Participants were examined at the research clinic using standardised protocols. Waist and hip circumferences and height and weight were measured and BMI calculated as weight in kilograms divided by height in metres squared. The mean of two seated resting blood pressure readings (mmHg) was used. Fasting blood samples were drawn and non-diabetic participants were given a 75 g oral glucose tolerance test [27]. Glucose and lipids were analysed using standardised assays. Prevalent diabetes was defined according to World Health Organisation criteria [28], self-report of physician diagnosed diabetes, or receipt of anti-diabetes medications.

A self-administered questionnaire included information on medical history, smoking status, alcohol intake, diet, physical activity, occupation, and socio-demographic background. Social class was classified according to the Registrar General’s occupation-based classification [29] into five main categories, and categorised into non-manual and manual occupations. Smoking status was derived from yes/no responses to the questions “Do you smoke cigarettes?” and “Have you ever smoked cigarettes regularly?” and categorised into never, former and current smokers. Alcohol intake was derived from a series of questions; “In the past 12 months have you taken an alcoholic drink:” followed by answer categories from twice a day to not at all; “How often have you had a drink of beer or cider during the last 12 months?” and “When you had a drink of beer or cider in the last 12 months, how many pints have you usually drunk on any occasion?” was used to derive the number of units consumed on average per week. This was done similarly for wine, fortified wine and spirits. A unit of alcohol was defined as half a pint of beer or cider, a glass of wine or fortified wine, or a single unit of spirits. Physical activity was derived from participant responses to questions about usual physical activity based on the Allied Dunbar National Fitness Survey questionnaire [30]. This included sections on occupational activity, commuting (walking and cycling), sports and other strenuous activity. Physical activity energy expenditure was calculated as a product of frequency, duration and intensity (in metabolic equivalent task units, METs) for each specific activity [31]. Frequency of fruit and vegetable consumption was assessed using a simple dietary questionnaire with response categories of ‘not in the last week’, ‘once in the last week’, ‘2–3 times in the last week’, and ‘on most days in the last week’ for the intake of “fresh green vegetables” and “fresh fruit”.

Four health behaviours (smoking, alcohol consumption, physical activity level and fruit and vegetable consumption) were categorised dichotomously into healthy and unhealthy as follows.

A healthy smoking behaviour was defined as not currently smoking. A moderate alcohol intake was defined as consuming between one unit per week and up to 14 units/week for women or 21 units/week for men based on past evidence on alcohol consumption and CVD outcomes among Europeans [5–7]. As alcohol consumption patterns differ between Europeans and South Asians, with abstention being more common among the latter [32], we performed an exploratory analysis to test the appropriateness of the currently used cut-off among both ethnic groups. When we compared abstention with our definition of moderate alcohol intake we found increased hazard of CVD in both Europeans and South Asians (HR: 1.53 (1.11–2.11) and 1.31 (1.00–1.71), respectively), independent of age, sex, BMI, hypertension, lipids and social class, supporting the use of the cut-off in both ethnic groups. Physical activity was summarised as the total weekly energy (MJ) expended in sporting activities, cycling, walking, and in other strenuous activity during leisure time. Physically active was defined as those undertaking moderate exercise (e.g. walking, cycling, gardening) for 5 h or more per week or vigorous exercise (e.g. jogging, swimming) for ≥2.5 h per week, which equals a daily activity energy expenditure of at least 3 kcal/kg/day. This cut-off was based on ethnic specific data [33] where physical activity was categorised as moderately to highly active when energy expenditure was ≥3 kcal/kg/day, a level adequate to obtain health benefits [34]. A frequent (healthy) fruit and vegetable intake was defined as consumption of fruit or vegetables on most days, with intakes of ≥5.5 times/week of either fruit, or vegetables, or both.

A combined health behaviour score was constructed ranging from a minimum score of zero to a maximum score of four. Participants scored one point if they complied with a healthy behaviour and zero if they did not.

Between 2008–2011, surviving participants were followed-up through a combination of completing health and lifestyle questionnaires, primary care medical record review and/or attendance at a study clinic visit [30]. All participants were flagged for mortality with the Office of National Statistics (ONS). For CHD, we included the first event after baseline identified from cause of death including any of the following: angina, myocardial infarction or its sequelae or atherosclerotic heart disease using International Classification of Disease (ICD) 9 codes 410–415 or ICD10 codes I200-I259 [35]. Primary care data were independently reviewed by two senior physicians blinded to participant ethnicity and identity. Adjudication by a third physician was conducted if required. For stroke, the first event after baseline was identified from cause of death including any of the following ICD9 codes 430–439 or ICD10 codes I600-I698. Primary care data were reviewed in a similar manner to CHD, with definite or probable diagnosis of stroke made according to pre-determined criteria based on symptoms, duration of symptoms and MR/CT imaging. We also included participant report of physician diagnosed stroke and duration of symptoms in excess of 24 hours [27]. We combined any first post-baseline fatal or non-fatal CHD or stroke event as CVD outcome.

Analyses were performed using Stata/SE 12.1 (Stata Corp, College Station, TX, USA). Baseline characteristics were summarised by the number of health behaviours for each ethnic group separately using means ± SD, medians (IQR) and frequencies and percentages as relevant. Differences were tested using ANOVA, Kruskal-Wallis, or χ² tests, as appropriate.

There was no interaction between sex and health behaviours on CVD events, but a significant interaction (p<0.001) was present between ethnicity and health behaviour score, justifying combined analyses for men and women but separate analyses by ethnicity. After checking that the proportional hazards assumption was not violated, we used Cox proportional hazards regression with calendar time (years) as the underlying timescale to examine the association between combined and individual health behaviours and incident CVD. We also examined incident CHD as an outcome. Each individual contributed person-time from the year of entry into the study until the year of first fatal or non-fatal event of either CHD or stroke, year of death from any other cause, or end of follow-up in November 2011, whichever came first.

For the combined health behaviours two models were constructed; Model 1 adjusted for age and sex; Model 2 was additionally adjusted for potential confounders and major CVD risk factors: systolic and diastolic blood pressure, hypertension treatment, total cholesterol, HDL-cholesterol, BMI, social class, employment status, and occupational physical activity. For the individual health behaviours a third model (Model 3) was constructed which was the same as Model 2 above but additionally mutually adjusted for the three remaining health behaviours.

A number of sensitivity analyses were conducted by separately including in model 2 each of: education, prevalent diabetes, triglycerides, waist circumference, and waist to hip ratio. We also excluded participants with events in the first two years and in the first five years of follow-up to minimise potential bias from reverse causality.

We applied several approaches to address potential bias due to missing data. To test whether the participants with and without follow-up data differed, we compared baseline characteristics between the total sample, those with follow-up data, and those without follow-up data. We also examined the association between health behaviours and fatal CVD in the total sample (n = 3,088), including those with baseline data (without prevalent CVD) on health behaviours. This approach allowed us to include also those lost to follow-up. Finally, using logistic regression we examined the association between missing follow-up data (yes/no) and number of baseline health behaviours in the total sample.

We calculated the PAF [36] to estimate the proportion of CVD and CHD cases within this population that could have been avoided had all participants adhered to the healthy behaviours, assuming a causal relation between exposure and outcome. To calculate the PAF we compared participants with a health behaviour score of 4 with the rest of the population in each ethnic group using adjusted Cox regression (Model 2) and the Rockhill et al. method [36]:

where pd_i is the proportion of total cases in the population in the ith exposure category and RR_i is the adjusted relative risk for the ith exposure category relative to the unexposed (those with all 4 health behaviours). This equation allows multi-category exposures to be considered, rather than comparing only the extremes.

Compared with the UK EPIC-Norfolk study, we used comparable measures for smoking, but used a sex-specific alcohol definition, a stricter definition of being physically active, and a more limited self-reported diet measure, yet our estimates for incident CVD and CHD for Europeans are consistent with the EPIC-Norfolk findings for incident stroke [18]. Similarly, our findings for Europeans in Southall are consistent with findings in three American studies of health professionals [14–16].

The population impact we observed of a PAF of 43% for CHD (and 28% for CVD) in Europeans is lower than previously reported [14, 15]. However, the US-based studies estimated a population attributable risk (PAR) of incident CHD, comparing those with adherence to five health behaviours with the rest of the population, translating to a PAR of 62% in male health professionals [15] and a PAR of 82% in The Nurses’ Health Study [14], while our method enabled us to account for each level of adherence to the health behaviour score.

The populations in other published studies are dominated by health professionals [14–16] whereas our findings show that these associations are also apparent in manual occupations. The CVD incidence rates we observed in both Europeans and South Asians from West London was higher than rates reported for US Health Professional men [15] and adults from the UK Norfolk-based study [37]. The greater magnitude of the disease burden in the current population implies the potential for risk reductions especially in South Asians, where 95% of the population could improve their health behaviours and if everyone adhered to the four health behaviours, 51–63% of CVD events could be prevented.

With regard to single health behaviours, we found smoking to be a strong risk factor both for incident CVD and CHD, consistent with previous evidence in Europeans [14–16, 19] and South Asians [22, 38]. Non-moderate alcohol intake was associated with increased risks of CVD in Europeans and CHD in South Asians. However, the distribution of adherence to this health behaviour varied substantially by ethnicity; non-moderate alcohol intake in Europeans was mostly due to high intakes (n = 273, 25% of total), whereas in South Asians, non-adherence was mostly due to abstaining (n = 420, 42% of total) as previously reported [32]. We found a tendency (not significant) towards increasing CVD and CHD incidence with being inactive among Europeans, but among South Asians there was a null association. Our finding of South Asians being less physically active than Europeans is consistent with previous reports [39, 40]. A comprehensive dietary score [14, 15] and plasma vitamin C levels [18, 19], as indicators of a healthy diet behaviour, have previously reported to be associated with decreased risk of CVD related outcomes, while the fruit and vegetable intake health behaviour in the current analysis was not.

The loss to follow-up (29%) was a potential source of bias in this study, but three separate sets of analyses confirmed this was not substantial. We compared baseline characteristics among participants with and without follow-up data, we restricted analyses to fatal CVD only (since everyone was traced for mortality), and we found no association between having missing follow-up data and number of health behaviours at baseline. With a single time-point assessment of self-reported health behaviours we cannot account for potential changes in health behaviours during the study period. By dichotomising each of the four health behaviours and weighting the healthier behaviours equally, we did not allow for differential importance or potential dose-response relationships of the individual health behaviours. Physical activity was measured using a questionnaire, which was limited in differentiation of intensity, and we were unable to include occupational physical activity in the activity health behaviour due to limited information, though we did account for occupational physical activity in our multivariable model. The cut-off for being classified as physically active corresponds to about 30 minutes of vigorous activity per day which is higher than the recommendation of 30 minutes of moderate to vigorous activity a day [41]. Nevertheless previous studies have used different definitions of being physically active [19, 33, 42, 43], and 37% of Europeans and 23% of South Asians met the cut-off of 3kcal/kg/day in our study, which is consistent with an earlier study comparing ethnic groups [33]. Our higher cut-off of qualifying as being physically active was expected to show a strong relationship between physical activity and CVD outcomes. The dietary questionnaire we used was limited by being simple, not validated and did not enable estimation of total energy intake. However, we were able to include fruit and vegetable intake frequency, widely used as a marker of healthy diet [10, 12, 13]. Our non-significant findings for the physical activity and the fruit and vegetable intake health behaviour, are possibly explained by measurement error, recall bias and social desirability bias. Misclassification of participants to the healthy behaviours would result in an underestimation of the true effect of the health behaviour and our estimates would therefore be conservative. Religious practice could be an important factor in determining some health behaviours (such as non-consumption of alcohol among Muslim participants or not smoking among Sikh participants); adjustment for religion did not materially influence our findings (results not shown). Our findings are based on first-generation immigrants, therefore they may not apply to all South Asians. However, emerging studies on UK born South Asians have indicated the importance of health behaviours in second generation South Asians [44, 45]. We note that our definition of CVD included CHD and stroke events in line with our previous work [26] and we did not include peripheral vascular disease events as these data were very limited on GP record review. Finally we cannot quantify the completeness of our case ascertainment with ONS linkage for mortality, but we would not expect completeness to differ by ethnicity and also any under-reporting would tend to bias the results towards the null, therefore our estimates might be conservative. Our use of the direct method of primary care data review for events by two senior physicians blinded to participant ethnicity and identity has, however, enhanced case ascertainment.

The large population-based cohort of both Europeans and South Asians and long follow-up period in this study provides unique prospective ethnicity-specific data on incident CVD outcomes. The estimation and measurement of a wide variety of variables at baseline allowed for adjustment for many possible covariates and our findings were robust to a range of adjustments and sensitivity analyses. The PAFs reported in this study provide novel information that can be used in public health messages for CVD prevention. Among Europeans, previously only population impacts for incident CHD, or for mortality but not incident CVD, have been reported. In South Asians, the impact of a composite health behaviour score on CVD or CHD incidence was previously unknown.

To our knowledge this study is the first to investigate the role of combined health behaviours on risk of CVD and CHD and their population impact in South Asian immigrants in the UK compared to a European population. Though South Asians in the UK have an excess risk of CVD and CHD compared with Europeans, the risk reduction by adherence to a healthy lifestyle is also of great importance in South Asians, as was shown previously in European populations but was unknown for South Asians. Increasing physical activity among South Asians has the greatest potential for prevention strategies underlined by the striking disparity in this health behaviour. From a public health perspective, to target and inform this specifically high-risk group about health behaviours should be prioritised in order to reduce the burden of CVD events.

All data necessary to replicate these analyses and to draw conclusions are contained in the anonymised supporting information files.