BMJ

Confidential: For Review Only

Childhoo d in relation to major causes of death: 68 years of follow-up of Scotland’s whole 1936-born population

Journal: BMJ

Manuscript ID BMJ.2016.037184.R1

Article Type: Research

BMJ Journal: BMJ

Date Submitted by the Author: 25Apr2017

Complete List of Authors: Calvin, Catherine; University of Oxford, Psychiatry Batty, David; UCL, Medicine Der, Geoff; University of Glasgow, Brett, Caroline; Liverpool John Moores University Taylor, Adele; University of , Pattie, Alison; , Čukić, Iva; University of Edinburgh, Psychology Deary, Ian; Dept of Psychology

Keywords: intelligence, IQ, mortality risk, causes of death, population study

https://mc.manuscriptcentral.com/bmj Page 1 of 60 BMJ

Childhood intelligence in relation to major causes of death: 68 years of followup of Scotland’s whole 1936 1 2 born population 3 4 5

6 a,b,c 7 Catherine M. Calvin , Postdoctoral Research Assistant 8 Confidential: For Review Only 9 G. David Batty c,d, Reader in 10 11 Geoff Der c,e, Senior Research Fellow 12 13 Caroline E. Brett c,f , Lecturer in Health Psychology Adele Taylor b, Research Assistant 14 15 b 16 Alison Pattie , Research Associate 17 b,c 18 Iva Čukić , Postdoctoral Research Assistant 19 20 Ian J. Deary b,c, Professor of Differential Psychology 21 22 23 24 a 25 Department of Psychiatry, University of Oxford, Warneford Hospital, Oxford OX3 7JX, UK

26 b 27 Department of Psychology, University of Edinburgh, 7 George Square, Edinburgh, EH8 9JZ, UK 28 29 c Centre for Cognitive and (CCACE), Department of Psychology, 30 31 University of Edinburgh, 7 George Square, Edinburgh, EH8 9JZ, UK 32 33 d Department of Epidemiology & Public Health, University College London, 119 Torrington Place, 34 35 36 London, WC1E 6BT, UK 37 e 38 MRC/CSO Social and Public Health Sciences Unit, University of Glasgow, 200 Renfield Street, Glasgow 39 40 G2 3QB 41 42 f Natural Sciences and Psychology, Liverpool John Moores University, Tom Reilly Building, Byrom Street, 43 44 45 Liverpool, L3 3AF, UK 46 47 48 49 CORRESPONDENCE TO: 50 51 Ian Deary, Centre for Cognitive Ageing and Cognitive Epidemiology (CCACE), Department of Psychology, 52 53 University of Edinburgh, 7 George Square, Edinburgh, EH8 9JZ, UK 54 55 56 T. 0044 131 650 3452. E. [email protected] , OR 57 58 Catherine Calvin, Department of Psychiatry, University of Oxford, Warneford Hospital, Oxford OX3 7JX, 59 60 UK https://mc.manuscriptcentral.com/bmj BMJ Page 2 of 60

T: 0044 1865 613103 E. [email protected] 1 2 3 4 5 Word count : 6525 6 7 8 Confidential: For Review Only 9 Keywords : intelligence, IQ, mortality risk, causes of death, population study 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/bmj Page 3 of 60 BMJ

Abstract 1 2 3 4 5 6 Objectives To examine the association between measured childhood intelligence and leading causes of 7 8 deathConfidential: in men and women over the life course. For Review Only 9 10 11 Design Prospective cohort study based on a whole population of participants born in Scotland in 1936 and 12 13 14 linked to mortality data across 68 years of followup. 15 16 17 Setting Scotland. 18 19 20 Participants 33 536 men and 32 229 women who were participants in the Scottish Mental Survey of 1947 21 22 23 and who could be linked to cause of death data up to December 2015. 24 25 26 Main outcome measures Causespecific mortality, including from coronary heart disease, stroke, specific 27 28 cancer types, respiratory disease, digestive disease, external causes, and dementia. 29 30 31 Results Childhood intelligence was inversely associated with all major causes of death. The age and sex 32 33 34 adjusted hazard ratios (and 95% confidence intervals) per standard deviation (~15 points) advantage in 35 36 intelligence test score were strongest for respiratory disease (0.72 ; 0.70 to 0.74), coronary heart disease 37 38 (0.75; 0.73 to 0.77), and stroke (0.76; 0.73 to 0.79). Other notable associations (all P<.001) were observed 39 40 for: accidental deaths (0.81; 0.75 to 0.86), smokingrelated cancers (0.82; 0.80 to 0.84), digestive disease 41 42 43 (0.82; 0.79 to 0.86), and dementia (0.84; 0.78 to 0.90). Weak associations were apparent for suicide (0.87; 44 45 0.74 to 1.02) and nonsmokingrelated cancer deaths (0.96; 0.93 to 1.00), and their confidence intervals 46 47 included unity. In sexspecific analyses the above inverse associations were observable in both men and 48 49 women. Childhood intelligence was related to selected cancer presentations, including lung (0.75; 0.72 to 50 51 52 0.77), stomach (0.77; 0.69 to 0.85), bladder (0.81; 0.71 to 0.91), oesophageal (0.85; 0.78 to 0.94), liver 53 54 (0.85; 0.74 to 0.97), colorectal (0.89; 0.83 to 0.95), and haematopoietic (0.91; 0.83 to 0.98). Sensitivity 55 56 analyses on a representative subsample of the cohort observed only small attenuation effects (by 10 to 26%) 57 58 on controlling for potential confounders, including three childhood socioeconomic status indicators. In a 59 60 https://mc.manuscriptcentral.com/bmj BMJ Page 4 of 60

replication sample from Scotland, of a similar birthyear cohort and followup period, smoking and adult 1 2 socioeconomic status partially attenuated (16 to 58%) the associations. 3 4 5 6 Conclusions In a whole national population yearofbirth cohort followed over the life course from age 11 7 8 to ageConfidential: 79 years, higher scores on a wellvalidated For childhood Review intelligence test were associated Only with lower 9 10 mortality risk ascribed to coronary heart disease and stroke, smokingrelated cancers (particularly lung and 11 12 stomach), respiratory diseases, digestive diseases, accidental death, and dementia. 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/bmj Page 5 of 60 BMJ

What is already known on this subject 1 2 3 4 5 Higher intelligence tested in childhood or early adulthood is related to later death in prospective cohort 6 7 studies. 8 Confidential: For Review Only 9 This relationship is replicated for cardiovascular diseaserelated and external causes of death. However, 10 11 evidence for the association with cancerrelated mortality is conflicting, and associations with other 12 13 leading causes of death (i.e. respiratory and digestive diseases) are largely unknown. 14 15 16 The literature has so far drawn heavily upon male conscript studies, followed up to middle adulthood. 17 18 19 20 21 Section 2: What this study adds 22 23 24 This is the first study to report on the association between childhood intelligence in a whole country’s 25 26 yearofbirth population (men and women) and specific causes of death in older adulthood. 27 28 Intelligence tested in childhood is inversely associated with all leading lifetime causes of death, 29 30 31 including coronary heart disease, stroke, cancer, external causes, respiratory disease, digestiverelated 32 33 diseases, and, dementia. However, the results for specific cancer sites are heterogeneous, and significant 34 35 only for smokingrelated cancers. 36 37 The causespecific mortality risk according to childhood intelligence is consistent in men and women, 38 39 with the exception of suicide, which is significant in men only. 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/bmj BMJ Page 6 of 60

Introduction 1 2 Findings from prospective cohort studies based on human populations from Australia, Sweden, Denmark, 3 4 5 the US, and the UK indicate that higher cognitive ability (intelligence) measured using standard tests in 6 7 childhood or early adulthood is related to a lower risk of total mortality by mid to lateadulthood (1). The 8 Confidential: For Review Only 9 association is evident in men and women (1,2); it is incremental across the full range of ability scores (2,3); 10 11 and does not appear to be confounded by socioeconomic status of origin nor perinatal factors (1,4,5). 12 13 Whereas similar gradients are also apparent for selected causes of death, such as cardiovascular disease 14 15 16 (3,6–10), suicide (7,10–15), and accidents (7,16,17), the association with other leading causes remains 17 18 inconclusive or little tested. In the present study, mortality surveillance for a whole country born in 1936 19 20 who had an assessment of childhood intelligence provides the valuable opportunity to examine littletested 21 22 intelligence–mortality associations and consider specificity by exploring the strengths of these associations 23 24 25 according to leading causes of death, in men and women, and over almost the entire life course. 26 27 28 29 Several hypotheses have been posited to explain associations between intelligence and later mortality 30 31 risk (18). The suggested causal mechanisms put forward, where cognitive ability is the exposure and disease 32 33 or death the outcome, include: mediation by adverse or protective health behaviours in adulthood (e.g., 34 35 36 smoking, physical activity), disease management and health literacy, and adult socioeconomic status 37 38 attainment (e.g. indicating occupational hazards) (18). Recent evidence of a genetic contribution to the 39 40 association between general cognitive ability and longevity (19), however, might support a system integrity 41 42 theory that posits a ‘latent trait of optimal bodily functioning’ proximally indicated by both cognitive test 43 44 45 performance and disease biomarkers (20). None of these possibilities are mutually exclusive. Whereas 46 47 cognitive epidemiology (18) makes a unique contribution to better understanding health inequalities in 48 49 populations, by its successful application of a wellvalidated behavioural trait that performs independently of 50 51 social gradients in its association with health indices (21,22), there remains a fundamental question 52 53 regarding how specific and multifaceted the link is between individual differences in intelligence and 54 55 56 longevity. Evidence for an association with several leading causes of death has either not been replicated 57 58 (dementia and respiratory disease) (7,23), is conflicting (different cancer sites) (5,7,8,24–27), or hitherto 59 60 untested (digestive system disease). At least six publications have compared associations between premorbid https://mc.manuscriptcentral.com/bmj Page 7 of 60 BMJ

intelligence and a selection of causespecific mortalities (5,7,8,10,24,28), including some wellcharacterised 1 2 and extremely large cohorts with extensive followup; however, overrepresentation of menonly samples 3 4 5 (7,10), and followup that was terminated in middleage (5,10) has limited the generalisability of findings. 6 7 Furthermore, low numbers of events for diseases common in older adult populations may have contributed 8 Confidential: For Review Only 9 to some apparently conflicting results (8,24,28). 10 11 The present study aims to investigate the magnitudes of association between childhood intelligence 12 13 and all major causes of death, using a whole yearofbirth population followed up to older age, therefore 14 15 16 capturing sufficient numbers of cases for each outcome. Secondly, it aims to investigate sex differences in 17 18 the associations. Thirdly, it conducts sensitivity analyses to test for some possible mechanisms of 19 20 association, including confounding and mediation by socioeconomic status. 21 22 23 24 25 26 27 Methods 28 29 30 31 Study population and data sources 32 33 In this prospective cohort study, all individuals born in Scotland in 1936, and registered at school in 34 35 36 Scotland in 1947, were targeted for tracing, and subsequent data linkage to death certificates. This was 37 38 carried out by the National Records of Scotland, with permission from the Registrar General of Scotland, 39 40 using the National Health Service (NHS) Central Register for members traceable in Scotland, and the MRIS 41 42 Integrated Database and Administration System for those in England and Wales. The Confidentiality 43 44 45 Advisory Group of the Health Research Authority gave support under section 251 of the NHS Act 2006 for 46 47 linkage without consent of mortality records including cause of death with age 11 intelligence scores from 48 49 the Scottish Mental Survey 1947 (SMS1947) (25,29,30). 50 51 52 53 Childhood intelligence 54 55 56 th 57 On 4 June 1947, ~94% of the Scottish population born in 1936 and who were registered as attending school 58 59 in Scotland (75 252) completed a test of general intelligence in the SMS 1947 (N=70 805) (25,29,31). This 60 https://mc.manuscriptcentral.com/bmj BMJ Page 8 of 60

involved administration of the Moray House Test No. 12, which has 71 items tapping verbal and nonverbal 1 2 reasoning ability (29). The test was introduced by school teachers who read aloud instructions before the 3 4 5 start of a 45minute test period. Each participant was allocated a score out of the maximum of 76 (some 6 7 items scored more than one point). The Moray House Test No. 12 had good concurrent validity (correlation 8 Confidential: For Review Only 9 coefficient =~0.8) with a wellstandardised individuallyadministered Stanford version of the Binet test of 10 11 intelligence in 1947 (25,29), and with performance on Raven’s Progressive Matrices —a widelyused non 12 13 verbal ability test—(correlation coefficient =~0.7) in a recent followup study (32). The test has high life 14 15 16 long stability of individual differences (33,34), and there is good evidence for its external validity across 17 18 several studies (28,35,36). 19 20 21 22 Cause-specific death outcomes 23 24 th th 25 Causes of death were coded according to the International Classification of Diseases (ICD), 6 to 10 26 27 Edition codes (see Supplementary material: Table S1 for ICD codes). The NHS (England and Wales) 28 29 recorded deaths linked to cause of death data were provided on dates up to and including 31 December 30 31 2015, and respective data for NHSCR (Scottish)recorded deaths were provided up to 30 June 2015. 32 33 Subsequent cause of death updates provided by NHSCR, on a quarterly basis up to 31 December 2015, were 34 35 36 provided without unique ID numbers for individuals, and therefore a matching process to link these to 37 38 Scottish Mental Survey data was applied (see Supplementary Materials for these details). 39 40 41 42 Patient Involvement 43 44 45 No cohort participants contributed to the development of the present research question or the outcome 46 47 measures, nor were they involved in the design, recruitment, or conduct of the study. There is no intention to 48 49 disseminate the results of this electronic data linkage study to its participants. 50 51 52 53 Statistical analysis 54 55 56 The analytic sample includes 65 765 sample members who were traced for linkage to mortality records and 57 58 who had a census date (i.e. a final GP posting date for those who emigrated or joined the Armed Forces) and 59 60 an intelligence test score from 1947 (see Figure 1). They represent 92.3% of those who took part in the https://mc.manuscriptcentral.com/bmj Page 9 of 60 BMJ

SMS1947, and 87.4% of Scottish schoolchildren born in 1936. Exclusions due to missing intelligence test 1 2 scores affected 6.8% (5103) of the total 1936 Scotland birth cohort─which has previously been reported 3 4 5 (25,31)─and 6.6% (4744) of those who were successfully linked and had a census date. 6 7 8 Confidential: For Review Only 9 Analyses were conducted using SPSS Statistics 21, unless otherwise stated. Cox proportional hazards 10 11 regression models produced hazard ratios with 95% confidence intervals to summarise the relation between 12 13 childhood intelligence and each cause of death, adjusting for age (in days) at cognitive testing and sex 14 15 16 (dummy variable), for which there were no missing data. We ran models where the cause of death was 17 18 cardiovascular disease, coronary heart disease, stroke, cancer by type, respiratory disease, digestive disease, 19 20 externally caused, accidental, suicide, or dementia. Given that suicide is the likely cause of death in the 21 22 majority of death certificates that state ‘open verdict’ or ‘undetermined intent’ (37), we not only report on 23 24 25 models to predict deaths formally recorded as suicide, but we add deaths of undetermined intent to models 26 27 of suicide and report these in the text of the Results. A cause of death was included if it was listed among 28 29 multiple causes of death (MCOD) on the death certificate, but we repeated each model where the endpoint 30 31 was the underlying cause of death (38), and these results are reported in Supplementary materials . 32 33 In these survival analyses, the time scale was calendar time (days) from the date of the intelligence 34 35 th 36 test (4 June 1947) to the census date, which was the earliest of: (i) date of death; (ii) latest posting date (for 37 st 38 migration or Armed Forces entrants); or (iii) 31 December 2015 (end of followup). Members who died 39 40 from a different cause to that being modelled were included in the denominator (persons at risk) and 41 42 censored at date of death, as is standard practice in epidemiological analyses. The proportional hazards 43 44 45 assumption was assessed by inspection of logminuslog plots and formally tested in Stata 14 using the 46 47 Schoenfeld residuals test. The assumption held for associations between intelligence and all major causes of 48 49 death according to both methods, with the exception of cancerrelated deaths using logminuslog plots and 50 51 accidental death under formal testing (p=.03). In order to visually inspect associations for linearity, hazard 52 53 ratios and their 95% confidence intervals were graphically plotted to show the risk of each event type in 54 55 56 accordance with intelligence test score decile groups (for >1000 numbers of deaths) or quartile groups (for 57 58 <1000 numbers of cases). For our main results, hazard ratios and their 95% confidence intervals were 59 60 estimated for causespecific mortality according to a one SD (~15 point) advantage in intelligence test score. https://mc.manuscriptcentral.com/bmj BMJ Page 10 of 60

With the suggestion that some intelligence–death associations are modified by gender, we formally tested 1 2 for sex differences by including an interaction term (Sex X Intelligence Score) in models predicting cause of 3 4 5 death. We then produced effect estimates separately for men and women. Finally, to adjust for multiple 6 7 testing, using R (39), we made False Discovery Rate correction to significance Pvalues resulting from all 8 Confidential: For Review Only 9 models (40), and report on statistically significant models (p<.05) that fail this correction. 10 11 12 13 14 15 16 Assessing selection bias caused by missing intelligence test data 17 18 We addressed whether 6.8% missingness on the intelligence test in the SMS1947 caused selection bias that 19 20 affected the magnitude of the intelligencemortality association. We conducted sensitivity analyses using a 21 22 representative subsample (1.7%; N=1208) of the SMS1947, the socalled ‘6Day Sample’ (25,37,38). This 23 24 25 subsample was affected by a similar degree of missingness as the whole SMS1947 on the MHT; however, 26 27 they all took an additional, individuallyadministered intelligence test on another occasion (see 28 29 Supplementary Materials for more details of the 6Day Sample). 30 31 32 33 Confounder-adjustment: controlling for school 34 35 36 In the absence of individuallevel data on socioeconomic status in the full SMS1947, the school attended 37 38 was used as a proxy measure to adjust for potential confounding by background socioeconomic status, given 39 40 that primary school has moderate correlation with paternal social class in a Scottish cohort study (41). 41 42 Schools were specified as strata in the models for major causes of death, and cancer subtypes, and fixed 43 44 45 effects models were fit whereby all characteristics shared by pupils from the same school are controlled for. 46 47 There is however the possibility that children are selected into some schools because of prior cognitive 48 49 ability, and therefore there might be overadjustment in this particular sensitivity analysis. 50 51 52 53 Confounder-adjustment: subgroup analyses 54 55 56 Models for leading causes of death in association with childhood intelligence scores were repeated in a 57 58 representative subsample of the SMS1947, the socalled ’30Day Sample’ (7.2%; N=5083), on whom more 59 60 background data were available to test for potential confounding (see Supplementary Materials for more https://mc.manuscriptcentral.com/bmj Page 11 of 60 BMJ

information on this subsample). These data included socioeconomic status variables (paternal occupational 1 2 status, home overcrowding, and school absenteeism) and physical status indicators (height and physical 3 4 5 disability). 6 7 8 Confidential: For Review Only 9 Adjustment for adult socioeconomic status and smoking: a replication sample 10 11 In the absence of data in the intermediary ‘lifetime’ period of the study with which to test for potential 12 13 mediation, we report on equivalent models run on a replication sample the ‘West of Scotland Twenty07 14 15 16 study’ – with and without adjustment for adult smoking status and occupational status. This cohort shares 17 18 similarities to the Scottish Mental Survey 1947, including: (1) Scotlandderived; (2) 1930s birth cohort; (3) 19 20 linkage to mortality records to at least 2015 (see Supplementary Materials for more details of this replication 21 22 sample). 23 24 25 26 27 Results 28 29 30 31 Characteristics of the sample 32 33 Among 75 252 members of the 1936 birth cohort, 3242 were untraceable, 630 were without census dates 34 35 36 (i.e. migrated without posting date), 4744 had no data on childhood intelligence, and 871 were mismatches 37 38 of the linkage (Figure 1). Among 65 765 individuals (32 229 of whom were women, and who attended 5182 39 40 different schools) comprising the analytic sample, 25 979 were deceased and 30 464 were confirmed to be 41 42 living in the UK at followup. Mean (and SD) age at death was 66.1 (10.6) years; mean time to followup 43 44 45 was 57.0 (18.4) years, and total personyears of followup was 3.54 million. There were no marked 46 47 differences in the characteristics of the analytic sample and the remainder of the Scottish Mental Survey 48 49 1947, except that, among those with a childhood mental test score, the analytic sample scored on average 50 51 higher on the Moray House Test (M=37.1, SD=15.7) than those excluded from the analyses (M=35.4, 52 53 SD=16.4; p<.001). 54 55 56 57 58 Childhood intelligence in relation to cause-specific mortality 59 60 https://mc.manuscriptcentral.com/bmj BMJ Page 12 of 60

Figure 2 shows graphs depicting the associations between childhood intelligence (by decile or quartile 1 2 group) and risk of the major causes of death. These show that, for the large majority of endpoints, 3 4 5 associations were inverse; that is, higher childhood intelligence was associated with a lower risk of cause 6 7 specific death. Risk of lifetime respiratory diseaserelated death fell by twothirds in the topperforming 8 Confidential: For Review Only 9 decile for childhood intelligence versus the bottom decile. Furthermore, mortality risk was halved for those 10 11 in the highest vs lowest decile of intelligence according to deaths caused by coronary heart disease, stroke, 12 13 smokingrelated cancers, digestive diseases, and external causes. The risk of dementiarelated mortality and 14 15 16 deaths by suicide were reduced by at least a third in the highest performing quartile of intelligence test score 17 18 versus the lowest quartile. There was no evident association between childhood intelligence and non 19 20 smokingrelated cancer mortality. 21 22 23 24 25 Figure 3 is a forest plot of the risk of major causes of death associated with a one SD higher score in 26 27 childhood intelligence. The effect sizes were similar whether the outcome was the primary cause of death or 28 29 a cause mentioned anywhere on the death certificate; therefore, the plot presents those for ‘any mention of’ 30 31 (see Supplementary Figure 1 for the equivalent plot where the outcome was underlying cause of death). All 32 33 hazard ratios revealed inverse associations, and the large majority showed narrow confidence intervals. The 34 35 36 strongest effect sizes were seen for respiratory disease (hazard ratio; 95% confidence interval: 0.72; 0.70 to 37 38 0.74), coronary heart disease (0.75; 0.73 to 0.77), and stroke (0.76; 0.73 to 0.79). All others fell within the 39 40 hazard ratio range from 0.80 to 0.87, except for nonsmokingrelated cancer deaths (0.96; 0.93 to 1.00)─the 41 42 statistical significance of this latter effect did not survive FDR correction. The relatively wide confidence 43 44 45 interval for deaths due to intentional selfharm reflects the lower number of cases for this endpoint (0.87; 46 47 0.74 to 1.02). When deaths of undetermined intent were included in the model for suicide, the confidence 48 49 interval fell below 1 (0.84, 0.74 to 0.96, cases: n=220). 50 51 52 53 Sex interaction effects were computed with childhood intelligence in the total sample, and sex 54 55 56 specific hazard ratios were also estimated for the associations between childhood intelligence and leading 57 58 causes of death (see Supplementary Table 2). In models of the total sample, sex interaction terms were 59 60 significant for risk of cardiovascular disease, coronary heart disease, smokingrelated cancers, respiratory https://mc.manuscriptcentral.com/bmj Page 13 of 60 BMJ

diseases, and dementia. Whereas the inverse patterns of intelligencemortality association observed for the 1 2 total sample were seen in both men and women, the magnitude of the associations were slightly greater 3 4 5 among women than among men for the majority of causespecific deaths. The greatest difference in effect 6 7 estimate was for dementia, whereby a one SD higher score in childhood intelligence was associated with a 8 Confidential: For Review Only 9 10% reduced dementia risk in men and a 24% reduced dementia risk in women. Otherwise, between men 10 11 and women there were equivalent hazard ratios for nonsmokingrelated cancer deaths and external causes 12 13 of death, except for deaths by intentional selfharm, which were associated with childhood intelligence in 14 15 16 men (0.80; 0.66 to 0.96) but not women (1.15; 0.82 to 1.60). This sex difference was also evident for suicide 17 18 when open verdict deaths were included (men: 0.76; 0.66 to 0.89; women: 1.05; 0.83 to 1.34; sex by 19 20 intelligence interaction effect in the total sample: P=.030). 21 22 23 24 25 Childhood intelligence in association with death by specific cancer type 26 27 Figure 4 presents a series of graphs for associations between childhood intelligence score groupings (decile 28 29 or quartile) and deaths related to 15 specific cancers. Approximately half of these showed inverse patterns of 30 31 association with a degree of linearity, including cancers of the oesophagus, colon or rectum, stomach, liver, 32 33 lung, kidney, bladder, and blood. The strongest association was evident for lung cancerrelated death; those 34 35 36 in the highest performing decile of childhood intelligence were at twothirds reduced risk compared to the 37 38 lowest performing decile. Cancers showing negligible or irregular associations with childhood mental ability 39 40 were: mouth, pancreas, skin, ovaries, breast (women only), prostate (men only), and brain or central nervous 41 42 system. 43 44 45 46 47 Figure 5 is a forest plot of hazard ratios showing the strengths of association between a one SD 48 49 higher score in childhood intelligence and risk of death related to 15 specific cancers, and total cancers as a 50 51 comparator (see Supplementary Figure 2 for equivalent forest plot reflecting primary cause cases only). 52 53 Whereas the effect estimate for death by any cancer clearly reveals an inverse association with narrow 54 55 56 confidence intervals (0.86; 0.84 to 0.88), there is obvious heterogeneity across the results for specific cancer 57 58 sites. Among the eight specific cancers that showed linearity in their inverse association with childhood 59 60 mental ability in Figure 4, the plot reveals robust associations for seven of these (oesophageal, colorectal, https://mc.manuscriptcentral.com/bmj BMJ Page 14 of 60

stomach, liver, lung, bladder, and, lymphoid or haematopoietic), with hazard ratios in the range 0.75 to 0.91. 1 2 The strongest associations with childhood mental ability were deaths associated with lung cancer (0.75; 0.72 3 4 5 to 0.77) and stomach cancer (0.77; 0.69 to 0.85). 6 7 8 Confidential: For Review Only 9 Sex interaction effects with childhood intelligence were modelled to predict specificcancer related 10 11 mortality in the total sample, and then sexspecific hazard ratios were calculated (see Supplementary Table 12 13 3). A significant sex interaction effect ( p=.013) with mental ability in predicting lung cancer in the total 14 15 16 sample may be a chance finding, and the direction of association was the same in men (0.77; 0.74 to 0.81) 17 18 and women (0.70; 0.66 to 0.75). Otherwise these analyses revealed no clear sex differences in direction or 19 20 magnitude. 21 22 23 24 25 Sensitivity analyses: assessment of selection bias caused by missing intelligence test data 26 27 In analyses of the socalled ‘6Day Sample’ of the Scottish Mental Survey 1947 (i.e. those born on the first 28 29 day of the evennumbered months of 1936), selection bias according to missing Moray House Test scores in 30 31 the SMS1947 was not found to affect the effect size of the association between intelligence test scores (on 32 33 the TermanMerrill, an individuallyadministered test) and total mortality (see Supplementary Materials for 34 35 36 these sensitivity analyses). 37 38 39 40 Sensitivity analyses: assessment of potential confounding 41 42 Firstly, we attempted to control for potential confounding by adjusting for the fixed effects of school 43 44 45 attended, in the full SMS1947. The HRs for causespecific mortality, and for deaths related to cancer 46 47 subtype, generally showed small attenuation effects of about 0.02 HR points (see Supplementary Figures 3 48 49 and 4 for forest plots), and the patterns of specificity were similar to those of the main analyses. 50 51 Secondly, in subgroup analysis of the SMS1947, including 2039 men and 1992 women of the 30 52 53 Day Sample, hazard ratios and their 95% confidence intervals demonstrated inverse associations between 54 55 56 childhood intelligence and allcause mortality, and mortality related to cardiovascular disease, cancer 57 58 (smokingrelated), respiratory disease, and digestive disease (but not nonsmoking cancer, external causes, 59 60 and, dementia, unlike in the total sample) (see Supplementary Table 4 for results). Respiratory disease held https://mc.manuscriptcentral.com/bmj Page 15 of 60 BMJ

the strongest association with childhood intelligence (0.73; 0.66 to 0.80), as was the case for the full sample. 1 2 In confounderadjusted models that included three childhood socioeconomic status indicators, the effect 3 4 5 estimates for these six predictive models attenuated by 7 to 26%. Adding physical status indicators to the list 6 7 of potential confounders in the model had negligible effect; the attenuation effect from the basic model 8 Confidential: For Review Only 9 ranged from 10 to 26%. 10 11 12 13 Sensitivity analyses: assessment of mediation 14 15 16 In a replication sample – The West of Scotland Twenty07 study – we observed very similar hazard ratios to 17 18 the SMS1947 for the associations between intelligence (tested in middle age) and specific underlying causes 19 20 of death (see Supplementary Figure 3), albeit they were slightly stronger and with less precision. The one 21 22 noticeable difference was for stroke as underlying cause of death, which showed a weak association with 23 24 25 midlife intelligence (0.90; 0.70 to 1.15). After adjustments were made for smoking, adult socioeconomic 26 27 status, and selfrated health, there were attenuation effects of 27 to 65% (see Supplementary Table 5), and 28 29 associations remained the most robust for respiratory disease (0.77; 0.59 to 0.99) and coronary artery disease 30 31 (0.79; 0.65 to 0.96). 32 33 34 35 36 Discussion 37 38 39 40 In this first wholepopulation birth cohort study linking childhood intelligence test scores to cause of death 41 42 data, in a followup spanning from 11 to 79 years of age, there were inverse associations for all major causes 43 44 45 of death, including coronary heart disease, stroke, cancer, respiratory disease, digestive disease, external 46 47 causes of death, and dementia. For specific cancer types the results were heterogeneous, with only smoking 48 49 related cancers showing an association with childhood ability. In general, the effect sizes were similar for 50 51 women and men (albeit marginally greater for women), with the exception of death by suicide that had an 52 53 inverse association with childhood ability in men but not women. In a representative subsample with 54 55 56 additional background data, there was evidence that childhood socioeconomic status and physical status 57 58 indicators had no more than a modest confounding impact on the observed associations. A replication study 59 60 with adult data on smoking and socioeconomic status showed less than a quarter to twothirds percentage https://mc.manuscriptcentral.com/bmj BMJ Page 16 of 60

attenuation of the equivalent effect estimates for intelligence test performance in association with cause 1 2 specific mortality. 3 4 5 6 7 Strengths and weaknesses of this study 8 Confidential: For Review Only 9 The absence of covariate data from childhood that could account for the effect estimates we observed in our 10 11 main analysis is a weakness of the study. However this lack of data should be balanced with evidence from 12 13 our sensitivity analyses, including a representative subsample of the cohort. Additionally, there are extensive 14 15 16 reports from the vast majority of studies in the literature that have attempted to control for potential 17 18 confounding. In our own analysis of the whole SMS1947, adjusting for school led to little attenuation of the 19 20 intelligencemortality associations. However, because this is only a proxy for background socioeconomic 21 22 status, and because it could lead to overcontrolling because some schools select on cognitive ability, we 23 24 25 undertook further sensitivity analyses. In our own analysis of a representative subsample of the Scottish 26 27 Mental Survey 1947, three indicators of background socioeconomic status, which are significant correlates 28 29 of intelligence test scores, explained a quarter or less of the intelligence versus causespecific mortality 30 31 associations. This is consistent with the main literature that has reported low attenuation effects in models of 32 33 premorbid intelligence and allcause (see Calvin et al., 2011 for a metaanalysis (1)) or causespecific 34 35 36 mortality, including cardiovascular disease (4,5,10,42–44), cancer (5,26), and external causes (5,10,45) 37 38 including suicide (10,15,46) and accidental death (10,16,17). Additional adjustments for perinatal factors 39 40 and/or physical status indicators also show marginal (if any) attenuation effects (4,5), as we observed here in 41 42 our adjustment of height and physical disability. Other psychological factors that have been shown to 43 44 45 correlate with intelligence test performance, including personality traits and mental health indicators, may 46 47 yet partially explain these associations. However, in a previous paper on the 6Day Sample of the SMS1947, 48 49 we found that the personality trait of ‘dependability’ (rated in adolescence and similar to the widelystudied 50 51 trait called conscientiousness), which was also predictive of mortality, did not account for the intelligence 52 53 association (47). In a different sample, the personality trait neuroticism was shown to moderate, but not 54 55 56 attenuate, the intelligencedeath association (48). The strength of some of the associations reported in the 57 58 present study, however, for example respiratory disease, are perhaps too strong to be explained by any 59 60 potential confounders we are aware of from the literature. https://mc.manuscriptcentral.com/bmj Page 17 of 60 BMJ

1 2 Another factor affecting validity of findings is historical inaccuracy of death certificates. Particular 3 4 5 causes of death, which may be underreported by coroners include dementia (49), stroke (50) and suicide 6 7 (37). Therefore, we judge that the greater chance of false negatives relative to false positives in the present 8 Confidential: For Review Only 9 study would act, if at all, to slightly attenuate the observed effect sizes. Indeed, in further analyses, including 10 11 ‘open verdict’ deaths in the suicide model increased the hazard ratio and narrowed its confidence intervals. 12 13 14 15 16 Strengths of the present study are in its analysing a nearcomplete year of birth cohort from a 17 18 nation’s population, with a very high return of linkage to all historical death records. The followup period 19 20 from childhood through to older age enables us to consider all major causes of death in the UK across the 21 22 life course. Not only can we reproduce more accurately, and with low bias, results from previous cohorts on 23 24 25 cardiovascular disease and external causes of death, but we have been able to make a substantial 26 27 contribution to the literature on premorbid intelligence and cancer risk, and report for the first time on 28 29 associations between childhood intelligence and death from digestive system diseases. Furthermore, our 30 31 whole population sample of men and women enables the reporting of sexspecific effects, which are often 32 33 lacking from a literature that draws heavily upon male conscript data. For example, we have been able to 34 35 36 report for the first time on the association between premorbid intelligence and the full range of cancer 37 38 subtypes in women. And, we report for the first time on greater effect sizes among women relative to men 39 40 for several causes of death, including respiratory disease, smokingrelated cancers, and dementia. 41 42 43 44 45 Comparison with other studies 46 47 An important aspect of the present study was to consider the role of premorbid intelligence in relation to 48 49 specific cancer sites, given the considerable heterogeneity in cancer aetiology, and conflicting findings from 50 51 previous studies of an association between premorbid intelligence and risk of total cancer mortality 52 53 (5,7,8,24–27). A cohort study of more than one million Swedish male conscripts─the largest to 54 55 56 date─reported that, among twenty incident cancers, only three had significant associations with premorbid 57 58 intelligence (27). The authors concluded that, with the likelihood of a few statistical false positives, cancer 59 60 was unlikely to explain the association between premorbid intelligence and total mortality risk. However, https://mc.manuscriptcentral.com/bmj BMJ Page 18 of 60

based upon evidence from the present study, and other comparatively smaller cohorts, we would caution 1 2 against overgeneralising this summary. Our results are consistent with previously reported effect sizes for 3 4 5 premorbid intelligence in association with lung cancer (5,7,8,24) and stomach cancer (8,24,27). 6 7 Furthermore, our study demonstrates, perhaps for the first time, that an association between prior 8 Confidential: For Review Only 9 intelligence and cancer risk may be specific to smokingrelated cancer types; significant inverse associations 10 11 were observed for five other smokingrelated cancers: oesophageal, colorectal, liver, bladder, and 12 13 haematopoietic. Whereas two previous studies reported null findings for colorectal and pancreatic cancer 14 15 16 risk in association with childhood intelligence, there were very low case numbers (8,24). Indeed in a recent 17 18 study of 728 165 Danish men intelligence in early adulthood was significantly inversely associated with 19 20 total gastrointestinal cancers (7). A pertinent question remains why more smokingrelated cancer types were 21 22 not related to premorbid intelligence in the one million Swedish males’ study (27). This may be due to their 23 24 25 much younger age at death ascertainment, and/or national differences in cigarette smoking prevalence. 26 27 Sweden has been exemplary in its reduction in cigarette use (particularly among men) since the late 1960s 28 29 (51), and maximum prevalence rates of smoking in Sweden at the time when the Swedish conscript study 30 31 was conducted would have been much lower (52) than equivalent rates in Scotland in the period of the 32 33 present study’s followup (53). If smoking is less prevalent in the Swedish population, and if smoking is an 34 35 36 important mediator of the association between premorbid intelligence and cancer risk, then this might 37 38 explain the reduced and nonsignificant effect sizes reported by Batty and colleagues (27). One notable and 39 40 significant finding from Sweden was a positive association between intelligence at conscription and incident 41 42 skin cancer, indicating that higher ability related to an increased risk (27). Although the pattern of this 43 44 45 relationship in our own data corroborates this result, we did not observe a significant effect, perhaps due to 46 47 our older cohort growing up in a time before cheap flights to hot countries became ubiquitous and the 48 49 proliferation in risk of skin cancer. The literature on premorbid intelligence and other cancer types is 50 51 relatively small and/or underpowered. We report for the first time on the association with ovarian cancer in 52 53 women, which we found to be null, and we replicate previous studies’ findings of null associations with 54 55 56 female breast cancer (5,8,24) and male prostate cancer (27). 57 58 59 60 https://mc.manuscriptcentral.com/bmj Page 19 of 60 BMJ

Preliminary evidence of an inverse association between childhood intelligence and risk of dementia 1 2 related death was reported from casecontrol studies of the 1932 Scottish Mental Survey that were linked to 3 4 5 incident dementia records (54,55). A recent prospective cohort study of this sample, followed up to 92 years, 6 7 validated the finding (23). Our findings are consistent with this, as is evidence of a greater association with 8 Confidential: For Review Only 9 dementiarelated mortality risk among women than men, also observed in the older Scottish cohort (23). 10 11 This sex differential is also supported by evidence from individual participant metaanalyses of early school 12 13 leaving in childhood associated with dementia death in women and not men (56). 14 15 16 17 18 Respiratory diseaserelated mortality held the strongest association with premorbid intelligence in 19 20 the present study (0.72; 0.70 to 0.74). In a recent study of male Danish conscripts, and the only previous 21 22 study to report on the association, respiratory disease mortality also held one of the strongest associations 23 24 25 with premorbid intelligence (0.62; 0.60 to 0.65) (7)─only homicide was stronger in its association. Previous

26 27 evidence that premorbid intelligence is associated with selfreported chronic lung disease by midlife (57,58), 28 29 and measured lung function in later life (59–62), supports these results. In contrast, a Scottish study 30 31 reporting a weak and nonsignificant inverse association with incident respiratory disease (8) may have been 32 33 underpowered. 34 35 36 37 38 Our own effect sizes for mortality risk in association with premorbid intelligence corroborate those 39 40 reported in the literature, for allcause mortality (0.80 vs aggregate HR = 0.80 in studies of ≥40 years’ 41 42 duration) (1), total cardiovascular disease (0.76 vs HRs: 0.74 to 0.93) (7–10,63), and, coronary heart disease 43 44 45 (0.75 vs HRs: 0.70 to 0.86) (3,4,7,9,42,43,64). The effect sizes for stroke risk in association with premorbid 46 47 intelligence have been less consistent in the literature, with significant inverse associations for incident or 48 49 fatalonly stroke observed in cohorts from Denmark, Scotland, Sweden and the US (3,4,7,65), and negligible 50 51 associations reported in separate cohorts from Denmark and Scotland (8,9,42). Weaker associations with 52 53 stroke in previous studies are probably influenced by low numbers of stroke cases in cohort studies driven in 54 55 56 part by historical underreporting of stroke in death certificates (50), and/or relatively low numbers of 57 58 ischaemic versus haemorrhagic strokes (7), particularly in cohort studies of younger age followup. In the 59 60 https://mc.manuscriptcentral.com/bmj BMJ Page 20 of 60

present study’s cohort followed to older age, the observed effect size for fatal stroke risk was equivalent in 1 2 magnitude to that for coronary heart disease. 3 4 5 6 7 Our findings accord with previous reports of inverse associations between premorbid intelligence 8 Confidential: For Review Only 9 and risk of externallycaused deaths, specifically from Danish and Scottish cohorts (5,7,45). In the Scottish 10 11 study that followed up on deaths from ages 15 to 57 years, external causes of death held the strongest 12 13 association with childhood intelligence relative to cancer, cardiovascular disease, and total mortality (5). In 14 15 16 contrast, we found that external causes of death held the weaker association when compared to 17 18 cardiovascular disease mortality. It may be that differences in the composition of specific types of death that 19 20 are externally caused in a younger versus older age cohort could influence the magnitude of this effect. 21 22 Moreover, the hazard ratio we report for accidental death in association with childhood intelligence (0.81; 23 24 25 0.75 to 0.86) is similar in magnitude to a study of Danish men born in 1953, intelligencetested at age 12, 26 27 and followed up to age 48 years (0.82; 0.78 to 0.86) (16). On the other hand, in studies where intelligence 28 29 was tested in late childhood or early adulthood, the effect size in relation to accidental death has been greater 30 31 (0.71 and 0.76 respectively) (7,16,17). Several highpowered prospective studies of male conscript cohorts 32 33 have reported inverse associations between premorbid intelligence and later suicide risk, including from 34 35 36 Australia (11), Denmark (7,12), and Sweden (10,13–15). Far less is reported on the association in women, 37 38 and with follow up to older age. However, our findings are very similar to the only previous study we are 39 40 aware of to report on sex differentials for incident suicide risk in association with premorbid intelligence in a 41 42 40 year followup study, where an association was evident in men (0.90; 0.83 to 0.99) but negligible in 43 44 45 women (1.04; 0.90 to 1.20) (66). 46 47 48 49 Explanations and implications for future research 50 51 The strongest associations observed in the present study were for natural causes of death related to coronary 52 53 heart disease, stroke, respiratory disease, lung cancer, and stomach cancer. Smoking is a modifiable risk 54 55 56 factor in each of these diseases, and higher premorbid intelligence has been related to lower likelihood of 57 58 current smoking (67) or past regular smoking (68), and the increased likelihood of quitting smoking (68,69). 59 60 Indeed, in a recent comparison of effect sizes for intelligence and causespecific mortalities in a Danish male https://mc.manuscriptcentral.com/bmj Page 21 of 60 BMJ

population, the authors made a similar summary of their own findings by implicating smoking as an 1 2 important mediator in the context of natural causes of death (7). Nevertheless, when tested, smoking is no 3 4 5 more than a partial mediator in associations between premorbid intelligence and total mortality (3), coronary 6 7 heart diseaserelated mortality (63), and incident coronary heart disease or stroke (64). In our replication 8 Confidential: For Review Only 9 study, the addition of adult smoking showed small attenuating effects (12 to 27%) on the associations 10 11 between midlife intelligence test performance and most causespecific mortalities, including coronary artery 12 13 disease and respiratory disease, and, had moderate attenuation effects on the associations with smoking 14 15 16 related cancers and lung cancer (40 and 50% respectively). Educational attainment or occupational status 17 18 indicators in adulthood have generally shown stronger attenuating effects (1). Whereas education is likely to 19 20 fall on the causal pathway, after intelligence and before death, given evidence that intelligence differences 21 22 predict later national school examination outcomes (70), education may alternatively be a proxy indicator of 23 24 25 intelligence, and yet have no direct causal association with longer life. This is an ongoing discussion about 26 27 possible overadjustment by education in cognitive epidemiology (71). It is likely that the influence of a 28 29 health behaviour (or pattern of behaviours) in explaining associations between premorbid function and 30 31 causespecific mortality will be more precisely measured by future studies with repeat assessment across the 32 33 life course, capturing cumulative risk. Indeed, followup of the Whitehall II study showed that four repeat 34 35 36 measures of lifestyle behaviours accounted for far more of the socialmortality gradient than baseline 37 38 measurement alone (72), albeit this difference was evident for specific behaviours (especially physical 39 40 activity), and not all (e.g., smoking). Alcohol consumption may have accounted for some of the association 41 42 we observed between premorbid intelligence and digestiverelated mortality, and this could be similarly 43 44 45 assessed. Alcoholrelated deaths were associated with premorbid intelligence in a 37 year follow up of the 46 47 19691970 Swedish Conscripts cohort (73). Childhood intelligence test performance was most strongly 48 49 associated with respiratory disease deaths. This might be attributable to a combination of smoking 50 51 prevalence and occupationalrelated exposures. For many in this birth cohort, those in the working classes 52 53 spent their early careers in mining and shipbuilding industries (men) or as cleaners and factory workers 54 55 56 (women). Therefore, it is possible that, with the decline of smoking rates and toxic exposures in the 57 58 workplace, the magnitude of IQdeath associations may be lower in more recent birth cohorts. Nevertheless, 59 60 when we adjusted for background social class in our subgroup analyses, and controlled for smoking and https://mc.manuscriptcentral.com/bmj BMJ Page 22 of 60

adult occupational status in our replication study, the risk of respiratorydisease deaths in relation to 1 2 intelligence scores was only partially attenuated (by 12% and 33% respectively). 3 4 5 6 7 Whereas the research literature has taken advantage of oftenrich lifestyle data to explore mediating 8 Confidential: For Review Only 9 effects on a pathway from premorbid intelligence to mortality risk, recent evidence from longitudinal twin 10 11 data suggests the association may be largely due to common genetic effects (19). With newly emerging data 12 13 from genomewide association studies (GWAS), there is likely to be a sizeable shift towards the testing of 14 15 16 potential genetic factors, alongside environmental factors, in cognitive epidemiology. Recent GWAS data 17 18 used in the context of the large UK Biobank sample has provided evidence for pleiotropy between midlife 19 20 (premorbid) cognitive performance and genetic variants of specific diseases, including coronary heart 21 22 disease, ischaemic stroke and Alzheimers’ disease (74). Whether this is evidence for biological pleiotropy 23 24 25 (thus supporting bodily system integrity theory), or, a causal pathway from genetic variant to disease 26 27 outcome, mediated by cognitive ability and subsequent healthrisk behaviours and/or occupational hazards, 28 29 is the subject of ongoing work using Mendelian Randomisation methods (75). There is the additional 30 31 prospect of interaction between healthrisk behaviours and genetic markers for disease, in explaining health 32 33 differentials attributed to intelligence variation, and none of these possibilities are mutually exclusive. 34 35 36 37 38 Due to the widely different causes of death in the present study, there is no assumption of a onesize 39 40 fitsall explanation of the observed effects. The association between premorbid intelligence and specific 41 42 external causes of death warrants in depth and separate coverage. The specific association between 43 44 45 premorbid intelligence and suicide risk in men, for example, may need inclusion in risk models alongside 46 47 other psychosocial factors, such as the modifying effects of psychosis (66) and highachieving parents (46), 48 49 and the potentially confounding and mediating effects of depression and anxiety disorders (76) and anti 50 51 social personality (13,15). Furthermore, suicide risk may be associated with premorbid intelligence via 52 53 reduced opportunities in the job market, putting a greater strain on men than women in our society. Other 54 55 56 causespecific mortalities, however, may be usefully studied together in their association with premorbid 57 58 intelligence. For example, the underlying mechanisms in the association between premorbid intelligence and 59 60 dementia mortality risk may have commonality with those of coronary heart disease and stroke. In support https://mc.manuscriptcentral.com/bmj Page 23 of 60 BMJ

of this, previous evidence indicates that the effect on total dementia is driven by associations with vascular 1 2 dementia, and not Alzheimer’s disease (54). Furthermore, that dementia mortality risk relates to premorbid 3 4 5 intelligence for lateonset but not earlyonset cases (55)─which tend to have a higher genetic 6 7 component─might implicate health behaviours and lifestyle factors. 8 Confidential: For Review Only 9 10 11 At last, we have been able to report on the strength of association between premorbid intelligence 12 13 and a range of specific causes of death in a full population, and with followup from childhood to near the 14 15 16 end of the life course. The field of cognitive epidemiology remains in a growth period; therefore, it is 17 18 premature to make recommendations to practitioners and policymakers, though we have made a start on that 19 20 elsewhere (18). Although we report that smoking and socioeconomic status are unlikely to fully mediate the 21 22 observed associations, future studies would benefit from measures of the cumulative load of such risk 23 24 25 factors over the life course. Finally, we highlight to researchers active in this field our robust findings of 26 27 premorbid intelligence in association with lifetime risk of mortality by coronary heart disease, stroke, 28 29 smokingrelated cancers, respiratory disease, digestiverelated disease, external causes, and, dementia, 30 31 specifically the relatively strong lifetime associations with smokingrelated diseases. We also emphasise the 32 33 marginally greater associated risks among women for coronary heart disease, smokingrelated cancers, 34 35 36 respiratory disease, and dementia, which might indicate sex differential effects of intelligence on modifiable 37 38 risk factors of disease. 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/bmj BMJ Page 24 of 60

FUNDING 1 2 3 4 This work was supported by a UK cross council Lifelong Health and Wellbeing Initiative [grant number 5 6 MRCG100140/1], which included a research fellowship for CMC and IC and research associate position for 7 8 Confidential: For Review Only 9 CEB, and for which IJD was the principal investigator. CMC, CB, GDB, IC and IJD are members of The 10 11 University of Edinburgh Centre for Cognitive Ageing and Cognitive Epidemiology, part of the cross council 12 13 Lifelong Health and Wellbeing Initiative [grant number MR/K026992/1]. Funding from the Biotechnology 14 15 and Biological Sciences Research Council (BBSRC) and Medical Research Council (MRC) is gratefully 16 17 18 acknowledged. 19 20 21 22 ACKNOWLEDGEMENTS 23 24 25 26 The authors wish to thank the NHS Central Register (NHSCR) for linkage to Scottish death records and the 27 28 29 Health and Social Care Information Centre (HSCIC) for linkage to Office of National Statistics (ONS) death 30 31 records from England and Wales. We are also indebted to Alisdair Tullo from University of Edinburgh for 32 33 custombuilding a program to match mortality cases without ID numbers to individuals in the Scottish 34 35 Mental Survey 1947 and to Paul Redmond from the University of Edinburgh for receiving the linked data. 36 37 38 39 40 COPYRIGHT 41 42 43 44 The Corresponding Author has the right to grant on behalf of all authors and does grant on behalf of all 45 46 authors, a worldwide licence to the Publishers and its licensees in perpetuity, in all forms, formats and media 47 48 49 (whether known now or created in the future), to i) publish, reproduce, distribute, display and store the 50 51 Contribution, ii) translate the Contribution into other languages, create adaptations, reprints, include within 52 53 collections and create summaries, extracts and/or, abstracts of the Contribution, iii) create any other 54 55 derivative work(s) based on the Contribution, iv) to exploit all subsidiary rights in the Contribution, v) the 56 57 58 inclusion of electronic links from the Contribution to third party material whereever it may be located; and, 59 60 vi) licence any third party to do any or all of the above. https://mc.manuscriptcentral.com/bmj Page 25 of 60 BMJ

1 2 DECLARATION OF COMPETING INTERESTS 3 4 5 6 7 All authors have completed the Unified Competing Interest form at www.icmje.org/coi_disclosure.pdf 8 Confidential: For Review Only 9 (available on request from the corresponding author) and declare: that IJD was the recipient of a Medical 10 11 Research Council grant for staff and consumables to complete the work, from which salaries were paid for 12 13 CMC, CEB, and IC; no other financial relationships with any organisations that might have an interest in the 14 15 16 submitted work in the previous three years; no other relationships or activities that could appear to have 17 18 influenced the submitted work. 19 20 21 22 TRANSPARENCY DECLARATION 23 24 25 26 27 The lead author affirms that this manuscript is an honest, accurate, and transparent account of the study 28 29 being reported; that no important aspects of the study have been omitted; and that any discrepancies from 30 31 the study as planned (and, if relevant, registered) have been explained. 32 33 34 35 36 ETHICAL APPROVAL 37 38 39 40 Ethical approval was not required. 41 42 43 44 45 DATA SHARING STATEMENT 46 47 48 49 The full dataset is available for ten years from the beginning of the study with permission of IJD at the 50 51 University of Edinburgh 52 53 54 55 56 CONTRIBUTORS 57 58 59 60 https://mc.manuscriptcentral.com/bmj BMJ Page 26 of 60

CC contributed to the design of the study, acquired the data, cleaned and analysed the data, interpreted the 1 2 data, and drafted and revised the paper. She is guarantor. GDB contributed to the design of the study, 3 4 5 interpreted the data, and critically revised the paper. CB managed the acquisition of data, interpreted the 6 7 data, and critically revised the paper. AP and AT acquired the data and critically revised the paper. IC 8 Confidential: For Review Only 9 acquired the data, interpreted the data, and critically revised the paper. GD proposed and analysed the 10 11 replication sample and critically revised the paper. IJD initiated the project, conceived and designed the 12 13 study, interpreted the data, and critically revised the manuscript. 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/bmj Page 27 of 60 BMJ

References 1 2 1. Calvin CM, Deary IJ, Fenton C, Roberts BA, Der G, Leckenby N, et al. Intelligence in youth and all 3 4 5 causemortality: Systematic review with metaanalysis. Int J Epidemiol. 2011;40(3):626–44. 6 7 2. Cukic I, Brett CE, Calvin CM, Batty GD, Deary IJ. Childhood IQ is associated with allcause 8 Confidential: For Review Only 9 mortality in a full yearofbirth cohort: 68year followup of 66 616 members of the Scottish Mental 10 11 Survey 1947. Submiss. 12 13 3. Batty GD, Wennerstad KM, Smith GD, Gunnell D, Deary IJ, Tynelius P, et al. IQ in Early Adulthood 14 15 16 and Mortality By Middle Age. Epidemiology. 2009;20(1):100–9. 17 18 4. Lawlor DA, David Batty G, Clark H, McIntyre S, Leon DA. Association of childhood intelligence 19 20 with risk of coronary heart disease and stroke: Findings from the Aberdeen Children of the 1950s 21 22 cohort study. Eur J Epidemiol. 2008;23(10):695–706. 23 24 25 5. Leon DA, Lawlor DA, Clark H, Batty GD, Macintyre S. The association of childhood intelligence 26 27 with mortality risk from adolescence to middle age: Findings from the Aberdeen Children of the 28 29 1950s cohort study. Intelligence. 2009;37(6):520–8. 30 31 6. Batty GD, Shipley MJ, Mortensen LH, Gale CR, Deary IJ. IQ in late adolescence/early adulthood, 32 33 risk factors in middleage and later coronary heart disease mortality in men: the Vietnam Experience 34 35 36 Study. Eur J Cardiovasc Prev Rehabil. 2008;15(3):359–61. 37 38 7. Christensen GT, Mortensen EL, Christensen K, Osler M. Intelligence in young adulthood and cause 39 40 specific mortality in the Danish Conscription Database – A cohort study of 728,160 men. 41 42 Intelligence2016 Nov;59:64–71. 43 44 45 8. Hart CL, Taylor MD, Whalley LJ, Starr JM, Hole DJ, Wilson V, et al. Childhoold IQ, Social Class, 46 47 Deprivation and their Relationship with Mortality and Morbidity Risk in Later Life: prospective 48 49 observational study linking the Scottish Mental Survey 1932 and the Midspan studies. Psychosom 50 51 Med2003;65:877–83. 52 53 9. Hart CL, Taylor MD, Smith GD, Whalley LJ, Starr JM, Hole DJ, et al. Childhood IQ and 54 55 56 cardiovascular disease in adulthood: Prospective observational study linking the Scottish Mental 57 58 Survey 1932 and the Midspan studies. Soc Sci Med. 2004;59(10 SPEC. ISS.):2131–8. 59 60 10. Hemmingsson T, Melin B, Allebeck P, Lundberg I. The association between cognitive ability https://mc.manuscriptcentral.com/bmj BMJ Page 28 of 60

measured at ages 1820 and mortality during 30 years of followup A prospective observational 1 2 study among Swedish males born 194951. Int J Epidemiol. 2006;35(3):665–70. 3 4 5 11. O’Toole BI, Cantor C. Suicide risk factors among Australian Vietnam era draftees. Suicide Life 6 7 Threat Behav. 1995;25(4):475–88. 8 Confidential: For Review Only 9 12. Osler M, Nybo Andersen A, Nordentoft M. Impaired childhood development and suicidal behaviour 10 11 in a cohort of Danish men born in 1953. J Epidemiol Community Health. 2008;62(1):23–8. 12 13 13. Allebeck P, Allgulander C, Fisher LD. Predictors of completed suicide in a cohort of 50,465 young 14 15 16 men: role of personality and deviant behaviour. BMJ. 1988;297(6642):176–8. 17 18 14. Batty GD, Whitley E, Deary IJ, Gale CR, Tynelius P, Rasmussen F. Psychosis alters association 19 20 between IQ and future risk of attempted suicide: cohort study of 1,109,475 Swedish men. BMJ. 21 22 2010;340:c2506. 23 24 25 15. Sörberg A, Allebeck P, Melin B, Gunnell D, Hemmingsson T. Cognitive ability in early adulthood is 26 27 associated with later suicide and suicide attempt: the role of risk factors over the life course. Psychol 28 29 Med. 2012;1–12. 30 31 16. Osler M, Nybo Andersen AM, Laursen B, Lawlor DA. Cognitive function in childhood and early 32 33 adulthood and injuries later in life: The Metropolit 1953 male birth cohort. Int J Epidemiol. 34 35 36 2007;36(1):212–9. 37 38 17. Batty GD, Gale CR, Tynelius P, Deary IJ, Rasmussen F. IQ in early adulthood, socioeconomic 39 40 position, and unintentional injury mortality by middle age: A cohort study of more than 1 million 41 42 Swedish men. Am J Epidemiol. 2009;169(5):606–15. 43 44 45 18. Deary IJ, Weiss a., Batty GD. Intelligence and Personality as Predictors of Illness and Death: How 46 47 Researchers in Differential Psychology and Chronic Disease Epidemiology Are Collaborating to 48 49 Understand and Address Health Inequalities. Psychol Sci Public Interes. 2010;11(2):53–79. 50 51 19. Arden R, Luciano M, Deary IJ, Reynolds CA, Pedersen NL, Plassman BL, et al. The association 52 53 between intelligence and lifespan is mostly genetic. Int J Epidemiol. 2016;45(1):178–85. 54 55 56 20. Deary IJ. Looking for “system integrity” in cognitive epidemiology. Gerontology. 2012;58(6):545– 57 58 53. 59 60 21. Lubinski D, Humphreys LG. Some bodily and medical correlates of mathematical giftedness and https://mc.manuscriptcentral.com/bmj Page 29 of 60 BMJ

commensurate levels of socioeconomic status. Intelligence. 1992;16(1):99–115. 1 2 22. Lubinski D. Cognitive epidemiology: With emphasis on untangling cognitive ability and 3 4 5 socioeconomic status. Intelligence. 2009;37(6):625–33. 6 7 23. Russ TC, Hannah J, Batty GD, Booth CC, Deary IJ, Starr JM. Childhood Cognitive Ability and 8 Confidential: For Review Only 9 Incident Dementia. Epidemiology. 2017;28(3):361–4. 10 11 24. Ian J Deary, LJ Whalley JMS. IQ at age 11 and longevity. In: CE Finch, JM Robine YC, editor. Brain 12 13 and Longevity: Perspectives in Longevity. Berlin: Springer; 2003. p. 153–64. 14 15 16 25. Ian J Deary, Lawrence J Whalley JMS. A lifetime of intelligence: followup studies of the Scottish 17 18 Mental Surveys of 1932 and 1947. Washington DC: American Psychological Association; 2009. 19 20 26. Batty GD, Mortensen LH, Gale CR, Shipley MJ, Roberts BA, Deary IJ. IQ in late adolescence/early 21 22 adulthood, risk factors in middle age, and later cancer mortality in men: The Vietnam experience 23 24 25 study. Psychooncology. 2009;18(10):1122–6. 26 27 27. Batty GD, Wennerstad KM, Davey Smith G, Gunnell D, Deary IJ, Tynelius P, et al. IQ in early 28 29 adulthood and later cancer risk: Cohort study of one million Swedish men. Ann Oncol. 30 31 2007;18(1):21–8. 32 33 28. Deary IJ, Whiteman MC, Starr JM, Whalley LJ, Fox HC. The impact of childhood intelligence on 34 35 36 later life: Following up the Scottish Mental Surveys of 1932 and 1947. J Pers Soc Psychol. 2004 37 38 Jan;86(1):130–47. 39 40 29. Maxwell J. The level and trend of national intelligence: the contribution of the Scottish mental 41 42 surveys (no. 48). London: University of London Press; 1961. 43 44 45 30. Brett CE. Realising health data linkage from a researcher’s perspective: following up the 6Day 46 47 Sample of the Scottish Mental Survey 1947. Longit Life Course Stud. 2014;5(3):283–98. 48 49 31. Thomson GH. The trend of Scottish intelligence. London: University of London Press; 1949. 50 51 32. Deary IJ, Whiteman MC, Starr JM, Whalley LJ, Fox HC. The impact of childhood intelligence on 52 53 later life: Following up the Scottish Mental Surveys of 1932 and 1947. J Pers Soc Psychol. 54 55 56 2004;86(1):130–47. 57 58 33. Deary IJ, Whalley LJ, Lemmon H, Crawford JR, Starr JM. The stability of individual differences in 59 60 mental ability from childhood to old age: Followup of the 1932 Scottish mental survey. Intelligence. https://mc.manuscriptcentral.com/bmj BMJ Page 30 of 60

2000;28(1):49–55. 1 2 34. Gow AJ, Johnson W, Pattie A, Brett CE, Roberts B, Starr JM, et al. Stability and change in 3 4 5 intelligence from age 11 to ages 70, 79, and 87: the Lothian Birth Cohorts of 1921 and 1936. Psychol 6 7 Aging. 2011;26(1):232–40. 8 Confidential: For Review Only 9 35. Deary IJ, Gow AJ, Pattie A, Starr JM. Cohort profile: The lothian birth cohorts of 1921 and 1936. Int 10 11 J Epidemiol. 2012;41(6):1576–84. 12 13 36. Deary ID, Whalley LJ, Starr JM. The Scottish Mental Surveys of 1932 and 1947. Washington DC: 14 15 16 American Psychological Association; 2009. 17 18 37. Gunnell D, Hawton K, Kapur N. Coroners’ verdicts and suicide statistics in England and Wales. BMJ 19 20 [Internet]. 2011;343:d6030. Available from: http://www.ncbi.nlm.nih.gov/pubmed/21980064 21 22 38. Redelings MD, Sorvillo F, Simon P. A comparison of underlying cause and multiple causes of death: 23 24 25 US vital statistics, 20002001. Epidemiology. 2006;17(1):100–3. 26 27 39. R Core Team. R: A language and environment for statistical computing. Vienna, Austria: R 28 29 Foundation for Statistical Computing; 2013. 30 31 40. Benjamini Y, Hochberg Y. Controlling the False Discovery Rate : A Practical and Powerful Approach 32 33 to Multiple Testing Author ( s ): Yoav Benjamini and Yosef Hochberg Source : Journal of the Royal 34 35 36 Statistical Society . Series B ( Methodological ), Vol . 57 , No . 1 Published by : J R Stat Soc. 37 38 1995;57(1):289–300. 39 40 41. Dundas R, Leyland AH, Macintyre S, Leon DA. Does the primary school attended influence self 41 42 reported health or its risk factors in later life? Aberdeen children of the 1950s Study. Int J Epidemiol. 43 44 45 2006;35(2):458–65. 46 47 42. Batty GD, Mortensen EL, Nybo Andersen AM, Osler M. Childhood intelligence in relation to adult 48 49 coronary heart disease and stroke risk: evidence from a Danish birth cohort study. Paediatr Perinat 50 51 Epidemiol. 2005 Nov;19(6):452–9. 52 53 43. Kajantie E, Räikkönen K, Henriksson M, Leskinen JT, Forsén T, Heinonen K, et al. Stroke Is 54 55 56 Predicted by Low Visuospatial in Relation to Other Intellectual Abilities and Coronary Heart Disease 57 58 by Low General Intelligence. PLoS One. 2012;7(11). 59 60 44. Modig Wennerstad K, Silventoinen K, Tynelius P, Bergman L, Rasmussen F. Association between https://mc.manuscriptcentral.com/bmj Page 31 of 60 BMJ

intelligence and typespecific stroke: a populationbased cohort study of early fatal and nonfatal 1 2 stroke in one million Swedish men. J Epidemiol Community Health. 2010;64:908–12. 3 4 5 45. Hodal Meincke R, Lykke Mortensen E, Avlund K, Rosthøj S, Jelling Sørensen H, Osler M. 6 7 Intelligence in early adulthood and mortality from natural and unnatural causes in middleaged 8 Confidential: For Review Only 9 Danish men. J Epidemiol Community Heal. 2014;68(2):130–6. 10 11 46. Gunnell D, Magnusson PKE, Rasmussen F. Low intelligence test scores in 18 year old men and risk 12 13 of suicide: cohort study. BMJ. 2005;330(7484):167. 14 15 16 47. Deary IJ, Batty GD, Pattie A, Gale CR. More Intelligent , More Dependable Children Live Longer. 17 18 2008;19(9):874–81. 19 20 48. Weiss A, Gale CR, Batty GD, Deary IJ. Emotionally Stable, Intelligent Men Live Longer: The 21 22 Vietnam Experience Study Cohort. Psychosom Med. 2009;71(4):385–94. 23 24 25 49. Thomas BM, Starr JM, Whalley LJ. Death certification in treated cases of presenile Alzheimer’s 26 27 disease and vascular dementia in Scotland. Age Ageing. 1997;26(5):401–6. 28 29 50. Corwin LE, Wolf PA, Kannel WB, McNamara PM. Accuracy of death certification of stroke: the 30 31 Framingham Study. Stroke. 1982;13:818–21. 32 33 51. WERSÄLL JP, EKLUND G. The decline of smoking among Swedish men. Int J Epidemiol. 1998 34 35 36 [cited 2016 Nov 23];27(1):20–6. 37 38 52. Rodu B, Stegmayr B, Nasic S, Cole P, Asplund K. Evolving patterns of tobacco use in northern 39 40 Sweden. J Intern Med. 2003;253(6):660–5. 41 42 53. Taulbut M, Gordon D, McKenzie K. Tobacco smoking in Scotland: an epidemiology briefing 43 44 45 [Internet]. 2008. Available from: http://www.scotpho.org.uk/publications/reportsandpapers/493 46 47 tobaccosmokinginscotlandanepidemiologybriefing 48 49 54. McGurn B, Deary IJ, Starr JM. Childhood cognitive ability and risk of lateonset Alzheimer and 50 51 vascular dementia. Neurology. 2008;71(14):1051–6. 52 53 55. Whalley LJ, Starr JM, Athawes R, Hunter D, Pattie a, Deary IJ. Childhood mental ability and 54 55 56 dementia. Neurology. 2000;55:1455–9. 57 58 56. Russ TC, Stamatakis E, Hamer M, Starr JM, Kivimäki M, Batty GD. Socioeconomic status as a risk 59 60 factor for dementia death: Individual participant metaanalysis of 86 508 men and women from the https://mc.manuscriptcentral.com/bmj BMJ Page 32 of 60

UK. Br J Psychiatry. 2013;203(1):10–7. 1 2 57. Der G, Batty GD, Deary IJ. The association between IQ in adolescence and a range of health 3 4 5 outcomes at 40 in the 1979 US National Longitudinal Study of Youth. Intelligence. 2009;37(6):573– 6 7 80. 8 Confidential: For Review Only 9 58. Wraw C, Deary IJ, Gale CR, Der G. Intelligence in youth and health at age 50. Intelligence. 10 11 2015;53:23–32. 12 13 59. Deary IJ, Whalley LJ, Batty GD, Starr JM. Physical fitness and lifetime cognitive change. Neurology. 14 15 16 2006;67(7):1195–200. 17 18 60. Carroll D, Batty GD, Mortensen LH, Deary IJ, Phillips AC. Low cognitive ability in early adulthood 19 20 is associated with reduced lung function in middle age: the Vietnam experience study. Thorax. 21 22 2011;66(10):884–8. 23 24 25 61. Richards M, Strachan D, Hardy R, Kuh D, Wadsworth M. Lung function and cognitive ability in a 26 27 longitudinal birth cohort study. Psychosom Med. 2005;67:602–8. 28 29 62. Vasilopoulos T, Kremen WS, Grant MD, Panizzon MS, Xian H, Toomey R, et al. Individual 30 31 differences in cognitive ability at age 20 predict pulmonary function 35 years later. J Epidemiol 32 33 Community Heal. 2015;69(3):261–5. 34 35 36 63. Batty GD, Shipley MJ, Gale CR, Mortensen LH, Deary IJ. Does IQ predict total and cardiovascular 37 38 disease mortality as strongly as other risk factors? Comparison of effect estimates using the Vietnam 39 40 Experience Study. Heart. 2008;94(12):1541–4. 41 42 64. Hemmingsson T, Essen J v., Melin B, Allebeck P, Lundberg I. The association between cognitive 43 44 45 ability measured at ages 1820 and coronary heart disease in middle age among men: A prospective 46 47 study using the Swedish 1969 conscription cohort. Soc Sci Med. 2007;65(7):1410–9. 48 49 65. Jokela M, Batty GD, Deary IJ, Silventoinen K, Kivimäki M. Sibling Analysis of Adolescent 50 51 Intelligence and Chronic Diseases in Older Adulthood. Ann Epidemiol. 2011;21(7):489–96. 52 53 66. Andersson L, Allebeck P, Gustafsson JE, Gunnell D. Association of IQ scores and school 54 55 56 achievement with suicide in a 40year followup of a Swedish cohort. Acta Psychiatr Scand. 57 58 2008;118(2):99–105. 59 60 67. Batty GD, Deary IJ, Schoon I, Gale CR. Mental ability across childhood in relation to risk factors for https://mc.manuscriptcentral.com/bmj Page 33 of 60 BMJ

premature mortality in adult life: the 1970 British Cohort Study. J Epidemiol Community Heal. 2007 1 2 Nov 1;61(11):997–1003. 3 4 5 68. Batty GD, Deary IJ, Macintyre S. Childhood IQ in relation to risk factors for premature mortality in 6 7 middleaged persons: the Aberdeen Children of the 1950s study. J Epidemiol Community Heal. 2007 8 Confidential: For Review Only 9 Mar 1;61(3):241–7. 10 11 69. Taylor MD, Hart CL, Davey Smith G, Starr JM, Hole DJ, Whalley LJ, et al. Childhood mental ability 12 13 and smoking cessation in adulthood: prospective observational study linking the Scottish Mental 14 15 16 Survey 1932 and the Midspan studies. J Epidemiol Community Heal. 2003 Jun 1;57(6):464–5. 17 18 70. Deary IJ, Strand S, Smith P, Fernandes C. Intelligence and educational achievement. Intelligence. 19 20 2007;35(1):13–21. 21 22 71. Deary IJ, Johnson W. Intelligence and education: Causal perceptions drive analytic processes and 23 24 25 therefore conclusions. Int J Epidemiol. 2010;39(5):1362–9. 26 27 72. Stringhini S, Sabia S, Shipley M, Brunner E, Nabi H, Kivimaki M, et al. Association of 28 29 socioeconomic position with health behaviors and mortality. Jama. 2010;303(12):1159–66. 30 31 73. Sjölund S, Allebeck P, Hemmingsson T. (IQ) in adolescence and later risk of 32 33 alcoholrelated hospital admissions and deaths37year followup of Swedish conscripts. Addiction. 34 35 36 2012;107(1):89–97. 37 38 74. Hagenaars SP, Harris SE, Davies G, Hill WD, Liewald DCM, Ritchie SJ, et al. Shared genetic 39 40 aetiology between cognitive functions and physical and mental health in UK Biobank (N=112 151) 41 42 and 24 GWAS consortia. Mol Psychiatry. 2016;(October 2015):31120. 43 44 45 75. Hagenaars SP, Gale CR, Deary IJ, Harris SE. Cognitive ability and physical health: a Mendelian 46 47 randomization study. Sci Report (in press). 48 49 76. Gale CR, Deary IJ, Boyle SH, Barefoot J, Mortensen LH, Batty GD. Cognitive ability in early 50 51 adulthood and risk of 5 specific psychiatric disorders in middle age: the Vietnam experience study. 52 53 Arch Gen Psychiatry. 2008;65(12):1410–8. 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/bmj BMJ Page 34 of 60

1 2 3 4 5 6 7 8 Confidential: For Review Only 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 Figure 1  : Derivation of the Scottish Mental Survey 1947 analytic sample. SMS1947=Scottish Mental   a 45 Survey of 1947. The majority with missing dates were those who had migrated or joined the Armed Forces for whom records contained no final posting date. bThe analytic sample represents 92% of the 46 SMS1947. c46.0% of those with vital status were deceased, which is approximate to Scottish population 47 statistics for 75-80 year olds. dIncludes: embarkees (n=5381; 2857 men); those cancelled from GP 48 registration (n=3391; 1383 men); Armed Forces recruits (n=454; 411 men); residents of Northern Ireland 49 or Isle of Man (n=96; 52 men)  50 51 162x180mm (300 x 300 DPI) 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/bmj Page 35 of 60 BMJ

1 2 3 4 5 6 7 8 Confidential: For Review Only 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 Figure 2  : Association between intelligence (Moray House Test score) at age 11 years and major causes 45 of death (age and sex-adjusted hazard ratios and 95% confidence intervals) to age 79 years, in the Scottish 46 Mental Survey 1947 47 48 166x186mm (300 x 300 DPI) 49 50 51 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/bmj BMJ Page 36 of 60

1 2 3 4 5 6 7 8 Confidential: For Review Only 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 Figure 3   : Hazard ratios (95% confidence intervals) for the association between a one SD higher score in intelligence test score at age 11 years and cause of death to age 79, N=65 765, in the Scottish Mental 31 Survey 1947. Effect sizes are adjusted for sex and age at cognitive testing 32 33 136x91mm (300 x 300 DPI) 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/bmj Page 37 of 60 BMJ

1 2 3 4 5 6 7 8 Confidential: For Review Only 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 Figure 4  : Association between intelligence (Moray House Test score) at age 11 years and cancer-specific deaths (age and school-adjusted hazard ratios and 95% confidence intervals) to age 79 years, in the 48 Scottish Mental Survey 1947 49 50 167x223mm (300 x 300 DPI) 51 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/bmj BMJ Page 38 of 60

1 2 3 4 5 6 7 8 Confidential: For Review Only 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 Figure 5: Hazard ratios (95% confidence intervals) for the association between a one SD higher score in 42 intelligence at age 11 years test score and cause of death by cancer type to age 79, in the Scottish Mental Survey 1947. Effect sizes are adjusted for age at intelligence testing and sex—with the exceptions of ovarian 43 and breast cancer (women only), and, prostate cancer (men only). 44 *Indicates non-smoking-related cancers (all others are smoking-related) 45 46 129x131mm (300 x 300 DPI) 47 48 49 50 51 52 53 54 55 56 57 58 59 60 https://mc.manuscriptcentral.com/bmj Page 39 of 60 BMJ

Supplementary Materials for: 1 2 3 4 5 Calvin et al. Childhood intelligence in relation to major causes of death: 68 years of followup of Scotland’s 6 7 whole 1936born population 8 Confidential: For Review Only 9 10 11 Contents 12 13 Pages 3-5: Matching death records to individuals in Scottish Mental Survey 1947 database 14 15 16 Page 6: Assessing selection bias due to missing MHT scores using the 6Day Sample, a representative 17 18 subsample of the Scottish Mental Survey 1947 19 20 Page 7: Supplementary Table 1, on International Classification of Diseases for major causes of death 21 22 Pages 8-10 : The 30Day Sample: A representative subsample of the Scottish Mental Survey 1947 23 24 Pages 11-12 : The West of Scotland Twenty07 study: A replication sample 25 26 Page 13: Supplementary Figure 1, on associations between intelligence and underlying causes of death in 27 28 29 the Scottish Mental Survey 1947 30 31 Page 14: Supplementary Figure 2, on associations between intelligence and cancer subtypes as underlying 32 33 causes of death in the Scottish Mental Survey 1947 34 35 Page 15: Supplementary Table 2, on sex differences in associations between intelligence and major causes 36 37 38 of death in the Scottish Mental Survey 1947 39 40 Page 16: Supplementary Table 3, on sex differences in associations between intelligence and death from 41 42 specific cancers in the Scottish Mental Survey 1947 43 44 Page 17: Supplementary Figure 3, on associations between intelligence and causes of death in the Scottish 45 46 Mental Survey 1947, adjusting for school 47 48 49 Page 18: Supplementary Figure 4, on associations between intelligence and cancer subtypes as causes of 50 51 death in the Scottish Mental Survey 1947, adjusting for school 52 53 Page 19: Supplementary Table 4, on confounderadjusted associations between intelligence and leading 54 55 causes of death in the 30Day subsample of the Scottish Mental Survey 1947 56 57 58 59 60 1 https://mc.manuscriptcentral.com/bmj BMJ Page 40 of 60

Page 20 : Supplementary Figure 5, a cohort comparison of associations between intelligence and underlying 1 2 causes of death: the Scottish Mental Survey 1947 and the West of Scotland Twenty07 study Page 21 : 3 4 5 Supplementary Table 5, on smoking and adult socioeconomic statusadjusted associations between 6 7 intelligence and underlying causes of death in the West of Scotland Twenty07 study 8 Confidential: For Review Only 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 2 https://mc.manuscriptcentral.com/bmj Page 41 of 60 BMJ

Matching death records to individuals in Scottish Mental Survey 1947 1 2 3 4 1. Matching death records without ID numbers to individuals in the Scottish Mental Survey 1947 database 5 6 7 8 Confidential: For Review Only 9 Scotlandrecorded deaths between 1 July 2015 and 31 December 2015 were provided to the study without 10 11 unique ID numbers for individuals. Therefore an electronic matching program was run to amalgamate 12 13 individuals’ death data with their respective childhood data. The program used key variables (first name, 14 15 surname, date of birth) to match individuals in the two datasets, utilising Python programming language 16 17 18 (https://www.python.org/), the NumPy library for scientific computing (http://www.numpy.org/), and, the 19 20 pandas library for data analysis (http://pandas.pydata.org/). In attempting to match 2453 individuals’ death 21 22 data with their childhood records, the program successfully matched 1939 individuals, whereas 514 failed, 23 24 due largely to first name variants between the two files. Of the failed matches, 315 were resolved manually, 25 26 27 leaving 199 nonmatches that were excluded from analyses. 28 29 30 31 2. Reconciling mismatches on sex and date of birth between individuals’ records in the mortality death and 32 33 the Scottish Mental Survey 1947 data 34 35 36 37 38 Following the merging of files, there were numbers of mismatches on sex and date of birth identified 39 40 between the historical records and the linkage data. There were 296 mismatches on sex and 3114 41 42 mismatches on date on birth. Manual checks on identifiable data were carried out to reconcile these 43 44 differences (details below) resulting in 987 exclusions due to mismatching (178 of these also met other 45 46 47 exclusion criteria). 48 49 50 51 Date of birth mismatches There were 3114 mismatches on date of birth (4.3% of the traced sample), 52 53 which is a similar proportion of mismatches to a previous linkage study in Scotland (Hart et al., Public 54 55 56 Health, 2003). The table below summarises the differences in dates of birth between the two records. Firstly, 57 58 we accepted mismatches as correct matches if there was a difference of 9 days or less in the dates, or, if 59 60 3 https://mc.manuscriptcentral.com/bmj BMJ Page 42 of 60

there was a difference of 10 days with only one digit different (n=1941). Secondly, we conducted manual 1 2 checks on samples of mismatches, and confirmed a match if at least one of the following criteria were met: 3 4 5 (1) where the name that matched was unusual; (2) where there was an unusual middle name that matched, 6 7 (3) when there was a confirmed maiden name. This process of manual checking found correct matches for 8 Confidential: For Review Only 9 all mismatches where the different dates were exactly one month different, and, where day and month 10 11 appeared in reverse to the other, and so we accepted these as true matches. 12 13 14 15 16 Reconciling differences in dates of birth between historical and linkage records 17 18 Difference in date of birth N Action Result of manual checking Final decision 19 1 day out 752 Accept N/A Accept 20 21 2 days out 330 Accept N/A Accept 22 23 3 days out 166 Accept N/A Accept 24 4 days out 115 Accept N/A Accept 25 26 5 days out 107 Accept N/A Accept 27 28 6 days out 74 Accept N/A Accept 29 7 days out 52 Accept N/A Accept 30 31 8 days out 53 Accept N/A Accept 32 9 days out 36 Accept N/A Accept 33 34 10 days out (1 digit out ) 256 Accept N/A Accept 35 36 10 days out (2 digits out) 1 Check the 1 record Uncertain Exclude 37 11 -30 days out 310 Check 20 records 80% correct; 20% uncertain Exclude 38 39 1 month out 18 1 Check 20 records 95% correct; 5% uncertain Accept 40 41 Different month 339 Check 20 records 53% correct; 47% uncertain Exclude 42 Different month (1 digit out) 19 Check all 19 records 84% correct; 16% uncertain Exclude 43 44 Different year 162 Check 20 records 75% correct; 15% uncertain Exclude 45 2 digit s out 1 Check all 3 records 1 correct; 2 uncertain Exclude 46 47 Day and month in reverse 5 Check all 5 records 100% correct Accept 48 49 Very different year 3 Check all 3 records 1 correct; 2 uncertain Exclude 50 Different day and month 152 Check 20 records 60% correct; 40% uncertain Exclude 51 52 Total 3114 53 54 55 56 57 58 59 60 4 https://mc.manuscriptcentral.com/bmj Page 43 of 60 BMJ

Sex mismatches Participants’ names were first checked to ensure compatibility between records, and then 1 2 the forename helped to determine sex. Where there were name ambiguities on sex (n=9), AP, AT, CMC, and 3 4 5 IJD, reached a consensus. One match was excluded from analyses due to a mismatch on both forename and 6 7 date of birth. 8 Confidential: For Review Only 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 5 https://mc.manuscriptcentral.com/bmj BMJ Page 44 of 60

Sensitivity analyses: Assessing selection bias due to missing MHT scores using the 6-Day Sample, a 1 2 representative subsample of the Scottish Mental Survey 1947 3 4 5 6 7 We conducted sensitivity analyses on the 6Day Sample (n=1208), a much smaller subsample, which is 8 Confidential: For Review Only 9 nevertheless representative of the Scottish Mental Survey 1947 (SMS1947) in terms of geographical spread 10 11 and average test performance on the Moray House Test. 1 The 6Day Sample were the 1936born 12 13 schoolchildren who were born on the 1 st day of the evennumbered months. The 6Day Sample completed, 14 15 16 on a different day, a second, individuallyadministered intelligence test during the same year as the 17 18 SMS1947 (“TermanMerrill”, which was a Stanford University revision of the Binet scale). The Terman 19 20 Merrill is highly correlated with scores on the Moray House Test, for those who had data on both (r=.80, 21 22 n=1112). The 6Day subsample had a rate of missingness on the SMS1947 of 8.0% (n=96), similar to that 23 24 25 of the total sample (6.8%). 26 27 The hazard ratio (HR) and 95% confidence interval for allcause mortality in association with a one SD 28 29 higher TermanMerrill score in the complete 6Day Sample was 0.71 (95% CI = 0.65 to 0.78), where there 30 31 were 490 deaths out of 1208 participants. Further, we estimated the HR for allcause mortality in association 32 33 with a one SD higher TermanMerrill score for members of the 6Day Sample for whom we also had data on 34 35 36 the SMS1947 intelligence test (444 deaths out of 1112 participants), as: 0.71 (0.65 to 0.79). Finally, we 37 38 estimated the equivalent HR for those 6Day Sample members who had missing SMS1947 intelligence test 39 40 data (46 deaths out of 96 participants), and the magnitude of the allcause mortality risk estimate in 41 42 association with a one SD higher TermanMerrill score was also unchanged: 0.71 (0.54 to 0.93). This effect 43 44 45 size is also similar (albeit slightly stronger) to that which we observed for the association between the Moray 46 47 House intelligence test and allcause mortality in the full SMS1947. 48 49 Members of the 6Day Sample who were absentees of the SMS1947 (n=96) had on average a lower 50 51 TermanMerrill score (M=98.9, SD=23.7) compared to those who completed the SMS1947 (M=102.55, 52 53 SD=20.09)─Cohen’s d was 0.16. We conclude that an absence of lower ability scorers would have made 54 55 56 little difference to our main results. 57 58 1. Macpherson JS. ElevenYearOlds Grow Up. London: University of London Press; 1958. 59 60 6 https://mc.manuscriptcentral.com/bmj Page 45 of 60 BMJ

Supplementary Table 1 1 2 International Classification of Diseases, 6th , 7 th , 8th , 9th and 10 th Edition (ICD6, ICD7, ICD8, ICD9 and 3 ICD10) codes for major causes of death 4 5 Primary Diagnosis ICD-6 and ICD-7 Codes ICD-8 Codes ICD-9 Codes ICD-10 Codes 6 7 8 CardiovascularConfidential: Disease 330-334, 410-456 For 393-448 Review 390-448 Only I01-I78, G45 9 Coronary Heart Disease 420 410 -414 402, 410 -414, 429.2 I11, I20 --I25 10 11 Cerebrovascular Disease (Stroke) 330 -334 430 -43 8 430 -434, 436 -438 I60 -I69, G45, G46.3 -.7 12 All malignant cancers 140-209 140-209 140-172, 174-209 C00-C43, C45-C97 13 14 Lip, oral cavity and pharynx 140-148 140-149 140-149 C01, C12, C00-C14 15 Oesophagus 150 150 150 C15 16 17 Colorectal 153-154 153-154 153-154 C18-C21 18 Stomach 151 151 151 C16 19 Liver and bile duct 155-156 155-156 155 C22 20 21 Pancreas 157 157 157 C25 22 Lung, bronchus 162-163 162 162 C33-C34 23 24 Skin 190-191 172-173 172-173 C43-C44 25 Ovarian (females only) 175 183 183 C56 26 27 Breast (females only) 170 174 174 -175 C50 28 Prostate (males only) 177 185 185 C61 29 Kidney, renal pelvis and ureter 180 189 189 C64 -C66 30 31 Urinary bladder 181 188 188-188 C67 32 Brain and central nervous system 193 191-192 191-192 C71 33 34 Lymphoid, haematopoietic and related 200-205 200-207 200, 204-208 C81-C96 35 tissue 36 All smoking-related cancers 140-165, 171, 175, 180- 140-151, 153-155, 157, 140-151, 153-155, 157, C00-C16, C18-C22, C25, 37 181, 201,204 160-162, 180, 183, 188, 160-162, 180, 183, 189, C30-C32, C34, C53, C56, 38 189, 200, 204, 208 188, 200, 204-208 C64-C67, C81-C86, C91- 39 C95 40 All non -smoking -related cancers Remaining malignant Remaining malignant Remaining malignant Remaining malignant 41 cancer codes after cancer codes after cancer codes after cancer codes after 42 deducting smoking deducting smoking deducting smoking deducting smoking 43 cancer codes cancer codes cancer codes cancer codes 44 Respiratory Diseases 470-527 460-519 460-519 J00-J99 45 46 Digestive System Diseases 530-587 520-577 520-579 K20-K93 47 External Causes of Death E800 -E999 E800 -E999 E800 -E999 V01 -Y99 48 All accidents E800 -E936 E800 -E929 E800 -E929 V01 -X59 49 50 Intentional self-harm (suicide) E970-E979 E950-E959 E950-E959 X60-X84, Y87 51 Open verdict - E980-E989 E980-E989 Y10-Y34, Y87.2, Y89.9 52 53 54 Dementia 304-306 290, 293 290, 294, 331 F00-F01, F03, F09, G30, 55 G31 56 57 58 59 60 7 https://mc.manuscriptcentral.com/bmj BMJ Page 46 of 60

The 30-Day Sample: A representative subsample of the Scottish Mental Survey 1947 1 2 3 4 Participants 5 6 The 36Day Sample (N=7380; 10.4%) members of the Scottish Mental Survey 1947 (SMS1947) were those 7 8 childrenConfidential: born on the first, second, and third days For of every month Review in 1936, who were selectedOnly for more 9 10 11 thorough investigation in 1947, and who were similar to the full SMS1947 in terms of geographical 12 1 13 representativeness and mean Moray House Test performance. After excluding members of the 6Day 14 15 Sample – those born on the first day of even numbered months – who took part in annual followup visits 16 17 into adulthood 2, we have selected the remainder (30Day Sample) for the present sensitivity analyses, which 18 19 3,4 20 have been the subject of previous investigations. Just as for the full SMS1947, the linkage of members of 21 22 the 30Day Sample to their mortality records was given approval by National Services Scotland NHS 23 24 Privacy Advisory Committee and the Confidentiality Advisory Group of the Health Research Authority. 5 25 26 27 28 Measures 29 30 31 32 33 Intelligence, socioeconomic status, and physical status 34 35 The Moray House Test No. 12 was the measurement of intelligence, described in detail in the main 36 37 manuscript. Covariate data that were recorded in 1947 around the same time as the intelligence test 6 38 39 40 included three variables indicating socioeconomic status, which can be used to account for potential 41 42 confounding in the present study. These were: 1) Father’s, or head of household, occupation , coded 43 44 according to levels (from low to high status) using the 1951 Classification of Occupations 7: unskilled jobs, 45 46 semiskilled, manual or nonmanual skilled, intermediate, professional; 2) Home occupancy , calculated by 47 48 dividing the number of individuals living in the child’s home by the number of rooms within the home; and 49 50 51 3) absent school days during the school year 1946-47 , calculated as a percentage. Confounding by physical 52 53 status can be assessed in the present study by further controlling for height (inches), and the presence of 54 55 physical disability (chorea, congenital paralysis, defective vision, deafness, encephalitis, epilepsy, and/or, 56 57 meningitis). 58 59 60 8 https://mc.manuscriptcentral.com/bmj Page 47 of 60 BMJ

1 2 Vital status 3 4 The procedures for tracing and linkage of the 30Day Sample to death certificates are those detailed in the 5 6 main manuscript for the SMS1947. Among 5083 of the historical sample with intelligence test scores, 4775 7 8 Confidential: For Review Only st 9 (93.9%) were traced for linkage to death records. By the followup date (31 December 2015), there were 10 11 1755 deceased, 2372 resident in Scotland, England or Wales, 357 embarked, and 23 Army conscripts. 12 13 14 15 16 17 Statistical analysis 18 19 20 21 The analytic sample included 4149 participants (81.6% of the 30Day Sample) who were linked to mortality 22 23 24 records, and who had complete data on intelligence test score, socioeconomic status and physical status 25 26 measures. Missing intelligence test scores affected 452, or 9.1%, of the 30-Day Sample , which is consistent 27 28 with the full SMS1947 as detailed in the main manuscript. There were 1531 deaths recorded in the analytic 29 30 sample 31 32 33 34 35 References 36 37 38 1 39 Maxwell J. The level and trend of national intelligence: the contribution of the Scottish mental surveys (no. 40 41 48). London: University of London Press; 1961. 42 43 44 2Deary IJ, Whalley LW, Starr JM. A lifetime of intelligence: followup studies of the Scottish Mental 45 46 Surveys of 1932 and 1947. Washington DC: American Psychological Association; 2009. 47 48 3Batty GD, Calvin CM, Brett CE, Čukić I, Deary IJ. Childhood body weight in relation to morbidity from 49 50 51 cardiovascular disease and cancer in older adulthood: 67year followup of participants in the 1947 Scottish 52 53 Mental Survey. American journal of epidemiology. 2015 Nov 1;182(9):77580. 54 55 56 57 58 59 60 9 https://mc.manuscriptcentral.com/bmj BMJ Page 48 of 60

4Batty GD, Calvin CM, Brett CE, Cukic I, Deary IJ. Childhood body weight in relation to causespecific 1 2 mortality: 67 year followup of participants in the 1947 Scottish Mental Survey. Medicine. 2016 Feb 3 4 5 1;95(6):e2263. 6 7 5Brett CE, Deary IJ. Realising health data linkage from a researcher’s perspective: following up the 6Day 8 Confidential: For Review Only 9 10 Sample of the Scottish Mental Survey 1947. Longitudinal and Life Course Studies. 2014 Oct 30;5(3):283 11 12 98. 13 14 15 6Scottish Council for Research in Education. The trend of Scottish intelligence: A comparison of the 1947 16 17 and 1932 surveys of the intelligence of elevenyearold pupils. London: University of London Press; 1949. 18 19 7General Register Office. Classification of occupations 1951. London: Her Majesty’s Stationary Office; 20 21 22 1956. 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 10 https://mc.manuscriptcentral.com/bmj Page 49 of 60 BMJ

The West of Scotland Twenty-07 study: A replication sample 1 2 3 4 5 Participants 6 7 8 Confidential: For Review Only 9 Participants are drawn from the West of Scotland Twenty07 Study. This is a longitudinal study designed to 10 11 investigate the social determinants of health and health inequalities. It comprises three narrow age cohorts 12 13 who were aged 15, 35 and 56 at baseline in 1987/88. Study participants were randomly sampled from the 14 15 16 Central Clydeside Conurbation, a mainly urban area centred on Glasgow city. They were interviewed in 17 18 their homes by trained interviewers at baseline and on up to four further occasions over the following 20 19 20 years. Further details of the study are given elsewhere.1 A comparison with data from the 1991 UK Census 21 22 showed it to be representative of the underlying population in terms of sex, social class, car ownership and 23 24 2 25 household tenure. Data used in the replication study pertain to the oldest of the three cohorts obtained at 26 27 baseline (N=1551) plus vital status obtained by record linkage. The birth years of the oldest cohort were: 28 29 1930 (n=60), 1931 (n=438), 1932 (n=825), and, 1933 (n=39). 30 31 32 33 Measures 34 35 36 37 38 Alice Heim 4 test of General Intelligence (AH4) 39 40 Part I of the AH4 3 was administered to the oldest cohort of the Twenty07 study at baseline. The AH4 was 41 42 designed for use with a cross section of the adult population and part I consists of 65 questions which 43 44 45 measure verbal and numeric cognitive abilities. It was administered according to the instructions in the test 46 47 manual. The result used here is the total number of questions answered correctly within the time limit of 10 48 49 minutes. For comparability this total score is standardised to zero mean and unit standard deviation (z 50 51 scored). 52 53 54 55 56 57 58 59 60 11 https://mc.manuscriptcentral.com/bmj BMJ Page 50 of 60

Vital Status 1 2 Vital status was ascertained via linkage to the UK’s NHS central register. Participants of the Twenty07 3 4 5 study were ‘flagged’ at the NHS central register at the outset of the study and since then regular notifications 6 7 of deaths, embarkations and reentries have been provided to the study. These notifications include date and 8 Confidential: For Review Only 9 cause of death. Primary and secondary causes are given in textual and coded forms, the latter being a 10 11 mixture of ICD versions 9 and 10. The primary outcome here is the underlying cause of death grouped as 12 13 shown in Supplementary Table 1. Of the analytic sample there were 820 deceased (men: 423) and 542 living 14 15 nd 16 by the census date (2 February 2017). 17 18 19 20 Covariates 21 22 Smoking status was classified as current smoker/exsmoker/never smoked. The Social Class measure is 23 24 25 derived from the occupation of the head of household coded to the six fold Registrar General’s 26 27 classification: Professional/Managerial and technical/Skilled nonmanual/Skilled manual/Semi 28 29 skilled/Unskilled.4 Selfrated health (over the last twelve months) was recorded as 30 31 excellent/good/fair/poor─this was dichotomised into excellent or good vs fair or poor. All covariates were 32 33 treated as categorical when adjusted for. 34 35 36 37 38 References 39 40 1Benzeval M, Der G, Ellaway A, et al. Cohort Profile: West of Scotland Twenty07 Study: Health in the 41 42 Community. International Journal of Epidemiology. 2009;38(5):12151223. 43 44 2 45 Der G. A comparison of the West of Scotland Twenty07 study sample and the 1991 census SARs. 46 47 Glasgow, 1998 48 49 3Heim AW. AH 4 Group Test of General Intelligence. Manual. By A (lice) W (inifred) Heim. Rev. Ed. 50 51 NFER Publishing Company; 1970. 52 53 4OPCS. Classification of Occupations. HMSO. London, 1980. 54 55 56 57 58 59 60 12 https://mc.manuscriptcentral.com/bmj Page 51 of 60 BMJ

Supplementary Figure 1 1 2 Hazard ratios (95% confidence intervals) for the association between a one SD higher score in intelligence 3 4 test score at age 11 years and underlying cause of death in the Scottish Mental Survey 1947, N=65 765. 5 6 7 Hazard ratios (HR) are adjusted for age at intelligence testing, and sex. 8 Confidential: For Review Only 9 10 11 12 Cause of death n HR (95% CI) 13 All causes 25,979 0.81 (0.80 to 0.82) 14 15 CVD 7128 0.76 (0.75 to 0.78) 16 CHD 4562 0.75 (0.73 to 0.78) 17 18 Stroke 1412 0.75 (0.71 to 0.79) 19 Cancer 8457 0.86 (0.84 to 0.88) 20 Smoking cancer 5862 0.82 (0.80 to 0.84) 21 22 Non-smoking cancer 2595 0.97 (0.93 to 1.01) 23 Respiratory disease 2204 0.66 (0.64 to 0.69) 24 Digestive disease 1170 0.80 (0.75 to 0.84) 25 26 External causes 867 0.83 (0.78 to 0.89) 27 Accidental 490 0.85 (0.78 to 0.93) 28 29 Self-harm 141 0.86 (0.73 to 1.01) 30 Dementia 523 0.86 (0.79 to 0.94) 31 32 0.40 0.50 0.60 0.70 0.80 0.90 1.00 1.10 1.20 1.30 33 34 HR (95% CI) 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 13 https://mc.manuscriptcentral.com/bmj BMJ Page 52 of 60

Supplementary Figure 2 1 2 Hazard ratios (95% confidence intervals) for the association between a one SD higher score in intelligence 3 4 test score at age 11 years and underlying cause of death by cancer type to age 79 years in the Scottish 5 6 7 Mental Survey 1947. Hazard ratios (HR) are adjusted for age at intelligence testing, and sex, with the 8 Confidential: For Review Only 9 exceptions of ovarian and breast cancer (women only), and, prostate cancer (men only). 10 11 12 13 Cancer type n HR (95% CI) 14 15 All malignant cancer 8457 0.86 (0.84 to 0.88) 16 17 Oral 154 0.97 (0.83 to 1.13) 18 19 Oesophagul 403 0.84 (0.76 to 0.92) 20 Colorectal 828 0.89 (0.83 to 0.95) 21 22 Stomach 330 0.76 (0.69 to 0.85) 23 24 Liver 178 0.86 (0.75 to 0.99) 25 26 Pancreatic 377 0.93 (0.84 to 1.03) 27 Lung 2602 0.75 (0.72 to 0.77) 28 29 Skin 92 1.03 (0.83 to 1.27) 30 31 Ovarian 246 0.99 (0.87 to 1.13) 32 33 Breast 613 1.02 (0.94 to 1.11) 34 Prostate 356 1.04 (0.93 to 1.15) 35 36 Kidney 208 0.89 (0.77 to 1.01) 37 38 Urinary bladder 209 0.77 (0.68 to 0.88) 39 40 Brain 224 0.96 (0.84 to 1.09) 41 Haematopoietic 495 0.93 (0.85 to 1.01) 42 43 44 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 1.10 1.20 1.30 1.40 1.50 45 46 HR (95% CI) 47 48 49 50 51 52 53 54 55 56 57 58 59 60 14 https://mc.manuscriptcentral.com/bmj Page 53 of 60 BMJ

Supplementary Table 2 1 2 Hazard Ratios (and 95% confidence intervals) for a one SD higher score in intelligence score at age 11 years 3 4 in relation to major causes of death in men and women separately in the Scottish Mental Survey 1947 5 6 7 Males Females p-value for sex 8 Confidential: For Review Onlyby MHT Cause of death 9 N HR (95% CI) N HR (95% CI) interaction 10 effect 11 All causes 15 271 0.83 (0.82 to 0.84) 10 708 0.78 (0.77 to 0.80) <.001 12 13 Cardiovascular disease 6070 0.79 (0.77 to 0.81) 3549 0.70 (0.68 to 0.73) <.001 14 15 Coronary heart disease 4045 0.78 (0.76 to 0.81) 1810 0.67 (0.64 to 0.70) <.001 16 17 Stroke 1114 0.78 (0.74 to 0.83) 939 0.72 (0.68 to 0.77) .074 18 19 Cancer 4890 0.87 (0.85 to 0.89) 4016 0.84 (0.82 to 0.87) .15 20 21 Smoking related 3614 0.84 (0.81 to 0.86) 2597 0.79 (0.76 to 0.82) .017 22 23 Non-smoking-related 1276 0.96 (0.91 to 1.02) 1419 0.96 (0.91 to 1.01) .87 24 Respiratory disease 3059 0.75 (0.73 to 0.78) 2254 0.68 (0.65 to 0.70) <.001 25 26 Digestive-system disease 1088 0.85 (0.80 to 0.90) 780 0.78 (0.73 to 0.84) .090 27 28 External cause 998 0.82 (0.78 to 0.87) 567 0.85 (0.78 to 0.93) .41 29 30 Accidental 505 0.82 (0.76 to 0.89) 278 0.77 (0.68 to 0.87) .54 31 32 Intentional self-harm 102 0.80 (0.66 to 0.96) 42 1.15 (0.82 to 1.60) .063 33 34 Dementia 414 0.90 (0.82 to 0.98) 373 0.76 (0.69 to 0.85) .021 35 36 N = 65 765. All HRs are adjusted for age at intelligence testing. MHT = Moray House Test. 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 15 https://mc.manuscriptcentral.com/bmj BMJ Page 54 of 60

Supplementary Table 3 1 2 Hazard Ratios (and 95% confidence intervals) for a one SD higher score in intelligence score at age 11 years 3 4 5 with specific cancerrelated deaths in men and women separately in the Scottish Mental Survey 1947 6 7 Males Females p-value for sex 8 Confidential: For Review Onlyby MHT Cancer type 9 N HR (95% CI) N HR (95% CI) interaction 10 effect 11 12 Oral 124 0.92 (0.78 to 1.09) 52 1.15 (0.85 to 1.56) .14 13 Oesophageal 293 0.87 (0.78 to 0.98) 130 0.81 (0.68 to 0.96) .72 14 15 Colorectal 558 0.92 (0.85 to 1.00) 349 0.83 (0.74 to 0.92) .089 16 17 Stomach 223 0.79 (0.69 to 0.89) 121 0.73 (0.61 to 0.88) .47 18 19 Liver 142 0.88 (0.75 to 1.04) 54 0.74 (0.57 to 0.98) .31 20 21 Pancreatic 203 0.98 (0.86 to 1.12) 186 0.87 (0.75 to 1.02) .36 22 23 Lung 1513 0.77 (0.74 to 0.81) 1089 0.70 (0.66 to 0.75) .013 24 25 Skin 58 1.05 (0.81 to 1.36) 44 0.99 (0.72 to 1.36) .73 26 Kidney 151 0.94 (0.81 to 1.10) 78 0.85 (0.67 to 1.06) .43 27 28 Bladder 162 0.82 (0.71 to 0.95) 86 0.77 (0.62 to 0.95) .53 29 30 Brain 136 0.92 (0.79 to 1.09) 93 1.00 (0.80 to 1.24) .67 31 32 Haematopoietic 341 0.87 (0.79 to 0.97) 208 0.98 (0.84 to 1.13) .18 33 34 35 N = 65 765. All HRs are adjusted for age at intelligence testing. Final column indicates pvalues for interaction effects 36 between sex and childhood intelligence in models that include the total sample. MHT = Moray House Test. 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 16 https://mc.manuscriptcentral.com/bmj Page 55 of 60 BMJ

Supplementary Figure 3 1 2 Hazard ratios (95% confidence intervals) for the association between a one SD higher score in intelligence 3 4 test score at age 11 years and cause of death in the Scottish Mental Survey 1947, adjusting for school (N=65 5 6 7 765). Hazard ratios (HR) are adjusted for age at intelligence testing, and sex. 8 Confidential: For Review Only 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 17 https://mc.manuscriptcentral.com/bmj BMJ Page 56 of 60

Supplementary Figure 4 1 2 Hazard ratios (95% confidence intervals) for the association between a one SD higher score in intelligence 3 4 test score at age 11 years and cause of death by cancer type in the Scottish Mental Survey 1947, adjusting 5 6 7 for school. Hazard ratios (HR) are also adjusted for age at intelligence testing, and sex, with the exceptions 8 Confidential: For Review Only 9 of ovarian and breast cancer (women only), and, prostate cancer (men only). 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 18 https://mc.manuscriptcentral.com/bmj Page 57 of 60 BMJ

1 2 Supplementary Table 4 3 4 Hazard ratios (95% confidence intervals) for the association between a one SD higher intelligence test score at age 11 years and cause of death to age 79, 5 Confidential: For Review Only 6 a 7 accounting for childhood socioeconomic factors and physical status indicators in the 30Day subsample of the Scottish Mental Survey 1947 8 9 Cause of death N cases Model 1: Model 2: Model 3: % % 10 attenuation d: attenuation: 11 Age, sex and school- Model 1 + background Model 2 + physical status adjusted socioeconomic factors b indicators c Model 1 to 2 Models 1 to 3 12 13 All causes 1531 0.82 (0.78 to 0.86) 0.84 (0.80 to 0.89) 0.85 (0.80 to 0.90) 15.8 18.9 14 15 Cardiovascular disease 700 0.80 (0.75 to 0.87) 0.82 (0.76 to 0.88) 0.82 (0.76 to 0.89) 7.1 10.1 16 17 Cancer 584 0.86 (0.79 to 0.94) 0.90 (0.82 to 0.98) 0.90 (0.82 to 0.98) 26.4 25.8 18 19 Smoking related cancer 407 0.83 (0.75 to 0.91) 0.86 (0.78 to 0.95) 0.86 (0.76 to 0.95) 21.3 20.6 20 21 Non-smoking cancer 177 0.95 (0.82 to 1.10) 0.98 (0.84 to 1.15) 0.98 (0.87 to 1.16) 70.1 69.4 22 23 Respiratory disease 367 0.73 (0.66 to 0.80) 0.76 (0.68 to 0.84) 0.75 (0.68 to 0.84) 12.7 11.5 24 Digestive-system disease 122 0.79 (0.67 to 0.95) 0.84 (0.70 to 1.01) 0.84 (0.70 to 1.01) 24.7 24.7 25 26 External cause 86 0.99 (0.80 to 1.23) 0.99 (0.79 to 1.24) 1.03 (0.82 to 1.30) - - 27 28 Dementia 52 1.02 (0.77 to 1.35) 1.00 (0.75 to 1.35) 1.03 (0.76 to 1.39) 77.2 32.5 29 30 a 31 N = 4031, with complete data on all covariates (79.3% of 30Day subsample).

32 b 33 Childhood socioeconomic status indicators include: (1) father’s or head of household’s occupational status, (2) home overcrowding, (3) school absenteeism during 1946 to 1947 34 cPhysical status indicators include: (1) height (cm) in 1947; (2) physical disability 35 36 d Percent attenuation = 100*((β confounderadjusted model – βmodel 1 )/(β model 1 )) 37 38 39 40 41 42 19 43 44 45 46 https://mc.manuscriptcentral.com/bmj 47 48 49 50 51 52 53 54 55 56 57 58 59 60 BMJ Page 58 of 60

Supplementary Figure 5 1 2 Hazard ratios (95% confidence intervals) for the association between a one SD higher score in premorbid 3 4 intelligence test score and underlying cause of death by older adulthood, in the Scottish Mental Survey 1947 5 6 and West of Scotland Twenty07 cohorts. Hazard ratios (HR) are adjusted for age at intelligence testing, and 7 8 Confidential: For Review Only 9 sex. 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 20 https://mc.manuscriptcentral.com/bmj Page 59 of 60 BMJ

1 2 Supplementary Table 5 3 4 Hazard ratios (95% confidence intervals) for the association between a one SD higher Alice Heim 4 test score at age ~55 years and cause of death by older 5 Confidential: For Review Only 6 a 7 adulthood, accounting for smoking and socioeconomic status, in the West of Scotland Twenty07 Study 8 9 10 Cause of death N cases Model 1: Sex Model 2: Model 3: Model 4: % attenuation b % attenuation % attenuation 11 12 model 1 + smoking model 2 + socioeconomic model 3 self-assessed Model 1 to 2 Model 1 to 3 Model 1 to 4 status health 13 14 All causes 820 0.76 (0.71 to 0.82) 0.81 (0.75 to 0.87) 0.82 (0.75 to 0.89) 0.84 (0.77 to 0.91) 23.2 27.7 36.5 15 16 Cardiovascular disease 272 0.74 (0.65 to 0.84) 0.78 (0.68 to 0.89) 0.79 (0.68 to 0. 91) 0.82 (0.70 to 0.95) 17.5 21.7 34.1 17 18 Coronary heart disease 163 0.71 (0.60 to 0.83) 0.74 (0.63 to 0.88) 0.75 (0.62 to 0.91) 0.79 (0.65 to 0.96) 12.1 16.0 31.2 19 20 Stroke 66 0.90 (0.70 to 1.15) 0.96 (0.74 to 1.24) 1.03 (0.77 to 1.37) 1.04 (0.78 to 1.39) 61.3 128.1 137.2 21 22 Cancer 290 0.84 (0.74 to 0.94) 0.88 (0.78 to 1.00) 0.87 (0.76 to 1.00) 0.88 (0.76 to 1.01) 26.7 20.1 26.7 23 24 Smoking related 188 0.78 (0.67 to 0.91) 0.86 (0.73 to 1.00) 0.86 (0.72 to 1.03) 0.87 (0.73 to 1.04) 39.3 39.3 44.0 25 26 Lung cancer 86 0.74 (0.59 to 0.92) 0.86 (0.68 to 1.10) 0.88 (0.68 to 1.14) 0.90 (0.69 to 1.18) 49.9 57.5 65.0 27 Non-smoking-related 42 0.94 (0.69 to 1.28) 0.87 (0.64 to 1.19) 0.81 (0.57 to 1.15) 0.80 (0.56 to 1.13) -125.1 -240.6 -260.6 28 29 Respiratory disease 95 0.64 (0.51 to 0.80) 0.70 (0.55 to 0.88) 0.74 (0.58 to 0.95) 0.77 (0.59 to 0.99) 20.1 32.5 41.4 30 31 Digestive-system disease 41 0.78 (0.56 to 1.07) 0.80 (0.57 to 1.11) 0.76 (0.52 to 1. 09) 0.75 (0.52 to 1.08) 10.2 -10.5 -15.8 32 33 Dementia 37 0.82 (0.58 to 1.15) 0.85 (0.60 to 1. 20) 0.95 (0.65 to 1.39) 0.99 (0.68 to 1.46) 18.2 74.2 94.9 34

35 aN = 1362 (men = 613) 36 b Percent attenuation = 100*((β confounderadjusted model – βmodel 1 )/(β model 1 )) 37 38 39 40 41 42 21 43 44 45 46 https://mc.manuscriptcentral.com/bmj 47 48 49 50 51 52 53 54 55 56 57 58 59 60 BMJ Page 60 of 60

1 2 3 4 5 6 7 8 Confidential: For Review Only 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 22 https://mc.manuscriptcentral.com/bmj