Study population

The COSM and the SMC are prospective cohorts of men and women, respectively, in central Sweden. The COSM was initiated in the late autumn of 1997 when all men who were born between 1918 and 1952 and resided in Västmanland and Örebro counties received a mailed questionnaire about diet, beverage consumption, lifestyle, and other risk factors for chronic diseases. A total of 48,850 men (49 % of the source population) answered the questionnaire. Simultaneously, an identical questionnaire (except for some sex-specific questions) was completed by 39,227 women (70 % response rate) who were born between 1914 and 1948 and lived in Västmanland and Uppsala counties. Those women had been enrolled in the SMC in 1987–90, but the baseline data did not include cardiovascular risk factors. We therefore used the 1997 questionnaire as the baseline for the present analyses. We excluded men and women with an incorrect/missing personal identification number (n = 297 men and n = 243 women), those with a diagnosis of AF (n = 1,496 men and n = 634 women) or cancer other than non-melanoma skin cancer (n = 2,592 men and n = 1,811 women) before baseline, those who died between the administration of the 1997 questionnaire and start of follow-up (1 January 1998; n = 55 men and n = 26 women), and those with missing data on coffee consumption (n = 2,529 men and n = 1,919 women). This left 41,881 men, 45–79 years of age, and 34,594 women, 49–83 years of age for analysis. The Regional Ethical Review Board at Karolinska Institutet in Stockholm, Sweden, approved the study. The completion of the self-administered questionnaire was considered to imply informed consent.

Assessment of coffee consumption and covariates

The questionnaire completed by participants in the autumn of 1997 inquired about education, smoking, body weight, height, physical activity, history of hypertension and diabetes, family history of myocardial infarction before 60 years of age, diet, and consumption of coffee and other beverages. The questionnaire assessed average consumption of 96 foods and beverages during the previous year. With regard to coffee consumption, participants were asked to report how many cups of coffee they consumed per day or per week. In a validation study (conducted in a subsample of 111 participants of the SMC) of a similar questionnaire, the questionnaire assessment of coffee correlated well with the mean of four 7-day dietary record assessments (Spearman correlation coefficient, r = 0.61) [16].

Cardiac disease was defined as a diagnosis of ischemic heart disease or heart failure in the Swedish National Inpatient Register. We classified participants as having hypertension and diabetes if they reported in the questionnaire that they had any of these diseases or if they had a diagnosis of hypertension or diabetes in the Swedish Hospital Discharge Register or the Swedish National Diabetes Register. Participants were asked to indicate their time spent on walking and bicycling during the last year. They could choose from six predefined categories: almost never; <20 min/d; 20–40 min/d; 40–60 min/d; 1–1.5 h/d; or ≥1.5 h/d. We calculated body mass index as self-reported weight (in kg) divided by the square of self-reported height (in m). Alcohol (ethanol) consumption, in the last year, was calculated by multiplying the frequency of consumption of beer, wine and liquor by the amount consumed at each occasion.

Ascertainment of cases

Via linkage of study participants (using the personal identification number assigned to each Swedish citizen) with the Swedish Hospital Discharge Register, we obtained information on dates of diagnosis of AF. Diagnoses in the Hospital Discharge Register are coded according to the Swedish International Classification of Disease (ICD) system. AF was defined as atrial fibrillation or atrial flutter (ICD-10 code I48). A previous validation study of AF diagnoses in the Hospital Discharge Register showed that the AF diagnosis is correct in 97 % of the cases [17].

Statistical methods

Participants were followed up from 1 January 1998 until the date of diagnosis of AF, date of death (information obtained from the Swedish Death Register) or end of follow-up (31 December 2009), whichever came first. We categorized participants into five groups (approximate quintiles) according to their coffee consumption: <2 cups/d (reference); 2 to <3 cups/d; 3 to <4 cups/d; 4 to <5 cups/d; and ≥5 cups/d. We used Cox proportional hazards regression models with age as the time scale to estimate relative risks (RR) with corresponding 95 % confidence intervals (CI). All multivariable models included education (less than high school; high school; university) and known risk factors for AF, including smoking (never; past and <20 pack-years; past and ≥20 pack-years; current and <20 pack-years; current and ≥20 pack-years), history of cardiac disease (yes or no), history of hypertension (yes or no), history of diabetes (yes or no), body mass index (in kg/m², continuous), walking/bicycling (almost never; <20 min/d; 20–40 min/d; 40–60 min/d; >1 h/d), family history of myocardial infarction (yes or no), and consumption of alcohol (nondrinkers; and <1, 1–6, 7–14 and >14 drinks/wk) and tea (cups/d, continuous). Analyses of men and women combined also adjusted for sex.

Tests for trend were conducted by modeling the median consumption of coffee in each category as a continuous variable. In addition, the continuous measure of coffee consumption (cups/d) was used to fit a restricted cubic spline model and to obtain a smooth representation of the RR as a function of coffee consumption. We used three knots to divide the continuous coffee consumption into four intervals.

We performed sensitivity analyses by excluding those with a history of cardiac disease or hypertension because they might have changed their coffee consumption before baseline. Furthermore, we examined the effect of excluding coffee abstainers from the reference group or excluding AF cases diagnosed during the first two years of follow-up. We also conducted an analysis using coffee abstainers as the reference group. All statistical analyses were conducted using SAS (version 9.3; SAS Institute, Cary, NC, USA). Two-sided P values <0.05 were considered statistically significant.

Meta-analysis

We conducted a dose–response meta-analysis that included results from our two prospective cohorts (COSM and SMC) as well as findings from previously published prospective studies of coffee consumption and AF. Studies were identified by a computerized search of PubMed and Embase through 22 July 2015, and by reviewing the reference lists of retrieved articles and previous meta-analyses on caffeine intake and AF [11, 12]. We used the search terms ‘coffee’ or ‘hot beverages’ or ‘diet’ combined with ‘atrial fibrillation’ or ‘flutter’ or ‘arrhythmia’. No restrictions were imposed. Studies were eligible for inclusion in the meta-analysis if they: 1) had a prospective design; 2) the exposure was coffee consumption; 3) the outcome was incidence of AF or AF and atrial flutter combined; and 4) RRs with 95 % CIs were reported.

From every study, we extracted the first author’s last name, publication year, country of origin, sex and age of study participants, years of follow-up, method to ascertain AF events, covariates adjusted for in the analysis, and RRs with 95 % CIs (from the most fully adjusted model) for each category of coffee consumption. We also extracted the total number of cases and participants as well as the number of cases and participants in each exposure category.

We combined the study-specific RRs for the highest versus the lowest category of coffee consumption using a random-effects model, which considers both within- and between-study variability. Subgroup analyses by country and sex were conducted. In a sensitivity analysis, we added studies that reported results on caffeine (but not coffee) intake in relation to risk of AF. We performed a random-effects dose–response meta-analysis using the same method as described in previous meta-analyses [4, 5] to compute the RRs per 2 cups/d increment of coffee consumption. Statistical heterogeneity was investigated using the P and I ² statistics [18]. Publication bias was assessed with Egger’s test [19]. We used Stata (version 12.0, StataCorp, College Station, TX, USA) for the statistical analyses.