Study population

We used data from two prospective population-based cohorts of men and women from three Swedish counties (Västmanland, Örebro and Uppsala): the Cohort of Swedish Men (COSM) and the Swedish Mammography Cohort (SMC). In the late autumn of 1997, 48 850 men and 39 227 women responded to a 350-item questionnaire that sought information on diet and other risk factors for chronic diseases. We omitted men and women with a missing or an erroneous personal registration number (n=297 men and n=243 women), those who died before start of follow-up (n=55 men and n=26 women), those with a history of cancer (n=2592 men and n=1811 women) or CVD (IHD, heart failure and stroke; n=5761 men and n=4319 women) and those with implausible total energy intake (ie, 3 standard deviations from the log_e-transformed mean energy intake; n=441 men and n=405 women) or with missing data on chocolate consumption (n=3169 men and n=3155 women). After these exclusions, 67 640 participants (36 535 men and 31 105 women) aged 45−83 years remained for the analysis. The Regional Ethical Review Board at Karolinska Institutet in Stockholm, Sweden approved the study. Completion of the self-administered questionnaire was considered to imply informed consent.

Assessment of chocolate consumption

Information on consumption of chocolate and other foods was obtained from a self-administered 96-item food-frequency questionnaire (FFQ) completed by participants at baseline in 1997. Participants were asked to report how often on average over the past year they had consumed chocolate. The FFQ had only one question on chocolate consumption (type of chocolate was not specified). The FFQ had eight predefined frequency consumption categories (never, 1–3/month, 1–2/week, 3–4/week, 5–6/week, 1/day, 2/day and ≥3/day). Participants were categorised into four groups according to their chocolate consumption: never, 1–3/month, 1–2/week and ≥3–4/week. The five highest frequency categories were collapsed because only 8.7% of participants consumed chocolate ≥3–4 times/week. The validity and reproducibility of the FFQ have been previously published.12

Assessment of covariates

Information on education, family history of MI before 60 years of age, smoking, aspirin use, physical activity (including walking/bicycling and leisure-time exercise), body weight, height, history of diabetes, hypertension and hypercholesterolaemia, and alcohol intake was obtained by a self-administered questionnaire. Information on history of diabetes and hypertension was also obtained from the Swedish National Diabetes Register (diabetes only) and the Swedish National Patient Register. Body mass index (BMI) was calculated as weight (in kg) divided by the square of height (in metres). Pack-years of smoking history were computed by multiplying the number of packs of cigarettes smoked per day by the number of years of smoking.

Case ascertainment

Incident MI cases (including fatal cases) were ascertained by linkage of participants (using the unique personal registration number assigned to each Swedish resident) to the Swedish National Patient Register and the Swedish Cause of Death Register at the National Board of Health and Welfare. MI was classified using the International Classification of Diseases 10th Revision code I21 for acute MI. Deaths were ascertained by linkage with the Swedish Cause of Death Register.

Statistical analysis

For each participant, person-years of follow-up accrued from 1 January 1998 until the date of diagnosis of MI, date of death (from any cause) or 31 December 2010, whichever came first. Cox proportional hazards regression models were used to estimate hazard ratios (hereafter referred to as relative risks (RR)) with 95% CI of MI according to categories of chocolate consumption. The first multivariable model was stratified by baseline age (in years) and sex and included education (less than high school, high school, university), family history of MI before 60 years of age (no, yes), smoking (never; past <20 or ≥20 pack-years; current <20 or ≥20 pack-years), aspirin use (never, 1–6 tablets/week, ≥7 tablets/week or users with unknown number of tablets), walking/bicycling (almost never, <20 min/day, 20–40 min/day, 40–60 min/day, >1 h/day), exercise (<1 h/week, 1 h/week, 2–3 h/week, 4–5 h/week, >5 h/week) and intakes of total energy (kcal/day; continuous), alcohol (g/day; quintiles), processed meat (servings/week; quintiles) and fruits and vegetables (servings/day; quintiles). In a second multivariable model, we included the above covariates plus potential intermediates: BMI (kg/m²; continuous) and history and diagnosis of diabetes (no, yes), hypertension (no, yes) and hypercholesterolaemia (no, yes) at baseline. A separate category for missing was used to handle missing covariate data (missing was ≤3%, with the exception for exercise (5.5%)). Further adjustment for intake of unprocessed red meat, fish, dairy products, sweetened beverages, coffee and tea did not alter the results; therefore, these variables were not included in the multivariable model. The proportional hazards assumption was tested using Schoenfeld residuals and was found to be satisfactory.

Tests for linear trend across categories of chocolate consumption were performed by creating a variable containing the median value of chocolate consumption for each category and treating this variable as a continuous variable in the Cox model. In a sensitivity analysis, we excluded cases diagnosed during the first 3 years of follow-up. We examined whether the association between chocolate consumption and MI risk was modified by sex, high educational level (high school or university), overweight (BMI ≥25 kg/m²) and history of diabetes, hypertension and hypercholesterolaemia at baseline. The statistical significance of the potential effect modifications was tested using the likelihood ratio test, comparing models with and without interaction terms. Analyses were performed with SAS V.9.3 (SAS Institute, Cary, North Carolina, USA). All p values were 2-tailed and the significance level was set at an α of 0.05.

Meta-analysis

We performed a meta-analysis that incorporated results from the current Swedish study into findings from previous prospective studies of the association of chocolate consumption with risk of MI or IHD. Inclusion criteria were prospective study and results for chocolate intake in relation to MI/IHD incidence or mortality. We excluded letters, abstracts, systematic reviews and meta-analyses; case–control, cross-sectional, ecological and experimental studies. We searched PubMed (http://www.ncbi.nlm.nih.gov/pubmed) and EMBASE (http://www.embase.com) from inception until 4 February 2016, using the search terms chocolate or cocoa combined with cardiovascular disease or myocardial infarction or heart disease. No restrictions were applied. Reference lists of pertinent articles were reviewed to identify further relevant studies. From each study, we extracted the following data: the first author's last name, publication year, the name of the study, study location, sex and age of participants, years of follow-up, sample size (number of events and total number of participants), chocolate intake categories, covariates adjusted for in the multivariable model and the most fully adjusted RRs.

We combined the RRs for the highest versus lowest category of chocolate consumption. We also performed a dose–response meta-analysis using the method by Greenland and Longnecker13 and Orsini et al14 to compute the trend from the correlated log RRs across categories of chocolate consumption. Studies were included in the dose–response meta-analysis if they reported RRs for at least three categories of chocolate consumption. If results were reported in servings of chocolate, we assumed that one serving equals 30 g chocolate, which is the approximate serving size of chocolate in Swedish men.12 Study-specific RRs were combined using a random-effects model,15 which considers both within-study and between-study variation (weighting was based on the inverse of the variance). The I² statistic16 was used to quantify heterogeneity between studies. Publication bias was assessed with Egger's test.17 We used Stata V.14.1 (StataCorp, College Station, Texas, USA) to analyse the data.