0·6–1·5) 0·96 1·2 (0·7–1·9) 0·53 PP 104/151 (69%) 94/135 (70%) 1·0 (0·6–1·7) 0·96 1·2 (0·7–2·0) 0·47 Depression response† mITT 117/189 (62%) 107/175 (61%) 1·0 (0·9–1·1) 0·73 0·9 (0·6–1·4) 0·75 PP 100/151 (66%) 87/135 (64%) 0·9 (0·9–1·0) 0·64 0·9 (0·5–1·4) 0·55 Data are n/N (%) or odds ratio (95% CI). CBT=cognitive behavioural therapy. BA=behavioural activation. SCID=Structured Clinical Interview for the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition. mITT=modified intention to treat. PP=per protocol.

) 87/135 (64%) 0·9 (0·9–1·0) 0·64 0·9 (0·5–1·4) 0·55 Data are n/N (%) or odds ratio (95% CI). CBT=cognitive behavioural therapy. BA=behavioural activation. SCID=Structured Clinical Interview for the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition. mITT=modified intention to treat. PP=per protocol. * Participants with Patient Health Questionnaire 9 scores of 9 or less. † Participants with a 50% reduction from baseline in Patient Health Questionnaire 9 score. Table 4 Economic data at 18 months BA CBT Difference p value Costs per participant (£) Intervention £974·81 (475·02); 159 £1235·23 (610·03); 168 −£262·29 (−381·40 to −143·19) <0·0001 Hospital £860·23 (1509·88); 159 £927·26 (1975·64); 168 −£75·67 (−451·75 to 300·42) 0·69 Community health and social care £644·36 (816·07); 159 £944·25 (1726·17); 168 −£15·14 (−304·90 to 274·62) 0·91 Medication £103·20 (197·92); 159 £117·64 (265·92); 168 £2·15 (−39·83 to 44·13) 0·92 Total £2596·62 (1846·72); 159 £3250·74 (3040·99); 168 −£343·24 (−857·62 to 171·13) 0·19 EQ-5D-3L utility score Baseline 0·548 (0·307); 159 0·474 (0·317); 168 .. .. 6 months 0·683 (0·310); 153 0·677 (0·310); 151 .. .. 12 months 0·684 (0·341); 147 0·671 (0·348); 156 .. .. 18 months 0·670 (0·311); 152 0·624 (0·335); 157 .. .. QALYs 0·985 (0·422); 152 0·935 (0·433); 157 0·050 (−0·046 to 0·145) 0·31 Data are mean (SD); n or mean difference (95% CI). BA=behavioural activation. CBT=cognitive behavioural therapy. EQ=EuroQol. QALY=quality-adjusted life-year.

Introduction At the end of the 20th century, concern for high rates of teenage conception in the UK compared with other western European countries, together with the strong association between early parenthood and deprivation, provided the impetus to public health efforts to prevent teenage pregnancy. In 1999, the UK Government launched a 10-year, nationwide Teenage Pregnancy Strategy1 in England with the dual aims of achieving a 50% reduction in conception rates in women younger than 18 years by 2010, and mitigating social exclusion in teenage parents by increasing their participation in education, employment, or training.2 A strong rationale for the strategy was the desire to halt the cycle of deprivation resulting from the increment of disadvantage conferred by early pregnancy additional to that experienced before conception.3, 4, 5, 6 The teenage pregnancy strategy gave rise to a multicomponent programme comprising a national media campaign; improvements to sex education and young people's sexual health services; support for young parents to increase participation in education, training, and employment; and joint action to ensure national and local coordination across statutory and voluntary agencies.7 Financial resources were allocated according to teenage pregnancy rates in each top-tier local authority area. Because of the concentration of high rates of teenage pregnancy in poorer areas of the country, financial resources were disproportionately invested in the more deprived areas of England. Scotland, Wales, and Northern Ireland decided their initiatives independently.

regnancy rates in each top-tier local authority area. Because of the concentration of high rates of teenage pregnancy in poorer areas of the country, financial resources were disproportionately invested in the more deprived areas of England. Scotland, Wales, and Northern Ireland decided their initiatives independently. Research in context Evidence before this study We searched PubMed for studies published between Jan 1, 2000, and March 31, 2016, with no language restrictions using the search terms “teenage pregnancy”, “teenage conception”, “trends”, and “public health policy” to identify studies describing trends in teenage pregnancy and policy-related factors associated with change, in the UK and elsewhere. We examined the titles and abstracts identified by the search, excluding those that were not relevant, and reviewed the full text of the remaining articles to assess appropriateness for inclusion.

studies describing trends in teenage pregnancy and policy-related factors associated with change, in the UK and elsewhere. We examined the titles and abstracts identified by the search, excluding those that were not relevant, and reviewed the full text of the remaining articles to assess appropriateness for inclusion. Few studies were able to measure direct indicators of intervention-related action, and none combined general population survey data and routinely collected statistics. Most studies were from the USA and the UK, settings in which the prevalence of teenage pregnancy is high compared with other high-income countries. Existing studies have near universally reported strong associations between higher rates of teenage pregnancy and lower socioeconomic level. They have, however, differed in their interpretation of whether the decline in the prevalence of teenage pregnancy in the 21st century can be attributed to broader societal changes, such as increased educational attainment, or to public health interventions such as increased use of reliable contraception. Added value of this study

Few studies were able to measure direct indicators of intervention-related action, and none combined general population survey data and routinely collected statistics. Most studies were from the USA and the UK, settings in which the prevalence of teenage pregnancy is high compared with other high-income countries. Existing studies have near universally reported strong associations between higher rates of teenage pregnancy and lower socioeconomic level. They have, however, differed in their interpretation of whether the decline in the prevalence of teenage pregnancy in the 21st century can be attributed to broader societal changes, such as increased educational attainment, or to public health interventions such as increased use of reliable contraception. Added value of this study Our study aimed to assess progress towards the goals of the Teenage Pregnancy Strategy mounted in England to reduce conception rates in women younger than 18 years, combining analyses of data from a population-based survey and from routinely collected national statistics. The scale of the decline in conception rates in women younger than 18 years, and its association with intervention-related investment and with both demographic and behavioural factors, suggest a combined influence of both public health intervention and secular trends on the decline in conceptions in women younger than 18 years in England. Implications of all the available evidence

Our study aimed to assess progress towards the goals of the Teenage Pregnancy Strategy mounted in England to reduce conception rates in women younger than 18 years, combining analyses of data from a population-based survey and from routinely collected national statistics. The scale of the decline in conception rates in women younger than 18 years, and its association with intervention-related investment and with both demographic and behavioural factors, suggest a combined influence of both public health intervention and secular trends on the decline in conceptions in women younger than 18 years in England. Implications of all the available evidence The trend towards postponement of key life events such as completion of education, leaving home, starting employment, and settling with a partner is now near universal. These factors seem to have contributed to a global trend towards fewer early pregnancies. Coincidental with this, focused and sustained efforts to lower the prevalence of early pregnancy by raising awareness, changing social norms, and increasing access to education and reliable contraception seem to have accelerated the trend towards fewer teenage pregnancies.

global trend towards fewer early pregnancies. Coincidental with this, focused and sustained efforts to lower the prevalence of early pregnancy by raising awareness, changing social norms, and increasing access to education and reliable contraception seem to have accelerated the trend towards fewer teenage pregnancies. Conception rates in women younger than 18 years fell by 51% between 1998 and 2014 in England. Of interest is the extent to which the decline seems to be related to the Teenage Pregnancy Strategy and other interventions; whether the decline has been seen equally in the most and least disadvantaged women; and what factors remain associated with conception in women younger than 18 years. In this study, we combined analyses of routinely available area-level data for conceptions, deprivation, and policy-related expenditure, and individual-level data from three decennial waves of the National Survey of Sexual Attitudes and Lifestyles (Natsal) to describe change in outcomes relating to key goals of the English Teenage Pregnancy Strategy—ie, conception rates in women younger than 18 years and the prevalence of participation in education, work, and training in women with a child conceived before the age of 18 years.

of Sexual Attitudes and Lifestyles (Natsal) to describe change in outcomes relating to key goals of the English Teenage Pregnancy Strategy—ie, conception rates in women younger than 18 years and the prevalence of participation in education, work, and training in women with a child conceived before the age of 18 years. Methods Study design and participants In this observational study, we analysed routine data for births and abortions for top-tier local authorities (top-tier local authorities consist of London boroughs, metropolitan borough councils, and unitary authorities, and are a level of local government with responsibilities including education and social services) in England from 1994 to 2013 and individual-level data from three waves of the National Survey of Sexual Attitudes and Lifestyles (wave 1: 1990–91 [Natsal-1], wave 2: 1999–2001 [Natsal-2], and wave 3: 2010–12 [Natsal-3]). We obtained data by calendar year for the resident population and births and abortions by age of mother from the Office for National Statistics for each of the top-tier local authorities in England. Because of their small resident populations, two authorities (City of London and the Scilly Isles) were combined with more populous neighbouring authorities (Hackney and Cornwall, respectively), resulting in a total of 148 local authorities for analysis.

Statistics for each of the top-tier local authorities in England. Because of their small resident populations, two authorities (City of London and the Scilly Isles) were combined with more populous neighbouring authorities (Hackney and Cornwall, respectively), resulting in a total of 148 local authorities for analysis. We also obtained data from the Department for Communities and Local Government at the top-tier local authority level for the 2004 Index of Multiple Deprivation, a score of socioeconomic deprivation based on a weighted average of 38 separate indicators across seven distinct domains (employment, income, health and disability, education skills and training, barriers to housing and other services, crime, and living environment).8 As an indicator of the extent of Teenage Pregnancy Strategy-related local activity, we obtained from the Department for Children, Schools and Families the Teenage Pregnancy Strategy annual Local Implementation Grant awarded to each top-tier local authority for the financial year(s) 1999–2000 and 2010–11, the amount reflecting the challenge in terms of conception rates in women younger than 18 years and the size of the population. From these data we calculated the total investment per head by dividing the amount by the 2001 Office of National Statistics estimate of resident women aged 13–17 years. We obtained ethics approval for Natsal-2 from University College Hospital, North Thames Multicentre, and all local research ethics committees in Britain, and for Natsal-3 from the Oxford Research Ethics Committee A.

viding the amount by the 2001 Office of National Statistics estimate of resident women aged 13–17 years. We obtained ethics approval for Natsal-2 from University College Hospital, North Thames Multicentre, and all local research ethics committees in Britain, and for Natsal-3 from the Oxford Research Ethics Committee A. Procedures The Natsals are probability sample surveys of British residents of whom 18 876 aged 16–59 years were interviewed between May, 1990, and November, 1991 (Natsal-1); 11 161 aged 16–44 years were interviewed between May, 1999, and February, 2001 (Natsal-2); and 15 162 aged 16–74 years between September, 2010, and August, 2012 (Natsal-3). Participants resident in London were oversampled in Natsal-2, and those aged 16–34 years were oversampled in Natsal-3. The unadjusted response rate in Natsal-1 was 64·7%, and the cooperation rate (of eligible addresses contacted) was 71·5%. The unadjusted response rate for Natsal-2 was 63·1% and the adjusted rate, taking account of oversampling in London, was 65·4%. The response rate for Natsal-3 was 57·7% and the cooperation rate was 65·8%.

adjusted response rate in Natsal-1 was 64·7%, and the cooperation rate (of eligible addresses contacted) was 71·5%. The unadjusted response rate for Natsal-2 was 63·1% and the adjusted rate, taking account of oversampling in London, was 65·4%. The response rate for Natsal-3 was 57·7% and the cooperation rate was 65·8%. In Natsal-1, paper questionnaires were self-completed and interviewer administered. In Natsal-2 and Natsal-3, participants were interviewed with a combination of computer-assisted personal and computer-assisted self-interviews. Experimental comparison of pencil and paper and computer-assisted interviewing revealed no important differences in responses.9 For all three surveys, after correcting for unequal selection probabilities, a non-response post-stratification weight was applied to ensure comparability with census data in terms of age, sex, and region. Further details of the methods and response calculations are described elsewhere.10, 11 Natsal-1 and Natsal-2 asked about the number and timing of all abortions and livebirths; Natsal-3 asked for timing and outcome of each pregnancy. We calculated conceptions that occurred before age 18 years among 18–24-year-old women by summing births that occurred before age 18 years 9 months and abortions occurring before age 18 years, excluding miscarriages and stillbirths, for consistency with Office of National Statistics procedures. We assessed change in the prevalence of participation in any of education, work, or training in young mothers.

by summing births that occurred before age 18 years 9 months and abortions occurring before age 18 years, excluding miscarriages and stillbirths, for consistency with Office of National Statistics procedures. We assessed change in the prevalence of participation in any of education, work, or training in young mothers. Variables used in the analyses presented in this paper were selected on the basis of pre-existing evidence of their association with early conception.3, 4, 5, 6 They include age at first heterosexual intercourse; use of a reliable method of contraception at first sex; consensuality in terms of being equally or more willing than the partner at time of first sex; autonomy of decision making—ie, not influenced by peer pressure or use of drugs or alcohol; and retrospective views on the timing of first sex—ie, whether it occurred sooner or later than ideal or at the right time. We also included the main source of information about sexual matters at the start of sexual activity, adequacy of that information, and ease of communication with parents about sex. Demographic measures included age at interview, family structure (whether the participant lived with both parents until age 16 years [Natsal-2] or age 14 years [Natsal-3]), educational attainment, and area-level deprivation measured with the Index of Multiple Deprivation.12 Full postcode data for use in estimating the Index of Multiple Deprivation were not available for Natsal-1.

re (whether the participant lived with both parents until age 16 years [Natsal-2] or age 14 years [Natsal-3]), educational attainment, and area-level deprivation measured with the Index of Multiple Deprivation.12 Full postcode data for use in estimating the Index of Multiple Deprivation were not available for Natsal-1. Statistical analysis Analyses using routinely collected data included tabulation and graphical representation of trends in conception rates of women younger than 18 years by quartiles of per head total Local Implementation Grant investment and quartiles of area-level deprivation. Random-effects meta-regression analyses of the change in conception rates of women younger than 18 years between 1994–98 (the pre-intervention baseline) and 2009–13 (the most recently available Office of National Statistics data) were based on aggregate top-tier local authority level data. Because of variability in how each top-tier local authority spent their grant, a random-effects meta-regression analysis was used. Absolute and percentage changes between these timepoints were analysed as outcomes separately and data from each top-tier local authority were weighted by the inverse of the variance of the outcome estimate. The explanatory factors considered were per head Teenage Pregnancy Strategy investment, region, and deprivation (2004 Index of Multiple Deprivation scores). For the two continuous covariates (Teenage Pregnancy Strategy investment and Index of Multiple Deprivation score), the linearity of their association with the outcome variable (change in conception rate in women younger than 18 years) was assessed by also including quadratic terms in the regression model.

tion scores). For the two continuous covariates (Teenage Pregnancy Strategy investment and Index of Multiple Deprivation score), the linearity of their association with the outcome variable (change in conception rate in women younger than 18 years) was assessed by also including quadratic terms in the regression model. Analyses using Natsal data included logistic regression to examine factors associated with conception in women younger than 18 years separately in 1999–2001 and 2010–12 in women aged 18–24 years, with and without adjustment for the effect of all variables except age at first intercourse. Logistic regression with interaction terms was also used to assess whether associations between factors and conception in women younger than 18 years had changed between 1999–2001 and 2010–12. Analyses using Natsal data also included bivariate tabulations and logistic regression to examine the prevalence in 1990–91, 1999–2001, and 2010–12 of participation in work, education, or training in women aged 18–24 years, according to whether or not they had a child conceived before age 18 years. Logistic regression with interaction terms was used to assess whether the association between early motherhood and participation changed over time. All analyses were done with Stata 12.1 (version 13; StataCorp LP, College Station, Texas), accounting for stratification, clustering, and weighting of the Natsal data.

ars. Logistic regression with interaction terms was used to assess whether the association between early motherhood and participation changed over time. All analyses were done with Stata 12.1 (version 13; StataCorp LP, College Station, Texas), accounting for stratification, clustering, and weighting of the Natsal data. Role of the funding source The funders had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Routinely collected national data for conceptions in women younger than 18 years show a steady decline from their peak in 1998, with an apparent acceleration in that decline from 2007 onwards (figure 1). The trend was initially driven by the decline in maternities, until 2007 when the plot lines for abortions and births converged. The annual rates of conception in women younger than 18 years by quartile of area-related deprivation between 1994 and 2013 declined across all deprivation levels but this decline was larger in the most deprived areas (figure 2). Between 1998 and 2013, for example, conception rate declined by 16 conceptions per 1000 women aged 15–17 years in the least deprived areas compared with 33 conceptions per 1000 women aged 15–17 years living in the most deprived areas. This differential decline has resulted in a partial convergence in conception rates in women younger than 18 years across the quartiles of deprivation.

conceptions per 1000 women aged 15–17 years in the least deprived areas compared with 33 conceptions per 1000 women aged 15–17 years living in the most deprived areas. This differential decline has resulted in a partial convergence in conception rates in women younger than 18 years across the quartiles of deprivation. The annual data for the same period by quartile of strategy-related expenditure at local level—ie, per head total Local Implementation Grant investment—showed partial gradual convergence in conception rates in women younger than 18 years across Local Implementation Grant quartiles, with the greatest decline in areas receiving the highest Local Implementation Grant award (figure 2). Between 1998 and 2013, areas receiving the highest Local Implementation Grant award had a decline of 34 conceptions per 1000 women aged 15–17 years, whereas in areas receiving the lowest level, the decline was 16 conceptions per 1000 women aged 15–17 years.

iving the highest Local Implementation Grant award (figure 2). Between 1998 and 2013, areas receiving the highest Local Implementation Grant award had a decline of 34 conceptions per 1000 women aged 15–17 years, whereas in areas receiving the lowest level, the decline was 16 conceptions per 1000 women aged 15–17 years. Regression analyses assessing the association between Teenage Pregnancy Strategy funding and decline in conception rates in women younger than 18 years showed an estimated reduction in the conception rate of 11·4 (95% CI 9·6–13·2; p<0·0001) per 1000 women aged 15–17 years for every £100 Teenage Pregnancy Strategy spend per head (table 1). After adjustment for socioeconomic deprivation and region, this trend retained statistical significance, with a reduction of 8·2 conceptions (5·8–10·5; p<0·0001) per 1000 women aged 15–17 years per £100 Teenage Pregnancy Strategy spend per head. This trend was also reflected when modelling the percentage change in conception rates in women younger than 18 years as the outcome; for every £100 Teenage Pregnancy Strategy spend per head, a 6·2% (95% CI 2·3–10·2) reduction in the conception rate in women younger than 18 years was reported after adjusting for region and deprivation. Testing of quadratic terms indicated that associations with Teenage Pregnancy Strategy spend per head (and Index of Multiple Deprivation score) were approximately linear over the range of values considered for both absolute and percentage change, although the two models are not strictly compatible. The value of I2 for both meta-regression analyses was almost 100% because the outcome estimates from the top-tier local authorities are all very precise as they are based on rates from tens of thousands of person-years, and hence the variability seen is interpreted as variability in the underlying change across top-tier local authorities.

both meta-regression analyses was almost 100% because the outcome estimates from the top-tier local authorities are all very precise as they are based on rates from tens of thousands of person-years, and hence the variability seen is interpreted as variability in the underlying change across top-tier local authorities. The prevalence and odds ratios (ORs) of conception in women younger than 18 years by selected characteristics of women aged 18–24 years in 1999–2001 (Natsal-2) and 2010–12 (Natsal-3) are shown (table 2). Comparisons between the two surveys must be cautiously made because of minor but possibly important differences in question formulation. However, with this proviso and within the limits of precision of the estimates for the 18–24-year-old age group, the results for the two survey waves are broadly similar with regard to the prevalence of behaviour likely to be proximally and causally associated with conception. The proportion with more than minimum academic qualifications was, however, appreciably higher in 2010–12 (Natsal-3) compared with 1999–2001 (Natsal-2), as was the proportion reporting school lessons as their main source of sex education. A more modest increase between the two timepoints in the proportion reporting use of reliable contraception at first intercourse was noted.

ons was, however, appreciably higher in 2010–12 (Natsal-3) compared with 1999–2001 (Natsal-2), as was the proportion reporting school lessons as their main source of sex education. A more modest increase between the two timepoints in the proportion reporting use of reliable contraception at first intercourse was noted. Lower socioeconomic status, lower educational attainment, earlier onset of sexual activity, receiving sex education from sources other than school and, more weakly, negative opinion about the timing of first intercourse were associated with higher conception in women younger than 18 years in both surveys. Interaction testing revealed little clear evidence of change in the association with conception in women younger than 18 years between surveys for any of the factors considered (all p values were greater than 0·05 and are not shown). However, the association between living in an area in the highest quintile of deprivation and conception in women younger than 18 years was somewhat weaker in Natsal-3 than in Natsal-2, (Natsal-2 1999–2001: adjusted OR [AOR] 5·18 [95% CI 1·91–14·05]); Natsal-3 2010–12: AOR 2·89 [1·25–6·67]). A similar finding was seen for educational level. Having minimum or no academic qualifications was strongly associated with conception in both surveys, but the AOR was somewhat lower in Natsal-3 than in Natsal-2 (Natsal-2 1999–2001: AOR 5·61 [2·97–10·62]; Natsal-3 2010–12: AOR 3·61 [2·29–5·67]). The weaker association between use of reliable contraception at first intercourse and conception in women younger than 18 years was further attenuated between Natsal-2 and Natsal-3.

e AOR was somewhat lower in Natsal-3 than in Natsal-2 (Natsal-2 1999–2001: AOR 5·61 [2·97–10·62]; Natsal-3 2010–12: AOR 3·61 [2·29–5·67]). The weaker association between use of reliable contraception at first intercourse and conception in women younger than 18 years was further attenuated between Natsal-2 and Natsal-3. Results from Natsal-1, Natsal-2, and Natsal-3 relating to the participation of women aged 18–24 years according to whether or not they had conceived a child before age 18 years are shown (table 3). Estimates should be treated with caution because of the comparatively small survey subsamples in the 18–24-year-old age group. The results for all three surveys show the proportion of women in education, work, or training at the time of interview was higher by a considerable order of magnitude in those who did not conceive a child before 18 years than in those who did (table 3). However, although the likelihood of participation of women who were not young mothers was unchanged across the three surveys (1999–2001 vs 1990–91: OR 1·21 [95% CI 0·93–1·58]; 2010–12 vs 1999–2001: 0·86 [0·67–1·11]), in young mothers it remained constant between Natsal-1 and Natsal-2, but doubled between Natsal-2 and Natsal-3 (1999–2001 vs 1990–91: 1·12 [0·56–2·25]; 2010–12 vs 1999–2001: 1·99 [0·99–4·00]). Consequently, the association between young motherhood and participation weakened substantially between Natsal-2 and Natsal-3.

in young mothers it remained constant between Natsal-1 and Natsal-2, but doubled between Natsal-2 and Natsal-3 (1999–2001 vs 1990–91: 1·12 [0·56–2·25]; 2010–12 vs 1999–2001: 1·99 [0·99–4·00]). Consequently, the association between young motherhood and participation weakened substantially between Natsal-2 and Natsal-3. Discussion We report a marked decline in conceptions in women younger than 18 years, which was greater in areas of greater deprivation and also in areas of higher Teenage Pregnancy Strategy investment. The steep deprivation gradient previously associated with conception in women younger than 18 years has been partly attenuated. Similarly, the association between conception before age 18 years and lower educational level remains significant but has weakened over the period. Young people increasingly learn about sexual matters mainly from school lessons, and the association between conception before age 18 years and receiving sex education from other sources remains strong. The prevalence of participation in education, work, and training in women with a child conceived before age 18 years, although low, increased between 1999–2001 and 2010–12, a trend not seen in other, same-aged women.

n between conception before age 18 years and receiving sex education from other sources remains strong. The prevalence of participation in education, work, and training in women with a child conceived before age 18 years, although low, increased between 1999–2001 and 2010–12, a trend not seen in other, same-aged women. Reductions in rates of teenage pregnancy—ie, those occurring before age 20 years—have been seen in other high-income countries,13 indicating a broader secular trend. Fewer comparative data exist specifically on pregnancies in women younger than 18 years, and hardly any include abortion data. However, data for the birth rate in women younger than 18 years in the European Union between 2004 and 2014 show a decline from 13·6 to 6·8 per 1000 women aged 15–17 years in the UK compared with the average of 7·7 to 6·0 per 1000 women for all 28 countries (Office of National Statistics, unpublished). Moreover, the weakening of the association between conception rates in women younger than 18 years and deprivation in our study reverses the previous trend. In the 1980s and 1990s, teenage pregnancy rates increased in women living in the most deprived areas but remained unchanged or decreased in those in more affluent areas.14, 15 The increase in economic participation of teenage mothers seen in our data after 2000, is also a reversal of a previous trend. In 2005, we reported data from Natsal-1 and Natsal-2,16 which showed a widening gap at the end of the 20th century in the life chances and material prosperity of women who became mothers at an early age and those who did not. The recent improvement has occurred despite the less favourable economic climate in the post-2008 recession period, when the disparity between rich and poor has increased.

ng gap at the end of the 20th century in the life chances and material prosperity of women who became mothers at an early age and those who did not. The recent improvement has occurred despite the less favourable economic climate in the post-2008 recession period, when the disparity between rich and poor has increased. The decline in teenage pregnancy rates has been differentially attributed to distal factors such as increased educational attainment and to more proximal factors such as improved use of contraception. Researchers finding a strong association between falling teenage pregnancy rates and increasing use of long-acting reversible contraception in England since the late 1990s have underlined the contribution of increased access to reliable contraception to the decline.17 Observers in the USA have reached similar conclusions18 supported by intervention studies examining the effect of providing highly effective contraception free of charge.19 By contrast, researchers using ecological analyses of routinely collected area-level data in England to examine the association between educational attainment, contraceptive use, and trends in conception in women younger than 18 years concluded that the larger association with education indicates that this is the key driver.20 Interpreting the associations between educational level, effective contraception, and early pregnancy is difficult because of the likelihood of reverse causality. Low educational attainment is both a cause and consequence of teenage pregnancy. The effect of use of highly reliable contraception in reducing early conception rates might be masked in research by the tendency for women who experience early pregnancy to subsequently use more reliable methods. Our findings suggest that contraceptive use and educational attainment each wield an independent influence on the likelihood of conception in women younger than 18 years. Because Natsal asks about contraception in the last year and ever, we are unable to measure use at the time of conception, and so reliable method use at first intercourse has been used as a proxy indicator in these analyses. However, unpublished Natsal data show a two-fold increase in use of long-acting reversible contraception methods by sexually active 16–17 year olds between Natsal-2 and Natsal-3.

measure use at the time of conception, and so reliable method use at first intercourse has been used as a proxy indicator in these analyses. However, unpublished Natsal data show a two-fold increase in use of long-acting reversible contraception methods by sexually active 16–17 year olds between Natsal-2 and Natsal-3. Both increases in educational attainment and in use of highly effective contraception are likely to have contributed to the falling teenage pregnancy rate. Educational aspirations provide the motivation and contraception the means by which to avert early pregnancy. Both are clearly policy related. Other studies have shown that multiple interventions (combining educational and behavioural interventions) lower the rate of teenage pregnancy21 and highlight the importance of education in planning policies with this as their aim.22 This was reflected in the decision by the English Government to locate the unit responsible for the Teenage Pregnancy Strategy jointly in two government ministries, the Department of Education and the Department of Health, and to make joint working between education and health agencies a key component of the strategy.

This was reflected in the decision by the English Government to locate the unit responsible for the Teenage Pregnancy Strategy jointly in two government ministries, the Department of Education and the Department of Health, and to make joint working between education and health agencies a key component of the strategy. A strength of this study lies in the combination of individual-level and area-level data, and in our capacity to show independent associations between conception rates in women younger than 18 years and possibly influential variables. However, slight changes in question wording between the surveys might have influenced responses relating to pregnancy in Natsal-2 and Natsal-3. In particular, we should note that abortions were slightly under-reported in Natsal-3 compared with the official UK figure in 2011.23

years and possibly influential variables. However, slight changes in question wording between the surveys might have influenced responses relating to pregnancy in Natsal-2 and Natsal-3. In particular, we should note that abortions were slightly under-reported in Natsal-3 compared with the official UK figure in 2011.23 Compared with experimental approaches, observational studies have limitations in assessing the relative effect of policy-related intervention and secular trends on health outcomes. Specifically in this case, Teenage Pregnancy Strategy funding was determined by the pre-intervention conception rate. This creates challenges in disentangling the effect of Teenage Pregnancy Strategy funding from the effects of the pre-intervention rate on the subsequent conception rate, including the potential for regression to the mean. Our finding of an association between the decline in conception rates in women younger than 18 years and Teenage Pregnancy Strategy resourcing is suggestive that government-linked efforts have contributed towards lowering conception rates in women younger than 18 years but should not be seen as conclusive.

to the mean. Our finding of an association between the decline in conception rates in women younger than 18 years and Teenage Pregnancy Strategy resourcing is suggestive that government-linked efforts have contributed towards lowering conception rates in women younger than 18 years but should not be seen as conclusive. Individual-level-data have also enabled us to show, where analyses of area-related data have not,24 the progress that has been made in reducing the previously strong link between deprivation and early conception. The comparatively greater decline in conception rate in women younger than 18 years in the most deprived areas is worthy of note. Although the potential for decline might have been greater where baseline conception rates were higher, these were often areas in which complex and multiple social and health problems competed for public health efforts and resources.

tion rate in women younger than 18 years in the most deprived areas is worthy of note. Although the potential for decline might have been greater where baseline conception rates were higher, these were often areas in which complex and multiple social and health problems competed for public health efforts and resources. Progress has been made towards halting the cycle of poverty and income inequality long associated with early pregnancy, and in improving the life chances of young mothers. Despite this success, England's teenage pregnancy prevalence remains high compared with other high-income countries,25 and there is more to be achieved. Some have suggested, because improvements in use of effective contraception seem to have contributed to the decrease in rates, that the policy emphasis should be placed where it has been successful in the past. Higher contraception discontinuation rates in young people26 favour their use of long-acting reversible contraception and additional progress could be made by accelerating the rate at which they are used after onset of sexual activity. Despite the increase in use of these methods, they are used by barely one in six teenage women. The strong and sustained association between conception in women younger than 18 years and earlier sexual activity suggests that sizeable additional gains might also be made by helping young women to become sexually active at a time that is right for them.27 A third of all young women, and 60% of those for whom first sex occurs before age 16 years, subsequently consider that to have been too early for them. Bringing actual timing of first sexual encounter in line with preference for timing of first sexual encounter could cut teenage pregnancy rates further, and an intensified policy focus aimed at achieving this might be warranted.

sex occurs before age 16 years, subsequently consider that to have been too early for them. Bringing actual timing of first sexual encounter in line with preference for timing of first sexual encounter could cut teenage pregnancy rates further, and an intensified policy focus aimed at achieving this might be warranted. Our data also underline the importance of continued efforts to prevent teenage pregnancy. The decrease in conception rates in women younger than 18 years was more substantial in the later years of the Teenage Pregnancy Strategy and before then, there was scope for detractors to dismiss it as having failed to reach its goals.28, 29 Our interim study of the assessment of the Teenage Pregnancy Strategy in 2006 showed modest falls in the first 4 years of the strategy, albeit more marked in poorer areas of the country30 in which government investment had been greatest. From 2007, however, pregnancy rates began to decrease more steeply and the downward trend continued after the strategy was mainstreamed in 2010. The acceleration is partly explained by the likelihood that strategy-related intervention took time to be implemented. It might also reflect the renewed efforts mounted in 2006 to mobilise additional resources and to persuade senior stakeholders to prioritise the issue in areas in which multiple socioeconomic problems were impeding progress, which is coincidental with a steeper decline in conception rates in women younger than 18 years in these areas in the later years of the Teenage Pregnancy Strategy. A further factor possibly contributing to the decrease in conception rates in women younger than 18 years is the synergy between strategy-related initiatives and other therapeutic and policy-related intervention—the availability of emergency contraception without prescription, for example, the recommendation by the National Institute for Health and Care Excellence that access to long-acting reversible contraception should be increased, and the rise in educational attainment in young people. Taken as a whole, the evidence underlines the importance of long-term, sustained, multifaceted prevention strategies to tackle the more intractable public health challenges.

d Care Excellence that access to long-acting reversible contraception should be increased, and the rise in educational attainment in young people. Taken as a whole, the evidence underlines the importance of long-term, sustained, multifaceted prevention strategies to tackle the more intractable public health challenges. Acknowledgments We thank the study participants, the team of interviewers and operational staff, and computing staff from NatCen Social Research. Natsal 1, 2, and 3 were supported by grants from the Medical Research Council and the Wellcome Trust, with contributions from the Economic and Social Research Council and Department of Health. Contributors KW, MJP, and PW conceived this Article. KW wrote the first draft, with further contributions from AC, RSG, LJG, MJP, and PW. KJ, RSG, LJG, and MJP did the statistical analysis, with guidance from AC, CHM, and PW. All authors interpreted data, reviewed successive drafts, and approved the final version of the Article. Declaration of interests AMJ is a Governor of the Wellcome Trust. All other authors declare no competing interests. Figure 1 Conception, maternity, and abortion rates and events of possible relevance to trends in women younger than 18 years between 1994 and 2013 (per 1000 women aged 15–17 years) Figure 2 Conception in women younger than 18 years by quartile between 1994 and 2013 by Index of Multiple Deprivation (A) and Local Implementation Grant investment for Teenage Pregnancy Strategy (B)

Figure 1 Conception, maternity, and abortion rates and events of possible relevance to trends in women younger than 18 years between 1994 and 2013 (per 1000 women aged 15–17 years) Figure 2 Conception in women younger than 18 years by quartile between 1994 and 2013 by Index of Multiple Deprivation (A) and Local Implementation Grant investment for Teenage Pregnancy Strategy (B) Table 1 Meta-regression analysis of the association between total Teenage Pregnancy Strategy funding per head and absolute and percentage change in the conception rate in women younger than 18 years between 1994–98 and 2009–13 Absolute change in conception rate in women <18 years per £100 per girl LIG spend (95% CI) p value Percentage change in conception rate in women <18 years per £100 per girl LIG spend (95% CI) p value Unadjusted −11·4 (−13·2 to −9·6) <0·0001 −8·6% (−11·9 to −5·4) <0·0001 Adjusted for region −11·5 (−13·5 to −9·5) <0·0001 −8·3% (−11·4 to −5·1) <0·0001 Adjusted for region and deprivation −8·2 (−10·5 to −5·8) <0·0001 −6·2% (−10·2 to −2·3) 0·0023 LIG=Local Implementation Grant. Table 2 Characteristics of women aged 18–24 years in England in Natsal-2 and Natsal-3 and factors associated with conception in women younger than 18 years

Absolute change in conception rate in women <18 years per £100 per girl LIG spend (95% CI) p value Percentage change in conception rate in women <18 years per £100 per girl LIG spend (95% CI) p value Unadjusted −11·4 (−13·2 to −9·6) <0·0001 −8·6% (−11·9 to −5·4) <0·0001 Adjusted for region −11·5 (−13·5 to −9·5) <0·0001 −8·3% (−11·4 to −5·1) <0·0001 Adjusted for region and deprivation −8·2 (−10·5 to −5·8) <0·0001 −6·2% (−10·2 to −2·3) 0·0023 LIG=Local Implementation Grant. Table 2 Characteristics of women aged 18–24 years in England in Natsal-2 and Natsal-3 and factors associated with conception in women younger than 18 years 1999–2001 2010–12 Number of participants (unweighted, weighted) Percentage of participants with characteristic* Percentage who had conception at age <18 years (95% CI) OR for <18 years conception Number of participants (unweighted, weighted) Percentage of participants with characteristic* Percentage who had conception at age <18 years (95% CI) OR for <18 years conception OR AOR (95% CI) p value OR AOR (95% CI) p value Total <18 years conception prevalence 967, 993 .. 13% (10·8–15·6) .. .. .. 1368, 817 .. 11% (9·1–12·4) .. .. .. Area-level deprivation 0·008 0·005 1 (least deprived) 113, 115 12% 3% (1·3–8·5) 1 1 .. 219, 131 16% 6% (3·1–9·9) 1 1 .. 2 115, 124 12% 10% (5·3–18·4) 3·23 3·4 (0·98–11·77) .. 227, 133 16% 8% (5·0–11·3) 1·37 1·18 (0·51–2·74) .. 3 137, 156 16% 8% (4·7–13·7) 2·56 2·2 (0·78–6·22) .. 259, 164 20% 7% (4·4–10·7) 1·25 1·37 (0·55–3·39) .. 4 229, 256 26% 14% (9·5–19·3) 4·58 3·48 (1·22–9·97) .. 307, 194 24% 13% (9·4–16·8) 2·43 2·15 (0·94–4·89) .. 5 (most deprived) 373, 342 34% 19% (14·7–24·3) 6·82 5·18 (1·91–14·05) .. 356, 195 24% 17% (13·5–21·8) 3·51 2·89 (1·25–6·67) .. Quintiles of baseline teenage pregnancy rate 0·387 0·3182 1 (low) 176, 203 21% 11% (7·1–16·6) 1 1 .. 249, 154 19% 8% (5·5–12·4) 1 1 .. 2 190, 188 19% 12% (7·7–18·7) 1·12 0·79 (0·34–1·80) .. 252, 139 17% 14% (10·1–18·4) 1·75 1·16 (0·58–2·31) .. 3 211, 211 21% 15% (10·6–21·5) 1·47 0·83 (0·38–1·80) .. 242, 134 16% 10% (6·9–15·0) 1·26 0·72 (0·33–1·60) .. 4 216, 252 25% 14% (8·8–20·7) 1·29 1·04 (0·50–2·18) .. 304, 185 23% 10% (7·4–14·1) 1·26 0·66 (0·32–1·36) .. 5 (high) 171, 136 14% 11% (6·9–18·1) 1·04 0·52 (0·23–1·18) .. 320, 204 25% 11% (7·8–15·0) 1·35 0·67 (0·31–1·43) .. Academic qualifications in participants ≥17 years <0·0001 <0·0001 Studying/gained further qualifications 435, 475 50% 4% (2·1–5·9) 1 1 .. 825, 523 66% 5% (3·4–6·1) 1 1 .. None, or those typically gained at 16 years† 500, 483 50% 23% (19·1–27·1) 7·99 5·61 (2·97–10·62) .. 510, 266 34% 23% (19·4–27·0) 6·24 3·61 (2·29–5·67) .. Heterosexual intercourse before age 16 years <0·0001 <0·0001 No 693, 713 72% 7% (4·8–9·0) 1 1 .. 884, 564 71% 5% (4·1–7·0) 1 1 .. Yes 270, 274 28% 29% (24·0–35·3) 5·86 3·33 (2·04–5·45) ..

lly gained at 16 years† 500, 483 50% 23% (19·1–27·1) 7·99 5·61 (2·97–10·62) .. 510, 266 34% 23% (19·4–27·0) 6·24 3·61 (2·29–5·67) .. Heterosexual intercourse before age 16 years <0·0001 <0·0001 No 693, 713 72% 7% (4·8–9·0) 1 1 .. 884, 564 71% 5% (4·1–7·0) 1 1 .. Yes 270, 274 28% 29% (24·0–35·3) 5·86 3·33 (2·04–5·45) .. 452, 233 29% 24% (20·2–28·6) 5·63 3·2 (2·03–5·04) .. Reliable contraception used at first sex‡ 0·0032 0·1906 Yes 701, 724 82% 12% (9·4–14·7) 1 1 .. 990, 584 87% 11% (9·0–12·9) 1 1 .. No 163, 156 18% 27% (19·7–35·3) 2·73 2·09 (1·28–3·40) .. 160, 87 13% 22% (16·2–28·8) 2·3 1·44 (0·83–2·48) .. View on timing of first sex 0·079 0·0693 About the right time/too late 508, 520 59% 10% (7·4–13·4) 1 1 .. 726, 439 65% 9% (7·2–11·4) 1 1 .. Too early 353, 357 41% 21% (16·9–25·9) 2·39 1·6 (0·95–2·70) .. 426, 232 35% 18% (15·0–22·4) 2·26 1·56 (0·96–2·52) .. Main reason for first sex 0·1782 0·7324 Autonomous reason 713, 723 85% 14% (11·8–17·5) 1 1 .. 932, 551 83% 11% (9·5–13·6) 1 1 .. Non-autonomous reason§ 128, 130 15% 15% (9·4–22·1) 1·01 0·58 (0·26–1·29) .. 205, 111 17% 17% (12·6–23·6) 1·64 1·09 (0·66–1·81) .. Partner more willing at first sex 0·356 0·1182 No 650, 673 77% 13% (10·4–15·8) 1 1 .. 970, 564 83% 12% (10·4–14·6) 1 1 .. Yes 211, 204 23% 20% (14·7–26·7) 1·7 1·27 (0·76–2·13) .. 203, 118 17% 14% (9·5–19·0) 1·12 0·62 (0·33–1·13) .. Sex education source 0·2297 0·0478 School lessons 253, 281 29% 9% (5·4–13·7) 1 1 .. 519, 314 38% 8% (6·0–10·3) 1 1 .. Parents 180, 193 20% 13% (8·4–18·5) 1·51 1·73 (0·75–3·99) .. 205, 119 15% 11% (7·3–15·4) 1·4 1·01 (0·48–2·11) .. Other 518, 502 51% 16% (12·7–19·7) 1·99 1·69 (0·92–3·10) .. 639, 383 47% 13% (10·3–15·7) 1·71 1·71 (1·06–2·75) .. Needed more information at first sex 0·1033 0·3937 No 223, 232 24% 7% (4·1–11·8) 1 1 .. 331, 207 28% 8% (5·9–11·6) 1 1 .. Yes 711, 727 76% 15% (12·7–18·7) 2·43 1·79 (0·89–3·61) .. 930, 535 72% 12% (10·4–14·7) 1·56 1·25 (0·75–2·10) .. Communication with parents about sex 0·5468 0·9317 Easy 328, 349 36% 12% (8·4–16·4) 1 1 .. 428, 256 32% 10% (7·6–13·4) 1 1 .. Difficult/not discussed/varied 616, 626 64% 13% (10·3–16·5) 1·12 1·22 (0·64–2·31) .. 896, 543 68% 10% (8·1–12·1) 0·98 0·98 (0·58–1·65) .. Lived with both natural parents¶ 0·2437 0·0088 Yes 683, 714 72% 11% (8·2–13·4) 1 1 .. 826, 527 65% 7% (5·2–8·6) 1 1 .. No 284, 279 28% 19% (15·0–24·6) 2·04 1·31 (0·83–2·08) .. 541, 289 35% 18% (14·8–21·3) 3·02 1·72 (1·15–2·57) ..

(10·3–16·5) 1·12 1·22 (0·64–2·31) .. 896, 543 68% 10% (8·1–12·1) 0·98 0·98 (0·58–1·65) .. Lived with both natural parents¶ 0·2437 0·0088 Yes 683, 714 72% 11% (8·2–13·4) 1 1 .. 826, 527 65% 7% (5·2–8·6) 1 1 .. No 284, 279 28% 19% (15·0–24·6) 2·04 1·31 (0·83–2·08) .. 541, 289 35% 18% (14·8–21·3) 3·02 1·72 (1·15–2·57) .. Conception before age 18 years was classified as livebirths occurring before age 18 years 9 months and abortions occurring at age 17 years and younger. Miscarriages were excluded because of common misreporting, in common with Office of National Statistics. Stillbirths excluded for comparability between Natsal-2 and Natsal-3. OR=odds ratio. AOR=adjusted odds ratio. * Percentages were calculated for weighted participants. † English General Certificate of Secondary Education or equivalent. ‡ Contraceptive pill or condom. § Non-autonomous reason for first sex defined as copying peers or being under the influence of alcohol in Natsal-2 and copying peers or being under the influence of alcohol or recreational drugs in Natsal-3.

* Percentages were calculated for weighted participants. † English General Certificate of Secondary Education or equivalent. ‡ Contraceptive pill or condom. § Non-autonomous reason for first sex defined as copying peers or being under the influence of alcohol in Natsal-2 and copying peers or being under the influence of alcohol or recreational drugs in Natsal-3. ¶ Until age 14 years in Natsal-2 and until age 16 years in Natsal-3. AOR is adjusted for whether participants lived with both natural parents until age 14 years or 16 years, ease of communication with parents about sex, Index of Multiple Deprivation, educational attainment, main source of sex education, whether more information was needed at first sex, contraceptive method use at first sex, whether both partners were equally willing at first sex, autonomy of the reason for first sex, and baseline conception rate at younger than 18 years (area-level), but not adjusted for first sex before age 16 years. Table 3 Current participation of women aged 18–24 years in England by experience of early motherhood (1990–91, 1999–2001, and 2010–12)

¶ Until age 14 years in Natsal-2 and until age 16 years in Natsal-3. AOR is adjusted for whether participants lived with both natural parents until age 14 years or 16 years, ease of communication with parents about sex, Index of Multiple Deprivation, educational attainment, main source of sex education, whether more information was needed at first sex, contraceptive method use at first sex, whether both partners were equally willing at first sex, autonomy of the reason for first sex, and baseline conception rate at younger than 18 years (area-level), but not adjusted for first sex before age 16 years. Table 3 Current participation of women aged 18–24 years in England by experience of early motherhood (1990–91, 1999–2001, and 2010–12) No experience of motherhood resulting from conception before age 18 years Experience of motherhood resulting from conception before age 18 years p value (from χ2 test) Denominators (unweighted, weighted numbers) Percentage participation in education, work, or training (95% CI) Denominators (unweighted, weighted numbers) Percentage participation in education, work, or training (95% CI) 1990–1991 1135, 1393 79% (76·1–81·7) 180, 134 20% (14·1–28·5) <0·0001 1999–2001 862, 910 82% (79·0–84·8) 106, 81 22% (14·4–33·1) <0·0001 2010–2012 1203, 738 80% (77·0–82·2) 104, 43 36% (27·1–47·0) <0·0001

Introduction The year 2015 marks the end of the Millennium Development Goals (MDGs) era, during which the under-5 mortality rate (U5MR) reduced by an impressive 53% globally, although still falling short of the MDG 4 target of a two-thirds reduction from 1990 to 2015.1, 2 Year 2016 marks the beginning of the implementation of the Sustainable Development Goals (SDGs).3 The SDGs target an U5MR of no more than 25 per 1000 livebirths in every country of the world in 2030.4 To plan how to eliminate preventable child deaths, information is needed about the current distribution of causes of child deaths and this has changed in recent decades. In this paper, we update our annual estimates of child mortality by cause to 2000–15; reflect on progress toward the MDG 4; and consider the implications for national and global priorities if the SDG target for child survival is to be achieved. Methods General estimation approaches We estimated the number of child deaths by cause for each of the 194 WHO member states for each year in 2000–15. This was done separately for neonates and children aged 1–59 months. The number of child deaths by cause was estimated as the product of the number of age-specific deaths due to all causes and age-specific and cause-specific mortality fractions. The age-specific all-cause death estimates were derived from age-specific child mortality estimates produced by the UN Inter-Agency Group for Child Mortality Estimation (UN-IGME).5 The livebirth estimates were produced by the UN Population Division.6

o all causes and age-specific and cause-specific mortality fractions. The age-specific all-cause death estimates were derived from age-specific child mortality estimates produced by the UN Inter-Agency Group for Child Mortality Estimation (UN-IGME).5 The livebirth estimates were produced by the UN Population Division.6 To generate cause-specific mortality fractions (CSMFs) for neonates and 1-59-month–olds, we applied our estimation framework with updates.7 The estimation framework comprises three components. Component one covers countries with adequate vital registration (VR) (67 for neonates, 69 for 1–59-month-olds) for which we used CSMFs derived from the country-specific VR data as is or with minor adjustments.8 Component two covers countries with inadequate VR and low U5MR (<35 per 1000 livebirths in 2000–15; 47 for neonates and 44 for 1–59-month-olds). For these countries, we modelled CSMFs using a multinomial logistic regression (MLR) applied to input CSMFs calculated from number of deaths by cause from VR countries (ie, component one) and their distal (eg, socioeconomic indicators) and proximate (eg, childhood life-saving intervention coverage values) determinates of child survival as model inputs.9 Component two is referred to as a VR based multicause model or VRMCM. Component three is for countries with inadequate VR and high U5MR (≥35 per 1000 livebirths in 2000–15; 80 for neonates and 81 for 1–59-month-olds). For these countries, we modelled CSMFs using MLR with empirical input CSMFs calculated from number of deaths by cause extracted from verbal autopsy (VA) studies and primarily proximate determinants as model inputs. Component three is referred to as a VA based multicause model or VAMCM. Details of the estimation framework including the use of MLR and information on the standard International Classification of Diseases codes by cause have been published elsewhere.7, 8 The distal and proximate determinants used as model inputs are listed in the appendix (p 5).

to as a VA based multicause model or VAMCM. Details of the estimation framework including the use of MLR and information on the standard International Classification of Diseases codes by cause have been published elsewhere.7, 8 The distal and proximate determinants used as model inputs are listed in the appendix (p 5). Research in context Evidence before this study Our study group, formerly referred to as the WHO and UNICEF's Child Health Epidemiology Reference Group (CHERG), has systematically reviewed, estimated, and published a series of child mortality by cause estimates since 2003, with the last publication presenting estimates for years 2000–13. To collect data published since 2013, we did an updated systematic review to identify quality child cause-of-death studies published between Jan 1, 2013, and Feb 12, 2015 in the following databases: PubMed, Embase, ISIS Web of Knowledge, Medline BIOSIS, Popline, WHOLIS (via Global Health Library, including the regional specific databases of LILACs, African Index Medicus, WPRIM, IMEMR, and PAHO) and IndMed without language limitation. Search strategies, search terms, and study inclusion and exclusion criteria were consistent with our previous studies. Other investigators have estimated distribution of mortality by cause among all age groups. Added value of this study

Our study group, formerly referred to as the WHO and UNICEF's Child Health Epidemiology Reference Group (CHERG), has systematically reviewed, estimated, and published a series of child mortality by cause estimates since 2003, with the last publication presenting estimates for years 2000–13. To collect data published since 2013, we did an updated systematic review to identify quality child cause-of-death studies published between Jan 1, 2013, and Feb 12, 2015 in the following databases: PubMed, Embase, ISIS Web of Knowledge, Medline BIOSIS, Popline, WHOLIS (via Global Health Library, including the regional specific databases of LILACs, African Index Medicus, WPRIM, IMEMR, and PAHO) and IndMed without language limitation. Search strategies, search terms, and study inclusion and exclusion criteria were consistent with our previous studies. Other investigators have estimated distribution of mortality by cause among all age groups. Added value of this study In this paper, we updated the estimates from years 2000–13 to 2000–15 to reflect on the progress toward the MDG 4 and draw implications for the SDG child survival target. Our updates are based on substantially more input data and several important methodological advances, including using adjusted empirical instead of modelled child cause-of-death estimates for China for the first time. Implications of all the available evidence

In this paper, we updated the estimates from years 2000–13 to 2000–15 to reflect on the progress toward the MDG 4 and draw implications for the SDG child survival target. Our updates are based on substantially more input data and several important methodological advances, including using adjusted empirical instead of modelled child cause-of-death estimates for China for the first time. Implications of all the available evidence Estimates presented here are the most up-to-date, and likely thus far the most valid ones of child mortality by cause at the global, regional, and national levels. Such information can and should be used to inform child survival policy making and resource allocation. Future research should further consider how to best incorporate increasing national empirical estimates and balance between empirical and modelled estimates. Continued investment in child cause-of-death data collection and estimation applying innovative approaches will further improve validity of such important information.

e research should further consider how to best incorporate increasing national empirical estimates and balance between empirical and modelled estimates. Continued investment in child cause-of-death data collection and estimation applying innovative approaches will further improve validity of such important information. We updated our database to include new VR data reported to WHO up to July 30, 2015, and new VA data identified through an updated systematic review of literature published between Jan 1, 2013, and Feb 12, 2015 (table 1; appendix p 3). Despite completing the systematic review in 2015, our estimates for year 2015 were intended to cover the entire year. In total, the input number of deaths increased by 43·0% for neonates (from 2·629 million to 3·760 million) and 22·7% for children aged 1–59 months (from 3·312 million to 4·063 million). For China, we replaced modelled estimates10 with adjusted empirical estimates from the China Maternal and Child Health Surveillance System (MCHSS), the details of which can be found elsewhere.11 A summary map of input data and estimation methods is presented in the appendix (p 3).

(from 3·312 million to 4·063 million). For China, we replaced modelled estimates10 with adjusted empirical estimates from the China Maternal and Child Health Surveillance System (MCHSS), the details of which can be found elsewhere.11 A summary map of input data and estimation methods is presented in the appendix (p 3). Updates on estimation methods A few methodological updates are shown here. First, we used the Plasmodium falciparum parasite rate (PfPR)12 in place of the more subjective malaria index8 as one of the candidate covariates to model the fraction of deaths due to malaria in countries with high transmission intensity. PfPR is the proportion of the population carrying asexual blood-stage parasites and is considered as an indicator of malaria transmission intensity (appendix p 4). In the VAMCM, we used post-hoc adjustment to consider the effects of recently scaled up interventions, which previously included insecticide treated bed nets (ITN). Since PfPR reflects the impact of ITN,12 we dropped ITN from the post-hoc adjustment. Second, we considered the impacts of pneumococcal conjugate vaccine13, 14 and rotavirus vaccine,15 in addition to the previously included Haemophilus influenzae type b (Hib) vaccine in the post-hoc adjustment. Specifically, we calculated pneumonia-specific, meningitis-specific, and diarrhoea-specific deaths averted due to each of these vaccines in the post-hoc adjustment and redistributed the cause-specific deaths averted to the remaining causes pro rata. The cause-specific deaths averted were calculated as the product of the following four quantities: 1) cause-specific deaths estimated by VAMCM before post-hoc adjustment, 2) the fractions due to vaccine-specific serotypes, 3) vaccine coverage, and 4) vaccine effectiveness, where available, or efficacy. Details of the parameters used in the post-hoc adjustment are available in the appendix (p 7). Lastly, a 7-year moving-average smoother was applied to the national-level prediction covariates used in the VAMCM to attenuate implausible spikes due to systematic errors in measurement or inconsistencies in covariate definitions.

s of the parameters used in the post-hoc adjustment are available in the appendix (p 7). Lastly, a 7-year moving-average smoother was applied to the national-level prediction covariates used in the VAMCM to attenuate implausible spikes due to systematic errors in measurement or inconsistencies in covariate definitions. Model selection and uncertainty estimation We selected the final model by cross validation. Specifically, we selected 10% of the observed cause-of-death data points, and fit each of the candidate models using the remaining 90%. We then predicted the CSMFs in the withheld selection, and determined the difference between the observed and predicted estimates. We repeated this process with 500 random subsets. We selected as final the model with the smallest average out of sample prediction error.13, 16 We estimated uncertainty in model coefficients by bootstrap resampling of input data sets from all estimation components and their respective distributions. This uncertainty was propagated through to the model predictions.13, 17 Uncertainty in the estimates of under-5 and neonatal deaths was also included using the UN-IGME methodology.1 We also accounted for potential variability due to post-hoc adjustment and the modelled number of deaths due to measles, pertussis, HIV, and malaria outside of sub-Saharan Africa. The 2·5 and 97·5 percentiles were taken as the lower and upper ranges of the uncertainty.

was also included using the UN-IGME methodology.1 We also accounted for potential variability due to post-hoc adjustment and the modelled number of deaths due to measles, pertussis, HIV, and malaria outside of sub-Saharan Africa. The 2·5 and 97·5 percentiles were taken as the lower and upper ranges of the uncertainty. Estimates reporting Estimates of deaths in the 1–59 month period due to preterm birth complications, intrapartum-related events, and congenital abnormalities were produced previously, but were collapsed into the category of other. They are reported separately here. Notably, the number of deaths due to, for example, congenital abnormalities among under-5 is then the sum of the numbers of deaths due to congenital abnormalities among neonates and 1–59-month-olds. Other conditions among children aged 1–59 months include causes originated during the perinatal period, cancer, severe malnutrition, and other specified causes. Since we estimated child mortality by cause for 2000–15 but not for 1990–99, we were not able to assess cause-specific progress toward the MDG 4 for the entire period of 1990–2015. However, we can benchmark cause-specific progress in 2000–15 with the 4·4% average annual rate of reduction (ARR) required to achieve the MDG 4.8, 18 We also present the aggregated cause-of-death profile by six U5MR strata using the 2015 estimates with the cutoffs of 10, 25, 50, 75, and 100 per 1000 livebirths, referred to as very low, low, medium, medium high, high, and very high mortality strata, respectively.1

f reduction (ARR) required to achieve the MDG 4.8, 18 We also present the aggregated cause-of-death profile by six U5MR strata using the 2015 estimates with the cutoffs of 10, 25, 50, 75, and 100 per 1000 livebirths, referred to as very low, low, medium, medium high, high, and very high mortality strata, respectively.1 To promote transparency and replicability of global health estimates, we have also included the GATHER reporting checklist in the appendix (p 9).19 Additional details of the input data, estimation methodology including statistical codes, and estimates are online and publicly available through the Maternal and Child Epidemiology Estimation's website. Role of the funding source The funder of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. All authors had full access to all the data in the study and the corresponding author had final responsibility for the decision to submit for publication.

To promote transparency and replicability of global health estimates, we have also included the GATHER reporting checklist in the appendix (p 9).19 Additional details of the input data, estimation methodology including statistical codes, and estimates are online and publicly available through the Maternal and Child Epidemiology Estimation's website. Role of the funding source The funder of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. All authors had full access to all the data in the study and the corresponding author had final responsibility for the decision to submit for publication. Results In 2015, among the 5·941 million children who did not live to age 5 years, 2·681 million (45·1%) died in the neonatal period (figure 1). The leading causes of deaths in children under 5 were preterm birth complications (1·055 million [95% UR 0·935–1·179]; 17·8% [UR 15·4–19·0]), pneumonia (0·921 million [0·812–1·117], 15·5% [13·9–17·6]), and intrapartum-related events (0·691 million [0·598–0·778], 11·6% [9·9–12·7]; table 2). Among neonates, the leading causes were preterm birth complications (0·944 million [UR 0·832–1·066], 15·9% [UR 13·8–17·3]), intrapartum related events (0·637 million [0·550–0·723], 10·7% [9·2–11·8]), and sepsis or meningitis (0·401 million [0·280–0·522], 6·8% [4·7–8·6]). Among children who died in the 1-59-month period, the leading causes were pneumonia (0·762 million [UR 0·651–0·943], 12·8% [UR 11·5–14·6]), diarrhoea (0·509 million [0·401–0·661], 8·6% [7·0–10·2]), and injuries (0·327 million [0·272–0·410], 5·5% [4·6–6·3]).

s or meningitis (0·401 million [0·280–0·522], 6·8% [4·7–8·6]). Among children who died in the 1-59-month period, the leading causes were pneumonia (0·762 million [UR 0·651–0·943], 12·8% [UR 11·5–14·6]), diarrhoea (0·509 million [0·401–0·661], 8·6% [7·0–10·2]), and injuries (0·327 million [0·272–0·410], 5·5% [4·6–6·3]). Sub-Saharan Africa and southern Asia remained the MDG regions with the highest numbers of under-5 deaths in 2015 (2·947 million and 1·891 million, respectively). The distribution of under-5 deaths by cause differed substantially by region (appendix p 11). For example, in sub-Saharan Africa, the leading causes of under-5 deaths were pneumonia (0·490 million [UR 0·417–0·631], 16·6% [UR 14·8–19·1]), preterm birth complications (0·356 million [0·283–0·433], 12·1% [9·3–13·6]), and intrapartum-related events (0·338 million [0·278–0·397], 11·5% [9·0–12·3]). Southern Asia, however, had a higher fraction of neonatal deaths (57·0%), with preterm birth complications being the leading cause (0·478 million [0·394–0·552], 25·3% [21·7–28·7]).

plications (0·356 million [0·283–0·433], 12·1% [9·3–13·6]), and intrapartum-related events (0·338 million [0·278–0·397], 11·5% [9·0–12·3]). Southern Asia, however, had a higher fraction of neonatal deaths (57·0%), with preterm birth complications being the leading cause (0·478 million [0·394–0·552], 25·3% [21·7–28·7]). Among the 194 countries estimated, the number of under-5 deaths varied between 1 death and 1·201 million deaths in 2015. The ten countries with the highest number of under-5 deaths were collectively responsible for three-fifths (60·4%, 3·587 million) of the global under-5 deaths. Their cause composition is presented in the appendix (p 12). The share of neonatal deaths in these countries varied from 30·9% (0·094 of 0·305 million in DR Congo) to 62·3% (0·074 of 0·119 million in Bangladesh). The leading cause among under-5s was pneumonia in Angola (0·029 million [UR 0·011–0·064], 17·4% [UR 13·8–21·9]), DR Congo (0·046 million [0·026–0·074], 15·2% [12·8–17·9]), Ethiopia (0·031 million [0·012–0·060], 17·1% [14·2–20·7]), Nigeria (0·133 million [0·087–0·209], 17·8% [15·9–20·6]), and Tanzania (0·014 million [0·007–0·028], 14·6% [12·8–17·5]). All of these countries are in sub-Saharan Africa. The leading cause was preterm birth complications in Bangladesh (0·024 million [UR 0·018–0·031], 19·8% [UR 15·6–24·0]), Indonesia (0·028 million [0·023–0·038], 18·9% [16·3–23·3]), India (0·330 million [0·269–0·387], 27·5% [23·4–31·5]), and Pakistan (0·102 million [0·071–0·137], 23·6% [18·4–28·7]). Most of these countries are in southern Asia. The leading cause was congenital abnormalities in China (0·036 million [UR 0·034–0·039], 19·7% [UR 18·0–21·5]). Malaria was an important cause in DR Congo and Nigeria, responsible for 0·036 million ([UR 0·017–0·062; 11·9% [UR 8·2–16·5]) and 0·102 million ([0·056–0·186]; 13·6% [10·2–17·8]) deaths, respectively.

rn Asia. The leading cause was congenital abnormalities in China (0·036 million [UR 0·034–0·039], 19·7% [UR 18·0–21·5]). Malaria was an important cause in DR Congo and Nigeria, responsible for 0·036 million ([UR 0·017–0·062; 11·9% [UR 8·2–16·5]) and 0·102 million ([0·056–0·186]; 13·6% [10·2–17·8]) deaths, respectively. The risk of dying in the first 5 years, the U5MR, ranged between 1·9 and 155·1 per 1000 livebirths among the 194 countries in 2015. The ten countries with the highest U5MR are all in sub-Saharan Africa and had U5MRs above 90 per 1000 livebirths. Three of these ten countries (Angola, Nigeria, and DR Congo) are among the ten countries with the most under-5 deaths mentioned above. In the remaining seven countries, the leading cause among under-5s was pneumonia in Benin, Central African Republic, Equatorial Guinea, Somalia, and Chad, and malaria in Mali and Sierra Leone (appendix p 13). Additional estimates of country-specific and cause-specific numbers of deaths are available in the appendix (p 14).

he remaining seven countries, the leading cause among under-5s was pneumonia in Benin, Central African Republic, Equatorial Guinea, Somalia, and Chad, and malaria in Mali and Sierra Leone (appendix p 13). Additional estimates of country-specific and cause-specific numbers of deaths are available in the appendix (p 14). Globally, more than 4 million (4·020 million) fewer under-5 deaths occurred in 2015 compared with in 2000. During this period, causes of child deaths changed gradually at the global level (appendix p 45). Although pneumonia and preterm birth complications were also leading causes of under-5 deaths in 2000 (1·738 million [UR 1·654–1·997]; 17·4% [UR 16·0–19·3] and 1·339 [1·169–1·464]; 13·4% [11·6–14·2], respectively), diarrhoea was replaced as the third leading cause in 2000 (1·213 million [1·115–1·451], 12·2% [11·0–14·3]) by intrapartum-related events in 2015. U5MR declined from 77·8 to 42·5 per 1000 livebirths over this period. Mortality rates for pneumonia, diarrhoea, neonatal intrapartum related events, malaria, and measles all reduced by more than 30% (figure 2). Collectively, reductions in these causes (21·7 less deaths per 1000 livebirths) were responsible for 61·6% of the total reduction in U5MR (35·3 less deaths per 1000 livebirths) in 2000–15.

rates for pneumonia, diarrhoea, neonatal intrapartum related events, malaria, and measles all reduced by more than 30% (figure 2). Collectively, reductions in these causes (21·7 less deaths per 1000 livebirths) were responsible for 61·6% of the total reduction in U5MR (35·3 less deaths per 1000 livebirths) in 2000–15. The global ARR of U5MR in 2000–15 was 4·0%, below the 4·4% required to achieve the MDG 4 in 1990–2015. All-cause neonatal mortality has been declining at a slower rate than that of children aged 1–59 months, at 3·1% versus 4·7% respectively. As a result, the fraction of neonatal deaths increased from 39·3% in 2000 to 45·1% in 2015. Nine causes achieved an ARR of at least 4·4% since 2000, with ARRs ranging from 13·1% [UR 6·8–16·1] for measles to 4·6% [3·4–5·4] for neonatal pneumonia (appendix p 46). By comparison, neonatal mortality due to congenital abnormalities only declined by 0·8% per year (UR 0·1–2·0).

ed from 39·3% in 2000 to 45·1% in 2015. Nine causes achieved an ARR of at least 4·4% since 2000, with ARRs ranging from 13·1% [UR 6·8–16·1] for measles to 4·6% [3·4–5·4] for neonatal pneumonia (appendix p 46). By comparison, neonatal mortality due to congenital abnormalities only declined by 0·8% per year (UR 0·1–2·0). Among the ten MDG regions, eastern Asia, which is composed mainly of China, saw the fastest reduction in U5MR in 2000–15 at an ARR of 8·2%. Mortality rates of almost all causes in this region have declined by at least 70% from 37·1 to 10·8 per 1000 livebirths (appendix p 49). The cause composition also changed, with congenital abnormalities (0·038 million [UR 0·036–0·041], 19·3% [UR 17·8–21·0]) and injuries (0·028 [0·027–0·030], 14·2% [13·0–15·3]) replacing pneumonia (0·024 [0·022–0·026], 20·1% [18·1–22·3]) and intrapartum-related events (0·028 million [0·023–0·030], 15·7% [14·2–17·3]) as the leading causes among under-5s in 2015 (appendix p 59). In sub-Saharan Africa, malaria, diarrhoea, and measles among children aged 1–59 months saw substantial reductions, contributing 18, 13, and 11 per 1000 livebirths to the reduction of U5MR, respectively (appendix p 55). The leading cause changed from malaria (0·699 million [UR 0·612–0·960], 16·4% [UR 13·5–20·7]%) in 2000 to pneumonia (0·490 million [0·417–0·631], 16·6% [14·8–19·1]%) in 2015 (appendix p 65). In southern Asia, U5MR due to preterm birth complications have only been declining at an ARR of 1·4% per annum (UR −0·3 to 2·2), a rate slower than the regional ARR of U5MR at 3·8%. As a result, the contribution of preterm birth complications to under-5 deaths has increased from 16·7% (UR [14·2–18·6], 0·598 million [0·497–0·664]) in 2000 to 25·3% ([21·7–28·7], 0·478 million [0·394–0·552]) in 2015. Despite a faster decline at an ARR of 5·4% [UR 3·7–7·3], pneumonia remained one of the leading killers of under-5 children in southern Asia, responsible for 0·687 million (0·626–0·771) deaths (19·2% [UR 17·7–21·6]) in 2000 and 0·285 million (0·215–0·372) deaths (15·1% [13·4–17·4]) in 2015. Additional regional and national trends can be viewed in the appendix (p 14).

·7–7·3], pneumonia remained one of the leading killers of under-5 children in southern Asia, responsible for 0·687 million (0·626–0·771) deaths (19·2% [UR 17·7–21·6]) in 2000 and 0·285 million (0·215–0·372) deaths (15·1% [13·4–17·4]) in 2015. Additional regional and national trends can be viewed in the appendix (p 14). Figure 3 shows the CSMFs and cause-specific mortality rates by six U5MR strata in 2015. Seven (Angola, Central African Republic, Chad, Mali, Nigeria, Sierra Leone, and Somalia) of the 194 countries fell in the very high mortality stratum. Collectively, they were responsible for a fifth (20·1%, 1·193 million of 5·942 million) of global under-5 deaths in 2015. Pneumonia, malaria, and diarrhoea were the leading causes in this stratum. The high, medium high, and medium strata each included 17, 20, and 36 countries. They were responsible for about a quarter (23·7%), a tenth (11·0%), and a third (34·9%) of global under-5 deaths in 2015, respectively. The leading causes in these three strata were very similar, being preterm birth complications, pneumonia, and intrapartum-related events. 54 countries were in the low mortality stratum. They contributed to 9% of the world's under-5 deaths. Their leading causes were congenital abnormalities, preterm birth complications, pneumonia, and intrapartum-related events. Another 60 countries fell in the very low mortality strata, responsible for 2% of the global under-5 deaths. The leading causes in this stratum were congenital abnormalities, preterm birth complications, and injuries. When moving from the very low to the very high mortality strata, the fractions and mortality rates of pneumonia, diarrhoea, and malaria increase. By contrast, the fractions of congenital abnormalities decrease, although its mortality rate is still the lowest in the very low mortality stratum (figure 3B).

and injuries. When moving from the very low to the very high mortality strata, the fractions and mortality rates of pneumonia, diarrhoea, and malaria increase. By contrast, the fractions of congenital abnormalities decrease, although its mortality rate is still the lowest in the very low mortality stratum (figure 3B). Discussion Child survival has improved substantially in the MDG era even though the targeted two-thirds reduction was not achieved.1, 20 This progress has been partly credited to the establishment of the MDGs, the ensuing increase in official development assistance, and the consequential scaling up of many life-saving interventions.20, 21, 22 However, the progress has been uneven and high levels of child mortality persist in many countries. In regard to cause-specific pace of reduction in 2000–15, measles and neonatal tetanus have seen tremendous progress, as have major causes such as diarrhoea and pneumonia though to a lesser degree. Injuries, neonatal preterm birth complications, neonatal intrapartum-related events, neonatal congenital abnormalities, and neonatal sepsis or meningitis are major causes among those with insufficient decline (ARR <4·4%) in 2000–15. To achieve the SDG child survival target, substantial progress is needed for these causes.

uries, neonatal preterm birth complications, neonatal intrapartum-related events, neonatal congenital abnormalities, and neonatal sepsis or meningitis are major causes among those with insufficient decline (ARR <4·4%) in 2000–15. To achieve the SDG child survival target, substantial progress is needed for these causes. Neonatal mortality has declined more slowly than that of the 1–59-month-olds. If neonatal causes had been declining at a rate achieved by the 1–59 month age group, the world would have attained the MDG 4 target ahead of time. Eastern Asia is an exception in that it managed to achieve faster decline among neonates than among older children. Case studies are valuable on how eastern Asia, primarily China, has achieved an impressive and balanced decline across age groups and causes.11 By contrast, sub-Saharan Africa had the largest disparity in progress between the two age groups, with neonatal survival only having been improving at an ARR of less than half of that of the 1–59 month age group. Five of the 12 post-neonatal causes achieved an ARR of at least 4·4%, yet only two of the eight neonatal causes did in this region.

Africa had the largest disparity in progress between the two age groups, with neonatal survival only having been improving at an ARR of less than half of that of the 1–59 month age group. Five of the 12 post-neonatal causes achieved an ARR of at least 4·4%, yet only two of the eight neonatal causes did in this region. Focusing on the high-burden regions, sub-Saharan African countries had a quarter of the world's livebirths in 2015.6 This figure is projected to increase to a third in 2030.6 Ensuring family planning needs of adolescent girls, women, and couples are satisfied with modern contraception is a key to reduce the number of child deaths in this region and globally in the next 15 years.21 Major infectious causes in sub-Saharan Africa had reductions in 2000–15, but infectious causes such as pneumonia, diarrhoea, malaria, and sepsis or meningitis remain important and should be a focus of child survival efforts going forward. Many neonatal causes and injuries have been declining more slowly in this region and interventions should be enhanced to address these conditions. In southern Asia, the contributions of preterm birth complications and congenital abnormalities to under-5 deaths have increased substantially in 2000–15. A major focus of child survival programmes in this region has to be on neonatal causes.

this region and interventions should be enhanced to address these conditions. In southern Asia, the contributions of preterm birth complications and congenital abnormalities to under-5 deaths have increased substantially in 2000–15. A major focus of child survival programmes in this region has to be on neonatal causes. Out of equity considerations, the SDG child survival target calls on all countries to reduce U5MR to 25 per 1000 livebirths or below by 2030. Country strategy formulation should consider their current U5MR and cause-of-death profile. When prioritising by child mortality strata, for the very high mortality countries, the focus should still be on the leading infectious causes, such as pneumonia, malaria, and diarrhoea. All of these can be addressed by highly effective and low cost preventive and therapeutic interventions, such as breastfeeding promotion, and Haemophilus influenzae type b and pneumococcal vaccines for pneumonia,23 improved water and sanitation, rotavirus vaccine, zinc supplementation, oral rehydration solutions, and community case management for diarrhoea,23 and insecticide treated bed-nets, intermittent preventive treatment in pregnancy, and artemisinin-based combination therapy for malaria.24 In addition, recent approval of a malaria vaccine holds potential for further malaria reduction, although major challenges exist for vaccine schedules.25

management for diarrhoea,23 and insecticide treated bed-nets, intermittent preventive treatment in pregnancy, and artemisinin-based combination therapy for malaria.24 In addition, recent approval of a malaria vaccine holds potential for further malaria reduction, although major challenges exist for vaccine schedules.25 Among countries with high, medium high, and medium child mortality, a clear child survival policy and programme focus should be to further invest in reducing deaths due to preterm birth complications, pneumonia, and intrapartum-related events. Relevant interventions for pneumonia are described above. Improved labour and delivery management is also important to reduce the causes of neonatal deaths.26 Antenatal corticosteroids and Kangaroo mother care are among the major recommended interventions to improve preterm birth outcomes.27 Neonatal resuscitation and comprehensive emergency obstetric care are among those recommended to reduce deaths due to intrapartum-related events.28, 29, 30

e causes of neonatal deaths.26 Antenatal corticosteroids and Kangaroo mother care are among the major recommended interventions to improve preterm birth outcomes.27 Neonatal resuscitation and comprehensive emergency obstetric care are among those recommended to reduce deaths due to intrapartum-related events.28, 29, 30 114 countries already have an U5MR of no more than 25 per 1000 livebirths in 2015.1 Effectively, they have achieved the SDG child survival target. However, this does not mean that they have homogeneous child cause-of-death profiles. Different from their higher mortality peers, countries in the low mortality stratum have congenital abnormalities as the leading cause. Clearly, congenital abnormalities should receive special attention in this stratum. Reducing the burden of congenital abnormalities will require better detection of some conditions and surgery for many. In addition to preterm birth complications and intrapartum-related events, pneumonia is still relatively important among low mortality countries compared with their very low mortality counterparts.

um. Reducing the burden of congenital abnormalities will require better detection of some conditions and surgery for many. In addition to preterm birth complications and intrapartum-related events, pneumonia is still relatively important among low mortality countries compared with their very low mortality counterparts. For countries in the very low mortality stratum, there is still room for improvement. Rapid reductions were seen between 2000 and 2015 for countries in this mortality stratum. For example, Portugal and Czech Republic had an U5MR around seven per 1000 livebirths in 2000. They both achieved an ARR of 4·5% for U5MR in 2000–15. Cause wise, they both had an accelerated decline, for example, in neonatal intrapartum-related events, which could be due to improved delivery care. These set examples for countries with very low child mortality to achieve rapid cause-specific reduction. In addition to congenital abnormalities, injuries also become increasingly important in this stratum. More rigorous research is needed to understand the epidemiology and effectiveness of injury interventions, such as barriers to prevent drowning, safer stoves to prevent burns, and car seats to prevent road traffic injury.31 Coordination and cooperation across sectors are essential. Injuries and other childhood conditions in settings of conflict and humanitarian crisis, such as Syria, should be a priority for global assistance.32

as barriers to prevent drowning, safer stoves to prevent burns, and car seats to prevent road traffic injury.31 Coordination and cooperation across sectors are essential. Injuries and other childhood conditions in settings of conflict and humanitarian crisis, such as Syria, should be a priority for global assistance.32 In addition to newly implemented vaccines, other new interventions could bear the potential to further improve child survival. New WHO guidelines on antibiotic management of neonatal infections have been released based on the results of the Simplified Antibiotic Therapy Trial.33, 34, 35, 36 These guidelines could further encourage community treatment and reduce mortality from neonatal infections. Based on preliminary results from the Strategic Timing of Antiretroviral Treatment study, which was done among adults,37 WHO revised its HIV treatment recommendation to “treat-all”, that is to treat anyone living with HIV with antiretroviral as soon as possible and to offer people at high risk with preventive antiretrovirals.38 This strategy could further avert under-5 HIV/AIDS deaths, although more evidence for children is needed. During the scale-up of antiretroviral treatment, precautions need be exercised to ensure that existing childhood interventions are not crowded out.39

to offer people at high risk with preventive antiretrovirals.38 This strategy could further avert under-5 HIV/AIDS deaths, although more evidence for children is needed. During the scale-up of antiretroviral treatment, precautions need be exercised to ensure that existing childhood interventions are not crowded out.39 We warn against attributing reductions of child cause-specific mortality rates to covariates used in our estimation due to circularity. For example, accelerated decline estimated in pneumonia, meningitis, and diarrhoea mortality, particularly in the past few years, is in part the result of directly taking account of the expected effect of Haemophilus influenzae type b vaccine, pneumococcal conjugate vaccine, and rotavirus virus vaccines in the post-hoc adjustment.

accelerated decline estimated in pneumonia, meningitis, and diarrhoea mortality, particularly in the past few years, is in part the result of directly taking account of the expected effect of Haemophilus influenzae type b vaccine, pneumococcal conjugate vaccine, and rotavirus virus vaccines in the post-hoc adjustment. Despite an increasing number of VA study data points and two additional countries now included as adequate VR countries, the data gap remains large for high burden countries and regions where 90% of under-5 deaths still occur in countries estimated by VAMCM, yet only 3% occur in countries with adequate VR in 2015 (appendix p 3). To improve internal validity of VA, innovative approaches such as the minimally invasive tissue sampling (MITS) can be applied.40 The establishment of the Child Health and Mortality Prevention Surveillance Network41 applying the MITS technique is welcome, although understanding the external validity of these approaches and estimates is crucial. In high burden countries where systematic collection of cause of death information lacks yet resources and technical capacity could become sustainably available, sample registration system should be attempted. Initiatives, such as the Countrywide Mortality Surveillance for Action by the Bill and Melinda Gates Foundation, can be useful efforts.

where systematic collection of cause of death information lacks yet resources and technical capacity could become sustainably available, sample registration system should be attempted. Initiatives, such as the Countrywide Mortality Surveillance for Action by the Bill and Melinda Gates Foundation, can be useful efforts. Our uncertainty ranges do not fully capture all the associated uncertainties. For example, we have not taken into consideration uncertainty associated with model inputs or covariates. The estimates for 2015 are particularly uncertain, because they were prepared using model inputs available up to the end of August, 2015. At the time, the most recent empirical estimates of model inputs were only available for up to year 2014, and those for 2015 were either assumed the same as those in 2014 or modelled through simplistic approaches. This could overestimate the burden of causes for which effective interventions are being rapidly scaled up. For example, in countries where pneumococcal conjugate vaccine and rotavirus vaccine are being rapidly scaled up, we might have overestimated the burden of pneumonia, meningitis, and diarrhoea in 2015 by assuming that the coverage of pneumococcal conjugate vaccine and rotavirus vaccine was the same in 2015 as in 2014. On the other hand, applying effectiveness and sometimes efficacy parameters from trials could have overestimated the life-saving effects of interventions, resulting in underestimated pneumonia, meningitis, and diarrhoea deaths. To what extent the above two sources of biases cancel each other out is unknown. Additional uncertainty is likely to be associated with U5MR and number of all-cause death estimates in the past few years as fewer empirical data points are available compared to earlier periods.42 Therefore, continued evaluation and reflection on MDG 4 beyond 2015 are important.

iases cancel each other out is unknown. Additional uncertainty is likely to be associated with U5MR and number of all-cause death estimates in the past few years as fewer empirical data points are available compared to earlier periods.42 Therefore, continued evaluation and reflection on MDG 4 beyond 2015 are important. In the next round, we plan to apply methods to better incorporate countries' transition from VAMCM to VRMCM, and those from VRMCM to VR. We will also investigate methods to better incorporate national empirical data recently collected from low-income and middle-income countries into our modelled estimates. Additional research is under way to better synthesise region-specific efficacy and effectiveness of pneumococcal conjugate vaccine as inputs to the post-hoc adjustment.

ears, and biennially thereafter up to and including age 74 years. Unless individuals have explicitly opted out of screening, all eligible people are sent invitation letters and screening information by their regional hub. Thus every individual scheduled to be invited during the study periods was eligible for inclusion. 8–10 days after the initial invitation letter, recipients are sent gFOBT kits and instructions. Individuals are asked to collect two samples from each of three separate bowel motions, and to return the completed kit to the regional hub, in a prepaid envelope, for processing. If kits are deemed to be spoilt, have a technical failure, or yield an unclear result, a repeat gFOBT kit is sent. If a recipient does not respond, a reminder letter is sent 4 weeks from the time of the initial invitation. If there is no response after a further 13 weeks, the individual's “screening episode” is closed for that period. If gFOBT yields an abnormal result, the person is referred to his or her local screening centre for diagnostic investigations.

igate methods to better incorporate national empirical data recently collected from low-income and middle-income countries into our modelled estimates. Additional research is under way to better synthesise region-specific efficacy and effectiveness of pneumococcal conjugate vaccine as inputs to the post-hoc adjustment. Much has been accomplished on child survival in 1990–2015, and particularly since 2000. However, accelerated investment in child survival is imperative post-2015 to achieve the SDG child survival target. US$25 billion have been pledged by governments over the next 5 years to improve the health of women, children, and adolescents.43 With the pledged resources and hopefully more to come, concerted efforts are needed across disease control and prevention programmes to maintain progress for countries which have accomplished rapid decline, and to accelerate progress for those that have had slower reductions in the past. Progress on child survival has benefited from a cohesive action plan in achieving the MDG 4 in the past two decades. United and continued actions are needed to achieve the SDG child survival target by 2030 and end preventable child deaths in a generation.44 For Maternal and Child Epidemiology Estimations see http://tinyurl.com/Hopkins-MNCH-cod-openaccess Supplementary Material Supplementary appendix

Much has been accomplished on child survival in 1990–2015, and particularly since 2000. However, accelerated investment in child survival is imperative post-2015 to achieve the SDG child survival target. US$25 billion have been pledged by governments over the next 5 years to improve the health of women, children, and adolescents.43 With the pledged resources and hopefully more to come, concerted efforts are needed across disease control and prevention programmes to maintain progress for countries which have accomplished rapid decline, and to accelerate progress for those that have had slower reductions in the past. Progress on child survival has benefited from a cohesive action plan in achieving the MDG 4 in the past two decades. United and continued actions are needed to achieve the SDG child survival target by 2030 and end preventable child deaths in a generation.44 For Maternal and Child Epidemiology Estimations see http://tinyurl.com/Hopkins-MNCH-cod-openaccess Supplementary Material Supplementary appendix Acknowledgments The study was supported by a grant from the Bill & Melinda Gates Foundation on Maternal and Child Epidemiology Estimation. Throughout the development of the estimates, technical inputs were provided by WHO staff, including Doris Ma Fat for causes of deaths in countries with adequate vital registration system, and Richard Cibulskis and Cristin Alexis Fergus for the estimates of Plasmodium falciparum parasite rate and general comments on malaria estimates. We thank Yipu Chen, Yujiang Chen, Mignote Solomon Haile, Hailun Liang, Ping Yeh, and Zihe Zheng for help with the systematic review of verbal autopsy studies used in the post-neonatal VAMCM; Yujiang Chen for her help in preparing some tables and figures for the paper; and all researchers who provided additional study information for their verbal autopsy studies.

lomon Haile, Hailun Liang, Ping Yeh, and Zihe Zheng for help with the systematic review of verbal autopsy studies used in the post-neonatal VAMCM; Yujiang Chen for her help in preparing some tables and figures for the paper; and all researchers who provided additional study information for their verbal autopsy studies. Contributors LL and YC did the analysis of the post-neonatal VAMCM and prepared estimates for China with help from JZ, and wrote the first draft of the paper. SO, JEL, and SC did the analysis of the neonatal VR data, VRMCM, and the VAMCM. DH prepared the national covariate time-series, did the analysis of post-neonatal VRMCM, and combined estimates from all models. JP did model selection for the post-neonatal models and generated uncertainty ranges for all estimates. REB and CM supervised all analyses. All co-authors provided feedback to the estimates and contributed to the subsequent versions of the manuscript. Declaration of interests We declare no competing interests. Figure 1 Global causes of under-5 deaths in 2015 Figure 2 Global trends in cause-specific mortality rates in neonates and children aged 1–59 months, 2000–15 *About 61% of the reduction comes from pneumonia, diarrhoea, malaria, and measles among 1-59-month olds and neonatal intrapartum related events. Figure 3 Cause-of-death mortality fractions (A) and cause-specific mortality rates (B) by U5MR strata, 2015

Figure 2 Global trends in cause-specific mortality rates in neonates and children aged 1–59 months, 2000–15 *About 61% of the reduction comes from pneumonia, diarrhoea, malaria, and measles among 1-59-month olds and neonatal intrapartum related events. Figure 3 Cause-of-death mortality fractions (A) and cause-specific mortality rates (B) by U5MR strata, 2015 U5MR=under-5 mortality rate. U5MR strata are defined as very low (<10 per 1000 livebirths), low (10–<25 per 1000 livebirths), medium (25–<50 per 1000 livebirths), medium high (50–<75 per 1000 livebirths), high (75–<100 per 1000 livebirths), and very high (≥100 per 1000 livebirths). Values less than 1 are not labelled. Table 1 New and total input data by estimation methods New input data Total input data Data points Deaths Countries Data points Deaths Countries Neonates VR 298 355 666 58 1628 1 621 273 66 VRMCM 674 772 904 64 2004 2 038 511 66 VAMCM 12 1897 5 124 100 119 37 1–59-month-olds VR 166 412 925 69 1104 2 043 763 69 VRMCM 147 288 535 58 1364 1 646 909 68 VAMCM 90 49 214 18 218 372 324 42 VR=vital registration. VRMCM=VR based multi-cause model. VAMCM=VA based multi-cause model. Table 2 Estimated numbers of deaths by cause and cause-specific mortality rate in 2015

ond, a reminder letter is sent 4 weeks from the time of the initial invitation. If there is no response after a further 13 weeks, the individual's “screening episode” is closed for that period. If gFOBT yields an abnormal result, the person is referred to his or her local screening centre for diagnostic investigations. Trial 1: gist leaflet The gist leaflet (appendix) was a simplified version of the screening information designed to be understood by readers with low literacy, numeracy, or both. This approach was informed by psychological theory21 and evidence gained predominately in the USA and in groups with low socioeconomic status or poor literacy. The data show that presenting complex information in simplified formats improves patients' satisfaction, comprehension, and decision making and leads to improved behavioural outcomes (eg, adherence to prescription regimens).14 During the development of this leaflet we were also mindful of guidelines on informed choice.22 We had previously done structured interviews to identify areas of the standard information booklet susceptible to being misunderstood and used this information to design the gist leaflet. The effects of the leaflet on comprehension were assessed in cognitive interviews and a trial in primary care, which showed improved comprehension compared with the standard information.19, 20

New input data Total input data Data points Deaths Countries Data points Deaths Countries Neonates VR 298 355 666 58 1628 1 621 273 66 VRMCM 674 772 904 64 2004 2 038 511 66 VAMCM 12 1897 5 124 100 119 37 1–59-month-olds VR 166 412 925 69 1104 2 043 763 69 VRMCM 147 288 535 58 1364 1 646 909 68 VAMCM 90 49 214 18 218 372 324 42 VR=vital registration. VRMCM=VR based multi-cause model. VAMCM=VA based multi-cause model. Table 2 Estimated numbers of deaths by cause and cause-specific mortality rate in 2015 Estimated number (UR; millions) Cause specific mortality rate (per 1000 livebirths) Children aged 0–59 months Preterm birth complications 1·055 (0·935–1·179) 7·556 (6·696–8·442) Pneumonia 0·921 (0·812–1·117) 6·594 (5·815–7·996) Intrapartum-related events 0·689 (0·598–0·778) 4·934 (4·283–5·569) Diarrhoea 0·526 (0·418–0·691) 3·768 (2·992–4·950) Sepsis/meningitis 0·517 (0·408–0·647) 3·699 (2·922–4·634) Congenital abnormalities 0·512 (0·455–0·606) 3·666 (3·256–4·338) Other conditions 0·841 (1·602–2·051) 6·020 (11·467–14·682) Neonates aged 0–27 days Preterm birth complications 0·943 (0·832–1·066) 6·753 (5·959–7·632) Intrapartum-related events 0·631 (0·550–0·723) 4·520 (3·937–5·177) Sepsis/meningitis 0·402 (0·280–0·522) 2·875 (2·005–3·739) Congenital abnormalities 0·305 (0·260–0·382) 2·183 (1·861–2·735) Pneumonia 0·161 (0·111–0·239) 1·154 (0·793–1·713) Tetanus 0·034 (0·018–0·084) 0·243 (0·129–0·600) Diarrhoea 0·018 (0·010–0·070) 0·126 (0·072–0·504) Other conditions 0·187 (0·142–0·240) 1·337 (1·017–1·719) Children aged 1–59 months Pneumonia 0·760 (0·651–0·943) 5·441 (4·661–6·752) Diarrhoea 0·509 (0·401–0·661) 3·642 (2·872–4·730) Injuries 0·326 (0·272–0·410) 2·337 (1·944–2·938) Malaria 0·306 (0·225–0·452) 2·193 (1·613–3·237) Congenital abnormalities 0·207 (0·165–0·259) 1·482 (1·178–1·851) Meningitis 0·115 (0·091–0·162) 0·824 (0·652–1·157) Preterm birth complications 0·112 (0·061–0·168) 0·802 (0·436–1·202) AIDS 0·086 (0·076–0·101) 0·614 (0·541–0·722) Measles 0·074 (0·038–0·268) 0·529 (0·274–1·920) Intrapartum-related events 0·058 (0·028–0·092) 0·414 (0·203–0·657) Pertussis 0·054 (0·053–0·060) 0·387 (0·377–0·427) Other conditions 0·654 (0·536–0·803) 4·683 (3·835–5·752) Uncertainty range (UR) is defined as the 2·5–97·5 centile· Other conditions among children aged 1–59 months included causes originated during the perinatal period, cancer, severe malnutrition, and other specified causes. Intrapartum-related events were formerly referred to as “birth asphyxia”.

Introduction Colorectal cancer is the fourth most common cause of cancer death worldwide,1 and the second most common in the UK.2 Screening by testing for occult blood in stools reduces mortality.3 In England, an organised colorectal cancer screening programme, the Bowel Cancer Screening Programme (BCSP), began in 2006, and offers guaiac faecal occult blood testing (gFOBT) every 2 years for people aged 60–74 years (previously up to 69 years). The UK cancer screening programmes are run by the National Health Service (NHS) with no financial costs to participants. This approach minimises inequity in the delivery of screening, but uptake for all the programmes shows a gradient by socioeconomic status.4, 5 The strongest gradient is for colorectal cancer screening: from the first 2·6 million gFOBT invitations in 2006–09, uptake was 61% in the least deprived quintile of residential areas and only 35% in the most deprived quintile.6

f screening, but uptake for all the programmes shows a gradient by socioeconomic status.4, 5 The strongest gradient is for colorectal cancer screening: from the first 2·6 million gFOBT invitations in 2006–09, uptake was 61% in the least deprived quintile of residential areas and only 35% in the most deprived quintile.6 Proposed explanations for reduced uptake of screening in more deprived groups include factors such as stress, low social support, and competing life demands.7 These factors are difficult to address through the screening programme. Literacy might also play an important part in uptake because information is delivered entirely through mailed written communications.8 Eligible adults are sent an invitation letter from their nearest regional screening hub, accompanied by a 13-page information booklet that covers complex issues such as risks and benefits (numerical information) that are designed to facilitate informed decision making. Comprehension of the screening offer, therefore, might be challenging in the most deprived areas in England, where up to half of people are either functionally illiterate or have only basic literacy and struggle with statistics.9 Research in Context Evidence before this study

Proposed explanations for reduced uptake of screening in more deprived groups include factors such as stress, low social support, and competing life demands.7 These factors are difficult to address through the screening programme. Literacy might also play an important part in uptake because information is delivered entirely through mailed written communications.8 Eligible adults are sent an invitation letter from their nearest regional screening hub, accompanied by a 13-page information booklet that covers complex issues such as risks and benefits (numerical information) that are designed to facilitate informed decision making. Comprehension of the screening offer, therefore, might be challenging in the most deprived areas in England, where up to half of people are either functionally illiterate or have only basic literacy and struggle with statistics.9 Research in Context Evidence before this study We searched the Ovid MEDLINE, PsycINFO, and Embase databases for reports on randomised controlled trials assessing interventions to increase uptake of cancer screening, published between 1980 and 2014, and addressing socioeconomic status. We used the search string “Neoplasm/ OR cancer OR neoplas* OR onco* OR carcinoma AND Mass screening/ OR screen* OR test OR detect* OR mass screening OR cancer screening AND Intervention studies/ OR intervention stud* OR stud* OR strateg* OR promot* OR initiative* OR behavio* OR behavio* change AND Patient acceptance of health care/ OR patient compliance/ OR attend* OR uptake OR utili?* OR particip* OR complian* OR accept* OR adher* AND breast neoplasms/ OR breast OR mammogra* OR uterine cervical neoplasms/ OR cervical OR cervix OR colorectal neoplasms/ OR colorectal OR bowel OR colon OR rectal OR CRC AND Healthcare Disparities/ OR Health Status Disparities/ OR disparit* OR education OR social class OR social status OR depriv* OR income OR socioeconomic OR socio economic”. The search retrieved 103 articles addressing socioeconomic inequality, but none used reduction of inequality as the primary endpoint.

ctal OR CRC AND Healthcare Disparities/ OR Health Status Disparities/ OR disparit* OR education OR social class OR social status OR depriv* OR income OR socioeconomic OR socio economic”. The search retrieved 103 articles addressing socioeconomic inequality, but none used reduction of inequality as the primary endpoint. Added value of this study

as of the standard information booklet susceptible to being misunderstood and used this information to design the gist leaflet. The effects of the leaflet on comprehension were assessed in cognitive interviews and a trial in primary care, which showed improved comprehension compared with the standard information.19, 20 Trial 2: narrative leaflet Narrative information is recognised as an effective communication aid for individuals with poor literacy.23 The narrative leaflet (appendix) was created on the basis of information obtained in interviews with people who had participated in the BCSP. Material was selected to reflect screening outcomes: a normal gFOBT result, polyp removal, and screen-detected colorectal cancer. We tested the efficacy of the leaflet in a trial in primary care, in which respondents reported being more inclined to take part in screening than with standard information.13

ctal OR CRC AND Healthcare Disparities/ OR Health Status Disparities/ OR disparit* OR education OR social class OR social status OR depriv* OR income OR socioeconomic OR socio economic”. The search retrieved 103 articles addressing socioeconomic inequality, but none used reduction of inequality as the primary endpoint. Added value of this study We identified no previous studies that used a randomised controlled design to test the efficacy of interventions designed specifically to reduce the socioeconomic status gradient in screening uptake without compromising overall uptake. Our trials assessed different ways to increase the visibility and salience of a national colorectal cancer screening programme, for which invitations are mailed every 2 years to all people aged 60–74 years in England. Because the goal was to identify interventions that could be easily implemented, all the trials were embedded in the routine call and recall system of the English Bowel Cancer Screening Programme (BCSP). This approach also eliminated any selection bias associated with a research study, and provided access to objective data on screening uptake. All four interventions were supported by evidence of efficacy in other contexts or pilot data obtained by our group. The gist and narrative leaflets, designed to provide information understandable for people with poor literacy, had no effects. By contrast, invitation letters showing general practice endorsement and enhanced reminder letters improved overall uptake. Enhanced reminder letters also significantly reduced the socioeconomic status gradient. The negative results with the leaflets might have been due to the increase in the total mass of information in the mailing. The changes that could be made to the official letters were limited. The effects of the enhanced reminder letter might, therefore, have been due to increased visibility because the reminder mailing included only one letter. These results highlight the importance of doing trials in a real-life context to discover what can be achieved with minimum intervention.

icial letters were limited. The effects of the enhanced reminder letter might, therefore, have been due to increased visibility because the reminder mailing included only one letter. These results highlight the importance of doing trials in a real-life context to discover what can be achieved with minimum intervention. Implications of all the available evidence Inequality in uptake is an important limitation of cancer screening programmes, even with systems that ensure equitable delivery of invitations or with home-based screening tests that avoid the barrier of clinic attendance. The main finding from the four ASCEND trials is that reduction in inequality is extremely difficult with use of downstream approaches in an organised screening programme. The significant reduction in socioeconomic status gradient achieved with the enhanced reminder letter might lead to improved equity at no further cost. Additionally, we found a high level of general practice support for the endorsement trial, which raised overall uptake, and, therefore, adding such endorsement to the reminder letter might increase this effect further. The negative results with two of our interventions highlight the challenges of communicating effectively with people with poor literacy who need to make decisions based on medical information.

which raised overall uptake, and, therefore, adding such endorsement to the reminder letter might increase this effect further. The negative results with two of our interventions highlight the challenges of communicating effectively with people with poor literacy who need to make decisions based on medical information. The screening hubs are required to use the standard information booklet, but each can provide a limited amount of extra material to improve presentation of the screening offer. Additionally, a few words can be added to the standard invitation and reminder letters. The ASCEND project was designed to test four different supplements to the screening information materials aimed at modifying inequality in screening uptake. Various studies have aimed to increase cancer screening uptake, but have been done across the whole population10, 11, 12 or in low-income groups,13 whereas ASCEND was specifically designed to assess the socioeconomic status gradient. The interventions tested in ASCEND had a strong theoretical rationale for use based on evidence of improving screening uptake in low socioeconomic status groups.14, 15 The materials developed for two of the trials had also been pretested in the early stages of the programme for effects on understanding and motivation.16, 17, 18, 19, 20 The four trials were embedded within the routine delivery of the screening programme. The protocol of this trial is available on the trial website.

15 The materials developed for two of the trials had also been pretested in the early stages of the programme for effects on understanding and motivation.16, 17, 18, 19, 20 The four trials were embedded within the routine delivery of the screening programme. The protocol of this trial is available on the trial website. Methods Study design and population We did four separate, two-arm, cluster-randomised controlled trials that involved individuals eligible to receive routine invitations from the BCSP for screening (Figure 1, Figure 2, Figure 3, Figure 4). The trial designs used time-defined cluster randomisation to usual care (standard information and letter [control]) or the intervention (supplemented information or letter). The interventions were designed to be implemented within the routine procedures of the BCSP, which are covered by Health and Social Care Information Centre (HSCIC) approval in relation to patient-identifiable data. We obtained ethics approval for all trials from the National Research Ethics Service Committee London-Harrow.

interventions were designed to be implemented within the routine procedures of the BCSP, which are covered by Health and Social Care Information Centre (HSCIC) approval in relation to patient-identifiable data. We obtained ethics approval for all trials from the National Research Ethics Service Committee London-Harrow. At any time, about 9 million adults in England are eligible for gFOBT through the BCSP. The programme is coordinated by five regional screening hubs. All five hubs were included in the study. People in each region and registered with a general practitioner are eligible for screening from age 60 years, and biennially thereafter up to and including age 74 years. Unless individuals have explicitly opted out of screening, all eligible people are sent invitation letters and screening information by their regional hub. Thus every individual scheduled to be invited during the study periods was eligible for inclusion.

d in the BCSP. Material was selected to reflect screening outcomes: a normal gFOBT result, polyp removal, and screen-detected colorectal cancer. We tested the efficacy of the leaflet in a trial in primary care, in which respondents reported being more inclined to take part in screening than with standard information.13 Trial 3: general practice endorsement International evidence shows that screening invitations sent by individuals' family doctors improve uptake in groups with low socioeconomic status.11, 14, 15, 16 BCSP invitation letters are sent from the hubs, but there is space on the letter to mention the support of the general practice with which an individual is registered, although not the named general practitioner. We created a general practice endorsement that appeared as a banner across the invitation letter (appendix). We sought consent from all practices in England (n=8142), in collaboration with a primary care advisory group and HSCIC, by sending each one a written invitation to be part of the trial, followed up by reminders 4 and 8 weeks later. Permission to link the practice address to the invitation was granted by 6480 (80%) of practices.

consent from all practices in England (n=8142), in collaboration with a primary care advisory group and HSCIC, by sending each one a written invitation to be part of the trial, followed up by reminders 4 and 8 weeks later. Permission to link the practice address to the invitation was granted by 6480 (80%) of practices. Trial 4: enhanced reminder letter Reminders have a slight impact on uptake, including in very-low-income groups.13, 15 They also provide an opportunity to restate the screening offer. We created an enhanced reminder letter that was aimed specifically at individuals who had not responded to the initial invitation. A simple restatement of the screening offer was made in a short paragraph added to the end of the standard letter, and a banner was added to the start of the letter that said “A reminder to you” (appendix). The offer restatement text had been refined through a series of focus groups and a review of reminder-related queries made by patients to the BCSP's telephone helpline.

a short paragraph added to the end of the standard letter, and a banner was added to the start of the letter that said “A reminder to you” (appendix). The offer restatement text had been refined through a series of focus groups and a review of reminder-related queries made by patients to the BCSP's telephone helpline. Randomisation Randomisation was based on day of invitation, with “day within hub” constituting the randomisation unit. Trials 1 and 2 were run over 10 consecutive days in November, 2012, and March, 2013, respectively. Trials 3 and 4 were run over 20 consecutive days in June to July, 2013, and July to August, 2013, respectively. 2 weeks before the start of each intervention, a random number sequence was generated with a continuous random number for each hub day, and numbers higher than the median were allocated to the intervention groups and those lower than the median to the control groups. For trials 1 and 2, randomisation schedules were created and sent to a printing company (Real Digital International, Croydon, UK), for the Southern, London, and Eastern hubs, and to the in-house invitation service for the North East and Midlands and North West hubs. For trials 3 and 4, randomisation was done through the Bowel Cancer Screening System, which identifies the eligible population for screening in each hub. For trial 3 this system was modified to enable selection of invitees who belonged to general practices that had agreed to endorse the BCSP before the creation of the invitation letters. Although masking of hubs was not possible, there was no direct contact between hub staff sending the invitations or reminders and invitees. Each study group was unaware of the materials received by the other study groups unless members of the same household received different invitation types or a person had been invited previously.

ng of hubs was not possible, there was no direct contact between hub staff sending the invitations or reminders and invitees. Each study group was unaware of the materials received by the other study groups unless members of the same household received different invitation types or a person had been invited previously. Outcome measures The primary outcome was socioeconomic status gradient in screening uptake over quintiles of Index of Multiple Deprivation (IMD). We defined screening uptake as the proportion of routinely invited individuals that returned gFOBT kits within 18 weeks of being sent invitations that led to an result of normal or abnormal (with clinical referral for prosepctive colonoscopy) by the date of data extraction (18 weeks after the last day of the intervention). We used the English IMD 2010 score associated with home postcodes to classify socioeconomic status.24 IMD is an area-based measure that combines income, employment, health and disability, education, skills and training, barriers to housing and services, crime, and living environment, into a deprivation score. The scores are assigned to small geographical areas termed lower-layer super output areas, of which there are 32 844 in England, each covering about 1500 individuals. Each recipient's postcode was linked to the relevant lower-layer super output area. The IMD scores were grouped into quintiles based on national distributions with use of predefined national cutoffs.

areas termed lower-layer super output areas, of which there are 32 844 in England, each covering about 1500 individuals. Each recipient's postcode was linked to the relevant lower-layer super output area. The IMD scores were grouped into quintiles based on national distributions with use of predefined national cutoffs. The age and sex of each recipient were obtained from the BCSP database. We gathered information on whether each individual was being invited for the first time (prevalent first-time episode), being sent a biennial invitation having previously not responded (prevalent episode), or being sent a biennial invitation having been screened before (incident episode). Secondary outcomes were the median number of days to return the gFOBT kit and the proportion of spoilt and undelivered test kits, by intervention and IMD quintile. We also assessed the marginal cost per additional person receiving each intervention, calculated with actual costs incurred during the study and valued according to market prices.

the median number of days to return the gFOBT kit and the proportion of spoilt and undelivered test kits, by intervention and IMD quintile. We also assessed the marginal cost per additional person receiving each intervention, calculated with actual costs incurred during the study and valued according to market prices. Statistical analysis Target sample sizes for each trial were estimated to detect an average increase in uptake of 3 percentage points, based on 1 percentage point increase in the least deprived quintile and 5 percentage points in the most deprived quintile, with 90% power and p<0·05. The final calculation was based on the demographic composition of the hub that required the largest sample size (Midlands and North West). Because invitations were randomised by day, but the number of invitations sent per day varies, we applied an inflation factor of 1·7 to ensure that the sample size would confer adequate statistical power. We calculated, therefore, that 46 000 individuals (23 000 per group) would be needed for each of trials 1 and 2. However, due to the volume of invitations sent out each week, this target would have been achieved within 5 days of invitations, and because small numbers of clusters increase the risk of bias25 we had specified 10-day intervention periods. The final sample size for each of these trials, therefore, was 140 000–160 000 individuals. The estimated sample size for trial 3 (84 000) assumed agreement from 30% of practices and, therefore, the required sample size was increased by a factor of 100/30 to a target of 280 000. To achieve this sample size would need 14 days of sending invitations, but we had allowed 20 days. The target sample size for trial 4 was 140 000, which reflected the fact that the daily number of reminders is substantially lower than first invitations.

e size was increased by a factor of 100/30 to a target of 280 000. To achieve this sample size would need 14 days of sending invitations, but we had allowed 20 days. The target sample size for trial 4 was 140 000, which reflected the fact that the daily number of reminders is substantially lower than first invitations. The intervention periods for trials 3 and 4 overlapped because of initial interest in using a factorial design to investigate the combined effect of adding the general practice endorsement to the enhanced reminder letter as well as the initial invitation. Owing to space constraints on the reminder letter, however, we could not proceed with this plan, but it meant that some individuals in trial 3 who did not respond to their invitation within 28 days were included in trial 4. The primary outcome was analysed by logistic regression. Odds ratios (ORs), p values, and 95% CIs were calculated with conservative variance estimation to allow for the potential clustering effects, and were controlled for hub, age, sex, and screening round.25, 26 The conservative variance analysis allowed for correlation of individuals within randomisation clusters but not between clusters, and used the Huber-White information sandwich method to estimate variance.27, 28 The primary outcome was tested by the two-factor interaction term between intervention group and IMD quintile. Analyses were done on an intention-to-treat basis. Analyses were done with SAS version 9.3 and Stata version 12.1. This study is registered, number ISRCTN74121020.

n sandwich method to estimate variance.27, 28 The primary outcome was tested by the two-factor interaction term between intervention group and IMD quintile. Analyses were done on an intention-to-treat basis. Analyses were done with SAS version 9.3 and Stata version 12.1. This study is registered, number ISRCTN74121020. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Baseline characteristics were well balanced for each trial and showed that the populations were representative of that served by the BCSP (table 1).6 Overall uptake per study was 57·4%, 57·7%, 57·9%, and 25·4% in trials 1, 2, 3, and 4, respectively; the proportion is low in trial 4 because it only targeted individuals who had not responded within 4 weeks of the invitation letter. In all trials, uptake was strongly and negatively associated with deprivation, with the difference between the least and most deprived quintiles in each control arm ranging from 20 to 24 percentage points (table 2).

al 4 because it only targeted individuals who had not responded within 4 weeks of the invitation letter. In all trials, uptake was strongly and negatively associated with deprivation, with the difference between the least and most deprived quintiles in each control arm ranging from 20 to 24 percentage points (table 2). In trial 1, uptake was similar in the intervention and control groups and the least and the most deprived quintiles (table 2). The socioeconomic status gradient in screening uptake did not differ by IMD quintile and no significant increase was seen in overall uptake (table 3, appendix). The median number of days to return the test kit was 23 (range 12–126) in the intervention group and 22 (11–126) in the control group. Median response times did not differ by IMD quintile (appendix). The proportions of spoilt and undelivered test kits were very small and were similar in the two groups and across IMD quintiles (appendix). For trial 2, uptake did not differ between groups or between least and most deprived quintiles (table 2). The socioeconomic status gradient in screening uptake did not differ across deprivation quintiles and no effect was seen on overall uptake (table 3). The median number of days to return the test kit was 26 (range 11–126) in the intervention group and 26 (10–126) in the control group, and timing did not differ by IMD quintile (appendix). The proportions of spoilt and undelivered test kits were similar in the two study groups and across IMD quintiles (appendix).

e 3). The median number of days to return the test kit was 26 (range 11–126) in the intervention group and 26 (10–126) in the control group, and timing did not differ by IMD quintile (appendix). The proportions of spoilt and undelivered test kits were similar in the two study groups and across IMD quintiles (appendix). In trial 3, general practice endorsement was associated with a slight percentage point differential in uptake between the least and most deprived quintiles (table 2). We also noted a slight gradient in uptake related to socioeconomic status, but this effect was not significant (table 3). Although the unadjusted OR indicated little effect on overall uptake (appendix), the effect became significant after adjustment (table 3). This change is mainly due to differences in effect sizes between study groups for screening episode (table 2). The median number of days taken to return the test kit was 22 (range 8–126) for the intervention group and 23 (11–126) for the control group, and timings were similar across IMD quintiles (appendix). The proportions of spoilt and undelivered test kits were similar in the two study groups and across IMD quintiles (appendix).

edian number of days taken to return the test kit was 22 (range 8–126) for the intervention group and 23 (11–126) for the control group, and timings were similar across IMD quintiles (appendix). The proportions of spoilt and undelivered test kits were similar in the two study groups and across IMD quintiles (appendix). In trial 4, the enhanced reminder letter was associated with a difference in uptake between the intervention and control groups and the lowest and highest deprivation quintiles (table 2). We found a significant interaction between uptake and IMD quintile, with a stronger effect seen in the most deprived than in the least deprived quintile (table 3). The unadjusted OR for overall uptake did not differ between study groups, but the effect became significant after adjustment (table 3). The median number of days to return test kits was 11 (range −4 to 89) in the intervention group and 11 (0–89) in the control group, and did not differ across IMD quintiles (appendix). The proportions of spoilt and undelivered test kits also did not differ (appendix).

ffect became significant after adjustment (table 3). The median number of days to return test kits was 11 (range −4 to 89) in the intervention group and 11 (0–89) in the control group, and did not differ across IMD quintiles (appendix). The proportions of spoilt and undelivered test kits also did not differ (appendix). Owing to the overlap in timing, the inclusion and randomisation statuses of trials 3 and 4 have been cross-tabulated (appendix). A larger proportion of recipients in the trial 3 intervention group was randomised to trial 4 than in the control group (49·4% vs 44·6%). Nevertheless, the unadjusted OR for participation within 4 weeks associated with general practice endorsement (before the reminder could have been received) was 1·06 (95% CI 0·99–1·04), which was higher than that for overall uptake (1·03, 0·95–1·11; appendix). Similarly, the unadjusted OR associated with the enhanced reminder letter for recipients who were not enrolled in trial 3 was 1·06 (95% CI 0·93–1·21), which is also higher than the unadjusted OR for overall uptake (1·04, 0·95–1·14; appendix). Furthermore, the OR associated with receiving the enhanced reminder letter adjusted for trial 3 status was 1·04 (95% CI 0·95–1·14), which matched the unadjusted OR.

s who were not enrolled in trial 3 was 1·06 (95% CI 0·93–1·21), which is also higher than the unadjusted OR for overall uptake (1·04, 0·95–1·14; appendix). Furthermore, the OR associated with receiving the enhanced reminder letter adjusted for trial 3 status was 1·04 (95% CI 0·95–1·14), which matched the unadjusted OR. The average marginal costs of providing the gist and narrative leaflets were, respectively, £0·04 and £0·05 per person screened. For the general practice endorsement and enhanced reminder letters, a one-off cost of £78 000 was incurred to modify both in the BCSP IT system. As this cost would not be incurred again if the interventions were implemented, there was no marginal cost per person screened.

pectively, £0·04 and £0·05 per person screened. For the general practice endorsement and enhanced reminder letters, a one-off cost of £78 000 was incurred to modify both in the BCSP IT system. As this cost would not be incurred again if the interventions were implemented, there was no marginal cost per person screened. Discussion Reducing socioeconomic inequalities in cancer mortality is a priority worldwide. Cancer screening is a major component of efforts to bring forward diagnosis to earlier, more treatable stages. Even in the UK, where screening incurs no financial cost to the individual, uptake declines with increasing socioeconomic deprivation.4, 5, 6 Our four trials enabled assessment of interventions designed to lessen inequalities in uptake in large study populations. An important strength of ASCEND was that the trials were powered to measure the effects of interventions in relation to socioeconomic status in the total eligible population, rather than merely focusing on disadvantaged groups. The interventions, therefore, had the potential to reach a large number of people who had not previously participated in the screening programme. Use of routinely collected data enabled us to include most of the potential study population in our analysis, except for a very small group of people without IMD scores for their postcodes. Each intervention was also based on a well established rationale and empirical data, and was developed through a structured, comprehensive process. Only the enhanced reminder letter, however, led to a reduction in the socioeconomic status gradient in uptake. The gist and narrative leaflet interventions had no effect on uptake. The general practice endorsed letter was associated with increased uptake overall, but did not modify the socioeconomic status gradient. None of our interventions promoted early response or a reduced number of spoilt test kits. The numbers of undelivered information packs also did not differ, by group or IMD quintile.

The general practice endorsed letter was associated with increased uptake overall, but did not modify the socioeconomic status gradient. None of our interventions promoted early response or a reduced number of spoilt test kits. The numbers of undelivered information packs also did not differ, by group or IMD quintile. The gist and narrative leaflets in trials 1 and 2 were designed to make the offer of screening more visible to people with poor literacy skills. Both leaflets showed this potential when their effects were assessed on the basis of knowledge, attitudes, or intention to participate in screening.16, 20 A possible explanation for lack of effect in these trials is that the determinants of intention can differ from the determinants of action, and that the leaflets only affected the former. Another possible explanation relates to the fact that the two leaflets had to be added to the existing invitation or information rather than being provided as an alternative. Consequently, although the leaflets were designed to be simple, they increased the total mass of written material and might have undermined the goal of making the screening offer more visible.

t that the two leaflets had to be added to the existing invitation or information rather than being provided as an alternative. Consequently, although the leaflets were designed to be simple, they increased the total mass of written material and might have undermined the goal of making the screening offer more visible. The general practice endorsement significantly increased overall uptake, but the effect size was smaller than in many previous studies.10, 11, 12, 13 This difference was probably due to previous studies mostly using letters sent directly from the general practitioner or with the individual doctors' signatures on the letters. We were unable to apply such alterations for logistical reasons, which might have diluted the efficacy of this intervention. Owing to general practice endorsement having previously shown effects in low-income groups,13 we had hypothesised that the effect in this study would have been stronger in lower than in higher socioeconomic groups. Previous studies, however, had not been powered to assess effects on the socioeconomic gradient. The large size of the ASCEND trial, though, means that our negative result is definitive, at least with the format of endorsement that we used. Nevertheless, in view of the high level of agreement by practices to endorse the screening programme and the absence of a marginal cost per person screened, we recommend that the BCSP considers adding the general practice endorsement banner to screening invitation letters.

t with the format of endorsement that we used. Nevertheless, in view of the high level of agreement by practices to endorse the screening programme and the absence of a marginal cost per person screened, we recommend that the BCSP considers adding the general practice endorsement banner to screening invitation letters. One intervention that reduced the socioeconomic gradient was the enhanced reminder letter, which also slightly increased overall uptake. The aim of this intervention was to offer anyone who had not engaged with the original materials an additional opportunity to see and consider the screening offer. Unlike the gist and narrative leaflets, this enhancement was incorporated into the one-page reminder letter and, therefore, might have had higher visibility. Although the change in the gradient was small (as was the effect on overall uptake), this intervention was also virtually cost-free and, therefore, offers a practical way for the screening programme to reduce the socioeconomic gradient in uptake. As this addition to the reminder letter was minor, investigating the cost-effectiveness of adding a second reminder or using alternative channels, such as text messaging, to reiterate the offer of screening could be worthwhile.

a practical way for the screening programme to reduce the socioeconomic gradient in uptake. As this addition to the reminder letter was minor, investigating the cost-effectiveness of adding a second reminder or using alternative channels, such as text messaging, to reiterate the offer of screening could be worthwhile. ASCEND had some limitations. People from deprived backgrounds are likely to be struggling with multiple social and economic challenges, making it difficult for them to prioritise cancer screening. These upstream issues, however, cannot be addressed by minor variations in the format of a screening offer. Nonetheless, ensuring that the screening offer is not only mailed to all eligible adults but is also appropriate for a wide range of levels of literacy should be a goal of NHS screening programmes. We did not address broader attitudes to cancer. For example, cancer fatalism and other negative attitudes are more prevalent in groups with low than with high socioeconomic status, and fatalism has been associated with delayed diagnosis.29 Negative attitudes are not easily modified with simple written materials. We did not address various other downstream barriers, of which the most well established is the unpleasantness associated with completing the test kit. If the BCSP implements the faecal immunochemical test for haemoglobin, which typically only requires one stool sample, inequalities in participation might be reduced.30, 31

id not address various other downstream barriers, of which the most well established is the unpleasantness associated with completing the test kit. If the BCSP implements the faecal immunochemical test for haemoglobin, which typically only requires one stool sample, inequalities in participation might be reduced.30, 31 The sampling timeframe for trials 3 and 4 overlapped because the original plan had been that individuals randomised to receive the general practice endorsement letter in trial 3 would have a similar banner on their reminder letters in trial 4. This approach, however, was logistically impossible because of space limitations on the page. Thus, we analysed the two trials separately but did supplementary analyses to test whether the overlap had resulted in overestimation of effects for either intervention. We found no evidence of bias and, therefore, conclude that the overlap, although not desirable, did not compromise our results. The inclusion of strategies in routine programme delivery provides a model for future research, and there might be scope to test changes to the interventions that could strengthen the effects. For instance, supplying all the necessary screening information in smaller instalments by integrating additional communication points into the screening pathway might improve the visibility and efficacy of the gist and narrative leaflets. The use of additional reminder letters might, through a process of elimination, help to target the most deprived populations.

reening information in smaller instalments by integrating additional communication points into the screening pathway might improve the visibility and efficacy of the gist and narrative leaflets. The use of additional reminder letters might, through a process of elimination, help to target the most deprived populations. In conclusion, the enhanced reminder letter was the only strategy to significantly reduce the socioeconomic gradient, and overall uptake was only increased by this and the general practice endorsement intervention. In view of the very low expense, these interventions could be implemented with minimum cost or disruption to the existing programme. Our findings suggest that tailoring of information delivery to the communities being served might be useful. A possibility in poor-literacy groups is to supplement mailed information with direct contact with health professionals. The results of our four trials illustrate the difficulty of addressing inequality in screening uptake within an organised programme, but highlight the importance of continuing to investigate new strategies. Supplementary Material Supplementary appendix

In conclusion, the enhanced reminder letter was the only strategy to significantly reduce the socioeconomic gradient, and overall uptake was only increased by this and the general practice endorsement intervention. In view of the very low expense, these interventions could be implemented with minimum cost or disruption to the existing programme. Our findings suggest that tailoring of information delivery to the communities being served might be useful. A possibility in poor-literacy groups is to supplement mailed information with direct contact with health professionals. The results of our four trials illustrate the difficulty of addressing inequality in screening uptake within an organised programme, but highlight the importance of continuing to investigate new strategies. Supplementary Material Supplementary appendix Acknowledgments We dedicate this Article in memory of Professor Jane Wardle (1950–2015). This study is funded by the National Institute for Health Research (NIHR) Programme Grants for Applied Research Programme (RP-PG-0609-10106) and partly supported by the NIHR Collaboration for Leadership in Applied Health Research and Care (CLAHRC) North Thames (MCT and RR). We thank the managers of the screening hubs and staff at Real Digital International, Croydon, UK, for their oversight and management of the interventions. We thank Paul Greliak and Cecily Palmer for their contributions to the trial and project management, respectively, and Helen Seaman for her comments on earlier drafts of this paper. We thank the Health and Social Care Information Centre for guidance on formatting of the general practice endorsement and enhanced reminder letters. We also thank our community partners (Age UK, Community Health and Learning Foundation, Beating Bowel Cancer, Social Action for Health) and our patients' representatives for their help with developing and pretesting the intervention materials. This is a summary of independent research funded by the NIHR. The views expressed are those of the authors and not necessarily those of the National Health Service, NIHR, or UK Department of Health.

n for Health) and our patients' representatives for their help with developing and pretesting the intervention materials. This is a summary of independent research funded by the NIHR. The views expressed are those of the authors and not necessarily those of the National Health Service, NIHR, or UK Department of Health. Contributors JW and RR were joint principal investigators and wrote the grant application in collaboration with CvW, SPH, GH, RFL, SR, SS, SWD, AH, SM, and WA. JW, CVW, WA, SWD, SPH, and RR designed the study. JW, CvW, IK-H, SGS, LMM, GV, MCT, WA, and RR led the development and testing of the four interventions. SWD generated the randomisation codes. SPH, GH, RFL, SR, and SS were responsible for identifying eligible individuals and delivering the intervention, assisted by IK-H and RH. IK-H and JS led the data extraction. NC, SWD, AH, and SM analysed the data. JW and CvW drafted the report and all authors contributed to reviewing and revision. All authors approved the final version. CvW and RR will act as guarantors. Declaration of interests SPH, JS, WA, and RR have received grants from the National Institute of Health Research. SM has received a grant from Queen Mary University, London. RR has received a grant from the Health Foundation and Legal and General and the Dunhill Medical. All other authors declare no competing interests. Figure 1 Trial 1 profile (gist leaflet) IMD=Index of Multiple Deprivation. Figure 2 Trial 2 profile (narrative leaflet) IMD=Index of Multiple Deprivation. Figure 3 Trial 3 profile (general practice endorsement) IMD=Index of Multiple Deprivation.

Declaration of interests SPH, JS, WA, and RR have received grants from the National Institute of Health Research. SM has received a grant from Queen Mary University, London. RR has received a grant from the Health Foundation and Legal and General and the Dunhill Medical. All other authors declare no competing interests. Figure 1 Trial 1 profile (gist leaflet) IMD=Index of Multiple Deprivation. Figure 2 Trial 2 profile (narrative leaflet) IMD=Index of Multiple Deprivation. Figure 3 Trial 3 profile (general practice endorsement) IMD=Index of Multiple Deprivation. Figure 4 Trial 4 profile (enhanced reminder letter) IMD=Index of Multiple Deprivation. Table 1 Baseline characteristics of trial population

IMD=Index of Multiple Deprivation. Figure 2 Trial 2 profile (narrative leaflet) IMD=Index of Multiple Deprivation. Figure 3 Trial 3 profile (general practice endorsement) IMD=Index of Multiple Deprivation. Figure 4 Trial 4 profile (enhanced reminder letter) IMD=Index of Multiple Deprivation. Table 1 Baseline characteristics of trial population Trial 1 Trial 2 Trial 3 Trial 4 Standard information and gist leaflet (n=84 283) Standard information (n=78 791) Standard information and narrative leaflet (n=73 450) Standard information (n=76 421) GP endorsed invitation letter (n=130 876) Standard letter (n=133 449) Enhanced reminder letter (n=77 739) Standard letter (n=90 002) Age (years) 66 (59–74) 66 (59–74) 65 (59–74) 65 (59–74) 65 (59–74) 65 (59– 74) 65 (59–74) 64 (59–74) Sex Female 43 195 (51·2%) 40 671 (51·4%) 37 937 (51·5%) 39 086 (51·0%) 66 986 (51·0%) 68 591 (51·2%) 37 747 (48·4%) 43 574 (48·2%) Male 41 226 (48·8%) 38 433 (48·6%) 35 785 (48·5%) 37 609 (49·0%) 64 437 (49·0%) 65 420 (48·8%) 40 320 (51·6%) 46 839 (51·8%) Screening episode Prevalent first time 13 034 (15·4%) 12 410 (15·7%) 15 281 (20·7%) 12 510 (16·3%) 22 287 (17·0%) 23 582 (17·6%) 14 483 (18·6%) 21 271 (23·5%) Prevalent 26 368 (31·2%) 24 551 (31·0%) 22 209 (30·1%) 22 892 (29·8%) 40 441 (30·8%) 40 295 (30·1%) 39 862 (51·1%) 43 329 (47·9%) Incident 45 019 (53·3%) 42 143 (53·3%) 36 232 (49·1%) 41 293 (53·8%) 68 695 (52·3%) 70 134 (52·3%) 23 722 (30·4%) 25 813 (28·6%) Programme administrative hub 1 22 469 (26·6%) 24 369 (30·8%) 21 421 (29·1%) 21 118 (27·5%) 35 993 (27·4%) 34 598 (25·8%) 22 051 (28·2%) 25 490 (28·2%) 2 20 651 (24·5%) 21 004 (26·6%) 20 667 (28·0%) 16 723 (21·8%) 31 760 (24·2%) 40 550 (30·3%) 19 131 (24·5%) 23 107 (25·6%) 3 7416 (8·8%) 6636 (8·4%) 8509 (11·5%) 8795 (11·5%) 11 818 (9·0%) 13 255 (9·9%) 10 809 (13·8%) 10 385 (11·5%) 4 13 614 (16·1%) 12 858 (16·3%) 13 053 (17·7%) 12 900 (16·8%) 21 272 (16·2%) 21 439 (16·0%) 12 291 (15·7%) 12 796 (14·2%) 5 20 271 (24·0%) 14 237 (18·0%) 10 072 (13·7%) 17 159 (22·4%) 30 580 (23·3%) 24 169 (18·0%) 13 785 (17·7%) 18 635 (20·6%) IMD quintile* 1 (least deprived) 19 055 (22·6%) 18 554 (23·5%) 17 027 (23·2%) 17 073 (22·3%) 30 350 (23·1%) 31 381 (23·4%) 15 933 (20·4%) 18 928 (20·9%) 2 19 787 (23·5%) 18 295 (23·2%) 16 517 (22·5%) 17 675 (23·1%) 30 952 (23·6%) 31 340 (23·4%) 16 594 (21·3%) 19 446 (21·5%) 3 18 320 (21·7%) 15 993 (20·3%) 15 287 (20·8%) 16 161 (21·1%) 27 950 (21·3%) 28 181 (21·0%) 16 092 (20·6%) 18 286 (20·2%) 4 14 747 (17·5%) 13 469 (17·1%) 12 897 (17·6%) 13 385 (17·5%) 22 450

20·9%) 2 19 787 (23·5%) 18 295 (23·2%) 16 517 (22·5%) 17 675 (23·1%) 30 952 (23·6%) 31 340 (23·4%) 16 594 (21·3%) 19 446 (21·5%) 3 18 320 (21·7%) 15 993 (20·3%) 15 287 (20·8%) 16 161 (21·1%) 27 950 (21·3%) 28 181 (21·0%) 16 092 (20·6%) 18 286 (20·2%) 4 14 747 (17·5%) 13 469 (17·1%) 12 897 (17·6%) 13 385 (17·5%) 22 450 (17·1%) 23 007 (17·2%) 14 679 (18·8%) 16 853 (18·6%) 5 (most deprived) 12 374 (14·7%) 12 660 (16·0%) 11 722 (16·0%) 12 127 (15·9%) 19 174 (14·6%) 19 540 (14·6%) 14 441 (18·5%) 16 489 (18·2%) Score missing 138 133 272 274 547 562 328 411 Data are median (range) or number (%). GP=general practice. IMD=Index of Multiple Deprivation. * Quintile categories were based on national cutoffs.26 Table 2 Proportions of people who took up screening

(17·1%) 23 007 (17·2%) 14 679 (18·8%) 16 853 (18·6%) 5 (most deprived) 12 374 (14·7%) 12 660 (16·0%) 11 722 (16·0%) 12 127 (15·9%) 19 174 (14·6%) 19 540 (14·6%) 14 441 (18·5%) 16 489 (18·2%) Score missing 138 133 272 274 547 562 328 411 Data are median (range) or number (%). GP=general practice. IMD=Index of Multiple Deprivation. * Quintile categories were based on national cutoffs.26 Table 2 Proportions of people who took up screening Number (%) in trial 1 Number (%) in trial 2 Number (%) in trial 3 Number (%) in trial 4 Standard information and gist leaflet (n=84 283) Standard information (n=78 791) Standard information and narrative leaflet (n=73 450) Standard information (n=76 421) GP endorsed invitation letter (n=130 876) Standard letter (n=133 449) Enhanced reminder letter (n=77 739) Standard letter (n=90 002) Adequately screened (overall) 48 653 (57·6%) 45 290 (57·3%) 41 822 (56·7%) 44 904 (58·5%) 76 520 (58·2%) 77 122 (57·5%) 20 166 (25·8%) 22 712 (25·1%) Age (years) 60–64 19 727 (54·9%) 18 200 (54·2%) 18 264 (53·3%) 19 014 (55·2%) 33 331 (55·9%) 33 480 (54·8%) 10 251 (26·7%) 12 229 (26·1%) 65–69 18 657 (60·8%) 17 346 (61·1%) 14 673 (60·9%) 16 673 (62·4%) 27 382 (61·0%) 27 466 (60·5%) 6674 (26·8%) 6898 (24·8%) 70–74 10 269 (57·7%) 9744 (56·9%) 8885 (57·9%) 9217 (59·2%) 15 807 (58·7%) 16 176 (58·8%) 3241 (21·9%) 3585 (22·6%) Sex Female 25 585 (59·2%) 24 017 (59·1%) 22 499 (59·3%) 23 811 (60·9%) 40 707 (60·8%) 41 290 (60·2%) 10 267 (27·2%) 11 511 (26·4%) Male 23 068 (56·0%) 21 273 (55·4%) 19 323 (54·0%) 21 093 (56·1%) 35 813 (55·6%) 35 832 (54·8%) 9899 (24·5%) 11 201 (23·9%) Screening episode Prevalent first time 6466 (49·6%) 5981 (48·2%) 7678 (50·2%) 6231 (49·8%) 11 465 (51·4%) 11 646 (49·4%) 3739 (25·8%) 5398 (25·4%) Prevalent 3836 (14·5%) 3479 (14·2%) 3113 (14·0%) 3284 (14·3%) 5675 (14·0%) 5357 (13·3%) 2394 (6·0%) 2329 (5·4%) Incident 38 351 (85·2%) 35 830 (85·0%) 31 031 (85·6%) 35 389 (85·7%) 59 380 (86·4%) 60 119 (85·7%) 14 033 (59·2%) 14 985 (58·1%) IMD quintile* 1 (least deprived) 12 547 (65·8%) 12 178 (65·6%) 11 005 (64·6%) 11 411 (66·8%) 19 792 (65·2%) 20 716 (66·0%) 5522 (34·7%) 6601 (34·9%) 2 12 305 (62·2%) 11 412 (62·4%) 10 253 (62·1%) 11 080 (62·7%) 19 530 (63·1%) 19 604 (62·6%) 5107 (30·8%) 5782 (29·7%) 3 10 732 (58·6%) 9335 (58·4%) 8911 (58·3%) 9601 (59·4%) 16 571 (59·3%) 16 336 (58·0%) 4316 (26·8%) 4578 (25·0%) 4 7663 (52·0%) 6987 (51·9%) 6535 (50·7%) 7083 (52·9%) 11 902 (53·0%) 11 839 (51·5%) 3104 (21·1%) 3436 (20·4%) 5 (most deprived) 5322 (43·0%) 5316 (42·0%) 4966 (42·4%) 5580 (46·0%) 8433 (44·0%) 8324 (42·6%) 2040 (14·1%) 2198 (13·3%) Missing IMD score 84 62 152 149 2

6 571 (59·3%) 16 336 (58·0%) 4316 (26·8%) 4578 (25·0%) 4 7663 (52·0%) 6987 (51·9%) 6535 (50·7%) 7083 (52·9%) 11 902 (53·0%) 11 839 (51·5%) 3104 (21·1%) 3436 (20·4%) 5 (most deprived) 5322 (43·0%) 5316 (42·0%) 4966 (42·4%) 5580 (46·0%) 8433 (44·0%) 8324 (42·6%) 2040 (14·1%) 2198 (13·3%) Missing IMD score 84 62 152 149 2 92 303 77 117 GP=general practice. IMD=Index of Multiple Deprivation. * Quintile categories were based on national cutoffs.26 Table 3 Adjusted odds for screening uptake, overall and by deprivation quintile Trial 1 (standard information and gist leaflet) Trial 2 (standard information and narrative leaflet) Trial 3 (GP endorsed invitation letter) Trial 4 (enhanced reminder letter) Overall uptake 1·03 (0·99–1·06) p=0·15 1·00 (0·96–1·03) p=0·80 1·07 (1·04–1·10) p<0·0001 1·07 (1·03–1·11) p=0·001 1 (least deprived) 1·06 (1·01–1·11) 0·98 (0·93–1·04) 1·04 (0·99–1·08) 1·00 (0·94–1·06) 2 1·02 (0·97–1·07) 1·00 (0·94–1·06) 1·06 (1·02–1·10) 1·04 (0·98–1·11) 3 1·00 (0·94–1·08) 1·05 (0·97–1·13) 1·08 (1·03–1·13) 1·13 (1·06–1·20) 4 1·01 (0·94–1·08) 1·00 (0·94–1·06) 1·09 (1·04–1·15) 1·09 (1·02–1·17) 5 (most deprived) 1·04 (0·96–1·12) pinteraction=0·68 0·92 (0·86–0·98) pinteraction=0·11 1·07 (1·01–1·13) pinteraction=0·49 1·11 (1·04–1·20) pinteraction=0·005 Data are odds ratios (95% CI). *Adjusted for hub, age, sex, and screening episode.

Introduction In most developed countries worldwide, life expectancy is increasing by at least 2 years every decade and does not seem to be slowing down, at least for life expectancy at age 60 years.1 Nevertheless, disability trends have not shown such clear improvement, with results of studies in the USA and Europe showing increases, decreases, or stagnation in disability prevalence over time.2, 3, 4, 5, 6, 7, 8, 9, 10, 11 Differences between age groups within countries have also been reported. Health expectancies such as disability-free life expectancy (DFLE) combine information about quantity and quality of remaining years, and provide an improved indication of different scenarios, including whether: the extra years of life are healthy (compression of morbidity)12 or unhealthy (expansion of morbidity);13 unhealthy years are increasing, but the proportion of life spent healthy is increasing (relative compression) or decreasing (relative expansion);14 or morbidity and disability are increasing but severity of disability is not (dynamic equilibrium).15

(compression of morbidity)12 or unhealthy (expansion of morbidity);13 unhealthy years are increasing, but the proportion of life spent healthy is increasing (relative compression) or decreasing (relative expansion);14 or morbidity and disability are increasing but severity of disability is not (dynamic equilibrium).15 Health expectancies are important as indicators to monitor population health trends and inequalities internationally, nationally, and regionally.16, 17 Since 2004, the preferred indicator for the EU is healthy life-years, a DFLE based on self-report global activity limitation indicator.18 The UK Office for National Statistics has regularly published time series of DFLE (based on limiting long-standing illness) and healthy life expectancy (HLE; based on self-perceived health [SPH]) at birth and age 65 years. These time series suggest that trends in DFLE and HLE are much less consistent than trends in life expectancy. Comparisons of the periods 2005–07 and 2008–10 suggest that compression of morbidity took place in the UK, with increases in DFLE at birth of 1·4 years for men and 1·3 years for women, compared with life expectancy of 0·9 years for men and 0·6 years for women, although at age 65 years only women's DFLE rose significantly.19 However, data for England, comparing 2006–08 and 2009–11, show a continued rise in life expectancy that is larger than the increase in DFLE for men at birth and for both men and women at age 65 years—DFLE at birth for women even shows a slight decrease (0·1 years).20

age 65 years only women's DFLE rose significantly.19 However, data for England, comparing 2006–08 and 2009–11, show a continued rise in life expectancy that is larger than the increase in DFLE for men at birth and for both men and women at age 65 years—DFLE at birth for women even shows a slight decrease (0·1 years).20 Trends in DFLE and HLE in other countries have been reviewed,21 but not all show that extra years of life are healthy or free from disability. Beyond a real increase in morbidity and disability, several reasons exist for the recorded increase in number of years with ill health or disability. First, all health expectancy trends are based on self-report and therefore might be affected by rising expectations of health (and thus a lowered threshold for reporting of ill health or disability). Additionally, questions within countries, such as the UK census question, might change over time, reducing comparability. Second, whether increased years with disability or ill health result from increased cognitive or physical functional limitations is impossible to know since mental ill health is only implicitly included (through its effects on disability or self-reported general health). Finally, all trends in healthy life-years and the between-census values for the UK are based on community-dwelling populations only; people in institutions are only surveyed in the 10-year census.

ow since mental ill health is only implicitly included (through its effects on disability or self-reported general health). Finally, all trends in healthy life-years and the between-census values for the UK are based on community-dwelling populations only; people in institutions are only surveyed in the 10-year census. Our aim was to investigate how health expectancies at age 65 years or older changed between 1991 and 2011, by use of health measures, identical design, and total population (inclusion of institutional residents) available in the Cognitive Function and Ageing Studies (CFAS I and CFAS II). Specifically, we investigated whether (absolute or relative) compression, expansion, or dynamic equilibrium of morbidity took place.

2011, by use of health measures, identical design, and total population (inclusion of institutional residents) available in the Cognitive Function and Ageing Studies (CFAS I and CFAS II). Specifically, we investigated whether (absolute or relative) compression, expansion, or dynamic equilibrium of morbidity took place. Methods Study design Data for CFAS I were taken from baseline interviews done between 1989 and 1994 in the population aged 65 years or older in three (Cambridgeshire, Newcastle, and Nottingham) of the six geographical areas of the UK Medical Research Council CFAS. CFAS II baseline interviews were done between 2008 and 2011 in the same three areas and with the same study design and methods as in CFAS I. The sampling base for both studies was primary care registers within the areas, each of which provided 2500 individuals aged 65 years or older, with stratification by age group (65–74 years vs ≥75 years; 1250 people per stratum per area). We used oversampling to allow for losses (death, incorrect registration, ineligibility, general practitioner refusals, or participant or gatekeeper refusals). The primary-care practices screened records of patients in selected samples regularly for deaths and terminal illness. Selected individuals were sent an introductory letter from their family doctor; this letter was followed by a visit from a named study interviewer. Full details of the study design, methods, and response rates have been published.22

ened records of patients in selected samples regularly for deaths and terminal illness. Selected individuals were sent an introductory letter from their family doctor; this letter was followed by a visit from a named study interviewer. Full details of the study design, methods, and response rates have been published.22 Health domains We used three health measures as a basis for calculation of health expectancies: SPH, cognitive impairment, and disability. SPH was measured by asking “would you say that for someone of your age, your health in general is excellent/good/fair/poor”, and participants were categorised as having excellent–good versus fair–poor health. Cognitive impairment was defined by a Mini-Mental State Examination (MMSE)23 score (maximum score 30) as: severe impairment (0–17), mild impairment (18–25), or no impairment (26–30).24 We used a measure of disability based on basic activities of daily living (BADL) and instrumental activities of daily living (IADL).25 Participants were classified as having moderate–severe disability if they were unable to do at least one of five activities without human help: transfer to and from a chair (from interviewer assessment); put on shoes and socks; prepare a hot meal; get around outside; or have a bath or all-over wash. Participants who were able to do all five activities without help from another person but who needed help with at least one of the two additional IADLs (shop, including carrying of heavy bags, and do heavy housework) were classified as having mild disability.

t meal; get around outside; or have a bath or all-over wash. Participants who were able to do all five activities without help from another person but who needed help with at least one of the two additional IADLs (shop, including carrying of heavy bags, and do heavy housework) were classified as having mild disability. Statistical analysis The age-specific and sex-specific prevalence estimates of each health measure were calculated for CFAS I and CFAS II with inverse probability weighting to account for non-response differences between studies and selection of study design. Details of the weighting methods have been published previously.22 Analysis of change over time in prevalence of each health measure was done by logistic regression. Models were fitted with time (0=1991, 1=2011), age (5-year age band), and sex, then further adjusted for Townsend deprivation index (in CFAS I tertiles),26 education (0–9 years, 10–11 years, or ≥12 years), and region. Because proportional odds assumptions were violated for all three measures, separate models were fitted to any morbidity (fair–poor SPH, MMSE 0–25, or any disability) and severe morbidity (poor SPH, MMSE 0–17, and moderate–severe disability). The proportion of missing data for any of the health measures was low: 2·9% (1991) versus 4·2% (2011) for SPH; 1·8% (1991) versus 3·7% (2011) for MMSE; and 1·1% (1991) versus 4·2% (2011) for disability.

r–poor SPH, MMSE 0–25, or any disability) and severe morbidity (poor SPH, MMSE 0–17, and moderate–severe disability). The proportion of missing data for any of the health measures was low: 2·9% (1991) versus 4·2% (2011) for SPH; 1·8% (1991) versus 3·7% (2011) for MMSE; and 1·1% (1991) versus 4·2% (2011) for disability. Three main health expectancies—HLE, cognitive-impairment-free life expectancy (CIFLE), and DFLE—were calculated for the combined three regions common to CFAS I and CFAS II by sex for both timepoints (1991 and 2011) by the Sullivan method.27 This method applies the age-specific and sex-specific prevalence of the health measure to a standard life table for the same period. To assess dynamic equilibrium, we separated years with fair and poor health, mild and severe cognitive impairment, and mild and severe disability, within each health expectancy calculation. For the standard abridged life table calculations, we used population mid-year estimates and vital statistics death data provided by the Office of National Statistics28 at the district level for the three regions. Instead of using an average ax (the fraction of interval lived by those dying in the interval) of 0·5 in life table calculations, we calculated a precise national ax (and stratified by sex) from national mortality data for both years and used this value in the abridged local life table calculations. We closed life tables at age 90 years. Prevalence modelling was undertaken in SAS, version 9.3, and all health expectancy calculations were done in R, version 3.0.3.

For the standard abridged life table calculations, we used population mid-year estimates and vital statistics death data provided by the Office of National Statistics28 at the district level for the three regions. Instead of using an average ax (the fraction of interval lived by those dying in the interval) of 0·5 in life table calculations, we calculated a precise national ax (and stratified by sex) from national mortality data for both years and used this value in the abridged local life table calculations. We closed life tables at age 90 years. Prevalence modelling was undertaken in SAS, version 9.3, and all health expectancy calculations were done in R, version 3.0.3. Role of the funding source The funders are represented on the CFAS Management Committee and the Biological Resource Advisory Committee but they had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all data in the study and had final responsibility for the decision to submit for publication. Results We report changes over time in the components of the health expectancies—ie, life expectancy and the prevalence of each health measure. In the three regions combined between 1991 and 2011, life expectancy rose by 4·5 years for men and 3·6 years for women at age 65 years and by 3·0 years for men and 2·5 years for women at age 70 years; from age 80 years, increases for women exceeded those for men (appendix).

cy and the prevalence of each health measure. In the three regions combined between 1991 and 2011, life expectancy rose by 4·5 years for men and 3·6 years for women at age 65 years and by 3·0 years for men and 2·5 years for women at age 70 years; from age 80 years, increases for women exceeded those for men (appendix). In the three regions combined, 7635 people participated in CFAS I and 7796 in CFAS II. Because of age stratification, proportions of men and women in each age group were similar in 1991 and 2011, but a smaller proportion of women was included in 2011 (55%) than in 1991 (60%) (table 1). Improved access to education (and change in minimum school leaving age by birth cohort) was evident, with more than twice as many participants reporting 12 or more years of education in 2011 as in 1991 (22% vs 9%), and decreased deprivation in 2011 than in 1991 (table 1).

ded in 2011 (55%) than in 1991 (60%) (table 1). Improved access to education (and change in minimum school leaving age by birth cohort) was evident, with more than twice as many participants reporting 12 or more years of education in 2011 as in 1991 (22% vs 9%), and decreased deprivation in 2011 than in 1991 (table 1). Prevalence of cognitive impairment, fair or poor health, and disability was higher in women than in men and increased with age, although less so for SPH than for the other health measures (appendix). After adjustment for age and sex, prevalence of fair or poor SPH (odds ratio [OR] 0·83, 95% CI 0·78–0·88), any cognitive impairment (0·53, 0·49–0·56), severe cognitive impairment (0·49, 0·43–0·56), and moderate–severe disability (0·76, 0·70–0·82) substantially decreased between 1991 and 2011, but that of any disability increased (1·22, 1·14–1·30). These differences remained, but were attenuated after adjustment for region, education, and deprivation. However, the reduction in prevalence of any disability increased further (1·36, 1·27–1·47) after adjustment, and the lowered prevalence of fair–poor SPH was not statistically significant (1·02, 0·89–1·17) after adjustment. The raised prevalence of any disability was not accounted for by changes in cognitive impairment (1·56, 1·44–1·68) or by vision or hearing problems (1·33, 1·21–1·44). However, lessened prevalence of moderate–severe disability was not significant after adjustment for changes in cognitive impairment (0·99, 0·90–1·09).

t. The raised prevalence of any disability was not accounted for by changes in cognitive impairment (1·56, 1·44–1·68) or by vision or hearing problems (1·33, 1·21–1·44). However, lessened prevalence of moderate–severe disability was not significant after adjustment for changes in cognitive impairment (0·99, 0·90–1·09). Between 1991 and 2011, we identified absolute compression of cognitive morbidity at age 65 years for women, with gains in CIFLE of 4·4 years (95% CI 4·3–4·5) and a drop in years with any cognitive impairment (CILE) of 0·7 years (0·2–1·3); the decrease in CILE consisted of a significant decrease in years with mild or moderate–severe cognitive impairment (figure, table 2). Although CIFLE for men aged 65 years rose by 4·2 years (4·2–4·3) and the proportion of life spent cognitive-impairment free increased, we identified no significant decrease in CILE of any severity (figure, table 2).

of a significant decrease in years with mild or moderate–severe cognitive impairment (figure, table 2). Although CIFLE for men aged 65 years rose by 4·2 years (4·2–4·3) and the proportion of life spent cognitive-impairment free increased, we identified no significant decrease in CILE of any severity (figure, table 2). For SPH, we identified a compression of morbidity for men and women, with significantly enhanced HLE (3·8 years for men [95% CI 3·5–4·1] and 3·1 years for women [2·7–3·4]), although these increases were less than those in life expectancy. Additionally, we identified a significantly increased proportion of healthy remaining life (4·2 percentage points [2·0–6·5] for men and 3·0 percentage points [1·0–4·9] for women; figure, table 3). For disability, we identified dynamic equilibrium, since the proportion of years of disability-free life fell substantially (5·3 percentage points [3·4–7·2] for men and 9·3 percentage points [7·5–11·1] for women), although the increase in number of years with any disability was higher for years with mild disability (1·3 years [1·1–1·6] for men and 2·5 years [2·2–2·8] for women) than for years with moderate or severe disability (0·5 years [0·3–0·8] for men and 0·6 years [0·3–0·9] for women; figure, table 4). Time patterns at age 85 years were similar but with absolute compression of cognitive morbidity and a reduction in the proportion of life spent disability free seen only in women (appendix). CILE and life expectancy with moderate–severe disability are fairly constant with age (figure).

0·9] for women; figure, table 4). Time patterns at age 85 years were similar but with absolute compression of cognitive morbidity and a reduction in the proportion of life spent disability free seen only in women (appendix). CILE and life expectancy with moderate–severe disability are fairly constant with age (figure). Sensitivity analyses were done for CIFLE, assuming non-responders had twice the risk of severe cognitive impairment as responders, but had no effect on changes in CIFLE over time.

0·9] for women; figure, table 4). Time patterns at age 85 years were similar but with absolute compression of cognitive morbidity and a reduction in the proportion of life spent disability free seen only in women (appendix). CILE and life expectancy with moderate–severe disability are fairly constant with age (figure). Sensitivity analyses were done for CIFLE, assuming non-responders had twice the risk of severe cognitive impairment as responders, but had no effect on changes in CIFLE over time. Discussion Whether people are living longer, healthier lives than previously and compressing morbidity into a shorter period is a key concern for government, for society as a whole, and for individuals and their families. Our findings show that the answer crucially depends on how health is measured, with absolute compression observed for cognitive morbidity, relative compression observed for SPH, and dynamic equilibrium observed for disability. The increase of 4·4 years for women at age 65 years in CIFLE between 1991 and 2011 was more than the increase in life expectancy (3·6 years), with measurable falls in years with mild (0·5 years [0·3–0·8]) and moderate–severe cognitive impairment (0·2 years [0·1–0·4]), supporting previous findings of a decrease in prevalence of dementia during this period (panel).22 Moreover, women spend on average around twice as many years cognitively impaired as do men, and these values are fairly constant with age, as shown by other studies.35, 36 Findings for disability are less positive than those for cognitive impairment, with a decrease in the proportion of remaining life spent disability free, although the severity of disability seems to be milder than previously.

red as do men, and these values are fairly constant with age, as shown by other studies.35, 36 Findings for disability are less positive than those for cognitive impairment, with a decrease in the proportion of remaining life spent disability free, although the severity of disability seems to be milder than previously. Cognitive impairment is one reason for difficulty in calculation of IADLs and BADLs, and our findings suggest that reduced moderate–severe disability was a result of decreased cognitive impairment, although increases in any disability were not a result of cognitive impairment. Rises in IADLs and BADLs might be due to increased prevalence of other specific diseases or physical or sensory functional limitations.37 Our analyses showed that problems with vision and hearing (self-report or interviewer observed) did not account for rises in disability. By contrast, analysis of health trends from 1992 to 2007 for the population aged 65 years and older from the Health Survey for England showed stability in self-care activities but increased obesity and mobility limitations (walking 200 yards and climbing stairs), which might contribute to gains in mild disability.3 This finding accords well with evidence of high prevalence of arthritis in later young-old (aged 65–69 years) cohorts in one centre in CFAS I.38

stability in self-care activities but increased obesity and mobility limitations (walking 200 yards and climbing stairs), which might contribute to gains in mild disability.3 This finding accords well with evidence of high prevalence of arthritis in later young-old (aged 65–69 years) cohorts in one centre in CFAS I.38 Our study has limitations because of non-response and the subjectivity of two of the measures. Non-response was higher in CFAS II than in CFAS I, as is common in population-based studies. Reasons for non-response did not differ between the studies, although some reasons for refusal (by others on behalf of frail individuals, and by very active individuals) became more prominent.22 Because these reasons affect both good and poor health, they are unlikely to substantially bias the overall estimate from CFAS II, as our sensitivity analyses and those of a previous study22 showed. Moreover, these reasons would probably result in differences in the same direction for all the health measures, which we did not find. The only difference in study design between CFAS I and CFAS II was the change from two-stage sampling (prevalence screen followed by assessment) in CFAS I to a combined screen and assessment in CFAS II. Since all our health measures are derived from the initial prevalence screen, our findings are not affected by this change in study design or by the attrition that occurred between screen and assessment in CFAS I. Both disability and SPH are self-reported and might be subject to temporal changes in health expectations and thresholds for admission of activity limitation, although self-report of IADLs and BADLs and objective performance measures in very old people (aged 85 years) are consistent in one of the CFAS regions.39 Finally, adjustment of health expectancies for education, deprivation, or both needs longitudinal data since life tables are not routinely available for these factors.

though self-report of IADLs and BADLs and objective performance measures in very old people (aged 85 years) are consistent in one of the CFAS regions.39 Finally, adjustment of health expectancies for education, deprivation, or both needs longitudinal data since life tables are not routinely available for these factors. One might question what additional information is provided by health expectancy over other measures, since the amount of ill health in a population is often measured by prevalence alone. Nevertheless, as populations age, with more people surviving to the oldest age groups, in which the prevalence of chronic disorders is highest, the overall population prevalence of ill health might increase without individuals being at higher risk of ill health than they were previously. We show this effect by the true age-specific population and mortality rates for women in 1991 and 2011 and a hypothetical prevalence of ill health in 1991 (appendix). Despite a 20% lower prevalence of ill health in each age group in 2011, the overall prevalence in 2011 (9·7%) is higher than that in 1991 (9·3%), and, because of the increased numbers of people at the greatest risk, the absolute number of unhealthy individuals is larger. Health expectancy provides an integrated approach because it takes into account both the changes in living with ill health (prevalence) and changes in mortality that are responsible for increased life expectancy. However, health expectancies are independent of population size. Thus, application of hypothetical prevalence in the example to the true life tables for 1991 and 2011 results in an increase of 3·2 healthy years at age 65 years. Improvement of population health in addition to population ageing results in an increase in the part of life expectancy spent healthy, even with an increase in the overall prevalence of ill health due to more people being at risk. Health expectancy is therefore a potent means to identify interactions between health, ill health, and mortality.

n health in addition to population ageing results in an increase in the part of life expectancy spent healthy, even with an increase in the overall prevalence of ill health due to more people being at risk. Health expectancy is therefore a potent means to identify interactions between health, ill health, and mortality. Although years free of ill health defined by all measures have increased during the past two decades, absolute compression was observed only for cognitive impairment, and women in particular spend a smaller proportion of remaining life disability free and more years with mild disability than they did previously. The paradox that women have worse health but better survival than men is dependent on the set of minor health deficits included,40 and our future work will investigate which diseases and disorders are responsible for increased mild disability and whether patterns prevail across all regions. Nevertheless, our findings have important implications for government, employers, and individuals, specifically for raising of the state pension age and extension of working life, and for community care services and family carers who predominantly support people with mild–moderate disability to enable them to continue living independently. Supplementary Material Supplementary appendix

Although years free of ill health defined by all measures have increased during the past two decades, absolute compression was observed only for cognitive impairment, and women in particular spend a smaller proportion of remaining life disability free and more years with mild disability than they did previously. The paradox that women have worse health but better survival than men is dependent on the set of minor health deficits included,40 and our future work will investigate which diseases and disorders are responsible for increased mild disability and whether patterns prevail across all regions. Nevertheless, our findings have important implications for government, employers, and individuals, specifically for raising of the state pension age and extension of working life, and for community care services and family carers who predominantly support people with mild–moderate disability to enable them to continue living independently. Supplementary Material Supplementary appendix Acknowledgments CFAS II was supported by the UK Medical Research Council (MRC; research grant G0601022) and received support from the UK National Institute for Health Research (NIHR) comprehensive clinical research networks in West Anglia and Trent, and the Dementias and Neurodegenerative Disease Research Network in Newcastle. MRC CFAS (including CFAS I areas) was funded by the MRC and the UK National Health Service (NHS). FEM is supported by the MRC (grant number U105292687). CJ is supported by the AXA Research Fund. TW and PW are supported by the Economic and Social Research Council (grant number RES-062-23-2970). We thank the participants, their families, the family doctors and their staff, and the primary care trusts for their cooperation and support. We thank the CFAS II fieldwork interviewers at Cambridge, Nottingham, and Newcastle for their valuable contribution. This research was done within the UK NIHR collaboration for leadership in applied health research and care for Cambridgeshire and Peterborough and the Cambridge Biomedical Research Centre infrastructures, Nottingham city and Nottinghamshire county NHS primary care trusts, and UK NIHR Biomedical Research Centre for Ageing and Age-related Disease Award to Newcastle Upon Tyne Hospital Foundation Trust.

health research and care for Cambridgeshire and Peterborough and the Cambridge Biomedical Research Centre infrastructures, Nottingham city and Nottinghamshire county NHS primary care trusts, and UK NIHR Biomedical Research Centre for Ageing and Age-related Disease Award to Newcastle Upon Tyne Hospital Foundation Trust. Medical Research Council Cognitive Function and Ageing Collaboration CFAS Cambridge core team and fieldwork support: E Green, L Gao, R Barnes, J Warwick, A Mattison. CFAS management committee membership: A Arthur, C Baldwin, L E Barnes, J Bond, C Brayne, A Comas-Herrera, T Dening, G Forster, S Harrison, P G Ince, C Jagger, J Lowe, A S Macdonald, F E Matthews, C F M McCracken, I G McKeith, C Moody, B Parry, L Robinson, B C M Stephan, S Wharton, R Wittenberg, B Woods. CFAS biological resource advisory committee: I Allen (chair). Contributors The CFAS management committee all contributed to all aspects of the study, including fund raising, design, supervision, and drafting. FEM, CB, AA, CJ, LR, and BCMS acquired the data, and supported and did fieldwork. LG and EP received and cleaned the data. LEB and CB had responsibility for fieldwork in Cambridgeshire, LEB, LR, and JB had responsibility for fieldwork in Newcastle, and LEB and AA had responsibility for fieldwork in Nottingham. CJ designed the analyses, and CJ, PW, and TF did the analyses. CJ, PW, and FEM wrote the first draft and all authors edited the paper. CJ is guarantor of the analysis. Declaration of interests We declare no competing interests.

of all-cause mortality, adjusted for: Age only 1·06 (1·02–1·11) 1·00 (0·97–1·03) Ref Age, characteristics†, and treatment for common health disorder§ 1·01 (0·97–1·06) 1·00 (0·97–1·03) Ref Analyses are limited to the 719 671 women without cancer, heart disease, stroke, or chronic obstructive airways disease at baseline. * Happy only sometimes, rarely, or never. † Region of residence, area deprivation, educational qualifications, body-mass index, strenuous exercise, smoking, alcohol consumption, living with a partner, parity, participation in religious or other group activities, and sleep duration (appendix p 5 gives analyses adjusted for each of these separately). ‡ Self-rated health at baseline, in three categories: poor or fair, good, and excellent (the numbers for all women include the few who did not answer the question on self-rated health at baseline). § Treatment at baseline for high blood pressure, diabetes, asthma, arthritis, depression or anxiety (appendix p 5 gives the corresponding result adjusted only for treatment for depression or anxiety). RR=rate ratio. Ref=reference group. Panel Research in context Systematic review

Contributors The CFAS management committee all contributed to all aspects of the study, including fund raising, design, supervision, and drafting. FEM, CB, AA, CJ, LR, and BCMS acquired the data, and supported and did fieldwork. LG and EP received and cleaned the data. LEB and CB had responsibility for fieldwork in Cambridgeshire, LEB, LR, and JB had responsibility for fieldwork in Newcastle, and LEB and AA had responsibility for fieldwork in Nottingham. CJ designed the analyses, and CJ, PW, and TF did the analyses. CJ, PW, and FEM wrote the first draft and all authors edited the paper. CJ is guarantor of the analysis. Declaration of interests We declare no competing interests. Figure Life expectancy and years lived with cognitive impairment, fair–poor self-perceived health (SPH), mild disability, and moderate–severe disability in 1991 and 2011, all regions combined Table 1 Sociodemographic characteristics of Cognitive Function and Ageing Survey (CFAS) I and CFAS II CFAS I (n=7635) CFAS II (n=7796) Sex Women 4590 (60%) 4246 (55%) Age group (years) 65–69 1981 (26%) 1939 (25%) 70–74 1776 (23%) 1873 (24%) 75–79 1725 (23%) 1624 (21%) 80–84 1308 (17%) 1290 (17%) ≥85 845 (11%) 1070 (14%) Education (years full time) 0–9 5529 (74%) 2052 (27%) 10–11 1238 (17%) 3923 (51%) ≥12 692 (9%) 1704 (22%) Townsend deprivation index (tertile)* Low deprivation 2467 (33%) 3737 (48%) Middle deprivation 2419 (33%) 2412 (31%) High deprivation 2522 (34%) 1619 (21%) Data are n (%). Numbers calculated as a percentage of the non-missing values (there were some missing values for education and deprivation).

≥12 692 (9%) 1704 (22%) Townsend deprivation index (tertile)* Low deprivation 2467 (33%) 3737 (48%) Middle deprivation 2419 (33%) 2412 (31%) High deprivation 2522 (34%) 1619 (21%) Data are n (%). Numbers calculated as a percentage of the non-missing values (there were some missing values for education and deprivation). * CFAS I tertiles. Table 2 Life expectancy, cognitive-impairment-free life expectancy (CIFLE), and proportion of life free of cognitive impairment at age 65 years in 1991 and 2011

≥12 692 (9%) 1704 (22%) Townsend deprivation index (tertile)* Low deprivation 2467 (33%) 3737 (48%) Middle deprivation 2419 (33%) 2412 (31%) High deprivation 2522 (34%) 1619 (21%) Data are n (%). Numbers calculated as a percentage of the non-missing values (there were some missing values for education and deprivation). * CFAS I tertiles. Table 2 Life expectancy, cognitive-impairment-free life expectancy (CIFLE), and proportion of life free of cognitive impairment at age 65 years in 1991 and 2011 1991 2011 Difference (2011–1991) Men Women Men Women Men Women Life expectancy (years) 13·0 16·7 17·5 20·3 4·5 3·6 CIFLE (MMSE 26–30) (95% CI) 9·4 (9·2 to 9·6) 10·1 (9·8 to 10·4) 13·6 (13·4 to 13·9) 14·5 (14·1 to 14·8) 4·2 (4·2 to 4·3) 4·4 (4·3 to 4·5) Proportion of life free of cognitive impairment (95% CI) 72·4% (70·6 to 74·3) 60·5% (58·6 to 62·3) 78·2% (76·6 to 79·8) 71·2% (69·5 to 72·9) 5·8% (3·3 to 8·2) 10·7% (8·2 to 13·2) CILE (MMSE 0–25) (95% CI) 3·6 (3·4 to 3·8) 6·6 (6·4 to 6·8) 3·8 (3·5 to 4·1) 5·9 (5·5 to 6·2) 0·2 (−0·3 to 0·8) −0·7 (−1·3 to −0·2) mildCILE (MMSE 18–25) (95% CI) 3·1 (2·7 to 3·6) 5·6 (5·2 to 6·0) 3·4 (2·8 to 3·9) 5·1 (4·5 to 5·6) 0·3 (0·0 to 0·4) −0·5 (−0·8 to −0·3) Proportion of life with mild cognitive impairment (95% CI) 24·3% (21·1 to 27·5) 33·5% (31·1 to 36·0) 19·3% (16·3 to 22·3) 25·0% (22·3 to 27·6) −5·0% (−9·4 to −0·6) −8·5% (−12·2 to −5·0) sevCILE (MMSE 0–17) (95% CI) 0·4 (0·3 to 0·5) 1·0 (0·9 to 1·1) 0·4 (0·3 to 0·5) 0·8 (0·7 to 0·9) 0·0 (−0·1 to 0·1) −0·2 (−0·4 to −0·1) Proportion of life with severe cognitive impairment (95% CI) 3·2% (−0·3 to 6·8) 6·0% (3·1 to 8·9) 2·5% (−1·0 to 6·1) 3·9% (0·7 to 7·0) −0·7% (−5·7 to 4·3) −2·1% (−6·4 to 2·1) CILE=years with cognitive impairment. mildCILE=years with mild cognitive impairment. sevCILE=years with moderate–severe cognitive impairment.

·1) Proportion of life with severe cognitive impairment (95% CI) 3·2% (−0·3 to 6·8) 6·0% (3·1 to 8·9) 2·5% (−1·0 to 6·1) 3·9% (0·7 to 7·0) −0·7% (−5·7 to 4·3) −2·1% (−6·4 to 2·1) CILE=years with cognitive impairment. mildCILE=years with mild cognitive impairment. sevCILE=years with moderate–severe cognitive impairment. Table 3 Life expectancy, healthy life expectancy (self-perceived health), and proportion of life spent healthy at age 65 years in 1991 and 2011

·1) Proportion of life with severe cognitive impairment (95% CI) 3·2% (−0·3 to 6·8) 6·0% (3·1 to 8·9) 2·5% (−1·0 to 6·1) 3·9% (0·7 to 7·0) −0·7% (−5·7 to 4·3) −2·1% (−6·4 to 2·1) CILE=years with cognitive impairment. mildCILE=years with mild cognitive impairment. sevCILE=years with moderate–severe cognitive impairment. Table 3 Life expectancy, healthy life expectancy (self-perceived health), and proportion of life spent healthy at age 65 years in 1991 and 2011 1991 2011 Difference (2011–1991) Men Women Men Women Men Women Life expectancy (years) 13·0 16·7 17·5 20·3 4·5 3·6 HLE (95% CI) 8·8 (8·6 to 9·1) 11·2 (11·0 to 11·5) 12·6 (12·4 to 12·9) 14·3 (14·0 to 14·6) 3·8 (3·5 to 4·1) 3·1 (2·7 to 3·4) Proportion of life spent healthy (95% CI) 68·2% (66·5 to 69·9) 67·3% (65·9 to 68·7) 72·4% (70·9 to 73·9) 70·3% (68·8 to 71·7) 4·2% (2·0 to 6·5) 3·0% (1·0 to 4·9) unHLE (95% CI) 4·1 (3·9 to 4·3) 5·5 (5·2 to 5·7) 4·8 (4·5 to 5·1) 6·0 (5·8 to 6·3) 0·7 (0·3 to 1·0) 0·5 (0·2 to 1·0) fairHLE (95% CI) 3·3 (2·8 to 3·7) 4·4 (3·9 to 4·8) 3·7 (3·2 to 4·3) 4·9 (4·4 to 5·5) 0·4 (−0·2 to 1·2) 0·5 (−0·1 to 1·3) Proportion of life with fair health (95% CI) 25·1% (21·9 to 28·2) 26·2% (23·6 to 28·8) 21·4% (18·4 to 24·4) 24·2% (21·5 to 26·9) −3·7% (−8·0 to 0·7) 2·0% (−1·8 to 5·7) poorHLE (95% CI) 0·9 (0·4 to 1·3) 1·1 (0·6 to 1·6) 1·1 (0·5 to 1·6) 1·1 (0·5 to 1·7) 0·2 (0·0 to 0·0) 0·0 (−0·8 to 0·8) Proportion of life with poor health (95% CI) 6·7% (3·2 to 10·2) 6·5% (3·6 to 9·4) 6·1% (2·8 to 9·4) 5·5% (2·5 to 8·6) −0·6% (−5·4 to 4·2) −1·0% (−5·2 to 3·2) HLE=healthy life expectancy. unHLE=years with fair or poor health. fairHLE=years with fair health. poorHLE=years with poor health.

(0·0 to 0·0) 0·0 (−0·8 to 0·8) Proportion of life with poor health (95% CI) 6·7% (3·2 to 10·2) 6·5% (3·6 to 9·4) 6·1% (2·8 to 9·4) 5·5% (2·5 to 8·6) −0·6% (−5·4 to 4·2) −1·0% (−5·2 to 3·2) HLE=healthy life expectancy. unHLE=years with fair or poor health. fairHLE=years with fair health. poorHLE=years with poor health. Table 4 Life expectancy, disability-free life expectancy (DFLE), and proportion of life free of disability at age 65 years in 1991 and 2011

(0·0 to 0·0) 0·0 (−0·8 to 0·8) Proportion of life with poor health (95% CI) 6·7% (3·2 to 10·2) 6·5% (3·6 to 9·4) 6·1% (2·8 to 9·4) 5·5% (2·5 to 8·6) −0·6% (−5·4 to 4·2) −1·0% (−5·2 to 3·2) HLE=healthy life expectancy. unHLE=years with fair or poor health. fairHLE=years with fair health. poorHLE=years with poor health. Table 4 Life expectancy, disability-free life expectancy (DFLE), and proportion of life free of disability at age 65 years in 1991 and 2011 1991 2011 Difference (2011–1991) Men Women Men Women Men Women Life expectancy (years) 13·0 16·7 17·5 20·3 4·5 3·6 DFLE (95% CI) 10·3 (10·2 to 10·5) 11·0 (10·8 to 11·2) 12·9 (12·7 to 13·2) 11·5 (11·3 to 11·8) 2·6 (2·3 to 2·9) 0·5 (0·2 to 0·9) Proportion of life disability free (95% CI) 79·7% (78·3 to 81·0) 66·1% (64·9 to 67·4) 74·4% (73·0 to 75·8) 56·8% (55·5 to 58·2) −5·3% (−7·2 to −3·4) −9·3% (−11·1 to −7·5) DLE (95% CI) 2·6 (2·5 to 2·8) 5·7 (5·4 to 5·9) 4·5 (4·3 to 4·8) 8·8 (8·5 to 9·0) 1·9 (1·6 to 2·2) 3·1 (2·8 to 3·5) mildDLE (95% CI) 1·1 (0·9 to 1·2) 2·7 (2·6 to 2·9) 2·4 (2·2 to 2·6) 5·2 (5·0 to 5·6) 1·3 (1·1 to 1·6) 2·5 (2·2 to 2·8) Proportion of life with mild disability (95% CI) 8·2% (7·3 to 9·2) 16·4% (15·4 to 17·5) 13·8% (12·6 to 15·0) 25·8% (24·5 to 27·2) 5·6% (4·1 to 7·1) 9·4% (7·7 to 11·1) sevDLE (95% CI) 1·6 (1·4 to 1·7) 2·9 (2·7 to 3·1) 2·1 (1·9 to 2·2) 3·5 (3·2 to 3·7) 0·5 (0·3 to 0·8) 0·6 (0·3 to 0·9) Proportion of life with moderate–severe disability (95% CI) 12·0% (10·9 to 13·1) 17·4% (16·5 to 18·4) 11·8% (10·7 to 12·9) 17·3% (16·2 to 18·4) −0·2% (−1·8 to 1·3) −0·1% (−1·6 to 1·3) DLE=years with any disability. mildDLE=years with mild disability. sevDLE=years with moderate–severe disability.

o 0·8) 0·6 (0·3 to 0·9) Proportion of life with moderate–severe disability (95% CI) 12·0% (10·9 to 13·1) 17·4% (16·5 to 18·4) 11·8% (10·7 to 12·9) 17·3% (16·2 to 18·4) −0·2% (−1·8 to 1·3) −0·1% (−1·6 to 1·3) DLE=years with any disability. mildDLE=years with mild disability. sevDLE=years with moderate–severe disability. Panel Research in context Systematic review We updated a previous review of trends in health expectancies21 by searching MEDLINE and Web of Science (from Jan 1, 2009, to March 30, 2014) with the search terms “healthy life years”, “free life expectancy”, “active life expectancy”, “healthy life expectancy”, “health expectancy”, and “years of healthy life”. We excluded studies with only one timepoint. We identified no UK studies that included the population in care settings or that reported time differences in cognitive-impairment-free life expectancy. Interpretation

We updated a previous review of trends in health expectancies21 by searching MEDLINE and Web of Science (from Jan 1, 2009, to March 30, 2014) with the search terms “healthy life years”, “free life expectancy”, “active life expectancy”, “healthy life expectancy”, “health expectancy”, and “years of healthy life”. We excluded studies with only one timepoint. We identified no UK studies that included the population in care settings or that reported time differences in cognitive-impairment-free life expectancy. Interpretation Time differences in health expectancies based on self-rated health and disability are more numerous than those based on cognitive function and are often available nationally. Our findings of compression in fair–poor self-perceived health are in line with others across Europe.29, 30 Evidence of expansion of mild disability has been reported elsewhere,31, 32 although some countries have experienced compression of disability.2, 33 For the UK, estimates of health expectancies for the total population, including people in institutions, are only available from the 10-year censuses, and because the underlying disability (limiting longstanding illness) question changed between 1991 and 2001, our time differences with identical questions and study design provide a more accurate view of change than does the census.

ation, including people in institutions, are only available from the 10-year censuses, and because the underlying disability (limiting longstanding illness) question changed between 1991 and 2001, our time differences with identical questions and study design provide a more accurate view of change than does the census. Reduced mortality from coronary heart disease and stroke will have contributed to increased life expectancy, but would probably increase years with cognitive impairment and moderate or severe disability, for which we showed no evidence. However, evidence suggests that prevalence of sensory impairments (hearing and sight) and cognitive impairment, have likewise reduced in the past two decades, together with measured markers of disease such as hypertension, high cholesterol, and C-reactive protein,3 all of which, through improved disease management, might have reduced the severity of disability. Nevertheless, increased obesity and mobility limitations, and increased prevalence of arthritis, might have contributed to increases in mild disability. We have previously elucidated the contribution of specific diseases and disorders to disability-free life expectancy.34 We need to quantify whether our findings are due to diseases and disorders becoming less disabling during the past two decades.

Introduction Breast cancer is the most common cancer in women worldwide, with an estimated 1·6 million new cases reported every year.1 The proportion of these that are diagnosed as ductal carcinoma in situ (DCIS) has substantially increased over the past few decades due to the introduction of mammographic screening. It is estimated that approximately a fifth of all screen-detected breast cancers are DCIS.2

d 1·6 million new cases reported every year.1 The proportion of these that are diagnosed as ductal carcinoma in situ (DCIS) has substantially increased over the past few decades due to the introduction of mammographic screening. It is estimated that approximately a fifth of all screen-detected breast cancers are DCIS.2 Management strategies for DCIS vary depending on histological grade, tumour characteristics, and extent of disease. Almost all aspects of treatment are controversial, including the need for any treatment for some screen-detected lesions,3 the extent of surgery,4 the use of radiotherapy,5, 6 and the use of adjuvant endocrine therapy.7, 8 The role of tamoxifen has been investigated in two large trials.7, 8 In the National Surgical Adjuvant Breast and Bowel Project (NSABP) B-24 trial,7 all women with DCIS received radiotherapy before being randomly assigned to tamoxifen or matching placebo. After a median of 6 years of follow-up, a significant 37% reduction in breast cancer recurrence was observed with tamoxifen compared with placebo.7 Retrospective evaluation of oestrogen receptors (ER) and progesterone receptors (PgR) in 732 patients from the original study showed that tamoxifen reduced subsequent breast cancer events by 51% for women with ER-positive DCIS.9 However, no significant benefit with tamoxifen was observed for women with ER-negative DCIS. In the UK/ANZ DCIS trial,8 1578 women with locally excised DCIS were randomly assigned to receive tamoxifen with or without radiotherapy. After a median of 12·7 years of follow-up, tamoxifen significantly reduced all new breast cancer events by 29%, with a significant impact on ipsilateral DCIS recurrence and contralateral tumours, but no effect on ipsilateral invasive recurrence.8

were randomly assigned to receive tamoxifen with or without radiotherapy. After a median of 12·7 years of follow-up, tamoxifen significantly reduced all new breast cancer events by 29%, with a significant impact on ipsilateral DCIS recurrence and contralateral tumours, but no effect on ipsilateral invasive recurrence.8 Research in Context Evidence before this study A PubMed search between Jan 1, 1990, and Dec 31, 2002 (with the terms “ductal carcinoma in situ”, “breast cancer”, “aromatase inhibitors”, and “endocrine therapy”) and discussion with colleagues yielded no clinical trials or large cohorts of women with ductal carcinoma in situ (DCIS) treated by aromatase inhibitors. There have been two previous trials of tamoxifen. In the National Surgical Adjuvant Breast and Bowel Project (NSABP) B-24 trial, all women with DCIS received radiotherapy before being randomly assigned to tamoxifen or matching placebo. After a median of 6 years of follow-up, a significant 37% reduction in breast cancer recurrence was observed with tamoxifen compared with placebo. In the UK/ANZ DCIS trial, 1578 women with locally excised DCIS were randomly assigned to receive tamoxifen with or without radiotherapy. After a median of 12·7 years of follow-up, tamoxifen significantly reduced all new breast cancer events by 29%, with a significant effect on ipsilateral DCIS recurrence and contralateral tumours, but no effect on ipsilateral invasive recurrence. A further PubMed search was performed in October, 2015, which found no further published articles except the NSABP B-35 trial conference abstract.

ed all new breast cancer events by 29%, with a significant effect on ipsilateral DCIS recurrence and contralateral tumours, but no effect on ipsilateral invasive recurrence. A further PubMed search was performed in October, 2015, which found no further published articles except the NSABP B-35 trial conference abstract. Added value of this study In combination with the B-35 trial, this trial provides the first evidence for the use of an aromatase inhibitor (here, anastrozole) compared with tamoxifen for postmenopausal women with locally excised hormone-receptor-positive DCIS after a median follow-up of 7·1 years. In this study, no clear efficacy differences were seen between the two treatments, although all available evidence supports a greater efficacy for anastrozole. Implications of all the available evidence Our results are consistent with the small benefit of anastrozole versus tamzifen as seen in the NSABP B-35 trial. This is also supported by direct evidence of greater efficacy for recurrence in adjuvant trials of women with early invasive cancer and indirect evidence of greater efficacy against new cancers in a preventive setting. Side-effect profiles between the drugs differed, but there was no clear overall advantage for either treatment. Anastrozole offers another treatment option for postmenopausal women with oestrogen-receptor-positive DCIS which might be more appropriate for some women with contraindications to tamoxifen.

eventive setting. Side-effect profiles between the drugs differed, but there was no clear overall advantage for either treatment. Anastrozole offers another treatment option for postmenopausal women with oestrogen-receptor-positive DCIS which might be more appropriate for some women with contraindications to tamoxifen. Until now no data have been available on the use of aromatase inhibitors for DCIS. Two trials of very similar design have been conducted. Both compared anastrozole with tamoxifen in postmenopausal women with ER-positive or PgR-positive DCIS. The NSABP B-35 results will be reported elsewhere.10 Here, we report the first results from the International Breast Cancer Intervention Study-II DCIS (IBIS-II DCIS).

rials of very similar design have been conducted. Both compared anastrozole with tamoxifen in postmenopausal women with ER-positive or PgR-positive DCIS. The NSABP B-35 results will be reported elsewhere.10 Here, we report the first results from the International Breast Cancer Intervention Study-II DCIS (IBIS-II DCIS). Methods Study design and participants We undertook a double-blind, randomised, placebo-controlled trial to compare anastrozole with tamoxifen for the prevention of locoregional and contralateral breast cancer. Participants were women aged 40–70 years, postmenopausal, and had DCIS diagnosed within 6 months before randomisation. Microinvasion of less than 1 mm was permitted. Patients treated by mastectomy were not eligible for this study but could be included in the IBIS-II breast cancer prevention trial.11 Radiotherapy was permitted according to local practice. Margin status was determined by the local pathologist and ER and PgR positivity was determined as greater than or equal to 5% positive cells (equivalent of Quick-score of three or above and H-score of ten or above). After a protocol amendment on Feb 24, 2009, women were also allowed to enter the trial if they had been diagnosed with atypical hyperplasia or lobular carcinoma in situ to allow treatment of these benign breast diseases known to respond to tamoxifen.12, 13

of Quick-score of three or above and H-score of ten or above). After a protocol amendment on Feb 24, 2009, women were also allowed to enter the trial if they had been diagnosed with atypical hyperplasia or lobular carcinoma in situ to allow treatment of these benign breast diseases known to respond to tamoxifen.12, 13 Exclusion criteria were: premenopausal at diagnosis; any previous diagnosis of breast cancer (including DCIS excised more than 6 months before randomisation or treated by mastectomy); diagnosis of any other cancer in the past 5 years (excluding non-melanoma skin cancer or in-situ cervical cancer); current treatment with anticoagulants; previous diagnosis of deep-vein thrombosis, transient ischaemic attack, or cerebrovascular accident; previous or current use of selective oestrogen receptor modulators; intention to use menopausal hormone therapy; unexplained postmenopausal bleeding; evidence of severe osteoporosis (T-score less than −4 at total hip or lumbar spine or more than two fragility fractures); history of lactose intolerance, glucose intolerance, or both; or life expectancy judged by the clinician to be less than 10 years. All women provided written informed consent and the study was approved by the ethics committees of all participating institutions. The study sponsor was Queen Mary University of London.

Exclusion criteria were: premenopausal at diagnosis; any previous diagnosis of breast cancer (including DCIS excised more than 6 months before randomisation or treated by mastectomy); diagnosis of any other cancer in the past 5 years (excluding non-melanoma skin cancer or in-situ cervical cancer); current treatment with anticoagulants; previous diagnosis of deep-vein thrombosis, transient ischaemic attack, or cerebrovascular accident; previous or current use of selective oestrogen receptor modulators; intention to use menopausal hormone therapy; unexplained postmenopausal bleeding; evidence of severe osteoporosis (T-score less than −4 at total hip or lumbar spine or more than two fragility fractures); history of lactose intolerance, glucose intolerance, or both; or life expectancy judged by the clinician to be less than 10 years. All women provided written informed consent and the study was approved by the ethics committees of all participating institutions. The study sponsor was Queen Mary University of London. Randomisation and masking Participants were randomly assigned in a 1:1 ratio to receive 1 mg/day oral anastrozole or 20 mg/day oral tamoxifen. Randomisation was stratified by major centre or hub. Randomised blocks (six, eight, or ten) were used to maintain balance and randomisation was performed centrally by electronic contact with the main trials centre. All treatment was given on a daily basis for 5 years and all women took two tablets per day (tamoxifen and anastrozole placebo, or anastrozole and tamoxifen placebo). All IBIS-II DCIS personnel, participants, and clinicians were masked to treatment allocation, except for the IBIS-II DCIS trial statistician, who had access to unblinded data, and the independent data monitoring committee, who reviewed interim data for safety purposes.

le placebo, or anastrozole and tamoxifen placebo). All IBIS-II DCIS personnel, participants, and clinicians were masked to treatment allocation, except for the IBIS-II DCIS trial statistician, who had access to unblinded data, and the independent data monitoring committee, who reviewed interim data for safety purposes. Procedures A dual-energy x-ray absorptiometry scan within 2 years before entry to the trial and two lateral spinal radiographs were required to assess bone density and vertebral fractures. Women were seen at 6 months, 12 months, and then annually up to the 5 year follow-up point at local clinics. Adherence to treatment was ascertained at each follow-up visit. After 5 years, follow-up was annual and either by a short postal questionnaire or clinic visit, depending on country. Clinical adverse events were recorded during the post-treatment follow-up period. Mammograms were performed at least every 2 years. Blood samples were taken at baseline, year 1, and year 5 for the evaluation of potential biomarkers.

nnual and either by a short postal questionnaire or clinic visit, depending on country. Clinical adverse events were recorded during the post-treatment follow-up period. Mammograms were performed at least every 2 years. Blood samples were taken at baseline, year 1, and year 5 for the evaluation of potential biomarkers. Outcomes The primary endpoint of this analysis was the development of histologically confirmed breast cancer, both invasive and new or recurrent DCIS. First events were further categorised as local recurrence (all ipsilateral disease), distant recurrence (including node-positive contralateral disease and recurrences at distant sites [eg, lung, bone, etc]), or isolated contralateral events. Secondary endpoints included ER status, breast cancer mortality, other cancers, cardiovascular disease, fractures, adverse events, and non-breast cancer deaths. Prespecified subgroup analyses of recurrence were for invasive versus DCIS, contralateral versus ipsilateral, and ER status (ER positive vs ER negative); other subgroup analyses were exploratory. Further post-hoc analyses included PgR and HER2 receptor status for invasive recurrence only. Future plans are to explore outcomes by ER levels and HER2 status of the primary tumour when tissue collection is complete, and to examine timing effects of treatment after the initial 5 year treatment period is completed.

ry. Further post-hoc analyses included PgR and HER2 receptor status for invasive recurrence only. Future plans are to explore outcomes by ER levels and HER2 status of the primary tumour when tissue collection is complete, and to examine timing effects of treatment after the initial 5 year treatment period is completed. Statistical analysis All analyses were done on a modified intention-to-treat basis, including all women who were enrolled, randomly assigned, and did not revoke consent for use of their data. Analyses of the efficacy endpoints were based on hazard ratios (HRs). Cox proportional hazard models14, 15 were used to derive these with corresponding 95% CIs. The analysis plan first tested non-inferiority of anastrozole (upper 95% CI of HR <1·25) and, if successful, then for the superiority of anastrozole. Survival curves were estimated using the Kaplan-Meier method.16 Secondary endpoints were compared using odds ratios (ORs), which closely approximate relative risk for rare events. Adverse events are presented if predefined or occurred in at least 5% of participants, and Fisher's exact tests were used to compare adverse events when appropriate. Adherence was calculated using the Kaplan-Meier method, censoring at breast cancer recurrence, death, or 5 years of follow-up. All p values were two-sided.

dverse events are presented if predefined or occurred in at least 5% of participants, and Fisher's exact tests were used to compare adverse events when appropriate. Adherence was calculated using the Kaplan-Meier method, censoring at breast cancer recurrence, death, or 5 years of follow-up. All p values were two-sided. We estimated a required sample size of 4000 on the basis of a 1·6% annual recurrence rate for tamoxifen-treated patients with a 16·7% relative reduction for anastrozole to show non-inferiority, and a 33% reduction to show superiority with 5-year median follow-up. Recruitment to the trial closed on Feb 8, 2012, after enrolment of 2980 of the 4000 planned participants. During the course of the trial, local recurrences occurred at less than half the rate anticipated in the analysis plan, due largely to improvements in the surgical treatment of DCIS. Consequently, the required numbers of events anticipated in the protocol would not be reached for a number of years, and the IBIS-II steering committee, with the agreement of the independent data monitoring committee, took the decision to analyse and report the results at this stage; data were collected up to the cutoff date of Sept 30, 2015. All analyses were done using Stata version 13.1. This trial is registered at the ISRCTN registry, number ISRCTN37546358.

We estimated a required sample size of 4000 on the basis of a 1·6% annual recurrence rate for tamoxifen-treated patients with a 16·7% relative reduction for anastrozole to show non-inferiority, and a 33% reduction to show superiority with 5-year median follow-up. Recruitment to the trial closed on Feb 8, 2012, after enrolment of 2980 of the 4000 planned participants. During the course of the trial, local recurrences occurred at less than half the rate anticipated in the analysis plan, due largely to improvements in the surgical treatment of DCIS. Consequently, the required numbers of events anticipated in the protocol would not be reached for a number of years, and the IBIS-II steering committee, with the agreement of the independent data monitoring committee, took the decision to analyse and report the results at this stage; data were collected up to the cutoff date of Sept 30, 2015. All analyses were done using Stata version 13.1. This trial is registered at the ISRCTN registry, number ISRCTN37546358. Role of the funding source The study funders had no role in design, data collection, data analysis, data interpretation, or writing of the report. IS had full access to all data in the study, and JC, JFF, and AH had final responsibility for the decision to submit for publication.

All analyses were done using Stata version 13.1. This trial is registered at the ISRCTN registry, number ISRCTN37546358. Role of the funding source The study funders had no role in design, data collection, data analysis, data interpretation, or writing of the report. IS had full access to all data in the study, and JC, JFF, and AH had final responsibility for the decision to submit for publication. Results Between March 6, 2003, and Feb 8, 2012, we recruited 2980 postmenopausal women with locally excised ER-positive or PgR-positive DCIS in 236 centres from 14 countries, and randomly assigned them to receive anastrozole (n=1471) or tamoxifen (n=1509; figure 1). A total of 42 women (22 in the anastrozole group, 20 in the tamoxifen group) withdrew their consent to use their data, leaving 2938 women for the primary analysis (figure 1). Baseline characteristics are presented in table 1. A further 26 women were found to be ineligible after randomisation (figure 1) but were included in the primary analysis. Median age was 60·3 years (IQR 56·1–64·6), and 658 (22%) were older than 65 years. Median body-mass index was 26·7 kg/m2 (IQR 23·6–30·4), with 903 (31%) of women being obese (>30 kg/m2) at baseline (table 1). Median age at menarche was 13 years (IQR 12–14) and at birth of first child was 24 years (IQR 21–27), and 814 (28%) women had had a hysterectomy before trial entry. 1336 (45%) women had used menopausal hormone therapy before trial entry and two-thirds were never-smokers (table 1). Only nine women (<1%) with atypical hyperplasia or lobular carcinoma in situ were entered into the trial.

st child was 24 years (IQR 21–27), and 814 (28%) women had had a hysterectomy before trial entry. 1336 (45%) women had used menopausal hormone therapy before trial entry and two-thirds were never-smokers (table 1). Only nine women (<1%) with atypical hyperplasia or lobular carcinoma in situ were entered into the trial. Baseline DCIS tumour characteristics are also shown in table 1. Median DCIS major diameter was 13 mm (IQR 7–22), median clear margin distance was 5 mm (IQR 2–10), and most women had either intermediate-grade (1224; 42%) or high-grade (1129; 38%) tumours. Radiotherapy was given to 2091 (71%) women. Again, we noted no significant differences between treatment groups. The cutoff date for this analysis was Sept 30, 2015. Median follow-up was 7·2 years (IQR 5·6–8·9) and 21 112 women-years of follow-up were accrued (10 670 women-years for anastrozole and 10 442 tamoxifen). 5 year adherence was estimated to be 67·6% (95% CI 65·1–70·0) in the anastrozole group compared with 67·4% (64·9–69·7) in the tamoxifen group (p=0·71; appendix). The main reasons for treatment cessation were adverse events and patient decision (data not shown).

(10 670 women-years for anastrozole and 10 442 tamoxifen). 5 year adherence was estimated to be 67·6% (95% CI 65·1–70·0) in the anastrozole group compared with 67·4% (64·9–69·7) in the tamoxifen group (p=0·71; appendix). The main reasons for treatment cessation were adverse events and patient decision (data not shown). A total of 144 breast cancer recurrences were reported; recurrences were mostly invasive (84 [58%]; table 2). Numerically fewer recurrences occurred with anastrozole (67 recurrences; annual rate 0·64% [95% CI 0·50–0·82]) than for tamoxifen (77; 0·72% [0·58–0·90]; HR 0·89 [95% CI 0·64–1·23]; figure 2). The non-inferiority of anastrozole was established (upper 95% CI <1·25), but its superiority to tamoxifen was not (p=0·49). Kaplan-Meier estimates of recurrence at 5 years were 2·5% (95% CI 1·8–3·5) for anastrozole and 3·0% (2·2–4·0) for tamoxifen. After 10 years of follow-up, recurrence was 6·6% (95% CI 4·9–8·8) and 7·3% (5·7–9·4), respectively. Among the 144 recurrences, 86 (60%) were ER-positive, 30 (21%) were ER-negative, and ER status was missing for 28 (19%). Among women with ER-positive recurrences, 30 (2%) were in the anastrozole group compared with 56 (4%) in the tamoxifen group (HR 0·55 [95% CI 0·35–0·86], p=0·008). Among women with ER-negative recurrences, 17 (1%) were in the anastrozole group compared with 13 (<1%) in the tamoxifen group (HR 1·34 [95% CI 0·65–2·75], p=0·43).

omen with ER-positive recurrences, 30 (2%) were in the anastrozole group compared with 56 (4%) in the tamoxifen group (HR 0·55 [95% CI 0·35–0·86], p=0·008). Among women with ER-negative recurrences, 17 (1%) were in the anastrozole group compared with 13 (<1%) in the tamoxifen group (HR 1·34 [95% CI 0·65–2·75], p=0·43). Analyses adjusted by age, body-mass index, menopausal hormone therapy use, grade, margins, and radiotherapy subgroups yielded similar HRs as in the univariate analyses (table 2). Similar numbers of DCIS recurrences were observed in each treatment group (29 for anastrozole vs 30 for tamoxifen; HR 0·99 [95% CI 0·60–1·65], p=0·98; table 2). A total of 69 deaths had been reported by the cutoff date (appendix). Overall, we noted no statistically significant difference between treatment arms (33 for anastrozole vs 36 for tamoxifen; HR 0·93 [95% CI 0·58–1·50], p=0·78) and no specific cause of death differed by treatment group. Only four deaths from breast cancer were recorded, one in the anastrozole group and three in the tamoxifen group. Overall, the frequency of cancers other than breast was not significantly different in the anastrozole and tamoxifen groups (61 vs 71; OR 0·88 [95% CI 0·61–1·26], p=0·47; table 3). However, endometrial, ovarian, and skin cancers were significantly more common with tamoxifen (table 2).

and three in the tamoxifen group. Overall, the frequency of cancers other than breast was not significantly different in the anastrozole and tamoxifen groups (61 vs 71; OR 0·88 [95% CI 0·61–1·26], p=0·47; table 3). However, endometrial, ovarian, and skin cancers were significantly more common with tamoxifen (table 2). We collected a comprehensive record of side-effects during the 5 years of treatment (table 4). The number of women reporting any event was similar between treatment groups overall (1323 for anastrozole vs 1379 for tamoxifen) but the specific profiles were different. Fractures were significantly higher in the anastrozole group (129 vs 100; OR 1·36 [95% CI 1·03–1·80], p=0·027) and musculoskeletal adverse events such as joint stiffness, paraesthesia, carpal tunnel syndrome, and osteoporosis were also significantly higher with anastrozole (table 4). Hypercholesterolaemia was furthermore significantly more common in women receiving anastrozole compared with those receiving tamoxifen, probably as a result of the cholesterol-reducing effects of tamoxifen. By contrast, anastrozole was associated with substantially fewer muscle spasms compared with tamoxifen (25 [2%] vs 106 [7%]; OR 0·23 [95% CI 0·14–0·36], p<0·0001).

y more common in women receiving anastrozole compared with those receiving tamoxifen, probably as a result of the cholesterol-reducing effects of tamoxifen. By contrast, anastrozole was associated with substantially fewer muscle spasms compared with tamoxifen (25 [2%] vs 106 [7%]; OR 0·23 [95% CI 0·14–0·36], p<0·0001). Apart from vaginal dryness, gynaecological symptoms were significantly higher with tamoxifen. Vasomotor symptoms were common in both treatment groups, but the frequency was significantly lower with anastrozole (818 [56%] vs 899 [60%]; OR 0·85 [95% CI 0·73–0·99], p=0·0310; table 4). We noted a significant decrease in pulmonary emboli and deep vein thromboses (seven vs 24; OR 0·30 [95% CI 0·11–0·71], p=0·0028). No statistically significant difference was seen for cardiovascular events overall or myocardial infarction in particular (table 4). However, transient ischaemic attacks (13 vs five; OR 2·69 [95% CI 0·90–9·65], p=0·05) and particularly cerebrovascular accidents (13 vs four; 3·36 [1·04–14·18], 0·025) were increased with anastrozole. Despite the differences in side-effect profiles, treatment adherence was virtually identical between treatment groups and was 67·6% for anastrozole and 67·4% for tamoxifen after 5 years (appendix).

Apart from vaginal dryness, gynaecological symptoms were significantly higher with tamoxifen. Vasomotor symptoms were common in both treatment groups, but the frequency was significantly lower with anastrozole (818 [56%] vs 899 [60%]; OR 0·85 [95% CI 0·73–0·99], p=0·0310; table 4). We noted a significant decrease in pulmonary emboli and deep vein thromboses (seven vs 24; OR 0·30 [95% CI 0·11–0·71], p=0·0028). No statistically significant difference was seen for cardiovascular events overall or myocardial infarction in particular (table 4). However, transient ischaemic attacks (13 vs five; OR 2·69 [95% CI 0·90–9·65], p=0·05) and particularly cerebrovascular accidents (13 vs four; 3·36 [1·04–14·18], 0·025) were increased with anastrozole. Despite the differences in side-effect profiles, treatment adherence was virtually identical between treatment groups and was 67·6% for anastrozole and 67·4% for tamoxifen after 5 years (appendix). In a post-hoc analysis, we assessed differences by subgroups of tumour for invasive recurrence (figure 3). The largest difference was noted for invasive ER-positive/HER2-negative tumours (10 recurrences with anastrozole vs 28 with tamoxifen; HR 0·37 [95% CI 0·18–0·75], p=0·0060; figure 3). HER2-positive tumours showed better efficacy with tamoxifen (HR 1·62 [95% CI 0·53–4·96]; heterogeneity p=0·05; figure 3).

ce (figure 3). The largest difference was noted for invasive ER-positive/HER2-negative tumours (10 recurrences with anastrozole vs 28 with tamoxifen; HR 0·37 [95% CI 0·18–0·75], p=0·0060; figure 3). HER2-positive tumours showed better efficacy with tamoxifen (HR 1·62 [95% CI 0·53–4·96]; heterogeneity p=0·05; figure 3). We did not find a differential effect on recurrence according to radiotherapy use at baseline (54 recurrences with radiotherapy vs 30 with no radiotherapy; HR 0·77 [95% CI 0·49–1·21], p=0·25). Furthermore, anastrozole was not more effective at reducing invasive recurrences in those women who had radiotherapy at baseline (HR 0·77 [95% CI 0·45–1·32], p=0·34) compared with those who did not (0·86 [0·42–1·77], p=0·69; heterogeneity p=0·79; figure 3). In a post-hoc analysis, we excluded 26 women who were found to be ineligible after randomisation (figure 1). Exclusion of these women from the primary analysis did not alter the results (data not shown). Furthermore, exclusions of the nine (<1%) women with atypical hyperplasia or lobular carcinoma in-situ from a post-hoc reassessment of the primary endpoint analysis had no effect on the results (data not shown).

n (figure 1). Exclusion of these women from the primary analysis did not alter the results (data not shown). Furthermore, exclusions of the nine (<1%) women with atypical hyperplasia or lobular carcinoma in-situ from a post-hoc reassessment of the primary endpoint analysis had no effect on the results (data not shown). Discussion In this large, randomised, double-blind, placebo-controlled trial comparing anastrozole with tamoxifen in women with ER-positive or PgR-positive DCIS treated by wide local excision with or without breast radiotherapy, the non-inferiority of anastrozole to tamoxifen was demonstrated, but a significant superiority efficacy was not, although we noted a slightly lower recurrence rate for anastrozole. However, the overall event rate was lower than anticipated, which might have contributed to non-significant results with wide confidence intervals, and as a result smaller effects of anastrozole might have been missed. This possible small benefit for anastrozole is consistent with the larger 27% reduction seen in the similar NSABP B-35 trial, which was statistically significant (p=0·03).10 Trials in the adjuvant setting have also indicated greater efficacy for anastrozole and other aromatase inhibitors versus tamoxifen.17, 18 Additionally, the reduction in contralateral breast cancer is consistent with the benefits of anastrozole compared with tamoxifen seen in the ATAC trial17 and compared with placebo in the IBIS-II breast cancer prevention trial.11 The greater efficacy of anastrozole for ER-positive or HER2-negative invasive recurrence has been seen elsewhere for invasive disease.19 HER2 status was not routinely collected for the baseline tumour, but tumour blocks are being collected retrospectively and its impact along with other markers will be reported at a later stage. Local recurrence rates were lower than those predicted on the basis of earlier trials. This is probably caused by greater attention to achieving clear surgical margins, and improvements in and more frequent use of radiotherapy.

d retrospectively and its impact along with other markers will be reported at a later stage. Local recurrence rates were lower than those predicted on the basis of earlier trials. This is probably caused by greater attention to achieving clear surgical margins, and improvements in and more frequent use of radiotherapy. Clear differences were seen in the side-effect profile. Many side-effects followed the expected pattern seen during treatment of invasive cancer,18 with a higher fracture rate and more musculoskeletal events with anastrozole, and more venous thromboembolic events, gynaecological events, and vasomotor symptoms with tamoxifen. The higher rate of strokes with anastrozole is surprising because this pattern was not seen in ATAC (62 strokes with anastrozole vs 80 with tamoxifen),17 or the prevention component of IBIS-II (three with anastrozole vs six with placebo),11 and these events were not lower for tamoxifen when compared with placebo in IBIS-I (ten with tamoxifen vs 12 with placebo).20 Tamoxifen has previously been reported to reduce headache occurrence,20, 21 so the higher rate of headache in the anastrozole group of our trial probably resulted from this rather than the effects of anastrozole. However, increased hypertension was seen for anastrozole in both ATAC17 and IBIS-II prevention;11 the small increase in our anastrozole group therefore seems to be a real treatment effect, although the mechanism is not understood.

le group of our trial probably resulted from this rather than the effects of anastrozole. However, increased hypertension was seen for anastrozole in both ATAC17 and IBIS-II prevention;11 the small increase in our anastrozole group therefore seems to be a real treatment effect, although the mechanism is not understood. Although occurrence of other cancers was similar overall, the incidence of specific cancers differed by treatment. A two-to-three-fold increase in endometrial cancer is well documented for tamoxifen,22, 23 by contrast with a reduced incidence compared with the general population anticipated for anastrozole in view of the strong hormone dependence for this tumour.24 In combination, these two effects account for the striking difference seen here. Ovarian cancer is not known to be affected by tamoxifen and the differences here probably result from a preventive effect of anastrozole, as previously seen in IBIS-II prevention (four cases with anastrozole vs seven with placebo) and indirectly in ATAC (ten with anastrozole vs 17 with tamoxifen), and supported by the increased risk associated with use of menopausal hormone therapy.25 A decrease of colorectal cancer has been reported in users of menopausal hormone therapy;26 although a small increase was reported in ATAC (39 cases with anastrozole vs 31 with tamoxifen), a lower risk was seen in IBIS-II prevention (three with anastrozole vs 11 with tamoxifen), so the role of aromatase inhibitors in affecting risk of colorectal cancer remains uncertain.

users of menopausal hormone therapy;26 although a small increase was reported in ATAC (39 cases with anastrozole vs 31 with tamoxifen), a lower risk was seen in IBIS-II prevention (three with anastrozole vs 11 with tamoxifen), so the role of aromatase inhibitors in affecting risk of colorectal cancer remains uncertain. The major strengths of this study include its multinational nature, large size, moderate length of follow-up, and detailed collection of side-effect data. The major limitation of this trial was the lower-than-expected event rate, which adds uncertainty about the lack of significance of some of the small differences seen. A few unexpected side-effects were also recorded, which require further validation in view of the amount of multiple testing. It is too early to assess the effect of these treatments on mortality and long-term follow-up; a full meta-analysis of all major endpoints with the B-35 study is planned to study these issues. In summary, anastrozole offers another option for postmenopausal women with ER-positive DCIS, and the choice between it and tamoxifen will probably depend more on previous history of other conditions (eg, osteoporosis and venous thrombosis) and short-term tolerability (musculoskeletal, vasomotor, and gynaecological symptoms) than differences in efficacy. Supplementary Material Supplementary appendix

In summary, anastrozole offers another option for postmenopausal women with ER-positive DCIS, and the choice between it and tamoxifen will probably depend more on previous history of other conditions (eg, osteoporosis and venous thrombosis) and short-term tolerability (musculoskeletal, vasomotor, and gynaecological symptoms) than differences in efficacy. Supplementary Material Supplementary appendix Acknowledgments This study was funded in part by Cancer Research UK (C569/A5032, C569/A16891, C8162/A26893), the National Health and Medical Research Council Australia (GNT300755, GNT569213), and the Breast Cancer Research Fund. Support was also provided by AstraZeneca, who also provided anastrozole, tamoxifen, and matching placebos, and Aventis Sanofi who supported the bone density study. The IBIS-II study was sponsored by Queen Mary University of London, UK. Contributors JFF, IS, AH, GvM, and JC designed the study. JFF, IS, AH, BB, NB, CL, GvM, WE, PN, MS, CH, REC, and JC collected data. IS and JC analysed the data. JFF, IS, AH, REC, and JC wrote the report. JFF, IS, AH, BB, CL, GvM, WE, PN, MS, CH, REC, LJ, IE, and JC interpreted the data. Declaration of interests JC and GvM have received funding from AstraZeneca to conduct clinical trials. MS has received travel support from Roche. The other authors declare no competing interests. Figure 1 Trial profile ER=oestrogen receptor. DCIS=ductal carcinoma in situ. Figure 2 Recurrence for all breast cancer according to treatment allocation Figure 3 Subgroup analyses for invasive breast cancer by cancer characteristics

Declaration of interests JC and GvM have received funding from AstraZeneca to conduct clinical trials. MS has received travel support from Roche. The other authors declare no competing interests. Figure 1 Trial profile ER=oestrogen receptor. DCIS=ductal carcinoma in situ. Figure 2 Recurrence for all breast cancer according to treatment allocation Figure 3 Subgroup analyses for invasive breast cancer by cancer characteristics Numbers do not add to totals because of missing values. The dotted line shows no effect point, and the bold line shows overall treatment effect point. ER=oestrogen receptor. Table 1 Baseline demographic and tumour characteristics according to treatment allocation Anastrozole (n=1449) Tamoxifen (n=1489) Age, years 60·4 (56·4–64·5) 60·3 (55·8–64·5) BMI, kg/m2 26·7 (23·5–30·4) 26·7 (23·7–30·2) Age at menarche, years 13·0 (12·0–14·0) 13·0 (12·0–14·0) Age at birth of first child 24·0 (21·0–27·0) 24·0 (21·0–27·0) Smoking Never 890 (61%) 934 (63%) Ever 496 (34%) 495 (33%) Missing 63 (4%) 60 (4%) Menopausal hormone therapy use 678 (47%) 658 (44%) Hysterectomy 406 (21%) 408 (22%) Radiotherapy 1027 (71%) 1064 (71%) Tumour size, mm 13 (7–22) 13 (7–22) Margins, mm 5 (2–10) 5 (2–10) Grade Low 293 (20%) 279 (19%) Intermediate 606 (42%) 618 (42%) High 542 (37%) 587 (39%) Missing 8 (<1%) 5 (<1%) Laterality Left 742 (51%) 789 (53%) Right 703 (49%) 696 (47%) Bilateral 5 (<1%) 4 (<1%) Data are median (IQR) or n (%). Characteristics given for the modified intention-to-treat population; numbers of individuals do not add to totals because of missing values. BMI=body-mass index.

37%) 587 (39%) Missing 8 (<1%) 5 (<1%) Laterality Left 742 (51%) 789 (53%) Right 703 (49%) 696 (47%) Bilateral 5 (<1%) 4 (<1%) Data are median (IQR) or n (%). Characteristics given for the modified intention-to-treat population; numbers of individuals do not add to totals because of missing values. BMI=body-mass index. Table 2 All breast cancer, invasive, and DCIS recurrences according to treatment allocation Anastrozole (n=1449) Tamoxifen (n=1489) Unadjusted analysis Adjusted analysis* HR (95% CI) p value HR (95% CI) p value All 67 (5%) 77 (5%) 0·89 (0·64–1·23) 0·49 0·83 (0·59–1·18) 0·31 Invasive† 37 (3%) 47 (3%) 0·80 (0·52–1·24) 0·32 0·72 (0·46–1·14) 0·16 Ipsilateral 20 (1%) 22 (1%) 0·93 (0·51–1·71) 0·82 0·77 (0·40–1·48) 0·44 Contralateral 17 (1%) 25 (2%) 0·69 (0·37–1·28) 0·24 0·68 (0·36–1·29) 0·24 DCIS 29 (2%) 30‡ (2%) 0·99 (0·60–1·65) 0·98 0·98 (0·57–1·69) 0·95 Ipsilateral 21 (1%) 23 (2%) 0·94 (0·52–1·69) 0·83 1·03 (0·55–1·91) 0·93 Contralateral 8 (<1%) 6 (<1%) 1·37 (0·47–3·94) 0·56 1·02 (0·33–3·18) 0·97 DCIS=ductal carcinoma in situ. HR=hazard ratio. * Adjusted for age, body-mass index, menopausal hormone therapy, grade, margins, and radiotherapy. † 1 missing for invasiveness. ‡ 1 missing data for laterality. Table 3 Frequency of cancers other than breast according to treatment allocation

Anastrozole (n=1449) Tamoxifen (n=1489) Unadjusted analysis Adjusted analysis* HR (95% CI) p value HR (95% CI) p value All 67 (5%) 77 (5%) 0·89 (0·64–1·23) 0·49 0·83 (0·59–1·18) 0·31 Invasive† 37 (3%) 47 (3%) 0·80 (0·52–1·24) 0·32 0·72 (0·46–1·14) 0·16 Ipsilateral 20 (1%) 22 (1%) 0·93 (0·51–1·71) 0·82 0·77 (0·40–1·48) 0·44 Contralateral 17 (1%) 25 (2%) 0·69 (0·37–1·28) 0·24 0·68 (0·36–1·29) 0·24 DCIS 29 (2%) 30‡ (2%) 0·99 (0·60–1·65) 0·98 0·98 (0·57–1·69) 0·95 Ipsilateral 21 (1%) 23 (2%) 0·94 (0·52–1·69) 0·83 1·03 (0·55–1·91) 0·93 Contralateral 8 (<1%) 6 (<1%) 1·37 (0·47–3·94) 0·56 1·02 (0·33–3·18) 0·97 DCIS=ductal carcinoma in situ. HR=hazard ratio. * Adjusted for age, body-mass index, menopausal hormone therapy, grade, margins, and radiotherapy. † 1 missing for invasiveness. ‡ 1 missing data for laterality. Table 3 Frequency of cancers other than breast according to treatment allocation Anastrozole (n=1449) Tamoxifen (n=1489) OR (95% CI) p value Total 61 71 0·88 (0·61–1·26) 0·47 Gynaecological 1 17* 0·06 (0·001–0·386) 0·0002 Endometrial 1 11 0·09 (0·002–0·64) 0·0044 Ovarian 0 5 0·00 (0·00–0·79) 0·027 Lung 11 7 1·62 (0·57–4·94) 0·32 Gastrointestinal 16 10 1·65 (0·70–4·08) 0·21 Colorectal 10 5 2·06 (0·64–7·71) 0·18 Lymphoma or leukaemia 8 5 1·65 (0·47–6·42) 0·44 Skin 12 23 0·53 (0·24–1·12) 0·07 Melanoma 4 4 1·03 (0·19–5·53) 0·97 Non-melanoma 8 19 0·43 (0·16–1·03) 0·040 Other 13 9 1·49 (0·59–3·96) 0·36 OR=odds ratio. * One cervical cancer. Table 4 Adverse events reported at any time according to treatment allocation

Anastrozole (n=1449) Tamoxifen (n=1489) OR (95% CI) p value Total 61 71 0·88 (0·61–1·26) 0·47 Gynaecological 1 17* 0·06 (0·001–0·386) 0·0002 Endometrial 1 11 0·09 (0·002–0·64) 0·0044 Ovarian 0 5 0·00 (0·00–0·79) 0·027 Lung 11 7 1·62 (0·57–4·94) 0·32 Gastrointestinal 16 10 1·65 (0·70–4·08) 0·21 Colorectal 10 5 2·06 (0·64–7·71) 0·18 Lymphoma or leukaemia 8 5 1·65 (0·47–6·42) 0·44 Skin 12 23 0·53 (0·24–1·12) 0·07 Melanoma 4 4 1·03 (0·19–5·53) 0·97 Non-melanoma 8 19 0·43 (0·16–1·03) 0·040 Other 13 9 1·49 (0·59–3·96) 0·36 OR=odds ratio. * One cervical cancer. Table 4 Adverse events reported at any time according to treatment allocation Anastrozole (n=1449) Tamoxifen (n=1489) OR (95% CI) p value Fractures 129 (9%) 100 (7%) 1·36 (1·03–1·80) 0·027 Pelvic or hip 11 (1%) 4 (<1%) 2·84 (0·84–12·25) 0·06 Spine 6 (<1%) 6 (<1%) 1·03 (0·27–3·85) 0·96 Musculoskeletal (any) 929 (64%) 811 (54%) 1·49 (1·28–1·74) <0·0001 Arthralgia 832 (57%) 729 (49%) 1·41 (1·21–1·63) <0·0001 Joint stiffness 74 (5%) 35 (2%) 2·24 (1·46–3·47) <0·0001 Paraesthesia 42 (3%) 23 (2%) 1·90 (1·11–3·33) 0·013 Carpal tunnel syndrome 35 (2%) 11 (1%) 3·33 (1·64–7·29) <0·0001 Osteoporosis 97 (7%) 54 (4%) 1·91 (1·34–2·73) <0·0001 Muscle spasm 25 (2%) 106 (7%) 0·23 (0·14–0·36) <0·0001 Vasomotor or gynaecological (any) 879 (61%) 1031 (69%) 0·69 (0·59–0·80) <0.0001 Hot flushes 818 (56%) 899 (60%) 0·85 (0·73–0·99) 0·031 Vaginal dryness 189 (13%) 159 (11%) 1·25 (1·00–1·58) 0·047 Vaginal haemorrhage 35 (2%) 80 (5%) 0·44 (0·28–0·66) <0·0001 Vaginal discharge 30 (2%) 136 (9%) 0·21 (0·14–0·32) <0·0001 Vaginal candidiasis 8 (1%) 42 (3%) 0·19 (0·08–0·41) <0·0001 Other Headache 82 (6%) 61 (4%) 1·40 (0·99–2·00) 0·049 Hypercholesterolaemia 43 (3%) 11 (1%) 4·11 (2·07–8·86) <0·0001 Major thromboembolic 7 (<1%) 24 (2%) 0·30 (0·11–0·71) 0·0028 Pulmonary embolism 5 (<1%) 8 (1%) 0·64 (0·16–2·23) 0·43 Deep vein thrombosis (without pulmonary embolism) 2 (<1%) 16 (1%) 0·13 (0·01–0·54) 0·0011 Any cardiovascular 93 (6%) 84 (6%) 1·15 (0·84–1·57) 0·38 Myocardial infarction 6 (<1%) 6 (<1%) 1·03 (0·27–3·85) 0·99 Cerebrovascular accident 13 (1%) 4 (<1%) 3·36 (1·04–14·18) 0·025 Transient ischaemic attack 13 (1%) 5 (<1%) 2·69 (0·90–9·65) 0·05 Hypertension 82 (6%) 73 (5%) 1·16 (0·83–1·63) 0·36 Any eye disease 230 (16% 209 (14%) 1·16 (0·94–1·42) 0·16 Cataract 72 (5%) 61 (4%) 1·22 (0·85–1·77) 0·26 OR=odds ratio.

Introduction Happiness and related measures of wellbeing are reportedly associated with reduced mortality, particularly from heart disease.1, 2, 3, 4 Postulated mechanisms to account for this association include the possibility that happiness might itself cause biological changes, such as in serum cortisol concentration or immune function, that could in turn affect mortality.2, 3 However, serious challenges exist in interpreting the association between happiness and reduced mortality as evidence for a protective biological mechanism for happiness. Unhappiness might, for example, be associated with lifestyle factors that can cause disease,3 such as smoking, high alcohol consumption, obesity, or physical inactivity. Perhaps more important is reverse causality whereby poor health, which is known to be associated with an increase in mortality, can also cause unhappiness. This results in a non-causal association between unhappiness and increased mortality—or, equivalently, between happiness and reduced mortality. Our aim was to establish whether, after appropriate allowance for reverse causality and for confounding by lifestyle and sociodemographic factors, any robust evidence remains that happiness itself, or related subjective measures of wellbeing such as being in control, relaxed, or not unduly stressed, are independently associated with reduced mortality.

after appropriate allowance for reverse causality and for confounding by lifestyle and sociodemographic factors, any robust evidence remains that happiness itself, or related subjective measures of wellbeing such as being in control, relaxed, or not unduly stressed, are independently associated with reduced mortality. Methods Study design and participants From May 1, 1996, to Dec 31, 2001, the Million Women Study recruited 1·3 million women aged 50–69 years through the national Breast Screening Programmes of England and Scotland, and has continued to follow them up by electronic record linkage, recording the causes of any deaths.5 At recruitment, and every 3–5 years subsequently, women were posted a questionnaire asking about sociodemographic factors, lifestyle, and health. Ethics approval was from the Anglia and Oxford multicentre research ethics committee. Access to hospital admission data was approved by the Information Centre for Health and Social Care (England) and the Information Services Division (Scotland). All study participants provided written consent.

Methods Study design and participants From May 1, 1996, to Dec 31, 2001, the Million Women Study recruited 1·3 million women aged 50–69 years through the national Breast Screening Programmes of England and Scotland, and has continued to follow them up by electronic record linkage, recording the causes of any deaths.5 At recruitment, and every 3–5 years subsequently, women were posted a questionnaire asking about sociodemographic factors, lifestyle, and health. Ethics approval was from the Anglia and Oxford multicentre research ethics committee. Access to hospital admission data was approved by the Information Centre for Health and Social Care (England) and the Information Services Division (Scotland). All study participants provided written consent. Procedures At baseline 3 years after recruitment, women were asked: “How often do you feel happy?” Possible responses were “most of the time”, “usually”, “sometimes”, or “rarely/never”. They were also asked about related subjective measures of wellbeing including how often they felt in control, relaxed, and stressed. In addition, women were asked whether they had had various common health disorders and to self-rate their current health as “excellent”, “good”, “fair”, or “poor”. In the questionnaire, self-rated health came before happiness and related measures. We used data from this 3 year survey as baseline for our investigation of any associations of unhappiness (or related factors) with cause-specific mortality, and our analyses are restricted to the women who answered this question on happiness. A random sample of women were re-sent the same questionnaire about 1 year after the first one to assess the repeatability of responses.6

estigation of any associations of unhappiness (or related factors) with cause-specific mortality, and our analyses are restricted to the women who answered this question on happiness. A random sample of women were re-sent the same questionnaire about 1 year after the first one to assess the repeatability of responses.6 All participants in the Million Women Study are routinely followed for death (or emigration), cancer registration, and hospital admission through electronic linkage to centrally held National Health Service (NHS) records, using a combination of name, date of birth, and NHS number. Underlying causes of death, cancers, and hospital admissions are coded according to WHO Tenth International Classification of Diseases (ICD-10). Follow-up time was from the date when the baseline questionnaire on happiness was answered to whichever was first of Jan 1, 2012, or date of death or emigration. Outcomes and exposures Our outcomes were mortality from all causes, from ischaemic heart disease (ICD-10 I20-I25), and from cancer (ICD-10 C00-C97). We classified women into three categories: unhappy (ie, sometimes, rarely, or never happy), usually happy, or happy most of the time. Because the latter two categories are similar, for some analyses we combined them into one category, called generally happy. Associations of mortality with other subjective measures of wellbeing (being in control, relaxed, and stressed) were also examined.

ely, or never happy), usually happy, or happy most of the time. Because the latter two categories are similar, for some analyses we combined them into one category, called generally happy. Associations of mortality with other subjective measures of wellbeing (being in control, relaxed, and stressed) were also examined. Statistical analysis For analyses examining which baseline factors were associated with happiness we used logistic regression (adjusted for various factors) to compare individuals who were unhappy with those who were generally happy (two-way classification). For the analyses of the association between unhappiness and mortality we used the three-way classification (with those happy most of the time as the reference group), but for clarity in the text, we report only the mortality rate ratios (RR) for unhappy versus happy most of the time. For analyses of all-cause mortality, ischaemic heart disease mortality, and cancer mortality, we used Cox proportional hazards models. We did sensitivity analyses to exclude the first 5 years of follow-up. We repeated such analyses for related measures of subjective wellbeing: being in control, being relaxed, and being stressed. To limit reverse causality, the main mortality analyses excluded women who had already had certain illnesses (heart disease, stroke, lung disease, or cancer, as done previously7); additional analyses assessed the effects of these exclusions.

Statistical analysis For analyses examining which baseline factors were associated with happiness we used logistic regression (adjusted for various factors) to compare individuals who were unhappy with those who were generally happy (two-way classification). For the analyses of the association between unhappiness and mortality we used the three-way classification (with those happy most of the time as the reference group), but for clarity in the text, we report only the mortality rate ratios (RR) for unhappy versus happy most of the time. For analyses of all-cause mortality, ischaemic heart disease mortality, and cancer mortality, we used Cox proportional hazards models. We did sensitivity analyses to exclude the first 5 years of follow-up. We repeated such analyses for related measures of subjective wellbeing: being in control, being relaxed, and being stressed. To limit reverse causality, the main mortality analyses excluded women who had already had certain illnesses (heart disease, stroke, lung disease, or cancer, as done previously7); additional analyses assessed the effects of these exclusions. RRs of death were first adjusted only for age, and then additionally adjusted for various combinations of self-rated health and sociodemographic and lifestyle characteristics. These characteristics were region of residence at recruitment (Scotland and the nine cancer registration regions covering England at that time); area deprivation (quintiles, based on the Townsend Index, a score incorporating census area data for employment, car ownership, home ownership, and household overcrowding8); educational achievement (college [after age 18 years], A-level qualifications [usually at age 18 years], O-level qualifications [usually at age 16 years], none of these); whether living with a partner (yes, no), parity (0, 1, ≥2), body-mass index (<25 kg/m2, 25 to <30 kg/m2, ≥30 kg/m2); strenuous exercise (0, <3 h per week, ≥3 h per week); smoking (never, past, current <15 cigarettes per day, current ≥15 cigarettes per day); alcohol consumption (0, <7 drinks per week, ≥7 drinks per week); hours of sleep (<7 h, 7 h, 8 h, ≥9 h); and participation in religious groups (yes, no) or other group activities (yes, no). All adjustment variables were from the baseline survey (ie, at the same time that happiness and related measures were recorded), except region, deprivation, education, and parity, which were recorded at recruitment, about 3 years earlier.

participation in religious groups (yes, no) or other group activities (yes, no). All adjustment variables were from the baseline survey (ie, at the same time that happiness and related measures were recorded), except region, deprivation, education, and parity, which were recorded at recruitment, about 3 years earlier. We used conventional 95% CIs or 99% CIs, except in figures that compared more than two groups. For these comparisons, the variance of the log risk was estimated for each group (including the reference group).9 We used these group-specific variances to calculate group-specific CIs, allowing valid comparisons between any two or more groups, whether or not one of them was designated as the reference group. Analyses were done with STATA version 13.1. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results At baseline, a total of 845 440 women (median age 59 years; IQR 55–63) responded to the question about how often they felt happy. Replies were: 39% (329 326 women) happy most of the time, 44% (369 738) usually happy, 16% (138 678) sometimes happy, and 1% (7698) rarely or never happy. In all analyses we combined women who reported being happy sometimes, rarely, or never, and describe them as unhappy.

o the question about how often they felt happy. Replies were: 39% (329 326 women) happy most of the time, 44% (369 738) usually happy, 16% (138 678) sometimes happy, and 1% (7698) rarely or never happy. In all analyses we combined women who reported being happy sometimes, rarely, or never, and describe them as unhappy. Among 10 143 women who completed the same happiness question twice, about 1 year apart, there was reasonably good reproducibility between the categorised responses (weighted κ for agreement 0·62). The two extreme categories had little crossover. Of 4003 women who reported being happy most of the time at baseline, only 2% (85) reported being unhappy 1 year later; conversely, of 1763 women who reported being unhappy at baseline, only 5% (81) reported being happy most of the time 1 year later. The strongest sociodemographic and lifestyle correlates of being generally happy were increasing age, having fewer educational qualifications, doing strenuous exercise, not smoking, living with a partner, and participating in religious and other group activities (figure 1). The relation between happiness and the number of hours of sleep was J-shaped, with women reporting about 8 h sleep most likely to be generally happy. Each of the indices of ill health at baseline was associated with unhappiness (figure 2). Of all factors shown in figure 1 or figure 2, the strongest associations with reported unhappiness were treatment for depression or anxiety and reporting only fair or poor general health (figure 2).

t likely to be generally happy. Each of the indices of ill health at baseline was associated with unhappiness (figure 2). Of all factors shown in figure 1 or figure 2, the strongest associations with reported unhappiness were treatment for depression or anxiety and reporting only fair or poor general health (figure 2). Women were followed for a mean of 9·6 years (SD 1·9) after completing the questionnaire about happiness. Including women with and without illness at baseline, 48 314 deaths were recorded in this time. Compared with those reporting being happy most of the time, women who had reported being unhappy had excess all-cause mortality when adjusted only for age (RR 1·36, 95% CI 1·33–1·40). Simultaneous adjustment for the sociodemographic and lifestyle factors in figure 1 and the indices of health in figure 2 completely eliminated this excess (fully adjusted RR 0·95, 0·93–0·98; appendix p 4). However, these analyses include women who already had life-threatening diseases at baseline. Hence, our subsequent analyses exclude the 125 769 women who at baseline already had heart disease, stroke, cancer, or chronic obstructive airways disease. These excluded women had three times the death rate of the women without any such illnesses (age-adjusted RR 2·91, 95% CI 2·85–2·96), and are omitted from the main analyses below.

equent analyses exclude the 125 769 women who at baseline already had heart disease, stroke, cancer, or chronic obstructive airways disease. These excluded women had three times the death rate of the women without any such illnesses (age-adjusted RR 2·91, 95% CI 2·85–2·96), and are omitted from the main analyses below. Of the remaining 719 671 women (median age 59 years, IQR 55–63), 4% (31 531) died during follow-up. In response to the question at baseline about how often they felt happy, 39% (282 619) reported being happy most of the time, 44% (315 874) usually happy, and 17% (121 178) unhappy. In crude analyses adjusted only for age, unhappiness remained associated with increased mortality (RR 1·29, 95% CI 1·25–1·33; table). This excess risk was partly accounted for by associations with various personal characteristics (appendix p 5). Self-rated health was, however, the key characteristic. Poor health at baseline was strongly associated with unhappiness at baseline (figure 2) and once we adjusted for self-rated health, unhappiness was no longer significantly associated with all-cause mortality (RR 1·02, 0·98–1·05; table). After simultaneous adjustment for all sociodemographic and lifestyle factors in figure 1 (personal characteristics) and all indices of health in figure 2, the association vanished (fully adjusted RR 0·98, 0·94–1·01, for all-cause mortality [table]; 0·97, 0·87–1·10, for ischaemic heart disease mortality; and 0·98, 0·93–1·02, for cancer mortality [appendix pp 8–9]).

ographic and lifestyle factors in figure 1 (personal characteristics) and all indices of health in figure 2, the association vanished (fully adjusted RR 0·98, 0·94–1·01, for all-cause mortality [table]; 0·97, 0·87–1·10, for ischaemic heart disease mortality; and 0·98, 0·93–1·02, for cancer mortality [appendix pp 8–9]). Further details of this multivariate adjustment are available (appendix p 5), showing that after adjusting for age, additional adjustment for each single personal characteristic in figure 1 changed the RR estimate only slightly (adjustment for smoking had the greatest effect). Simultaneous adjustment for all personal characteristics, however, halved the association between unhappiness and mortality (RR 1·14, 95% CI 1·11–1·18). Adjustment just for being treated for common health disorders (hypertension, diabetes, asthma, arthritis, depression, or anxiety), particularly depression or anxiety, also weakened the relationship (RR 1·21, 1·17–1·25; appendix p 5). The main findings were essentially unchanged in sensitivity analysis that ignored the first 5 years of follow-up (appendix p 6).

on health disorders (hypertension, diabetes, asthma, arthritis, depression, or anxiety), particularly depression or anxiety, also weakened the relationship (RR 1·21, 1·17–1·25; appendix p 5). The main findings were essentially unchanged in sensitivity analysis that ignored the first 5 years of follow-up (appendix p 6). Because self-rated health was so strongly associated with both happiness and mortality, we examined the associations between happiness and mortality in women separately by self-rated health (table, figure 3, appendix pp 7–9, pp 13–15). All-cause mortality was substantially greater for the 20% (134 727 of 685 464) of women who reported that their health was fair or poor than for the remaining 80% (550 737) of women who reported good or excellent health (RR 1·67, 95% CI 1·63–1·71). Within each category of self-rated health there was no significant excess mortality in individuals who reported being unhappy (table, figure 3). For women who reported only fair or poor health, mortality was actually lower in those who reported being unhappy compared with those who were happy most of the time, but these findings might be biased by some unhappy women tending to rate their general health worse than it was, thus producing a spuriously low mortality associated with being unhappy. Women reporting being in good or excellent health are less liable to such a potential bias, and we give results for these women only in figure 4 and figure 5, with results for women reporting fair or poor health in the appendix.

worse than it was, thus producing a spuriously low mortality associated with being unhappy. Women reporting being in good or excellent health are less liable to such a potential bias, and we give results for these women only in figure 4 and figure 5, with results for women reporting fair or poor health in the appendix. Among 550 737 women reporting good or excellent health, 1253 died from ischaemic heart disease and 12 943 from cancer; among these women, unhappiness was not associated with mortality from either cause (figure 4). Nor was unhappiness associated with mortality from these causes for women who reported being in poor or fair health (appendix pp 8–9, pp 14–15). Being treated for depression or anxiety was also strongly associated with self-reported unhappiness, so we did analyses separately for women who were and were not being treated for depression or anxiety at baseline (appendix pp 16–17). Again, unhappiness did not seem to be related to mortality in any of the subgroups of self-rated health. We examined the associations of mortality with three other related subjective measures of wellbeing (being in control, relaxed, or not stressed). Correlates of each measure with sociodemographic, lifestyle, and health indices at baseline were similar to those found for happiness (appendix pp 18–19), even though the different measures were themselves not strongly correlated (appendix p 10). Among women reporting good or excellent health, not being in control, not being relaxed, and being stressed were all not associated with increased mortality (figure 5; appendix pp 11–12).

found for happiness (appendix pp 18–19), even though the different measures were themselves not strongly correlated (appendix p 10). Among women reporting good or excellent health, not being in control, not being relaxed, and being stressed were all not associated with increased mortality (figure 5; appendix pp 11–12). Discussion Poor health can cause unhappiness and poor health increases mortality, so unhappiness is associated with increased mortality. Additionally, unhappiness might correlate with some adverse lifestyle choices. After allowance for these associations our large prospective study shows no robust evidence that happiness itself reduces cardiac, cancer, or overall mortality.

alth increases mortality, so unhappiness is associated with increased mortality. Additionally, unhappiness might correlate with some adverse lifestyle choices. After allowance for these associations our large prospective study shows no robust evidence that happiness itself reduces cardiac, cancer, or overall mortality. There is no perfect or generally agreed way to measure happiness or related subjective indices of wellbeing. Different approaches thus limit comparability between studies. We used a single question about happiness with a four-point scale, whereas other investigators have used different measures.10, 11, 12, 13, 14 Nevertheless, we are able to show the validity of our measure in three ways. First, personal factors found to be associated with happiness in this study (figure 1, figure 2) were similar to those reported by others who used either single-item or multi-item measures of happiness—ie, women were more likely to report feeling happy if they were older,15, 16, 17 less deprived,13, 15, 16, 18 physically active,11, 13, 19, 20, 21 did not smoke,3, 12, 20, 21, 22 had a partner,17, 18, 23 belonged to a religious group or participated in social activities,17, 18 and had adequate sleep3, 21 (but not too much). Women were also less likely to be happy if they had poor self-rated health or were being treated for various common health disorders, particularly depression or anxiety.16, 17, 18, 21, 24 Second, in analyses that were adjusted only for age and not corrected for other factors, our measure of unhappiness was correlated with increased mortality. Third, the response to our single question on happiness was reasonably repeatable on resurvey 1 year later (weighted κ=0·62), a level of repeatability comparable with that reported by other researchers using both single-item and multiple-item measures.3, 13, 25 Also, there was minimal cross-over in responses 1 year apart between the two extremes of happy most of the time or unhappy.

nably repeatable on resurvey 1 year later (weighted κ=0·62), a level of repeatability comparable with that reported by other researchers using both single-item and multiple-item measures.3, 13, 25 Also, there was minimal cross-over in responses 1 year apart between the two extremes of happy most of the time or unhappy. Crude analyses, adjusting only for age, showed some excess mortality to be associated with unhappiness, but this excess was completely eliminated after additional adjustment for personal characteristics and for poor health at baseline (table). This was true for all-cause mortality and, separately, for mortality from ischaemic heart disease and from cancer. Far fewer women died from heart disease than from cancer and confounding was greater, so our null findings are less definite for heart disease than for cancer.

tics and for poor health at baseline (table). This was true for all-cause mortality and, separately, for mortality from ischaemic heart disease and from cancer. Far fewer women died from heart disease than from cancer and confounding was greater, so our null findings are less definite for heart disease than for cancer. By far the most important adjustment factor was self-rated health. A systematic review of previous studies has confirmed that self-rated health predicts an increased risk of death, in agreement with our findings.26 Self-rated poor health was also strongly associated with unhappiness. Hence we examined the effects of happiness in categories of self-rated health, giving most weight to the findings in the many women who reported that they were in good or excellent health. In these women, unhappiness was not associated with an increased mortality. In women who reported that they were in fair or poor health, being unhappy was associated with a slightly lower mortality than being happy most of the time, but this finding could well be biased by unhappy women tending to rate their health as slightly worse than it actually was.

associated with an increased mortality. In women who reported that they were in fair or poor health, being unhappy was associated with a slightly lower mortality than being happy most of the time, but this finding could well be biased by unhappy women tending to rate their health as slightly worse than it actually was. Unhappiness might cause some people to do things known to affect mortality adversely—eg, smoke or be inactive.2, 3 Hence, such variables might be mediators rather than confounders of the unhappiness–mortality association. Furthermore, some lifestyle factors such as inactivity and morbid obesity could cause unhappiness. However, adjustment for most of the behavioural factors, except smoking, resulted in little or no attenuation of the RR estimates for mortality, suggesting that even if these factors were mediating the association, their contribution is small. Adjustment for smoking caused a greater attenuation of the RR estimates than did adjustment for any other personal characteristics, so it is possible that part of the association between unhappiness and mortality, particularly for cancer mortality, might be mediated by smoking.

association, their contribution is small. Adjustment for smoking caused a greater attenuation of the RR estimates than did adjustment for any other personal characteristics, so it is possible that part of the association between unhappiness and mortality, particularly for cancer mortality, might be mediated by smoking. Some, but not all, other prospective studies have reported that happiness or related subjective measures of wellbeing are associated with lower all-cause mortality (panel).4, 13, 14, 19, 20, 22, 27, 28, 29 However, few of those reports excluded people with life-threatening illnesses at baseline and adjusted for self-rated health (or related measures of ill health) at baseline. Self-rated health was the most important confounding factor in our analyses; where other investigators adjusted for self-rated health, any apparent excess mortality associated with unhappiness was attenuated or disappeared completely.27 If there is inadequate allowance for ill health at baseline, any associations between happiness and lower mortality are likely to be artefactual. Participants in the Million Women Study were slightly less likely to be from deprived areas than were the general UK population.5 However, at recruitment the cohort included about one in four women in England and Scotland in the eligible age range, indicating that findings should be generally applicable to middle-aged women in the UK. We provide no data about men or women of other ages.

be from deprived areas than were the general UK population.5 However, at recruitment the cohort included about one in four women in England and Scotland in the eligible age range, indicating that findings should be generally applicable to middle-aged women in the UK. We provide no data about men or women of other ages. It has been suggested that related subjective measures of wellbeing, including being in control, not being unduly stressed, or having positive or negative attitudes to life, could independently affect mortality.4, 12, 14, 30 However, just as for happiness, these associations were wholly accounted for by personal characteristics and ill health at baseline—after adjusting for these factors, any association with mortality was eliminated. We conclude that happiness and unhappiness have no material direct effect upon mortality. Supplementary Material Supplementary appendix Acknowledgments This study was supported by Cancer Research UK, the British Heart Foundation, and the UK Medical Research Council. BL is supported by a fellowship from the Australian National Health and Medical Research Council. We thank the Million Women Study participants, the National Health Service breast screening centres, and the staff from the Million Women Study coordinating centre. Contributors All authors contributed to study conception, design, analysis, and manuscript writing or revision. Declaration of interests We declare no competing interests. Figure 1 Correlates of being generally happy—relevance of personal and lifestyle characteristics at baseline

Acknowledgments This study was supported by Cancer Research UK, the British Heart Foundation, and the UK Medical Research Council. BL is supported by a fellowship from the Australian National Health and Medical Research Council. We thank the Million Women Study participants, the National Health Service breast screening centres, and the staff from the Million Women Study coordinating centre. Contributors All authors contributed to study conception, design, analysis, and manuscript writing or revision. Declaration of interests We declare no competing interests. Figure 1 Correlates of being generally happy—relevance of personal and lifestyle characteristics at baseline Analysis for whole population (N=845 440), including women later excluded for life-threatening health disorders. *ORs are adjusted for age, region, area deprivation, body-mass index, qualifications, strenuous exercise, smoking, alcohol, living with a partner, parity, participation in group activities, and sleep duration. OR=odds ratio. g-s CI=group-specific confidence interval. Figure 2 Correlates of being generally happy—relevance of various indices of health at baseline Analysis for whole population (N=845 440), including women later excluded for life-threatening health disorders. ORs are adjusted for age, region, area deprivation, body-mass index, qualifications, strenuous exercise, smoking, alcohol, living with a partner, parity, participation in group activities, and sleep duration. OR=odds ratio. Figure 3 RR of all-cause mortality by self-rated health and happiness

Analysis for whole population (N=845 440), including women later excluded for life-threatening health disorders. ORs are adjusted for age, region, area deprivation, body-mass index, qualifications, strenuous exercise, smoking, alcohol, living with a partner, parity, participation in group activities, and sleep duration. OR=odds ratio. Figure 3 RR of all-cause mortality by self-rated health and happiness Includes 719 671 women (31 531 deaths). Excludes women with cancer, heart disease, stroke, or chronic obstructive airways disease at baseline. RRs are adjusted for age, region, area deprivation, body-mass index, qualifications, strenuous exercise, smoking, alcohol, living with a partner, parity, participation in group activities, and sleep duration. Women who reported being in good or excellent health and happy most of the time are the reference group (RR=1·0). RR=rate ratio. g-s CI=group-specific confidence interval. Figure 4 Risk of ischaemic heart disease mortality and cancer mortality by happiness in women who rated their health as good or excellent at baseline

Includes 719 671 women (31 531 deaths). Excludes women with cancer, heart disease, stroke, or chronic obstructive airways disease at baseline. RRs are adjusted for age, region, area deprivation, body-mass index, qualifications, strenuous exercise, smoking, alcohol, living with a partner, parity, participation in group activities, and sleep duration. Women who reported being in good or excellent health and happy most of the time are the reference group (RR=1·0). RR=rate ratio. g-s CI=group-specific confidence interval. Figure 4 Risk of ischaemic heart disease mortality and cancer mortality by happiness in women who rated their health as good or excellent at baseline Includes 550 737 women (1253 ischaemic heart disease deaths, 12 943 cancer deaths). Excludes women with cancer, heart disease, stroke, or chronic obstructive airways disease at baseline, and women who rated their health as poor or fair at baseline. RRs are adjusted for age, region, area deprivation, body-mass index, qualifications, strenuous exercise, smoking, alcohol, living with a partner, parity, participation in group activities, and sleep duration. Women who reported being happy most of the time are the reference group (RR=1·0). RR=rate ratio. g-s CI=group-specific confidence interval. Figure 5 All-cause mortality by happiness and other measures of wellbeing in women who rated their health as good or excellent at baseline

Includes 550 737 women (1253 ischaemic heart disease deaths, 12 943 cancer deaths). Excludes women with cancer, heart disease, stroke, or chronic obstructive airways disease at baseline, and women who rated their health as poor or fair at baseline. RRs are adjusted for age, region, area deprivation, body-mass index, qualifications, strenuous exercise, smoking, alcohol, living with a partner, parity, participation in group activities, and sleep duration. Women who reported being happy most of the time are the reference group (RR=1·0). RR=rate ratio. g-s CI=group-specific confidence interval. Figure 5 All-cause mortality by happiness and other measures of wellbeing in women who rated their health as good or excellent at baseline Includes 550 737 women (20 073 deaths). Excluding women with cancer, heart disease, stroke, or chronic obstructive airways disease at baseline and women who rated their health as poor or fair at baseline. RRs are adjusted for age, region, area deprivation, body-mass index, qualifications, strenuous exercise, smoking, alcohol, living with a partner, parity, participation in group activities, and sleep duration. The referenced groups (RR=1·0) were women who reported being happy most of the time (A); in control most of the time (B); relaxed most of the time (C); and rarely or never stressed (D). g-s CIs that are not visible are smaller than the solid circle. RR=rate ratio. g-s CI=group-specific confidence interval.

ep duration. The referenced groups (RR=1·0) were women who reported being happy most of the time (A); in control most of the time (B); relaxed most of the time (C); and rarely or never stressed (D). g-s CIs that are not visible are smaller than the solid circle. RR=rate ratio. g-s CI=group-specific confidence interval. Table Effects of adjustment for personal characteristics and various indices of health on the association between all-cause mortality and how often women reported being happy

ep duration. The referenced groups (RR=1·0) were women who reported being happy most of the time (A); in control most of the time (B); relaxed most of the time (C); and rarely or never stressed (D). g-s CIs that are not visible are smaller than the solid circle. RR=rate ratio. g-s CI=group-specific confidence interval. Table Effects of adjustment for personal characteristics and various indices of health on the association between all-cause mortality and how often women reported being happy Unhappy* Usually happy Happy most of the time All women Number of women 121 178 315 874 282 619 Number of deaths 6052 13 720 11 759 RR (95% CI) of all-cause mortality, adjusted for: Age only 1·29 (1·25–1·33) 1·05 (1·03–1·08) Ref Age and personal characteristics† 1·14 (1·11–1·18) 1·04 (1·02–1·07) Ref Age and self-rated health‡ 1·02 (0·98–1·05) 0·97 (0·95–1·00) Ref Age, characteristics†, and self-rated health‡ 0·97 (0·94–1·00) 0·98 (0·96–1·01) Ref Age, characteristics†, self-rated health‡, and treatment for common health disorders§ 0·98 (0·94–1·01) 0·99 (0·96–1·01) Ref In women reporting poor or fair health at baseline Number of women 46 547 56 447 31 733 Number of deaths 3193 4049 2364 RR (95% CI) of all-cause mortality, adjusted for: Age only 0·99 (0·94–1·04) 0·97 (0·92–1·02) Ref Age, characteristics†, and treatment for common health disorders§ 0·93 (0·88–0·99) 0·97 (0·93–1·03) Ref In women reporting good or excellent health at baseline Number of women 68 762 244 488 237 487 Number of deaths 2509 8852 8712 RR (95% CI) of all-cause mortality, adjusted for: Age only 1·06 (1·02–1·11) 1·00 (0·97–1·03) Ref Age, characteristics†, and treatment for common health disorder§ 1·01 (0·97–1·06) 1·00 (0·97–1·03) Ref Analyses are limited to the 719 671 women without cancer, heart disease, stroke, or chronic obstructive airways disease at baseline.

‡ Self-rated health at baseline, in three categories: poor or fair, good, and excellent (the numbers for all women include the few who did not answer the question on self-rated health at baseline). § Treatment at baseline for high blood pressure, diabetes, asthma, arthritis, depression or anxiety (appendix p 5 gives the corresponding result adjusted only for treatment for depression or anxiety). RR=rate ratio. Ref=reference group. Panel Research in context Systematic review In November, 2014, we searched MEDLINE and PubMed for relevant reports with the terms “happiness”, “positive affect”, “well-being”, and “mortality” with no language or date restrictions. At least three comprehensive reviews had been done in the previous 6 years on the relation between happiness or related measures and mortality. We therefore did not do a new systematic review. The reviews and individual studies have reported inconsistent findings. Some suggest that happiness or related subjective measures of wellbeing are independently associated with decreased mortality, and some suggest no effect after adjustment for potential confounding factors such as poor health. Interpretation Some previous reports have confused cause and effect. Our findings show that unhappiness is associated with poor health mainly because poor health causes unhappiness and partly because unhappiness is associated with lifestyle factors such as smoking. After adjustment for these factors, no robust evidence remains that unhappiness or stress increase mortality or that being happy, relaxed, or in control reduces mortality.

Introduction In the past 15 years, renewed investment in malaria control has led to substantial reductions in malaria worldwide.1, 2 Despite this investment, malaria remains a leading cause of morbidity and mortality.1 Investigators have completed phase 3 testing of the RTS,S/AS01 Plasmodium falciparum vaccine candidate in two age groups at 11 centres in sub-Saharan Africa.3 Efficacy against clinical malaria was 20·3% (95% CI 13·6–26·5) in infants aged 6–12 weeks and 35·2% (30·5–39·5) in children aged 5–17 months during 32 months of follow-up. Vaccine efficacy against clinical malaria declined over time, from 45·1% (95% CI 41·4–48·7%) during months 0–20 to 16·1% (8·5–23·0%) during months 21–32 in children, and from 27·0% (21·1–32·5) to 7·6% (−1·4 to 15·9%), respectively, in infants. In the group given a fourth dose of vaccine 18 months after the initial course, efficacy against clinical malaria was 43·9% (95% CI 39·7–47·8) in children and 27·8% (21·7–33·4) in infants over 32 months.

3·0%) during months 21–32 in children, and from 27·0% (21·1–32·5) to 7·6% (−1·4 to 15·9%), respectively, in infants. In the group given a fourth dose of vaccine 18 months after the initial course, efficacy against clinical malaria was 43·9% (95% CI 39·7–47·8) in children and 27·8% (21·7–33·4) in infants over 32 months. Levels of malaria incidence and seasonality profiles varied across the 11 trial sites, and malaria incidence ranged from 0·03 to 4·27 clinical episodes per infant per year in the control group, which is broadly representative of settings in Africa.4 However, these estimates were obtained in cohorts with a high coverage of long-lasting insecticide-treated nets (75–80% in both intervention and control groups) and in the presence of high levels of good-quality access to care. Thus, mortality was lower in children in the control group than in the general population outside the trial and elsewhere in Africa, as shown for one of the study sites.5 Of note was the very small number of malaria deaths (19 [<1%] of 8922 children died over a median of 48·1 months [IQR 39–50] of follow-up3), probably because of the high level of care for trial participants. Estimates of the public health impact and cost-effectiveness of the vaccine in different African settings with more typical access to care and current intervention coverage are needed to inform global and national decisions about vaccine introduction. In particular, estimations should be made of the impact on both morbidity and mortality in the absence of high levels of treatment to ensure that appropriate comparisons can be made between the cost-effectiveness of this vaccine and other childhood vaccines.

orm global and national decisions about vaccine introduction. In particular, estimations should be made of the impact on both morbidity and mortality in the absence of high levels of treatment to ensure that appropriate comparisons can be made between the cost-effectiveness of this vaccine and other childhood vaccines. Research in context Evidence before this study

orm global and national decisions about vaccine introduction. In particular, estimations should be made of the impact on both morbidity and mortality in the absence of high levels of treatment to ensure that appropriate comparisons can be made between the cost-effectiveness of this vaccine and other childhood vaccines. Research in context Evidence before this study The phase 3 trial of the RTS,S/AS01 vaccine provides estimates of cases averted at the 11 trial sites. However, these estimates need to be translated into full population impact to calculate the cost-effectiveness of the vaccine compared with other malaria interventions. Mathematical models have previously estimated the public health impact and cost-effectiveness of malaria vaccines in different settings. We searched PubMed, the Cochrane Library, and other relevant data sources between April 1, 2015, and June 8, 2015, for studies of predictions and the cost-effectiveness of RTS,S malaria vaccine. The literature covered the period 2006, to June, 2015. We searched PubMed with the MeSH terms “RTS,S” [All Fields] AND “malaria” [All Fields] AND “model” [All Fields] OR “simulation” [All Fields] OR “prediction” [All Fields]. For the Cochrane Library and other data sources, we used the search terms “RTS,S”, AND “predictions”. The 29 manuscripts identified included 14 reports that referred to RTS,S or pre-erythrocytic vaccines and model analysis or predictions. Ten of these reports did not use trial data to parameterise models or were limited to data from phase 2 trials of RTS,S with a different adjuvant (AS02) or for vaccination in a different age group, whereas the other four reports made predictions of RTS,S or similar malaria vaccine candidates, but with vaccines parameterised from either intermediate phase 3 results or phase 2 trial data for RTS,S/AS01. Of these 14 publications, 12 were based on models included in the present study. A search for “malaria model” [All Fields] and “comparison” [All Fields] identified no publications that compared malaria transmission models.

ines parameterised from either intermediate phase 3 results or phase 2 trial data for RTS,S/AS01. Of these 14 publications, 12 were based on models included in the present study. A search for “malaria model” [All Fields] and “comparison” [All Fields] identified no publications that compared malaria transmission models. Added value of this study This is the first study to estimate the public health impact of the vaccine and its cost-effectiveness in populations beyond the trial with multiple mathematical models to address structural model uncertainty. This is also the first modelling study to use final site-specific results of the RTS,S phase 3 clinical trial and to systematically compare the malaria transmission models. Predictions are made for the full range of Plasmodium falciparum parasite prevalence settings in Africa. Implications of all available evidence

This is the first study to estimate the public health impact of the vaccine and its cost-effectiveness in populations beyond the trial with multiple mathematical models to address structural model uncertainty. This is also the first modelling study to use final site-specific results of the RTS,S phase 3 clinical trial and to systematically compare the malaria transmission models. Predictions are made for the full range of Plasmodium falciparum parasite prevalence settings in Africa. Implications of all available evidence The RTS,S malaria vaccine candidate provided modest protection against clinical malaria in children across different disease burden settings in the phase 3 clinical trial. A WHO recommendation for vaccine use in endemic Africa will additionally need to consider the public health impact and cost-effectiveness informed by our model predictions. This study reports on the major collaborative exercise coordinated by WHO undertaken for this purpose. Our results quantify the potential number of cases and deaths that could be averted by the RTS,S/AS01 vaccine when implemented in moderate to high transmission with full vaccine coverage and in the presence of existing malaria control interventions. Decisions about implementation will need to consider levels of malaria burden, the cost-effectiveness and coverage of other malaria interventions, health priorities, financing, and the capacity of the health system to deliver the vaccine.

l vaccine coverage and in the presence of existing malaria control interventions. Decisions about implementation will need to consider levels of malaria burden, the cost-effectiveness and coverage of other malaria interventions, health priorities, financing, and the capacity of the health system to deliver the vaccine. Mathematical models of malaria dynamics can be useful for estimation of the potential public health impacts of malaria vaccination beyond the efficacy estimates obtained from the individually randomised clinical trial. An important feature of these models is that they have been extensively parameterised to routine data from endemic settings. Additionally, the models capture potential age-shifting of cases to older ages, as shown for severe malaria in the phase 3 trial, but that would not be expected to be shown in 4 years of follow-up. Furthermore, via use of data relating clinical disease to death, models can be used to estimate effects on mortality based on realistic assumptions about treatment in non-trial settings. In line with WHO recommendations about use of cost-effectiveness information within the decision-making process for introduction of new vaccines,6, 7 and for impact estimates from GAVI, the Vaccine Alliance,8 we systematically compared four malaria transmission models to assess the public health impact and cost-effectiveness of routine use of the RTS,S/AS01 vaccine in African settings.

ormation within the decision-making process for introduction of new vaccines,6, 7 and for impact estimates from GAVI, the Vaccine Alliance,8 we systematically compared four malaria transmission models to assess the public health impact and cost-effectiveness of routine use of the RTS,S/AS01 vaccine in African settings. Methods Models and harmonisation assumptions Modelling groups were contacted by WHO to participate in this study. Four groups agreed to contribute: the Institute for Disease Modeling (EMOD-DTK),9 GSK Vaccines (GSK),10 Imperial College London (Imperial),11 and the Swiss Tropical and Public Health Institute (OpenMalaria).12 The four models encompass a range of structures and levels of complexity, with all simulating malaria transmission and vaccine impact for defined geographic areas, taking into account local exposure and population demographics. Two models (Imperial and OpenMalaria) capture any herd effect of vaccination, whereas the other two models (EMOD-DTK and GSK) are implemented as a fixed-exposure cohort. Appendix pp 5–20 provide a full comparison of the models.

or defined geographic areas, taking into account local exposure and population demographics. Two models (Imperial and OpenMalaria) capture any herd effect of vaccination, whereas the other two models (EMOD-DTK and GSK) are implemented as a fixed-exposure cohort. Appendix pp 5–20 provide a full comparison of the models. We used trial data3 for follow-up of 32 months or longer to parameterise vaccine efficacy against infection for the group aged 5–17 months. EMOD-DTK, GSK, and OpenMalaria13 used site-specific incidence or efficacy data aggregated into 3 month periods, and Imperial14 used individual-level data. We assumed no partial protection from the initial two doses, consistent with primary reporting from the trial.3 With use of these estimates derived from the trial data, we then made projections for wider transmission settings representative of normal access to care. Demographics, Plasmodium falciparum parasite prevalence in 2–10 year olds (PfPR2–10; a standard metric used to describe the intensity of malaria transmission2), seasonality, and case management were aligned across the models (table 1). We assumed that access to artemisinin-based combination therapy was 45%, at the top of the range reported in 2013,1 but significantly lower than in the trial.

(PfPR2–10; a standard metric used to describe the intensity of malaria transmission2), seasonality, and case management were aligned across the models (table 1). We assumed that access to artemisinin-based combination therapy was 45%, at the top of the range reported in 2013,1 but significantly lower than in the trial. We predicted the public health impact and cost-effectiveness of the vaccine for a range of PfPR2–10 levels (3–65%), assuming vaccine implementation was in addition to existing levels of malaria control interventions and treatment under the assumptions in table 1. We considered two immunisation schedules: three doses at age 6, 7·5, and 9 months (6–9 months with a three-dose schedule), and with an additional fourth dose at 27 months of age (6–9 months with a four-dose schedule). We chose these schedules because two visits (months 6 and 9) align with routine health-care visits (vitamin A supplement at 6 months and measles complex vaccine at 9 months). We focus on the cohort aged 5–17 months because efficacy was higher in this group than in the group aged 6–12 weeks.3 For both schedules, we defined a fully vaccinated child as a child who had received at least three doses.

e health-care visits (vitamin A supplement at 6 months and measles complex vaccine at 9 months). We focus on the cohort aged 5–17 months because efficacy was higher in this group than in the group aged 6–12 weeks.3 For both schedules, we defined a fully vaccinated child as a child who had received at least three doses. We assumed 90% vaccine coverage (similar to diphtheria-tetanus-pertussis [DTP3] in Africa16) for 6–9 month implementation, with 80% of these individuals receiving the fourth dose (72% coverage) based on a drop-off in coverage from DTP3 to measles vaccination at 9 months (table 1).16 We summarise outputs as clinical cases, severe cases, deaths, and undiscounted disability-adjusted life-years (DALYs; driven mainly by mortality) averted per 100 000 fully vaccinated children. Outputs are cumulative over a 15 year time horizon, chosen to be comparable to the timeframe of the new Global Technical Strategy for malaria17 and over a length of time that is long enough to capture any potential detrimental effect of shifting of cases to older ages. We report health outcomes for the entire population and for children younger than 5 years. We did this study in accordance with Good Clinical Practice guidelines and the Declaration of Helsinki. The trial protocol was approved by the ethical review board at each study centre and partner institution and by the national regulatory authority in each country.18 Because this work involves data simulations and analysis, informed consent was not required.

Clinical Practice guidelines and the Declaration of Helsinki. The trial protocol was approved by the ethical review board at each study centre and partner institution and by the national regulatory authority in each country.18 Because this work involves data simulations and analysis, informed consent was not required. Statistical analysis We estimated the incremental cost per clinical case or DALY averted compared with no vaccination with standard levels of access to treatment and other malaria interventions (with costs discounted at 3%). We evaluated incremental cost for assumed vaccine purchase prices of US$2, $5 and $10, corresponding to a cost per dose of $2·69, $6·52, and $12·91, respectively (table 1, appendix pp 33–34). We calculated costs from the perspective of the health-care provider, which include the cost of consumables (vaccines and immunisation supplies) for RTS,S introduction, diagnostics, and antimalarial drugs, and related commodities for malaria case management (table 1). We undertook a univariate sensitivity analysis of vaccine properties, health systems, and economic parameters for the OpenMalaria model (appendix pp 48–50). We present outputs from each model as medians and 95% prediction intervals. We present summary outputs across the four models as the median of all models plus ranges across the medians of the individual models.

Statistical analysis We estimated the incremental cost per clinical case or DALY averted compared with no vaccination with standard levels of access to treatment and other malaria interventions (with costs discounted at 3%). We evaluated incremental cost for assumed vaccine purchase prices of US$2, $5 and $10, corresponding to a cost per dose of $2·69, $6·52, and $12·91, respectively (table 1, appendix pp 33–34). We calculated costs from the perspective of the health-care provider, which include the cost of consumables (vaccines and immunisation supplies) for RTS,S introduction, diagnostics, and antimalarial drugs, and related commodities for malaria case management (table 1). We undertook a univariate sensitivity analysis of vaccine properties, health systems, and economic parameters for the OpenMalaria model (appendix pp 48–50). We present outputs from each model as medians and 95% prediction intervals. We present summary outputs across the four models as the median of all models plus ranges across the medians of the individual models. Role of the funding sources The PATH Malaria Vaccine Initiative contributed to study design and data interpretation. The Health Economics Group of GSK Vaccines was involved in data analysis, data interpretation, and writing of the report (as one model group in the comparison), but had no role in the opinions and results presented by the other groups. All other funders of the study had no role in study design, data analysis, data interpretation, or writing of the report. All authors had full access to the data and were jointly responsible for the decision to submit the manuscript.

omparison), but had no role in the opinions and results presented by the other groups. All other funders of the study had no role in study design, data analysis, data interpretation, or writing of the report. All authors had full access to the data and were jointly responsible for the decision to submit the manuscript. Results Figure 1 shows the trial and model-estimated cumulative vaccine efficacy by trial site at study end (≥32 months). Model-estimated efficacy generally falls within the confidence intervals of the trial data for both clinical malaria and severe disease (figure 1). The estimated underlying protection against infection was similar across models during the first 18 months, but diverged as the projections extended beyond the trial period (appendix pp 21–31).

acy generally falls within the confidence intervals of the trial data for both clinical malaria and severe disease (figure 1). The estimated underlying protection against infection was similar across models during the first 18 months, but diverged as the projections extended beyond the trial period (appendix pp 21–31). With the four mathematical models and our assumptions about implementation of the vaccine beyond the trial sites, we estimated that the absolute vaccine impact would increase with increasing levels of malaria transmission, averting from 15–32% of all clinical cases at a PfPR2–10 of 10%, to 5–22% of clinical cases at a PfPR2–10 of 65% (figure 2, table 2). Similarly, we estimated that 8–35% of malaria deaths in children younger than 5 years would be averted at a PfPR2–10 of 10%, and 5–24% would be averted at a PfPR2–10 of 65% (figure 2, table 2). The predicted lower proportion of events averted in higher versus lower transmission areas is partly a result of age-shifting of disease in higher transmission areas (figure 3, appendix pp 36–47). Across the models this finding translates to a median of 116 480 (range 31 450—160 410) cases of clinical malaria averted and 484 (189–859) deaths averted per 100 000 vaccinated children with the four-dose schedule for a PfPR2–10 of 10–65% (figure 2, table 2). Public health impact at a PfPR2–10 of less than 3% is expected to be small (figure 2, appendix pp 36–47).

median of 116 480 (range 31 450—160 410) cases of clinical malaria averted and 484 (189–859) deaths averted per 100 000 vaccinated children with the four-dose schedule for a PfPR2–10 of 10–65% (figure 2, table 2). Public health impact at a PfPR2–10 of less than 3% is expected to be small (figure 2, appendix pp 36–47). There was a lack of consensus between the models for the additional public health impact of the four-dose versus the three-dose schedule. Three models (GSK, EMOD-DTK, Imperial) predicted a substantial additional impact of the four-dose schedule (16–43% extra deaths averted and 21–55% extra clinical cases averted; table 2), whereas the fourth model (OpenMalaria) predicted a marginal impact. This variance is due to differences in the estimated impact of the fourth dose against infection and the estimated waning of protection against infection between the models (appendix pp 21–31).

averted and 21–55% extra clinical cases averted; table 2), whereas the fourth model (OpenMalaria) predicted a marginal impact. This variance is due to differences in the estimated impact of the fourth dose against infection and the estimated waning of protection against infection between the models (appendix pp 21–31). As expected with partially effective malaria control interventions,19, 20 we predicted a shift in cases to older ages due to reduced exposure and hence a delay in the development of naturally acquired immunity. Combined with the estimated waning of protection against infection of the vaccine, some of the initial impact of the vaccine on cases averted in very young children is therefore offset by predicted higher relative incidence at older ages (figure 3). This effect is predicted to be further delayed with the four-dose schedule (appendix pp 36–47). Similar effects are predicted for severe disease, with the age-shift occurring earlier than for clinical cases (appendix pp 36–47). Despite these findings, the overall cumulative impact of the vaccine on clinical cases, severe disease, and mortality over 15 years is predicted to be positive with the four-dose vaccine schedule (figure 3). Our analyses also assume that malaria exposure remains static for the length of follow-up. Thus, this age-shifting effect could be mitigated if transmission declines over this time horizon (eg, as a result of continued scale-up of other interventions21).

ed to be positive with the four-dose vaccine schedule (figure 3). Our analyses also assume that malaria exposure remains static for the length of follow-up. Thus, this age-shifting effect could be mitigated if transmission declines over this time horizon (eg, as a result of continued scale-up of other interventions21). The estimated incremental cost-effectiveness ratios (ICERs) for clinical cases and DALYs averted with both vaccination schedules compared with no vaccination were lowest at intermediate levels of PfPR2–10, but were generally less than $100 per DALY averted for a PfPR2-10 of more than 10% for a vaccine price of $5 per dose (table 2, figure 4). At a PfPR2–10 of less than 10% the vaccine is estimated to become substantially less cost-effective due to fewer cases and deaths being averted for the same overall cost of a vaccine programme (figure 4). Furthermore, consensus between the models is lower at a PfPR2–10 of less than 10% than at a PfPR2–10 of more than 10% (figure 4). The predicted ICER of a four-dose schedule varied between the models because of the different public health impact projections of the fourth dose. However, overall we estimated that the ICERs (compared with no vaccination) for the four-dose schedule were similar to those estimated for the three-dose schedule (table 2). This similarity is because the extra deaths averted with the four-dose schedule are offset by the extra costs estimated for delivery of the fourth dose, and because of our assumption that only 80% of children who receive the first three doses will return for the fourth dose. In a sensitivity analysis, the cost per DALY averted over the 15 year time horizon at most doubled from the baseline estimate when we considered a range of factors including lower vaccination coverage, lower estimates of vaccine efficacy, and higher vaccine price, with the greatest impact due to a price increase from $5 to $10 (appendix pp 48–50).

e cost per DALY averted over the 15 year time horizon at most doubled from the baseline estimate when we considered a range of factors including lower vaccination coverage, lower estimates of vaccine efficacy, and higher vaccine price, with the greatest impact due to a price increase from $5 to $10 (appendix pp 48–50). Discussion We predict a positive public health impact of the introduction of RTS,S/AS01 in settings with a PfPR2–10 between 3% and 65% over a 15 year time horizon with treatment levels representative of the current status in Africa. The absolute impact is substantial, with an estimated 116 500 (range 30 900–160 000) cases of clinical malaria and 484 (195–838) deaths averted per 100 000 vaccinated children with a four-dose schedule for a PfPR2–10 between 10% and 65%. This finding translates to roughly one malaria death prevented for every 200 children fully vaccinated, which, at a vaccine price per dose of $5, equates to $87 (range 43–238) per DALY averted. WHO estimates that 528 000 (range 315 000–689 000) malaria deaths occurred in Africa in 2013.1 Depending on the area of implementation, we estimate that 6–30% of deaths in children younger than 5 years could potentially be averted by RTS,S, when added to existing coverage of long-lasting insecticide-treated nets (68%1, 22) and with moderate levels of malaria treatment.

malaria deaths occurred in Africa in 2013.1 Depending on the area of implementation, we estimate that 6–30% of deaths in children younger than 5 years could potentially be averted by RTS,S, when added to existing coverage of long-lasting insecticide-treated nets (68%1, 22) and with moderate levels of malaria treatment. There was no statistically significant impact against severe disease measured in the trial for the three-dose schedule (4·5%, 95% CI −20·6 to 24·5 at 32 months' follow-up in the group aged 5–17 months), although there was sustained significant efficacy with the four-dose schedule (32·2%, 13·7–46·9).3 Moreover, there was no statistically significant impact on malaria mortality and all-cause mortality, although numbers of deaths were small and the trial was not powered to assess this outcome.3 By contrast, we predicted a net positive impact on severe disease cases and hence deaths averted with the three-dose schedule, with an incremental benefit of the fourth dose of 22% extra deaths averted depending on the setting. There are several possible reasons for this discrepancy. First, severe disease incidence in the trial was low, with substantial differences between sites in the point estimates and wide confidence intervals. Second, inferences about severe disease made by the models are based mainly on data from sites with poorer quality of care than in the trial and broader case definitions.23 Third, the model projections assume that the fourth dose reaches only 80% of the third-dose recipients, and consequently only 72% of the eligible population receive four doses (compared with near 100% coverage in the trial). Finally, these projections are for a 15 year time horizon whereas the trial data include a maximum of 4 years of follow-up. Partially protective malaria interventions reduce an individual's exposure to malaria infection resulting in a delay in the acquisition of natural immunity19, 20 and a shift of clinical and severe disease to older ages.19, 20, 21 This shift is predicted to be greatest in scenarios with high initial vaccine impact. Much of the age-shift effect is predicted to happen beyond 3·5–4 years and at a PfPR2–10 at the upper end of or higher than the PfPR2–10 in the trial, and is therefore not expected to be shown in the trial. As such, longer-term follow-up of trial participants is needed to fully understand any shifting of cases to older age groups, with this shift partly shown in the 4 year follow-up of one of the phase 2 trials.24

pper end of or higher than the PfPR2–10 in the trial, and is therefore not expected to be shown in the trial. As such, longer-term follow-up of trial participants is needed to fully understand any shifting of cases to older age groups, with this shift partly shown in the 4 year follow-up of one of the phase 2 trials.24 For regions where PfPR2–10 is more than 10%, RTS,S is predicted to be cost-effective compared with standard norms and thresholds. Broadly similar ICER estimates were obtained for the three-dose and four-dose schedules, with the additional public health benefit of the boosting schedule offset by the incremental cost of implementation of the additional dose. The higher cost of the four-dose versus the three-dose schedule is partly due to our assumption (based on the difference between DTP3 and first-dose measles vaccination coverage in Africa16) that only 80% of children who receive the first three doses would return for the fourth dose.16 At low prevalence (PfPR2–10 3%), when cases typically occur in older children and adults, we consistently noted that the vaccine was not cost-effective. Except for at very low prevalence, our estimated ICERs are lower than the national gross domestic products per person (median $842 [IQR 531–1668] across 43 malaria-endemic African countries with PfPR2–10>10% in 20142). Furthermore, in settings with PfPR2–10 levels of 20% or greater, ICERs at a vaccine price of $5 are less than $100 per DALY averted. A 2011 review25 summarised average incremental costs per DALY averted (2009 prices) for long-lasting insecticide-treated nets of $27 (range 8·15–110), insecticide residual spraying of $143 (135–150), and intermittent preventative treatment of $24 (1·08–44·24). Additionally, economic assessment of long-lasting insecticide-treated nets from large-scale field studies26, 27 estimated that the cost per DALY averted was between $13 and $89. However, there was wide variation in the costing methodologies used and economies of scale captured by these studies; therefore, these figures should be interpreted as indicative ranges rather than directly corresponding to our estimates. With these caveats, RTS,S is somewhat less cost-effective than long-lasting insecticide-treated nets, which are regarded as one of the most cost-effective interventions available for malaria control; this is an important point for countries to bear in mind because use of long-lasting insecticide-treated nets has also led to proven reductions in all-cause mortality.27

ffective than long-lasting insecticide-treated nets, which are regarded as one of the most cost-effective interventions available for malaria control; this is an important point for countries to bear in mind because use of long-lasting insecticide-treated nets has also led to proven reductions in all-cause mortality.27 Despite the differences between the models, all four groups reached consensus about the expected public health impact and cost-effectiveness of RTS,S/AS01 administration to children aged 6–9 months, and their consensus predictions help to define priority areas by levels of PfPR2–10 for policy recommendation. The systematic and harmonised comparison of multiple models aided interpretation of differences between the models, with these differences relating mostly to model characteristics that were evident in baseline model associations and baseline incidence in absence of vaccine. These characteristics include differences in relations between parasite prevalence and clinical incidence; case definitions; assumptions about rates of immune acquisition, immune decay, mechanisms of immunity; and differences due to datasets used for parameterisation.

tions and baseline incidence in absence of vaccine. These characteristics include differences in relations between parasite prevalence and clinical incidence; case definitions; assumptions about rates of immune acquisition, immune decay, mechanisms of immunity; and differences due to datasets used for parameterisation. Our analysis has some limitations. First, although the estimated vaccine efficacy and waning profile was similar across the models, these profiles diverged after 18 months of follow-up. The predicted long-term impact of the vaccine will inevitably depend on the pattern of these profiles and hence they should be updated with extended follow-up data. Second, whereas the models reproduce vaccine efficacy estimates from the trial, the projections of impact on disease and mortality are based on previous model-fitting to historical data relating clinical and severe incidence to mortality, in the assumption that associations in settings with more restricted access to health care are similar to those in settings with good access. Unlike the trials of insecticide-treated nets,27 which showed statistically significant reductions in malaria mortality, there was no significant impact on malaria mortality or all-cause mortality in the RTS,S trial. This outcome is likely to be due to the significantly lower levels of overall mortality in trial participants with high access to care than in those with low access. Nevertheless, the modelled estimates of RTS,S impact on mortality need further validation, and future studies done after implementation should include monitoring of impact on mortality. Third, we assumed that vaccination would be implemented at 6–9 months with three or four doses, yet actual implementation schedules and coverage might differ (including more than four doses). Over this age range, naturally acquired immunity develops rapidly and physiological effects can also affect the maturity of the vaccine-induced antibody response.28 Fourth, only two of the models allowed for possible indirect effects on transmission, with one model predicting indirect effects in settings with low malaria prevalence. Although the main aim of the vaccine is to provide direct protection, assessment of any indirect effects will be important in the post-licensure phase. Finally, we did not include productivity losses to households, and costs of immunisation and disease management were considered from a health-system perspective in our economic analysis.

aim of the vaccine is to provide direct protection, assessment of any indirect effects will be important in the post-licensure phase. Finally, we did not include productivity losses to households, and costs of immunisation and disease management were considered from a health-system perspective in our economic analysis. According to WHO guidelines,29 the societal perspective is generally preferred; however, in practice, this approach is rarely taken because not all costs are available. In general, addition of costs beyond the health system would increase the cost-effectiveness of vaccination, making our estimates lower. Despite these limitations, our results can help inform decisions about where to implement RTS,S to have the greatest impact.

roach is rarely taken because not all costs are available. In general, addition of costs beyond the health system would increase the cost-effectiveness of vaccination, making our estimates lower. Despite these limitations, our results can help inform decisions about where to implement RTS,S to have the greatest impact. WHO coordinated this model comparison process as an important contributory factor to the broader assessment processes related to this vaccine. These models did not incorporate safety aspects, as is standard for cost-effectiveness and impact models at the pre-licensure stage.7 These safety aspects include immediate reactogenicity including fever (noted with many vaccines), the identified risk of febrile convulsion within 7 days of vaccination (also not unique to this vaccine), and the potential risk of meningitis (which has not been causally related to vaccination), and are outlined in the June, 2014, report of the WHO Global Advisory Committee on Vaccine Safety.30 The WHO policy recommendation processes underway take all these elements into account as part of an overall assessment of benefits and harms, resource use and value for money, equity impacts, feasibility, and acceptability. Supplementary Material Supplementary appendix

WHO coordinated this model comparison process as an important contributory factor to the broader assessment processes related to this vaccine. These models did not incorporate safety aspects, as is standard for cost-effectiveness and impact models at the pre-licensure stage.7 These safety aspects include immediate reactogenicity including fever (noted with many vaccines), the identified risk of febrile convulsion within 7 days of vaccination (also not unique to this vaccine), and the potential risk of meningitis (which has not been causally related to vaccination), and are outlined in the June, 2014, report of the WHO Global Advisory Committee on Vaccine Safety.30 The WHO policy recommendation processes underway take all these elements into account as part of an overall assessment of benefits and harms, resource use and value for money, equity impacts, feasibility, and acceptability. Supplementary Material Supplementary appendix Acknowledgments Imperial College London and Swiss Tropical and Public Health Institute received funding from the PATH Malaria Vaccine Initiative and the Bill & Melinda Gates Foundation. Work done by the Institute for Disease Modeling was funded by the Global Good Fund of Bellevue, WA, USA. The London School of Hygiene & Tropical Medicine (SF, MJ) received funding from WHO and GAVI, the Vaccine Alliance. Imperial College London received additional fellowship (JTG, MTW) and centre (ACG) funding from the Medical Research Council (MRC) and the UK Department for International Development (DFID) under the MRC/DFID concordat agreement. No funding was provided by GlaxoSmithKline Biologicals SA to the non-GSK modelling groups or the comparison process as a whole. We thank Vasee Moorthy and Raymond Hutubessy (WHO, Geneva, Switzerland) for their role in bringing this consortium together and initiating the work outlined in this report; members of the WHO Joint Technical Expert Group on malaria vaccine modelling subgroup for their constructive comments; and Alison Reynolds for her support in undertaking this work. Calculations for OpenMalaria were done at the Centre for Scientific Computing (sciCORE) at the University of Basel, Switzerland, and we thank the many volunteers who made their computers available via www.malariacontrol.net for the simulations.

ive comments; and Alison Reynolds for her support in undertaking this work. Calculations for OpenMalaria were done at the Centre for Scientific Computing (sciCORE) at the University of Basel, Switzerland, and we thank the many volunteers who made their computers available via www.malariacontrol.net for the simulations. Contributors MAP, RV, and ACG conceived the study and designed the model comparison. MAP, RV, KG, MTW, JTG, ACG, CAB, EAW, and CS designed individual model group analysis. KG provided costing data. KG, JTG, and CS provided data. MAP, RV, CAB, CS, KG, MTW, NVdV, JTG, and PP-R did the analyses. SF, EAW, TAS, PAE, and ACG supported the analyses. MAP, SF, EAW, TAS, FM, MJ, and ACG supported interpretation and policy contextualisation. MAP wrote the first draft of the manuscript. MAP, ACG, RV, SF, FM, and TAS wrote the manuscript. MAP, RV, CAB, CS, KG, SF, MTW, FM, and ACG compiled the information in the appendix. All authors discussed the results and contributed to revision of the final manuscript. Declaration of interests FM is employed by PATH. CS and NVdV are employed by the GSK group of companies. All other authors declare no competing interests. Figure 1 Observed and model-predicted vaccine efficacy against clinical and severe malaria from month 0 to study end (≥32 months) by study site in the 5–17 month age category

Declaration of interests FM is employed by PATH. CS and NVdV are employed by the GSK group of companies. All other authors declare no competing interests. Figure 1 Observed and model-predicted vaccine efficacy against clinical and severe malaria from month 0 to study end (≥32 months) by study site in the 5–17 month age category Vaccine efficacy against all episodes of clinical malaria (primary case definition) in the three-dose group (A) and the four-dose group (B), and against severe malaria (primary case definition) in the three-dose group (C) and the four-dose group (D). Error bars show 95% CIs estimated from the trial data. *Intention-to-treat analysis. Figure 2 Model predictions of clinical cases and deaths averted per 100 000 children fully vaccinated with a three-dose or four-dose immunisation schedule for a range of baseline PfPR2–10 levels Results for the three-dose (A, C) and four-dose (B, D) schedules are cumulative following 15 years of routine use of RTS,S. Bars show median estimates and error bars show 95% credible intervals. Negative cases averted at low transmission are due to stochastic variation between model runs at low prevalence, rather than to any modelled biological mechanism. PfPR2–10=parasite prevalence in 2–10 year olds. Figure 3 Clinical cases averted in the vaccinated cohort at low (A) and high (B) endemicity after 15 years of routine use of RTS,S in a four-dose immunisation schedule

Results for the three-dose (A, C) and four-dose (B, D) schedules are cumulative following 15 years of routine use of RTS,S. Bars show median estimates and error bars show 95% credible intervals. Negative cases averted at low transmission are due to stochastic variation between model runs at low prevalence, rather than to any modelled biological mechanism. PfPR2–10=parasite prevalence in 2–10 year olds. Figure 3 Clinical cases averted in the vaccinated cohort at low (A) and high (B) endemicity after 15 years of routine use of RTS,S in a four-dose immunisation schedule Bars show median predictions over all four models and error bars show the range of predictions. Uncertainty within each model is not shown (see appendix pp 5–7). PfPR2–10=parasite prevalence in 2–10 year olds. Figure 4 Cost per clinical case or DALY averted as a function of baseline parasite prevalence in 2–10 year olds (PfPR2–10) Results assume a vaccine price of $2 (A, D), $5 (B, E), or $10 (C, E) per dose. Solid lines represent a three-dose immunisation schedule and dashed lines represent a four-dose immunisation schedule. Similar estimates of incremental cost-effectiveness ratios were obtained for the three-dose and four-dose schedules because the additional public health benefit of the boosted schedule is offset by the incremental cost of implementation of the additional dose. Uncertainty estimates surrounding the models' predictions are omitted for readability but overlap. DALY=disability-adjusted life-year. Table 1 Assumed demographics, implementation coverage, and vaccine efficacy profiles

Results assume a vaccine price of $2 (A, D), $5 (B, E), or $10 (C, E) per dose. Solid lines represent a three-dose immunisation schedule and dashed lines represent a four-dose immunisation schedule. Similar estimates of incremental cost-effectiveness ratios were obtained for the three-dose and four-dose schedules because the additional public health benefit of the boosted schedule is offset by the incremental cost of implementation of the additional dose. Uncertainty estimates surrounding the models' predictions are omitted for readability but overlap. DALY=disability-adjusted life-year. Table 1 Assumed demographics, implementation coverage, and vaccine efficacy profiles Harmonised comparison Demographics Constant population size and demography based on the life table for Butajira, Ethiopia, with an average life expectancy at birth of 46·6 years.15 Transmission intensity Parasite prevalence in 2–10 year olds between 3% and 65%, representing current transmission levels in Africa.2 Seasonality Perennial transmission (no seasonality). Case management Effective coverage (ie, treatment with parasitological cure) for clinical malaria is 45%. Access to care for severe malaria varied by model. Other interventions Predictions assume that current interventions in place at the start of vaccination remain at static levels. Vaccine efficacy Model estimates of RTS,S efficacy against infection profiles based on fitting to phase 3 trial data (appendix pp 22–30). Vaccine schedule Three doses of vaccine given at 6, 7·5, and 9 months (6–9 month implementation) with a fourth dose at month 27 (6–9 month with fourth dose). The first two doses of the primary series are assumed to have 0% efficacy. Vaccine coverage 90% coverage assumed for the three-dose schedule; we assumed a 20% drop-off in coverage for the fourth dose (72% coverage). Cost of RTS,S vaccination Vaccine and immunisation supplies including freight and wastage. The same costs were applied to all settings. These costs were estimated at US$6·52 per dose at vaccine price of $5, $2·69 per dose at vaccine price of $2, and $12·91 at vaccine price of $10 (appendix pp 33–34). Cost of malaria case management Costs are estimated by severity of illness and cover first-line antimalarial drugs, diagnostics, and related supplies including freight and wastage. We assumed full compliance and adherence with the age dosage. The same costs were applied to all settings, ranging from $1·07 to $2·27 per uncomplicated case, and from $21·78 to $55·58 per severe case (appendix pp 33–34). Table 2 Predictions of public health impact and cost-effectiveness of RTS,S for the 6–9 month three-dose and four-dose immunisation schedules at 15 years of follow-up in regions with a parasite prevalence in 2–10 year olds of 10–65%

per uncomplicated case, and from $21·78 to $55·58 per severe case (appendix pp 33–34). Table 2 Predictions of public health impact and cost-effectiveness of RTS,S for the 6–9 month three-dose and four-dose immunisation schedules at 15 years of follow-up in regions with a parasite prevalence in 2–10 year olds of 10–65% Three-dose schedule Four-dose schedule Proportion of clinical cases averted in children younger than 5 years 16·2% (7·3–24·1) 21·1% (7·9–30·6) Proportion of deaths averted in children younger than 5 years 13·8% (5·3–21·4) 18·0% (6·0–29·1) Clinical cases averted per 100 000 fully vaccinated children 93 940 (20 490–126 540) 116 480 (31 450–160 410) Deaths averted per 100 000 fully vaccinated children 394 (127–708) 484 (189–859) Incremental benefit* Clinical cases .. 22% (3 to 49) Deaths .. 31% (−1 to 53) ICER per clinical case averted (in US$) $2 per dose $13 (7–88) $10 (6–93) $5 per dose $30 (18–211) $25 (16–222) $10 per dose $61 (31–415) $51 (28–437) ICER per DALY averted (in US$) $2 per dose $35 (16–112) $38 (18–97) $5 per dose $80 (44–279) $87 (48–244) $10 per dose $147 (90–556) $154 (99–487) Data are median (range) across the models' medians. ICER=incremental cost-effectiveness ratios. * Proportion of additional events averted with four-dose versus three-dose immunisation schedule.

Introduction HIV is a disease of major importance in the UK, with an estimated 107 800 individuals with HIV at the end of 2013.1 Prognosis is excellent, but treatment is lifelong with an inexorable increase in costs to the National Health Service.2 Gay, bisexual, and other men who have sex with men are the most at risk of acquiring HIV in the UK.1 There has been no decrease in the numbers of new diagnoses reported each year for the past decade (3250 in 2013), and estimates suggest that HIV incidence has increased in this population.3 These trends have occurred despite increased HIV testing and a move towards earlier initiation of antiretroviral therapy, which renders most patients non-infectious.4, 5 Although HIV testing and promotion of condom use will always be core strategies for reducing risk, a more radical approach is needed for people who do not have HIV and whose condom use is inconsistent. One such approach is pre-exposure prophylaxis (PrEP), the provision of antiretroviral drugs before HIV exposure to prevent infection.

sting and promotion of condom use will always be core strategies for reducing risk, a more radical approach is needed for people who do not have HIV and whose condom use is inconsistent. One such approach is pre-exposure prophylaxis (PrEP), the provision of antiretroviral drugs before HIV exposure to prevent infection. The biological efficacy of daily oral tenofovir-based regimens used as PrEP to reduce HIV acquisition has been established through randomised placebo-controlled trials including men who have sex with men,6 heterosexual individuals,7, 8 and intravenous drug users.9 One purpose of using placebo in these studies was to avoid confounding bias due to risk compensation, which occurs if individuals perceive themselves to be protected by PrEP and so become more likely to engage in riskier sexual practices.10, 11 If this effect exists, it could undermine the biological protection conferred by PrEP and its value as a public health intervention.10, 12, 13, 14 Research in context Evidence before this study We reviewed all randomised controlled trials of pre-exposure prophylaxis (PrEP) listed in the HIV Prevention Research & Development Database, which has comprehensive information on biomedical clinical trials of HIV prevention that are planned, ongoing, or completed. We identified several completed and ongoing placebo-controlled trials designed to assess biological efficacy and demonstration projects designed to facilitate implementation, but no open-label randomised trials that assess real-life effectiveness. Added value of this study

We reviewed all randomised controlled trials of pre-exposure prophylaxis (PrEP) listed in the HIV Prevention Research & Development Database, which has comprehensive information on biomedical clinical trials of HIV prevention that are planned, ongoing, or completed. We identified several completed and ongoing placebo-controlled trials designed to assess biological efficacy and demonstration projects designed to facilitate implementation, but no open-label randomised trials that assess real-life effectiveness. Added value of this study PROUD is the first-open-label randomised controlled trial of PrEP, and used a pragmatic schedule and procedures to represent how PrEP would be used in routine clinical practice. Our results refute concerns that the effectiveness of PrEP would be compromised when used in clinical practice, and the reduction in HIV incidence exceeded that reported from any placebo-controlled trial. The incidence of HIV infection among men not on PrEP was high (nine cases per 100 person-years), implying that the offer of PrEP is likely to attract individuals who are most likely to benefit from it. Implications of all the available evidence A public health programme of PrEP could have a major role in preventing a condition that requires lifelong treatment and curtailing the HIV epidemic. Structural and financial barriers that might impede its implementation should be urgently addressed.

PROUD is the first-open-label randomised controlled trial of PrEP, and used a pragmatic schedule and procedures to represent how PrEP would be used in routine clinical practice. Our results refute concerns that the effectiveness of PrEP would be compromised when used in clinical practice, and the reduction in HIV incidence exceeded that reported from any placebo-controlled trial. The incidence of HIV infection among men not on PrEP was high (nine cases per 100 person-years), implying that the offer of PrEP is likely to attract individuals who are most likely to benefit from it. Implications of all the available evidence A public health programme of PrEP could have a major role in preventing a condition that requires lifelong treatment and curtailing the HIV epidemic. Structural and financial barriers that might impede its implementation should be urgently addressed. We designed the PROUD study (appendix) to assess the effectiveness of PrEP. The effectiveness was the net effect of efficacy, adherence, and any change in sexual behaviour as a result of PrEP. Here, we report the pilot phase, in which we assessed recruitment and retention to test the feasibility of a large-scale trial. However, the unexpectedly large number of HIV infections enabled us to present findings on the effectiveness of PrEP, as well as safety, adherence, and risk compensation.

as a result of PrEP. Here, we report the pilot phase, in which we assessed recruitment and retention to test the feasibility of a large-scale trial. However, the unexpectedly large number of HIV infections enabled us to present findings on the effectiveness of PrEP, as well as safety, adherence, and risk compensation. Methods Study design and participants We did this pragmatic, open-label, randomised controlled trial at 13 sexual health clinics in England. Eligible participants were male at birth, were aged 18 years or older, had previously attended the enrolling clinic, had been screened for HIV and other sexually transmitted infections, were HIV negative by a routinely used assay in the previous 4 weeks or on the day of enrolment, and had reported anal intercourse without a condom in the previous 90 days and likely in the opinion of the participant to have anal intercourse without a condom in the next 90 days. We excluded participants with acute viral illness possibly due to HIV seroconversion, any contraindication to tenofovir disoproxil fumarate or emtricitabine, and those being treated with or with treatment indicated for hepatitis B infection. The study was reviewed and approved by London Bridge Research Ethics Committee. The study protocol is available online. All patients provided written informed consent.

contraindication to tenofovir disoproxil fumarate or emtricitabine, and those being treated with or with treatment indicated for hepatitis B infection. The study was reviewed and approved by London Bridge Research Ethics Committee. The study protocol is available online. All patients provided written informed consent. Randomisation and masking We randomly assigned participants (1:1) to receive PrEP either starting at the enrolment visit (immediate group) or after a deferral period of 1 year (deferred group). The computer-generated randomisation list with variable block sizes (of four, six, and eight; stratified by clinical site) was prepared by one of the trial statisticians (DID) and incorporated within the database held at the coordinating centre. Randomisation was web-based and done by approved members of the research team at each clinic. Regular sexual partners were encouraged to enrol together and both partners allocated to the same group to minimise the possibility of drug sharing. Neither patients nor investigators were masked to the treatment allocation.

. Randomisation was web-based and done by approved members of the research team at each clinic. Regular sexual partners were encouraged to enrol together and both partners allocated to the same group to minimise the possibility of drug sharing. Neither patients nor investigators were masked to the treatment allocation. Procedures We used procedures that we envisaged for a public health PrEP programme, including the lack of a screening visit, and the use of HIV and sexually transmitted infection results collected at other clinics and during non-study visits. All laboratory investigations were done locally with routine assays in compliance with the UK standards for the management of sexually transmitted infections.15 These guidelines recommend urethral, rectal, and pharyngeal nucleic acid amplification tests for Chlamydia trachomatis and Neisseria gonorrhoeae, with culture for N gonorrhoeae as indicated; serology for syphilis; serology or nucleic acid assays for hepatitis B and C as indicated. The protocol did not stipulate the collection and storage of a baseline sample for HIV although this was routine practice in some clinics.

amydia trachomatis and Neisseria gonorrhoeae, with culture for N gonorrhoeae as indicated; serology for syphilis; serology or nucleic acid assays for hepatitis B and C as indicated. The protocol did not stipulate the collection and storage of a baseline sample for HIV although this was routine practice in some clinics. At the enrolment visit, baseline demographic, clinical, and sexual behavioural data were recorded. Participants were screened for sexually transmitted infections if they reported a new partner since their previous screen, and assessed for hepatitis B immunisation status. A rapid antibody point-of-care HIV test was done if no HIV antigen–antibody test had been done in the previous 4 weeks, and all participants had an HIV antigen–antibody test after randomisation. Interventions to reduce risk were offered according to routine practice at the clinic.

d for hepatitis B immunisation status. A rapid antibody point-of-care HIV test was done if no HIV antigen–antibody test had been done in the previous 4 weeks, and all participants had an HIV antigen–antibody test after randomisation. Interventions to reduce risk were offered according to routine practice at the clinic. The PrEP regimen was a single daily tablet containing 245 mg of tenofovir disoproxil fumarate and 200 mg of emtricitabine (Truvada; Gilead Sciences, Foster City, CA, USA). Participants allocated to the immediate group were initially prescribed 30 tablets together with information about dosing and potential side-effects, including that maximum protection against HIV would be achieved only after reaching steady state concentrations (roughly 2 weeks, estimated from five half-lives of the intracellular drug concentration).16 A blood sample was obtained to measure serum creatinine. An appointment was made within 1 month, primarily as a safety and tolerability check, and to prescribe 90 tablets. The same procedures were followed when participants in the deferred group started PrEP. Follow-up is scheduled to continue until the final enrolled participant has completed 2 years in the study.

eatinine. An appointment was made within 1 month, primarily as a safety and tolerability check, and to prescribe 90 tablets. The same procedures were followed when participants in the deferred group started PrEP. Follow-up is scheduled to continue until the final enrolled participant has completed 2 years in the study. All participants were asked to attend clinic every 3 months. These visits included an HIV test and a screen for bacterial sexually transmitted infections. Hepatitis C screening was indicated if the participant reported injecting or snorting drugs, fisting, or the use of sex toys. Sufficient PrEP was prescribed to extend 1 month beyond the next quarterly appointment. A subsequent protocol amendment allowed 6 months of PrEP to be prescribed in exceptional circumstances—eg, travel overseas. Serum creatinine was checked yearly, but additional tests were triggered at intervening visits if more than a trace of protein was detected by urine dipstick and could not be explained by infection.17 Potential side-effects of study drug and discontinuations for a medical event were asked about at each visit. In the event of HIV seroconversion, the earliest available HIV-positive sample was tested for genotypic drug resistance, in accordance with UK guidelines.18

e dipstick and could not be explained by infection.17 Potential side-effects of study drug and discontinuations for a medical event were asked about at each visit. In the event of HIV seroconversion, the earliest available HIV-positive sample was tested for genotypic drug resistance, in accordance with UK guidelines.18 Participants were asked to complete monthly questionnaires and daily diaries about sexual behaviour and adherence to PrEP, either online or on paper. A more detailed questionnaire, including information on the number and type of sexual partners in the previous 90 days, was administered at enrolment and at yearly visits. Plasma concentrations of tenofovir were measured in a sample of 52 participants who reported that they were taking PrEP and who attended one of five sites on a day the laboratory was able to process the samples. We attempted to identify additional HIV and sexually transmitted infection results in participants lost to follow-up by searching electronic clinic records in other PROUD clinics. Outcomes The primary outcome was time to accrual of 500 participants and retention. The secondary outcomes were HIV infection, safety, adherence, and risk compensation (see protocol). HIV infection was defined as a reactive HIV antigen–antibody test result (confirmed by the detection of HIV RNA), in participants without HIV infection at enrolment. Although retrospective testing of enrolment samples for HIV RNA was possible at some sites these results were not considered.

compensation (see protocol). HIV infection was defined as a reactive HIV antigen–antibody test result (confirmed by the detection of HIV RNA), in participants without HIV infection at enrolment. Although retrospective testing of enrolment samples for HIV RNA was possible at some sites these results were not considered. We included data up to and including the first test after 48 weeks or the closure date of the deferred group on Oct 13, 2014, whichever was earlier (the deferral phase). We censored person-years of observation at the date of the first reactive HIV test for participants who became infected, or the date of the last test for those who did not. We calculated expected person-years of observation assuming that participants had attended all study visits, as per the study protocol.

ral phase). We censored person-years of observation at the date of the first reactive HIV test for participants who became infected, or the date of the last test for those who did not. We calculated expected person-years of observation assuming that participants had attended all study visits, as per the study protocol. Statistical analysis PROUD was designed with a sample size of 5000 participants, powered to detect a 50% reduction in HIV incidence from 2·5 to 1·25 infections per 100 person-years. For the pilot study, we used an arbitrary 10% sample size of 500. Because of the unlikeliness of showing the effectiveness of PrEP in a pilot study, data were initially monitored by a single independent expert not masked to allocation. As it emerged that HIV incidence was much higher than anticipated, an independent data monitoring committee was set up in June, 2014. The committee regarded the difference between groups in rate of HIV infection (rate difference) as the key measure for public health policy, and adopted a lower 2·5% confidence limit greater than two infections per 100 person-years as a threshold for notifying the steering committee, although this was not a formal stopping rule.

tee regarded the difference between groups in rate of HIV infection (rate difference) as the key measure for public health policy, and adopted a lower 2·5% confidence limit greater than two infections per 100 person-years as a threshold for notifying the steering committee, although this was not a formal stopping rule. Analyses included all participants according to their randomised allocation (intention to treat) apart from the exclusion of individuals with a reactive HIV antigen–antibody result at enrolment in analyses of HIV incidence (modified intention to treat). We compared incidence rates between the two groups by both the rate difference and the rate ratio. We calculated exact 90% CIs rather than 95% CIs because we were primarily interested in the lower confidence limit—ie, the minimal estimate of the effectiveness.19 We derived the number-needed-to-treat to directly avert (prevent or delay) one HIV infection from the reciprocal of the rate difference.20 All analyses used data collected during the deferral phase of the trial, up to the date of extraction on June 10, 2015.

idence limit—ie, the minimal estimate of the effectiveness.19 We derived the number-needed-to-treat to directly avert (prevent or delay) one HIV infection from the reciprocal of the rate difference.20 All analyses used data collected during the deferral phase of the trial, up to the date of extraction on June 10, 2015. We planned to assess individual-level adherence and longitudinal sexual behaviour, but the low proportion of participants who completed the monthly questionnaire and diary prevented us from doing so. We therefore reported overall prescriptions of trial drug and cross-sectional analyses of sexual behaviour on the basis of baseline and 1 year questionnaires only. We compared the number of different anal sex partners at 1 year in each group using a stratified test for trend,21 according to the category at the enrolment visit. We used logistic regression to analyse the probability of detecting a sexually transmitted infection during follow-up, adjusting for the number of screens (as a linear term). We did the statistical analyses with Stata (version 13.1). The trial is registered with ISRCTN (ISRCTN94465371) and ClinicalTrials.gov (NCT02065986). Role of the funding source Employees of the funders had a role in the design of the study, data collection, analysis, and interpretation, and writing the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

The trial is registered with ISRCTN (ISRCTN94465371) and ClinicalTrials.gov (NCT02065986). Role of the funding source Employees of the funders had a role in the design of the study, data collection, analysis, and interpretation, and writing the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results We randomly assigned 544 participants between Nov 29, 2012, and April 30, 2014: 275 to the immediate group and 269 to the deferred group. Two participants had enrolled twice to access PrEP and were analysed in the deferred group (figure 1). The data monitoring committee considered the results of an interim analysis on Oct 6, 2014, and alerted the steering committee to a significantly increased risk of HIV infection in the deferred group. On Oct 13, 2014, the principal investigators at sites were requested by the steering committee to offer PrEP to all participants in the deferred group who had not yet had this opportunity (n=163).

n Oct 6, 2014, and alerted the steering committee to a significantly increased risk of HIV infection in the deferred group. On Oct 13, 2014, the principal investigators at sites were requested by the steering committee to offer PrEP to all participants in the deferred group who had not yet had this opportunity (n=163). Baseline characteristics were well balanced between the two groups (table 1). Median age was 35 years (29–43), 327 (61%) of 540 participants were university graduates, 217 (40%) of 540 were born outside of the UK, and 160 (30%) of 540 were living with a partner. In the previous 12 months, 331 (64%) of 517 had been diagnosed with a sexually transmitted infection (172 [33%] with rectal gonorrhoea or chlamydia), 184 (36%) of 510 had received at least one course of post-exposure prophylaxis, and the median number of HIV tests done was 3 (IQR 2–4). 231 (44%) of 525 participants had used one or more drugs associated with sexual disinhibition (γ-hydroxybutyrate, 4-methylmethcathinone, or methamphetamine) in the past 90 days.

dia), 184 (36%) of 510 had received at least one course of post-exposure prophylaxis, and the median number of HIV tests done was 3 (IQR 2–4). 231 (44%) of 525 participants had used one or more drugs associated with sexual disinhibition (γ-hydroxybutyrate, 4-methylmethcathinone, or methamphetamine) in the past 90 days. 14 (5%) of 275 participants in the immediate group were prescribed no further study drug after the initial prescription. Overall, sufficient study drug was prescribed for 88% of the total follow-up time. Tenofovir was detected in plasma of all 52 sampled participants (range 38–549 ng/mL) who reported that they were taking PrEP. 21 (8%) of 275 participants interrupted or missed doses because of 28 adverse event episodes. 13 of the episodes were considered related to study drug (table 2). 20 of 21 participants restarted study drug. The most common drug-related symptoms were nausea, headache, and arthralgia. Three of 21 participants interrupted study drug because of high creatinine concentration; two had comorbidities and were taking concomitant prescription drugs but a relationship to study drug could not be excluded, and one was thought to be due to recreational drugs. 29 serious adverse events (including one death) were reported in 27 participants, but none were attributed to study drug (appendix p 5).

tion; two had comorbidities and were taking concomitant prescription drugs but a relationship to study drug could not be excluded, and one was thought to be due to recreational drugs. 29 serious adverse events (including one death) were reported in 27 participants, but none were attributed to study drug (appendix p 5). Use of post-exposure prophylaxis (the recommended regimen at the time of the study was a 28-day course of tenofovir disoproxil fumarate–emtricitabine plus lopinavir) was common in the deferred group. 174 courses were prescribed to 85 participants during the deferral phase: 36 participants received one course, 27 received two courses, and 22 received three or more courses. Post-exposure prophylaxis was also prescribed to 12 participants (14 prescriptions) in the immediate group in this period.

in the deferred group. 174 courses were prescribed to 85 participants during the deferral phase: 36 participants received one course, 27 received two courses, and 22 received three or more courses. Post-exposure prophylaxis was also prescribed to 12 participants (14 prescriptions) in the immediate group in this period. Three participants (two in the immediate group, one in the deferred group) had a reactive HIV antigen–antibody test at baseline (all were non-reactive by an antibody point-of-care test; figure 1). A further 18 participants had no recorded HIV tests after the enrolment visit leaving 523 (96%) of 544 who contributed to the analysis of HIV incidence. We had 243 person-years of follow-up for the immediate group (94% of the expected 259 person-years) and 222 person-years for the deferred group (90% of the expected 245 person-years). 20 patients had new incident HIV infections in the deferred group (figure 2), of whom six had been prescribed a total of 12 courses of post-exposure prophylaxis during follow-up. In six patients, the last negative antigen–antibody test was at the enrolment visit. By contrast, only three incident HIV infections occurred in the immediate group. One participant had a reactive HIV test at the 4-week visit, and infection is thought to have pre-dated the start of PrEP, based on the history provided. The second participant was HIV reactive at 61 weeks and had been prescribed no study drug since the enrolment visit. The third participant presented with a seroconversion illness at 53 weeks; his last clinic attendance was at the 12-week visit when he was prescribed 90 tablets of study drug. These findings suggest that there were no breakthrough HIV infections in participants who were taking PrEP.

tudy drug since the enrolment visit. The third participant presented with a seroconversion illness at 53 weeks; his last clinic attendance was at the 12-week visit when he was prescribed 90 tablets of study drug. These findings suggest that there were no breakthrough HIV infections in participants who were taking PrEP. HIV incidence was significantly lower in the immediate group (1·2 cases per 100 person-years, 90% CI 0·4–2·9) than in the deferred group (9·0 per 100 person-years, 90% CI 6·1–12·8; p=0·0001). This difference corresponds to a proportionate reduction of 86% (90% CI 64–96) and a rate difference of 7·8 per 100 person-years (90% CI 4·3–11·3). 13 men (90% CI 9–23) in a similar population would need access to 1 year of PrEP to avert one HIV infection. HIV diagnoses in the deferred group were fairly evenly distributed over follow-up (appendix p 7). All five participants in the immediate group who had HIV infection were tested for resistance. Two of the three participants with a reactive test at enrolment or the 4-week visit developed mutations at codon 184 in reverse transcriptase (Met184Ile/Met, Met184Ile/Val/Met), probably selected by exposure to emtricitabine. No resistance was detected in the two participants with later infections, which was not surprising given their apparent non-adherence to PrEP. No participant had mutations associated with tenofovir disoproxil fumarate treatment (Lys65Arg, Lys70Glu).

t, Met184Ile/Val/Met), probably selected by exposure to emtricitabine. No resistance was detected in the two participants with later infections, which was not surprising given their apparent non-adherence to PrEP. No participant had mutations associated with tenofovir disoproxil fumarate treatment (Lys65Arg, Lys70Glu). Questionnaires about sexual behaviour in the previous 90 days were completed and returned by 534 participants at baseline (271 in the immediate group vs 263 in the deferred group) and by 406 participants at 1 year (212 vs 194). Total number of different anal sex partners varied widely at the two timepoints, and we detected no significant difference between groups at 1 year (p=0·57; appendix p 8). However, a larger proportion of participants allocated to immediate PrEP than allocated to deferred PrEP reported receptive anal sex with ten or more partners without a condom (21% vs 12%; p=0·03, test for trend).

two timepoints, and we detected no significant difference between groups at 1 year (p=0·57; appendix p 8). However, a larger proportion of participants allocated to immediate PrEP than allocated to deferred PrEP reported receptive anal sex with ten or more partners without a condom (21% vs 12%; p=0·03, test for trend). 152 (57%) of 265 participants in the immediate group versus 124 (50%) of 247 in the deferred group were diagnosed with one or more bacterial sexually transmitted infection during follow-up, most commonly gonorrhoea and chlamydia (table 3). The randomised comparison was biased by the greater number of screens for sexually transmitted infections in the immediate group versus the deferred group (mean 4·2 vs 3·6), a consequence of more regular study clinic attendance to collect prescriptions in the immediate group. After adjustment for the number of screens, we found no significant difference between the groups, either for individual sexually transmitted infections or overall (table 3). Particularly, the proportion of participants diagnosed with rectal gonorrhoea or chlamydia, which is an indicator of receptive anal intercourse without a condom, was much the same in the two groups (table 3). Six incident hepatitis C infections occurred, three in each group. Injecting drug use was the possible route of transmission in three of these participants (two in the immediate group vs one in the deferred group), and a fourth participant (in the deferred group) acquired hepatitis C virus infection around the time or shortly after HIV infection.

ections occurred, three in each group. Injecting drug use was the possible route of transmission in three of these participants (two in the immediate group vs one in the deferred group), and a fourth participant (in the deferred group) acquired hepatitis C virus infection around the time or shortly after HIV infection. Discussion Our findings refute concerns that the effectiveness of PrEP would be compromised in a real-world setting. Indeed, the reduction in HIV incidence we recorded exceeds that reported in any placebo-controlled trial.22 The proportion of sexually transmitted infections, including rectal gonorrhoea or chlamydia, did not differ significantly between groups despite a suggestion of risk compensation among a small proportion of PrEP recipients.

reduction in HIV incidence we recorded exceeds that reported in any placebo-controlled trial.22 The proportion of sexually transmitted infections, including rectal gonorrhoea or chlamydia, did not differ significantly between groups despite a suggestion of risk compensation among a small proportion of PrEP recipients. The study has strengths and weaknesses. First, the open-label rather than placebo-controlled design enabled us to capture the outcome that is most relevant for assessing PrEP within a public health prevention programme: the combination of the direct biological efficacy of the drug and the indirect effect of altered sexual behaviour among individuals who knew they were taking PrEP. Placebo-controlled trials may underestimate actual adherence because there is less incentive to take a tablet when the participant knows that it might be a placebo.11 Second, the lack of data on adherence to PrEP and sexual behaviour is a limitation. However, the measured drug concentrations validated the reports of participants who said they were taking study drug, by contrast with placebo-controlled trials.6, 23, 24 The absence of longitudinal data for sexual behaviour is frustrating, as we cannot assess precisely how participants matched adherence to risk, and insights into risk compensation are limited to a single timepoint at 1 year. However, we were able to use the information in the routine clinic records to capture the results of screens for HIV and sexually transmitted infection in the study database, and achieve a high level of follow-up for these endpoints. A larger study would have given more precise estimates of the effect of PrEP on sexually transmitted infections. Third, two men enrolled twice to access PrEP, and it is possible that others in the deferred group co-enrolled without detection or accessed PrEP from other sources, resulting in the effectiveness of PrEP being underestimated. Finally, because the trial stopped early the probability of type I error is increased.

tions. Third, two men enrolled twice to access PrEP, and it is possible that others in the deferred group co-enrolled without detection or accessed PrEP from other sources, resulting in the effectiveness of PrEP being underestimated. Finally, because the trial stopped early the probability of type I error is increased. An important issue in PrEP implementation programmes is eligibility.22 We included participants who reported at least one anal sex act without a condom in the preceding 90 days; consequently, the reported sexual risk behaviour at enrolment was diverse. Despite the broad eligibility criteria and extensive use of post-exposure prophylaxis, we recorded a high HIV incidence of nine cases per 100 person-years in the deferred group. This finding was the main determinant of the highly favourable estimate of 13 similar men who would need access to PrEP for 1 year to avert one HIV infection. Additional infections that would have occurred further down the transmission chain are not represented in this value. The incidence was roughly seven times higher than the national estimate of 1·34 cases per 100 person-years reported for men who have sex with men attending sexual health clinics in 2012, derived from avidity assay data.25 Although participants in PROUD were much more likely to have had rectal infections and to have used post-exposure prophylaxis than was the overall population of men who have sex with men attending sexual health clinics,26 the size of the difference in HIV incidence was nonetheless surprising. The difference suggests that the PROUD study population was highly selective, despite broad eligibility, and that the offer of PrEP generally attracts those men who are most likely to benefit from it. This finding is highly encouraging for PrEP implementation, although quantifying the likely demand in the UK remains challenging.

ence suggests that the PROUD study population was highly selective, despite broad eligibility, and that the offer of PrEP generally attracts those men who are most likely to benefit from it. This finding is highly encouraging for PrEP implementation, although quantifying the likely demand in the UK remains challenging. A potential disadvantage of PrEP is the generation of drug-resistant viruses and the resulting loss of treatment options.27 As was the case in the placebo-controlled trials,22 patients who had acute infection when PrEP was initiated had the highest risk of developing drug resistance. Acute infection can only be excluded if HIV testing follows a period of no potential exposure to HIV, which is not practical in people who have sex often and a delay in initiation of PrEP carries the greater risk of an HIV infection that could be avoided.

as initiated had the highest risk of developing drug resistance. Acute infection can only be excluded if HIV testing follows a period of no potential exposure to HIV, which is not practical in people who have sex often and a delay in initiation of PrEP carries the greater risk of an HIV infection that could be avoided. An economic assessment28 based on a mathematical model adapted to the UK epidemic in men who have sex with men suggests that providing targeted PrEP to this group from 2016 would be cost-effective at current prices, or without targeted implementation if tenofovir disoproxil fumarate–emtricitabine was halved in price. The investigators in the IPERGAY trial29 reported the same 86% reduction in HIV incidence using an on-demand regimen of tenofovir disoproxil fumarate–emtricitabine: two tablets taken 2–24 h before sex, one taken 24 h later, and one taken 48 h later. The median number of pills taken each month was 16, which would cost roughly half of a daily regimen. As well as fewer pills, other advantages of the on-demand regimen include the greater ease with which PrEP can be interrupted during periods of decreased or no risk.

sex, one taken 24 h later, and one taken 48 h later. The median number of pills taken each month was 16, which would cost roughly half of a daily regimen. As well as fewer pills, other advantages of the on-demand regimen include the greater ease with which PrEP can be interrupted during periods of decreased or no risk. In the UK, the standard of prevention is already high, with free walk-in services providing screening for HIV and sexually transmitted infections, treatment for sexually transmitted infections, condoms and encouragement to use them, post-exposure prophylaxis, and support for behaviour change. Nonetheless, there remains a substantial burden of new HIV diagnoses in men who have sex with men already attending sexual health clinics and thus accessing this standard of prevention. The impressive reduction in HIV incidence in people taking PrEP, without a measurable increase in other sexually transmitted infections, is reassuring for clinical, community, and public health stakeholders. National health services are under financial constraints, but they cannot afford to ignore the results of PROUD and IPERGAY, which strongly support the addition of PrEP to the current standard of prevention for men who have sex with men at risk of HIV infection. Supplementary Material Supplementary appendix

In the UK, the standard of prevention is already high, with free walk-in services providing screening for HIV and sexually transmitted infections, treatment for sexually transmitted infections, condoms and encouragement to use them, post-exposure prophylaxis, and support for behaviour change. Nonetheless, there remains a substantial burden of new HIV diagnoses in men who have sex with men already attending sexual health clinics and thus accessing this standard of prevention. The impressive reduction in HIV incidence in people taking PrEP, without a measurable increase in other sexually transmitted infections, is reassuring for clinical, community, and public health stakeholders. National health services are under financial constraints, but they cannot afford to ignore the results of PROUD and IPERGAY, which strongly support the addition of PrEP to the current standard of prevention for men who have sex with men at risk of HIV infection. Supplementary Material Supplementary appendix Acknowledgments We dedicate this paper to the late Professor Martin Fisher, who championed treatment and prevention in the UK for two decades, and whose leadership and companionship we miss. The study was supported by ad hoc funding from the MRC Clinical Trials Unit at University College London and an innovations grant from Public Health England, and most clinics received support through the UK National Institute of Health Research Clinical Research Network. Gilead Sciences provided Truvada, distributed drug to clinics, and awarded a grant for the additional diagnostic tests including drug concentrations in plasma. We thank the PROUD participants who recognised the need for the study design, the dedication of the clinic teams, and the oversight that the governance committees provided as the study evolved, including the community engagement group who helped to boost recruitment and advised us on dissemination. A full list of contributors is given in the appendix.

ho recognised the need for the study design, the dedication of the clinic teams, and the oversight that the governance committees provided as the study evolved, including the community engagement group who helped to boost recruitment and advised us on dissemination. A full list of contributors is given in the appendix. Contributors DTD and SM designed the study with ONG and AN. MD was the trial physician. DID did the analysis with advice from DTD, ANP, and SM. MG coordinated the collection of qualitative data, and the community engagement and participant involvement meetings. RG, AKS, AC, IR, GS, NM, CB, CJL, VA, MB, JF, ST, SA, MF, AMc, and SM were site principal investigators or ran study clinics. AM led the community engagement group. SHK advised on and managed drug concentration testing. BG, ONG, AMJ, SM, and ANP were members of the steering committee; MD, DID, MF, MG, AM, AN, JR, and AKS attended as observers. SM and DTD wrote the first draft of the report and subsequent versions. All authors commented on the report and approved the final version.

SHK advised on and managed drug concentration testing. BG, ONG, AMJ, SM, and ANP were members of the steering committee; MD, DID, MF, MG, AM, AN, JR, and AKS attended as observers. SM and DTD wrote the first draft of the report and subsequent versions. All authors commented on the report and approved the final version. Declaration of interests JR is an employee of Gilead Sciences and owns stock in the company. SA became an employee on June 17, 2015. Gilead Sciences provided a grant to the institution of DID, DTD, MG, and SM for this work, and fees for attendance at a PrEP advisory board and talks by SM. The UK Medical Research Council provided financial support to SHK. MF and MB received personal fees from Gilead Sciences for attendance at a pre-exposure prophylaxis advisory board relevant to this work. Grants have been provided to authors' institutions from the following companies: Gilead (AC, MD, MF, JF, BG, RG, JF, SHK, ST); Bristol-Myers Squibb (BG, SHK); ViiV Pharmaceuticals (JF, BG, RG, SHK); Janssen (JF, SHK); Merck (SHK); and Pfizer (RG) for other studies. Financial support for conference attendance was provided to authors by Gilead (AC, SHK, ST); Janssen (AC); Abbvie (SHK); Roche (RG); and ViiV (SHK). Personal fees for attendance at advisory boards were provided by Gilead (MB, MF, BG, NM, ANP, IR); Janssen (MD, MF, RG); Abbvie (MF, BG); ViiV (MF, BG, NM); and Bristol-Myers Squibb (MF, BG). SM is named on HIV vaccine patents with any profits going directly to Imperial College London. Figure 1 Trial profile

Declaration of interests JR is an employee of Gilead Sciences and owns stock in the company. SA became an employee on June 17, 2015. Gilead Sciences provided a grant to the institution of DID, DTD, MG, and SM for this work, and fees for attendance at a PrEP advisory board and talks by SM. The UK Medical Research Council provided financial support to SHK. MF and MB received personal fees from Gilead Sciences for attendance at a pre-exposure prophylaxis advisory board relevant to this work. Grants have been provided to authors' institutions from the following companies: Gilead (AC, MD, MF, JF, BG, RG, JF, SHK, ST); Bristol-Myers Squibb (BG, SHK); ViiV Pharmaceuticals (JF, BG, RG, SHK); Janssen (JF, SHK); Merck (SHK); and Pfizer (RG) for other studies. Financial support for conference attendance was provided to authors by Gilead (AC, SHK, ST); Janssen (AC); Abbvie (SHK); Roche (RG); and ViiV (SHK). Personal fees for attendance at advisory boards were provided by Gilead (MB, MF, BG, NM, ANP, IR); Janssen (MD, MF, RG); Abbvie (MF, BG); ViiV (MF, BG, NM); and Bristol-Myers Squibb (MF, BG). SM is named on HIV vaccine patents with any profits going directly to Imperial College London. Figure 1 Trial profile *First to deferred and subsequently to immediate; considered in the deferred group for analyses but continued on pre-exposure prophylaxis. †19 pairs of partners were allocated to the same group (14 to immediate, five to deferred) including six pairs (all assigned to the immediate group) not enrolled concurrently. ‡One participant who was allocated to the deferred group was prescribed immediate pre-exposure prophylaxis in error; he was included in the deferred group for analyses but continued on pre-exposure prophylaxis. §Includes unable to contact, moved away, and non-attendance as no longer at risk. ¶HIV status ascertained if confirmed HIV-positive or HIV-negative test after 48 weeks or after Oct 13, 2014.

pre-exposure prophylaxis in error; he was included in the deferred group for analyses but continued on pre-exposure prophylaxis. §Includes unable to contact, moved away, and non-attendance as no longer at risk. ¶HIV status ascertained if confirmed HIV-positive or HIV-negative test after 48 weeks or after Oct 13, 2014. Figure 2 Incident HIV infections Left bound for each HIV case represents last non-reactive HIV test; right bound represents first reactive HIV test. The dotted line represents time when participants in the deferred group became eligible for pre-exposure prophylaxis under the original protocol. *Had a stored enrolment sample that tested positive for HIV RNA but was retained in the analysis. Table 1 Baseline characteristics

Left bound for each HIV case represents last non-reactive HIV test; right bound represents first reactive HIV test. The dotted line represents time when participants in the deferred group became eligible for pre-exposure prophylaxis under the original protocol. *Had a stored enrolment sample that tested positive for HIV RNA but was retained in the analysis. Table 1 Baseline characteristics Immediate group (n=273) Deferred group (n=267) Age (years) 35 (30–43) 35 (29–42) Ethnicity White 220 (81%) 219 (83%) Asian 14 (5%) 15 (6%) Black 11 (4%) 10 (4%) Other 28 (10%) 21 (8%) University degree 161 (59%) 166 (62%) Unemployed 24 (9%) 20 (8%) Born outside the UK 110 (40%) 107 (40%) Relationship status Partner, living together 87 (32%) 73 (27%) Partner, living separately 40 (15%) 46 (17%) No partner 146 (53%) 147 (55%) Circumcised 77 (28%) 79 (30%) Chemsex* in past 90 days 115 (43%) 116 (45%) Sexually transmitted infection diagnosed in past 12 months Any 164 (63%) 167 (65%) Bacterial† 150 (58%) 155 (60%) Rectal gonorrhoea or chlamydia 89 (34%) 83 (32%) Number of HIV tests in past 12 months 3 (2–4) 3 (2–4) Used post-exposure prophylaxis in past 12 months 91 (35%) 93 (37%) Data are median (IQR) or n (%). Two participants in each group did not return the questionnaire. Data were missing for ethnicity (none in the immediate group vs two in the deferred group), education (one vs none), employment status (none vs two), born outside UK (one vs none), relationship status (none vs one), circumcision status (two vs two), chemsex use (seven vs eight), history of sexually transmitted infection (13 vs ten), previous HIV tests (ten vs ten), and use of postexposure prophylaxis (15 vs 15).

cation (one vs none), employment status (none vs two), born outside UK (one vs none), relationship status (none vs one), circumcision status (two vs two), chemsex use (seven vs eight), history of sexually transmitted infection (13 vs ten), previous HIV tests (ten vs ten), and use of postexposure prophylaxis (15 vs 15). * Use of either γ-hydroxybutyrate, 4-methylmethcathinone, or methamphetamine to facilitate or enhance sex. † Gonorrhoea, chlamydia, or syphilis. Table 2 Interruptions to treatment because of clinical or laboratory adverse events, by participant

cation (one vs none), employment status (none vs two), born outside UK (one vs none), relationship status (none vs one), circumcision status (two vs two), chemsex use (seven vs eight), history of sexually transmitted infection (13 vs ten), previous HIV tests (ten vs ten), and use of postexposure prophylaxis (15 vs 15). * Use of either γ-hydroxybutyrate, 4-methylmethcathinone, or methamphetamine to facilitate or enhance sex. † Gonorrhoea, chlamydia, or syphilis. Table 2 Interruptions to treatment because of clinical or laboratory adverse events, by participant Weeks since enrolment Signs and symptoms Grade* Relation to study drug* A 44 Hospital-acquired pneumonia Potentially life threatening Unlikely B 43 Chest pain musculoskeletal Potentially life threatening Unrelated C 4 Headache Severe Probable D 2 Fall Severe Unrelated E 35 Anxiety or panic attack Severe Unrelated F 43 Depression Severe Unrelated G 52 Manic depression Severe Unrelated H 0 Nausea, abdominal pain Moderate Probable C 0 Headache Moderate Probable I 5 Nausea Moderate Probable J 24 Polyarthralgia Moderate Probable K 49 Nausea Moderate Probable L 0 Influenza-like illness Moderate Possible M 4 High creatinine concentration Moderate Possible H 1 Breathlessness, palpitations, chest pain Moderate Unlikely N 1 Anxiety or depression Moderate Unlikely O 1 Gastroenteritis Moderate Unlikely H 2 Chest pain Moderate Unlikely P 46 Loin pain Moderate Unlikely B 47 Central chest pain Moderate Unlikely Q 6 Headache Moderate Unrelated O 6 Intermittent nausea Mild Definite A 39 High creatinine concentration Mild Probable R 12 Lipoatrophy Mild Possible R 28 Fatigue, arthralgia Mild Possible S 47 Arthralgia Mild Possible T 5 High creatinine concentration Mild Unlikely U 14 Abnormal liver function Mild Unlikely Events in participants in the immediate group during the deferral phase of follow-up. All participants other than participant B restarted study drug.

d Possible R 28 Fatigue, arthralgia Mild Possible S 47 Arthralgia Mild Possible T 5 High creatinine concentration Mild Unlikely U 14 Abnormal liver function Mild Unlikely Events in participants in the immediate group during the deferral phase of follow-up. All participants other than participant B restarted study drug. * As assessed by participant's clinician. Table 3 Bacterial sexually transmitted infections Immediate Deferred Unadjusted odds ratio Adjusted odds ratio (90% CI)* p value Any 152/265 (57%) 124/247 (50%) 1·33 1·07 (0·78–1·46) 0·74 Gonorrhoea† 103/261 (39%) 89/242 (37%) 1·12 0·86 (0·62–1·20) 0·46 Chlamydia† 77/261 (30%) 54/242 (22%) 1·46 1·27 (0·89–1·80) 0·27 Syphilis 30/263 (11%) 22/247 (9%) 1·32 1·29 (0·79–2·10) 0·39 Rectal gonorrhoea or chlamydia 93/258 (36%) 77/238 (32%) 1·18 1·00 (0·72–1·38) 0·99 Infections diagnosed during deferral phase of follow-up. Analysis based on participants with at least one screen. * Adjusted for the number of screens for specific infection. † Detected in throat, urethra, or rectum.

Introduction Haemorrhoids result from enlargement of the haemorrhoidal plexus and pathological changes in the anal cushions, a normal component of the anal canal. They are common, affecting about a third of the population.1 Approximately 23 000 haemorrhoidal operations were done in England in 2004–05.2 Repeated visits to hospital for therapy represent an important disruption to personal and working lives. Treatment depends on the degree of symptoms and prolapse, ranging from dietary advice, outpatient rubber band ligation (RBL), to operation requiring anaesthesia. Although RBL is cheap and serious complications rare, recurrence is common, particularly where prolapse is substantial.3 Patients often require further banding.3 Although variations exist (eg, ligasure haemorrhoidectomy), surgery is usually traditional haemorrhoidectomy or a stapled haemorrhoidopexy, both requiring anaesthesia. Traditional haemorrhoidectomy is associated with considerable postoperative discomfort, sometimes necessitating admission to hospital and delayed return to normal activity, but recurrence is low. Stapled haemorrhoidopexy has a slightly higher recurrence rate but patients return to normal activity more quickly than with traditional haemorrhoidectomy.4 Research in context Evidence before this study

Treatment depends on the degree of symptoms and prolapse, ranging from dietary advice, outpatient rubber band ligation (RBL), to operation requiring anaesthesia. Although RBL is cheap and serious complications rare, recurrence is common, particularly where prolapse is substantial.3 Patients often require further banding.3 Although variations exist (eg, ligasure haemorrhoidectomy), surgery is usually traditional haemorrhoidectomy or a stapled haemorrhoidopexy, both requiring anaesthesia. Traditional haemorrhoidectomy is associated with considerable postoperative discomfort, sometimes necessitating admission to hospital and delayed return to normal activity, but recurrence is low. Stapled haemorrhoidopexy has a slightly higher recurrence rate but patients return to normal activity more quickly than with traditional haemorrhoidectomy.4 Research in context Evidence before this study Haemorrhoidal artery ligation (HAL) is a relatively new procedure that has become increasingly established as a treatment for haemorrhoids. A NICE overview in 2010 and four systematic reviews published between 2009 and 2015 highlight the lack of good quality data as evidence for the advantages of the HAL technique. The reviews included five randomised trials, two comparative cohort trials, and 21 cohort studies. From these reviews, the pooled recurrence rate for HAL ranged from 11% to 17·5%. A commonly used technique for treatment of early grade haemorrhoids is rubber band ligation (RBL). This technique is simple and easy to carry out, requiring no anaesthetic and with rapid recovery. It is the obvious comparator for treatment of early grade haemorrhoids. To date, there have been no randomised trials that have compared HAL with RBL.

or treatment of early grade haemorrhoids is rubber band ligation (RBL). This technique is simple and easy to carry out, requiring no anaesthetic and with rapid recovery. It is the obvious comparator for treatment of early grade haemorrhoids. To date, there have been no randomised trials that have compared HAL with RBL. Added value of this study We did a multicentre, parallel-group randomised controlled trial of 370 patients comparing HAL with RBL. The recurrence rate for HAL was significantly lower than for RBL (30% vs 49%, p=0·001) at 12 months. Further treatment was required in 31% of the RBL group and 15% of the HAL group (adjusted odds ratio [aOR] for further procedure 2·86, 95% CI 1·65–4·93; p=0·0002). 18% of the RBL group required a second banding session within the year. Excluding these patients as recurrence if they reported improvement or cure at 1 year resulted in a larger reduction of our recurrence rate for RBL and no statistical difference between the groups (HAL 30% vs RBL 37·5%, aOR 1·35, 0·85–2·15; p=0·20). Quality of life, symptom severity score, continence score, and complications occurred at a similar frequency. Pain was greater and lasted longer after an HAL procedure. The health-care cost analysis was striking. In the base case results, HAL was around £1000 more expensive and is highly unlikely to be cost-effective at the £20 000–30 000 threshold. Implications of all the available evidence

We did a multicentre, parallel-group randomised controlled trial of 370 patients comparing HAL with RBL. The recurrence rate for HAL was significantly lower than for RBL (30% vs 49%, p=0·001) at 12 months. Further treatment was required in 31% of the RBL group and 15% of the HAL group (adjusted odds ratio [aOR] for further procedure 2·86, 95% CI 1·65–4·93; p=0·0002). 18% of the RBL group required a second banding session within the year. Excluding these patients as recurrence if they reported improvement or cure at 1 year resulted in a larger reduction of our recurrence rate for RBL and no statistical difference between the groups (HAL 30% vs RBL 37·5%, aOR 1·35, 0·85–2·15; p=0·20). Quality of life, symptom severity score, continence score, and complications occurred at a similar frequency. Pain was greater and lasted longer after an HAL procedure. The health-care cost analysis was striking. In the base case results, HAL was around £1000 more expensive and is highly unlikely to be cost-effective at the £20 000–30 000 threshold. Implications of all the available evidence The results of this study suggest both procedures have a higher recurrence rate than previously reported. Although the recurrence rate at 1 year is lower for HAL compared with a single RBL, many clinicians would consider RBL as a course of treatment. If those with repeat RBL within the year are excluded as recurrence if they reported cure or improvement at 1 year, the recurrence rate was similar. Other outcomes were also similar or worse (in terms of pain) after HAL. HAL is more expensive and not cost-effective.

inicians would consider RBL as a course of treatment. If those with repeat RBL within the year are excluded as recurrence if they reported cure or improvement at 1 year, the recurrence rate was similar. Other outcomes were also similar or worse (in terms of pain) after HAL. HAL is more expensive and not cost-effective. An alternative treatment is haemorrhoidal artery ligation (HAL). Although requiring anaesthesia, evidence suggests a recovery similar to RBL, but an effectiveness that approaches the more intensive surgical options. As a consequence, the HAL procedure has gained popularity, with more than 5000 procedures carried out in the UK per year (manufacturer communication). Estimates of HAL efficacy come from several randomised trials, four systematic reviews,5, 6, 7, 8 and one overview by the National Institute for Health and Care Excellence (NICE).9 All publications highlight the lack of good quality data as evidence for the advantages of the technique. To our knowledge, there are no existing randomised trials comparing HAL with RBL. We aimed to establish the clinical and cost-effectiveness of HAL compared with RBL in the treatment of symptomatic second-degree and third-degree haemorrhoids. The primary objective was to compare patient-reported symptom recurrence 12 months after intervention. Methods The protocol was published in 2012;10 protocol amendments subsequent to trial commencement are provided in the appendix.

We aimed to establish the clinical and cost-effectiveness of HAL compared with RBL in the treatment of symptomatic second-degree and third-degree haemorrhoids. The primary objective was to compare patient-reported symptom recurrence 12 months after intervention. Methods The protocol was published in 2012;10 protocol amendments subsequent to trial commencement are provided in the appendix. Study design and participants This multicentre, parallel-group randomised controlled trial took place in 17 acute UK NHS hospitals. Delegated study staff at these 17 hospitals identified and consented potential participants. Eligible participants were aged 18 years or older with symptomatic second-degree or third-degree haemorrhoids.11 We excluded patients who had previously received any haemorrhoid surgery, more than one injection treatment for haemorrhoids, or more than one RBL procedure within 3 years before recruitment. We also excluded patients with perianal sepsis, inflammatory bowel disease, colorectal malignancy, pre-existing sphincter injury, and immunodeficiency, hypercoagulability disorders, and patients who were unable to have general or spinal anaesthetic. Sheffield CTRU coordinated follow-up and data collection in collaboration with these centres. The study was approved by the NRES Committee South Yorkshire (REC reference 12/YH/0236). All participants provided written informed consent.

bility disorders, and patients who were unable to have general or spinal anaesthetic. Sheffield CTRU coordinated follow-up and data collection in collaboration with these centres. The study was approved by the NRES Committee South Yorkshire (REC reference 12/YH/0236). All participants provided written informed consent. Randomisation and masking Participants were individually randomly assigned (in a 1:1 ratio) to receive either HAL or RBL. Randomisation was computer-generated and stratified by centre using permuted blocks of random sizes two, four, and six. Allocation concealment was achieved using a centralised web-based randomisation system in which the participant identifier was entered before the allocation was revealed. The study was open-label with no blinding of participants, clinicians, or research staff. Procedures The pre-randomisation and baseline questionnaires included EQ-5D, pain visual analogue scale (VAS), Vaizey faecal incontinence score, and the Haemorrhoid Symptom Severity (HSS) score. Baseline data collected before the procedure (usually on the day of surgery) included ethnic origin, smoking history, height, weight, comorbidities, grade of haemorrhoid, and previous treatments for haemorrhoids.

l analogue scale (VAS), Vaizey faecal incontinence score, and the Haemorrhoid Symptom Severity (HSS) score. Baseline data collected before the procedure (usually on the day of surgery) included ethnic origin, smoking history, height, weight, comorbidities, grade of haemorrhoid, and previous treatments for haemorrhoids. RBL is performed with a device that applies a rubber band to each haemorrhoid via a proctoscope. This band constricts the blood supply causing it to become ischaemic before being sloughed off 1–2 weeks later. The resultant fibrosis reduces any prolapse that might be present. The procedure is a basic surgical skill that all senior staff within the NHS are familiar with and competent in performing. Bands were applied at the discretion of the surgeon but with a view to resolution of all disease.

g sloughed off 1–2 weeks later. The resultant fibrosis reduces any prolapse that might be present. The procedure is a basic surgical skill that all senior staff within the NHS are familiar with and competent in performing. Bands were applied at the discretion of the surgeon but with a view to resolution of all disease. HAL is performed with a proctoscope modified to incorporate a Doppler transducer. There are two types of equipment in common use, the HALO device (AMI HAL Doppler system, CJ Medical, Truro, UK) and the THD device (THD Lab, Correggio, Italy). Both devices enable accurate detection and targeted suture ligation of the haemorrhoidal arteries. When combined with a so-called pexy suture, both bleeding and haemorrhoidal prolapse are addressed. All surgeons participating in the trial ensured the need for a pexy suture due to prolapse was routinely assessed and recorded. The procedure is simple, uses existing surgical skills, and has a short learning curve, with the manufacturers recommending at least five mentored cases before independently practising. All surgeons involved in the study had completed this training and had carried out an additional five independent procedures before recruitment. Day 1, 7, and 21 questionnaires were given to the participants following their procedure and data were either collected over the telephone or the questionnaire was returned by post or handed in at the 6 week visit. These questionnaires included EQ-5D and a pain VAS.

HAL is performed with a proctoscope modified to incorporate a Doppler transducer. There are two types of equipment in common use, the HALO device (AMI HAL Doppler system, CJ Medical, Truro, UK) and the THD device (THD Lab, Correggio, Italy). Both devices enable accurate detection and targeted suture ligation of the haemorrhoidal arteries. When combined with a so-called pexy suture, both bleeding and haemorrhoidal prolapse are addressed. All surgeons participating in the trial ensured the need for a pexy suture due to prolapse was routinely assessed and recorded. The procedure is simple, uses existing surgical skills, and has a short learning curve, with the manufacturers recommending at least five mentored cases before independently practising. All surgeons involved in the study had completed this training and had carried out an additional five independent procedures before recruitment. Day 1, 7, and 21 questionnaires were given to the participants following their procedure and data were either collected over the telephone or the questionnaire was returned by post or handed in at the 6 week visit. These questionnaires included EQ-5D and a pain VAS. Questionnaires at 6 weeks were collected at the clinic visit (or over the telephone if there was no visit); these included EQ-5D, pain VAS, Vaizey score, HSS, and questions regarding further treatment.

Day 1, 7, and 21 questionnaires were given to the participants following their procedure and data were either collected over the telephone or the questionnaire was returned by post or handed in at the 6 week visit. These questionnaires included EQ-5D and a pain VAS. Questionnaires at 6 weeks were collected at the clinic visit (or over the telephone if there was no visit); these included EQ-5D, pain VAS, Vaizey score, HSS, and questions regarding further treatment. The clinical assessment form was completed at the 6 week visit by the consultant or from patient notes. If a proctoscopy was completed, this information was recorded here, along with a recurrence question (same as the primary outcome question) data on any complications, further treatment and planned treatment. To identify the proportion of patients with recurrent haemorrhoids at 12 months after intervention, since no validated patient-reported symptom score exists, we asked participants a question, 12 months post-intervention:12 “At the moment, do you feel your symptoms from your haemorrhoids are: (1) cured or improved compared with before treatment; or (2) unchanged or worse compared with before treatment?”

intervention, since no validated patient-reported symptom score exists, we asked participants a question, 12 months post-intervention:12 “At the moment, do you feel your symptoms from your haemorrhoids are: (1) cured or improved compared with before treatment; or (2) unchanged or worse compared with before treatment?” Patients were considered to have recurrent haemorrhoids when any of the following were recorded: “unchanged or worse compared with before starting treatment” at 12 months, patient reported; “any subsequent procedure” (RBL, HAL, haemorrhoidectomy, haemorrhoidopexy, haemorrhoidal injection or other relevant procedure) over the 12 months (general practitioner [GP] and/or hospital records); or presence of any symptoms or events that strongly indicate recurrence (among patients not meeting the two previous criteria), as adjudicated by two blinded trial investigators (JPT, SRB; appendix). Other data collected from questionnaires at 12 months were: EQ-5D, pain VAS, Vaizey score, HSS, and questions regarding further treatment. Postoperative assessment was included in the protocol, but was not carried out universally. If patients said they were better, many surgeons did not re-examine. This practice is in line with current majority clinical practice. Complementary, adjunctive treatments (eg, dietary counselling, stool hygiene and habits, use of fibre, use of local therapies such as vasoconstrictors) were not specifically included in the trial and were prescribed at surgeon's discretion along the lines of the pragmatic study design.

Postoperative assessment was included in the protocol, but was not carried out universally. If patients said they were better, many surgeons did not re-examine. This practice is in line with current majority clinical practice. Complementary, adjunctive treatments (eg, dietary counselling, stool hygiene and habits, use of fibre, use of local therapies such as vasoconstrictors) were not specifically included in the trial and were prescribed at surgeon's discretion along the lines of the pragmatic study design. Outcomes The primary outcome was the proportion of patients with recurrent haemorrhoids at 12 months after procedure, derived from the patient's self-reported assessment in combination with resource use from their GP and hospital records. Secondary endpoints assessed at 6 weeks and 12 months were: symptom severity (assessed with an HSS adapted from Nyström and colleagues),13 incontinence inventories (assessed using the validated Vaizey faecal incontinence score),10 pain (assessed using a 10 cm VAS), surgical complications, need for further treatment, persistent symptoms at 6 weeks, and health-state utility based on the EQ-5D.10

(assessed with an HSS adapted from Nyström and colleagues),13 incontinence inventories (assessed using the validated Vaizey faecal incontinence score),10 pain (assessed using a 10 cm VAS), surgical complications, need for further treatment, persistent symptoms at 6 weeks, and health-state utility based on the EQ-5D.10 Statistical analysis Assuming the proportion of patients who experience recurrence after RBL is 30% and after HAL is 15%, the sample size required for 80% power and 5% significance was 121 individuals per group. To account for any between-surgeon variation and loss to follow-up, this number was increased to 175 per group, on the basis of a 10% attrition and a conservative assumption that there would be 14 surgeons in the trial and an intraclass correlation (ICC) of 2·5% in keeping with typical ICCs.14 However, we considered it likely that each site would have at least two surgeons, in which case the power to detect this difference was 85%; 90% power if there was no between-surgeon variation. Because the surgical procedure was well developed and standardised, ICC was expected to be virtually zero and we expected the proposed sample size to have closer to 90% power.

ould have at least two surgeons, in which case the power to detect this difference was 85%; 90% power if there was no between-surgeon variation. Because the surgical procedure was well developed and standardised, ICC was expected to be virtually zero and we expected the proposed sample size to have closer to 90% power. We did the primary analyses in individuals who had undergone one of the interventions and been followed up for at least 1 year (defined as the study population in text). We did additional analyses in the per-protocol (PP) population, restricted to those individuals who complied with the protocol. Deviations from the protocol that were not considered in relation to PP analysis were related to the consent process, missed windows for the assessments, eligibility (two leading to amending the exclusion criteria and two ineligible participants were withdrawn), and one participant was given general anaesthetic for the RBL procedure.

the protocol that were not considered in relation to PP analysis were related to the consent process, missed windows for the assessments, eligibility (two leading to amending the exclusion criteria and two ineligible participants were withdrawn), and one participant was given general anaesthetic for the RBL procedure. We did the analysis of recurrence using a random intercept logistic regression model in which covariates were treatment allocation, sex, age at surgery, and history of previous intervention as fixed effects; the surgeon was included as a random effect. Further sensitivity analyses assessed whether other baseline characteristics (symptom score, EQ-5D-5L, body-mass index) altered the strength or appeared to modify the treatment effect. We compared the severity of haemorrhoidal symptoms between groups using a generalised least squares regression model, with the same covariates as the primary outcome. We did some sensitivity analyses that adjusted for severity at randomisation (where available), at baseline, and the average of the two. The difference in symptom severity was compared separately for the 6-week and 12-month timepoints. We analysed EQ-5D-5L, incontinence, and pain in the same manner as symptoms. We compared descriptively the secondary outcome of complications elicited during the complications review interview or from the patient notes at 6 weeks and 1 year after intervention. A planned analysis of the time to recurrence was dropped because of the difficulty of eliciting the time of patient-reported recurrence.

We compared descriptively the secondary outcome of complications elicited during the complications review interview or from the patient notes at 6 weeks and 1 year after intervention. A planned analysis of the time to recurrence was dropped because of the difficulty of eliciting the time of patient-reported recurrence. All confidence intervals were two-sided 95% intervals comparing HAL to RBL and all statistical hypotheses were two-sided tests. We did a cost-utility analysis in terms of incremental cost per quality-adjusted life-years (QALYs) gained. We calculated the costs (including repeat procedures) following the standard three stage approach: identification of resource use, measurement, and valuation using the National NHS reference costs. We did a secondary cost-effectiveness analysis, which estimated incremental cost per recurrence avoided. All costs were estimated from the NHS and personal social perspective as per NICE recommendations.15 Analyses were undertaken using the R and Stata programs. This project will be published in full in the Public Health Research journal series. This trial is registered (ISRCTN41394716). Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

This project will be published in full in the Public Health Research journal series. This trial is registered (ISRCTN41394716). Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Between Sept 9, 2012, and May 6, 2014, 372 participants (of the 969 screened) were randomly assigned to receive RBL or HAL; 187 participants were allocated to receive RBL, and 185 were allocated to receive HAL (figure 1). Two of these participants (both randomly assigned to RBL) were removed from the trial completely due to ineligibility: one before the procedure, and the second after the procedure. Of the 370 randomised participants who were followed up, 340 received treatment (figure 1); their baseline characteristics are shown in table 1.

f these participants (both randomly assigned to RBL) were removed from the trial completely due to ineligibility: one before the procedure, and the second after the procedure. Of the 370 randomised participants who were followed up, 340 received treatment (figure 1); their baseline characteristics are shown in table 1. 340 participants received treatment (figure 1). Primary outcome data were available for 337 participants (161 in the HAL group and 176 in the RBL group). At 12 months, 256 fully completed patient questionnaires, 236 GP forms, and 337 consultant forms were returned. Follow-up was completed at sites on Aug 28, 2015. The median time from surgery to follow-up was 367 days (365–385) for the RBL group and 367 days (365–374) for the HAL group. There were 183 participants for whom all three of the 12 months forms were fully completed and returned (98 in the RBL group and 88 in the HAL group). The analysis population included all 176 participants in the RBL group and 161 participants in the HAL group for whom recurrence data were available, from either the patient, clinician, or GP. Four participants received HAL despite being randomly assigned to RBL, whereas three participants assigned to HAL received RBL. Since the findings in the primary analysis population and the PP population were similar, the reporting is restricted to the primary analysis population, with the exception of adverse events and complications which are by treatment received.

omly assigned to RBL, whereas three participants assigned to HAL received RBL. Since the findings in the primary analysis population and the PP population were similar, the reporting is restricted to the primary analysis population, with the exception of adverse events and complications which are by treatment received. The number of participants with a recurrence at 12 months was 87 (49%) in the RBL group compared with 48 (30%) in the HAL group (adjusted odds ratio [aOR] 2·23, 95% CI 1·42–3·51, ICC=0·000; p=0·0005). The breakdown of recurrences, overall and by criteria is presented in table 2. The proportion of participants who reported recurrence was similar between groups, with 29% of respondents in both groups stating they believed symptoms from their haemorrhoids were unchanged or worse (aOR for self-reported recurrence 1·06, 0·60–1·85; p=0·85). The increased recurrence associated with RBL was mainly attributable to the high rate of additional procedures undertaken following initial intervention (32%), compared with 14% in the HAL group by 1 year follow-up (aOR for further procedure 2·86, 1·65–4·93; p=0·0002). A further three (2%) participants in the RBL group were considered to have symptoms consistent with recurrent haemorrhoids following review of medical contacts and procedures over the 12 month follow-up, which were undertaken blind to treatment group: in two cases the participants were recorded as possibly requiring further treatment at their 6-week visit but were subsequently lost to follow-up; a third had been admitted to hospital twice for excessive bleeding but had not undergone treatment.

r the 12 month follow-up, which were undertaken blind to treatment group: in two cases the participants were recorded as possibly requiring further treatment at their 6-week visit but were subsequently lost to follow-up; a third had been admitted to hospital twice for excessive bleeding but had not undergone treatment. At 6 weeks, data were available for 293 patients (figure). 43 (29%) of 150 participants in the RBL group reported their haemorrhoids as unchanged or worse, compared with 12 (8%) of 143 participants in the HAL group; additionally, one participant in each group had subsequently undergone RBL. Thus the overall number of patients with persistent symptoms was 44 (29%) versus 13 (9%); adjusted odds ratio 4·35 (95% CI 2·19–8·65; p<0·0001). At 6 weeks, HSS scores were higher in the RBL group, indicating short-term symptoms were less pronounced following HAL (appendix). The mean scores were 4·0 (SD 3·5) in the RBL group and 3·0 (3·1) in the HAL group, with an adjusted difference in means of 1·0 (95% CI 0·3 to 1·8; p=0·010). No difference was apparent at 12 months, with the mean being 3·6 (3·2) for RBL and 3·6 (3·3) for HAL (adjusted difference 0·0, 95% CI −0·8 to 0·8; p=0·98).

endix). The mean scores were 4·0 (SD 3·5) in the RBL group and 3·0 (3·1) in the HAL group, with an adjusted difference in means of 1·0 (95% CI 0·3 to 1·8; p=0·010). No difference was apparent at 12 months, with the mean being 3·6 (3·2) for RBL and 3·6 (3·3) for HAL (adjusted difference 0·0, 95% CI −0·8 to 0·8; p=0·98). Before intervention, the mean health utility (EQ-5D-5L) was around 0·9 in both groups but declined at days 1 and 7 in the HAL group (figure 2). For RBL the mean at day 1 was 0·84 (SD 0·19) and at day 7 it was 0·92 (0·15); in other words, health state was reduced for the first day but had reverted back at 1 week. By contrast, the mean health state for HAL had not returned to baseline values by day 7, with the mean being 0·76 (0·22) at day 1 and 0·83 (0·18) at day 7. The adjusted difference in means were 0·08 (95% CI 0·04–0·13; p<0·001) at day 1 and 0·08 (0·05–0·12; p=0·001) at day 7. The mean health utility was nearly similar with no statistical differences between the two groups (and above baseline values) at all timepoints from day 21 onwards. The Vaizey faecal incontinence score was similar between groups (appendix). An improvement of around one unit was noted in both groups at 6 weeks, with a difference between groups of −0·1 (95% CI −1·3 to 1·0; p=0·86). The improvement was maintained at 1 year, with a difference of 0·5 (95% CI −0·7 to −1·8; p=0·38).

Before intervention, the mean health utility (EQ-5D-5L) was around 0·9 in both groups but declined at days 1 and 7 in the HAL group (figure 2). For RBL the mean at day 1 was 0·84 (SD 0·19) and at day 7 it was 0·92 (0·15); in other words, health state was reduced for the first day but had reverted back at 1 week. By contrast, the mean health state for HAL had not returned to baseline values by day 7, with the mean being 0·76 (0·22) at day 1 and 0·83 (0·18) at day 7. The adjusted difference in means were 0·08 (95% CI 0·04–0·13; p<0·001) at day 1 and 0·08 (0·05–0·12; p=0·001) at day 7. The mean health utility was nearly similar with no statistical differences between the two groups (and above baseline values) at all timepoints from day 21 onwards. The Vaizey faecal incontinence score was similar between groups (appendix). An improvement of around one unit was noted in both groups at 6 weeks, with a difference between groups of −0·1 (95% CI −1·3 to 1·0; p=0·86). The improvement was maintained at 1 year, with a difference of 0·5 (95% CI −0·7 to −1·8; p=0·38). Patients rated their current pain due to haemorrhoids at baseline and at four timepoints over the subsequent 6 weeks using a 10-point VAS. HAL was associated with more short-term pain than was RBL. The mean pain 1 day after procedure was 3·4 (SD 2·8) in the RBL group and 4·6 (2·8) in the HAL group (difference −1·2, 95% CI −1·8 to −0·5; p=0·0002); at day 7 the mean scores were 1·6 (2·3) in the RBL group and 3·1 (2·4) in the HAL group (difference −1·5, −2·0 to −1·0; p<0·0001). The mean pain was similar between groups at 21 days (1·3 [2·0] in the RBL group vs 1·4 [1·9] in the HAL group; difference −0·1, −0·6 to 0·3; p=0·44) and 6 weeks (1·2 [2·1] in the RBL group vs 1·0 [1·8] in the HAL group; difference 0·2, −0·2 to 0·7; p=0·32).

4) in the HAL group (difference −1·5, −2·0 to −1·0; p<0·0001). The mean pain was similar between groups at 21 days (1·3 [2·0] in the RBL group vs 1·4 [1·9] in the HAL group; difference −0·1, −0·6 to 0·3; p=0·44) and 6 weeks (1·2 [2·1] in the RBL group vs 1·0 [1·8] in the HAL group; difference 0·2, −0·2 to 0·7; p=0·32). 15 individuals reported serious adverse events requiring hospital admission (table 3). One patient experienced several episodes of bleeding after RBL; further investigations revealed a rectal tumour. This serious adverse event was classified as pre-existing and was not included. Of the remaining 14 serious adverse events, 12 (7%) were among participants treated with HAL (one of whom had been switched from the RBL group) and two (1%) were in those treated with RBL. Six patients had pain requiring prolonged hospital stay (five treated with HAL, one treated with RBL), three had bleeding (not requiring transfusion, two treated with HAL, one treated with RBL), two had urinary retention, two had vasovagal upset, and one had possible sepsis (treated with antibiotics); all 14 events were prespecified as expected in the study protocol.

l stay (five treated with HAL, one treated with RBL), three had bleeding (not requiring transfusion, two treated with HAL, one treated with RBL), two had urinary retention, two had vasovagal upset, and one had possible sepsis (treated with antibiotics); all 14 events were prespecified as expected in the study protocol. The main findings of within trial cost-utility analysis suggest that the HAL procedure appeared not to be cost-effective compared with RBL at a cost-effectiveness threshold of £20 000–30 000 per QALY. In the base-case results, the difference in mean total costs was £1027 higher for HAL than for RBL (table 4). QALYs were higher for HAL than for RBL; however, the difference was very small (0·010), resulting in an incremental cost-effectiveness ratio (ICER) of £104 427 per additional QALY. At £20 000 per QALY threshold, HAL has zero probability of being cost-effective; at £30 000 threshold, it has 0·05 probability of being cost-effective. The mean total cost per patient for HAL was £1750 (95% CI 1333–2167) compared with £723 (551–896) for RBL (table 3).

The main findings of within trial cost-utility analysis suggest that the HAL procedure appeared not to be cost-effective compared with RBL at a cost-effectiveness threshold of £20 000–30 000 per QALY. In the base-case results, the difference in mean total costs was £1027 higher for HAL than for RBL (table 4). QALYs were higher for HAL than for RBL; however, the difference was very small (0·010), resulting in an incremental cost-effectiveness ratio (ICER) of £104 427 per additional QALY. At £20 000 per QALY threshold, HAL has zero probability of being cost-effective; at £30 000 threshold, it has 0·05 probability of being cost-effective. The mean total cost per patient for HAL was £1750 (95% CI 1333–2167) compared with £723 (551–896) for RBL (table 3). Among the 80 participants who required a further procedure, the majority (15 of 23 in the HAL group and 45 of 57 in the RBL group) underwent a single procedure and in most cases this was RBL, although some variation was noted across centres: the cost of an additional RBL procedure in this trial was £523·16. As RBL is a brief outpatient procedure with (relatively) minimal inconvenience to the patient, it could be argued that a repeat RBL is not itself indicative of a recurrence. Consequently, we did an additional post-hoc analysis to investigate the extent to which recurrence differed between an outpatient course of RBL treatment and HAL (ie, excluding second bandings in the RBL group). Of the 31 patients in the RBL group who underwent repeat RBL, 21 were reclassified as non-recurrences since they reported being cured or improved at 1 year. This analysis changed the number of recurrences to 66 (37·5%) in the RBL group, and 48 (30%) in the HAL group (adjusted odds ratio 1·35, 95% CI 0·85–2·15; p=0·20).

Of the 31 patients in the RBL group who underwent repeat RBL, 21 were reclassified as non-recurrences since they reported being cured or improved at 1 year. This analysis changed the number of recurrences to 66 (37·5%) in the RBL group, and 48 (30%) in the HAL group (adjusted odds ratio 1·35, 95% CI 0·85–2·15; p=0·20). A further (post-hoc) analysis looked at the proportion of participants whose symptom score was either zero or one, since this number corresponds to the definition of cure used by Nyström.13 The proportions suggest that although there were more “cured” patients in the HAL group at 6 weeks (31% RBL vs 38% HAL) and 1 year (27% RBL vs 31% HAL), there was no statistical difference between the groups (adjusted OR 0·73, 0·44–1·22; p=0·23 at 6 weeks and OR 0·79, 0·46 to 1·38; p=0·42 at 1 year). The trial design meant that there was a difference between interventions in terms of dates of randomisation and surgery (RBL was often carried out immediately whereas HAL patients went onto a waiting list; see table 1). Because of this, baseline data were recorded both at randomisation and at the time of surgery. The extent of agreement was generally similar regardless of the severity.

ions in terms of dates of randomisation and surgery (RBL was often carried out immediately whereas HAL patients went onto a waiting list; see table 1). Because of this, baseline data were recorded both at randomisation and at the time of surgery. The extent of agreement was generally similar regardless of the severity. Discussion Recurrence 12 months after HAL was significantly lower than after RBL. Haemorrhoidal disease is a benign condition with treatment primarily aimed at addressing symptoms. In the absence of a validated symptom scoring system, we felt the most important determinant of treatment success was patient-reported outcome of improvement and the need to avoid additional procedures. Where patients had undergone further intervention for haemorrhoids, they were considered to have recurred. Based on this premise, HAL appears superior. This apparent superiority should be put in practical context. 18% of the participants in the RBL group underwent repeat banding. This is common practice and patients might find this re-banding a more palatable option than having an operation if it has the same potential for improvement. Indeed some clinicians deem RBL as a course of treatment. Including these patients as a success (if they reported improvement at 12 months) resulted in a reduction in recurrence and no statistical difference between the groups.