The Copenhagen General Population Study (CGPS) is an ongoing prospective study of the Danish general population initiated in 2003. To be eligible, participants had to be of Danish descent (i.e., the participant and both parents were born in Denmark and were Danish citizens). Among individuals living in counties of greater Copenhagen aged 20 to 40 years, 25% were invited, whereas all eligible people aged >40 years were invited. The response rate was 42%. When the present study was initiated, 84,720 individuals had been included.

The Copenhagen City Heart Study (CCHS) was initiated in 1976 where a random sample of the Danish general population above age 20 years living in the Center of Copenhagen was invited to participate (number of participants, 14,223; response rate, 74%). This examination was followed by three examinations; a second examination in 1981 to 1983, where all previously invited plus 500 new individuals aged 20 to 24 years were invited (n=12,698, 70%); a third examination in 1991 to 1994 where all previously invited plus 3,000 new individuals aged 20 to 49 years were invited (n=10,135, 61%); and a fourth examination in 2001 to 2003 where all previously invited plus an additional sample of 1,040 individuals aged 20 to 29 years were invited (n=6,238, 50%). Eligibility criterion was Danish citizenship, and the participants does not reflect the ethnic admixture of Copenhagen (for instance, the proportion of inhabitants with foreign citizenship was 8% in 1994). However, even a few participants of foreign ethnicity could confound our results since the ADH1B*2 allele has shown considerable population stratification [25]. Hence, information on birth place as recorded in the Danish Civil Registration System [26] was obtained, and participants who were born in Asia, Africa, the Middle East, South America, and Greenland were excluded (n=243).

We used data from CCHS 1991 to 1994 and 2001 to 2003 examinations (n=11,365) and the CGPS 2003 to 2012 examination (n=84,720) [27-29]. Data were collected following similar principles. Participants filled in a questionnaire that was reviewed by an investigator on the day of attendance, had a physical examination performed, and had blood samples drawn for biochemical measurements and DNA extraction. Written informed consent was obtained. The Danish Ethics Committee approved the studies (No. 100.2039/91 and H-01-144/01).

Participants reported their consumption of beer (in bottles), wine (in glasses), and spirits (in units) with response categories of “never/ hardly ever,” “monthly,” “weekly,” or “daily” and number of drinks per week. The total number of drinks per week was calculated by summing up the individual types of alcohol and was categorised as in previous studies [30]: 1 to 6, 7 to 13, 14 to 20, 21 to 27, 28 to 34, and ≥35 drinks per week with 1 drink equalling 12 g alcohol. For women, the highest category was ≥28 drinks per week.

Previously, we have shown that stepwise increments of the self-reported alcohol consumption consistently corresponded to stepwise increments in factors expected to correlate with amount of alcohol consumption (alanine aminotransferase, γ-glutamyl transpeptidase, erythrocyte volume, and alkaline phosphatase), speaking in favour of the validity of self-reported alcohol consumption [18].

DNA was isolated from full blood and stored at -45℃. Participants were genotyped for the ADH1B genotype (rs1229984; Arg47His) and ADH1C genotype (rs698; Ile349Val) by Nanogen Technology (Carson City, NV, USA) [31,32] and TaqMan assays (Applied Biosystems, Foster City, CA, USA) [19]. Genotypes were in Hardy-Weinberg equilibrium in both studies (P>0.05). Genotypes ADH1B* 2/1 and ADH1B*2/2 were grouped because of few participants with these genotype. ADH1B and ADH1C genotypes were combined and ranked according to expected enzyme activity, ranging from most to least active (1, ADH1B*2/x+ADH1C*1/1; 2, ADH1B*1/1+ADH1C*1/1; 3, ADH1B*1/1+ADH1C*1/2; 4, ADH1B*1/1+ADH1C*2/2). Additionally, we used regression analysis with alcohol intake as a dependent variable and sex and genotypes as independent variables to test how much of the variance in alcohol consumption was explained by the ADH1B and ADH1C genotypes.

Information on diagnosis of stroke was obtained from the Danish National Patient Register that holds records of all Danish hospital admissions since 1978 and further outpatient and emergency ward contacts since 1995, and from the Danish Causes of Death Registry which contains information on causes of death for all deceased in Denmark [33-35]. In both registers, records are classified according to World Health Organization (WHO) International Classification of Diseases (ICD). Records with the following codes implying stroke were extracted: ICD-8 431 to 438 and ICD-10 I60 to I68 and G45.

All diagnoses of stroke were validated individually. Records were requested from general practitioners and hospitals and validated by two experienced medical doctors independently, blinded to the test results [36,37]. Ischemic stroke was defined according to WHO criteria as rapidly developing signs of focal or global disturbance of cerebral function lasting >24 hours (unless interrupted by death), with no apparent nonvascular cause [38]. Haemorrhagic stroke and subarachnoidal hemorrhage were excluded from the ischemic stroke group. To distinguish among infarction, intracerebral hemorrhage, and subarachnoidal hemorrhage, either computed tomography or magnetic resonance imaging scan, spinal fluid examination, autopsy, or surgical description were required. If the scan did not show infarction or hemorrhage, but symptoms in accordance with stroke criteria, that event was diagnosed as ischemic stroke. The diagnosis of ischemic and haemorrhagic stroke was not applied in cases in which a scan revealed signs of prior cerebrovascular disease, but without history of any symptoms, if symptoms were nonfocal, or if symptoms lasted <24 hours [39]. Alcoholic liver cirrhosis was ascertained similarly from the Danish registries using ICD-8 code 571.0 and ICD-10 code k70.3.

Of the initial sample of 89,675 individuals, we excluded those with stroke prior to baseline (n=2,737), missing information on alcohol consumption (n=4,499) and covariates (smoking [n=2,573]), school education (n=247), body mass index (BMI) (n=511), and leisure time physical activity (n=562). The final analytic sample therefore consisted of 78,546 men and women (where genotype information was available for 74,632 individuals) (Table 1).

Variables considered being potential confounders included the following: age; school education (<8, 8 to 11, and >11 years of education, corresponding to lower primary school, higher primary school, and secondary school); leisure time physical activity (sedentary, light, moderate, and vigorous); smoking (never smoker, former smoker, and current smoker of 1 to 14, 15 to 24, and >24 g of tobacco/day); BMI (linearly); family history of cardiovascular disease (either biological parent having had myocardial infarction or stroke); diabetes (by self-report), angina pectoris (by self-report), hypertension (by self-report); ischemic heart disease (hospital diagnosis before baseline, yes/no); congestive heart failure (hospital diagnosis before baseline, yes/no); angina pectoris (by self-report); heart medication (by self-report, yes/no); and use of cholesterollowering medication (by self-report, yes/no). Analyses were conducted using Stata version 12.1 (StataCorp, College Station, TX, USA).

P-values for trend between genotypes and baseline characteristics were calculated using chi-square test for categorical variables and Kruskal-Wallis test for continuous variables. For this purpose, genotypes were assigned with values reflecting the effect of genotypes on alcohol consumption.

Participants accrued person-time from baseline until time of stroke (n=2,335), death (n=7,524), emigration (n=468), or end of follow-up (n=68,161), whichever occurred first. Data from the CGPS and CCHS were combined after having ensured that results were similar in the two studies separately. This was tested formally in a nested log-likelihood test (P=0.16 for any stroke). Power calculation for the MR analysis were conducted using Stata procedure power logrank. Analyses were conducted following the four steps below, in accordance with the Mendelian randomization study principle [17,40].

Our result and the dose-response meta-analysis suggest that a potential nonlinear relationship is present between alcohol intake and risk of any stroke, and that an alcohol intake of 1 to 20 drinks/week is associated with a reduced risk of any stroke. Hence, the association seems to be j-shaped. The divergent dose-response relationships between alcohol consumption and risk of ischemic and haemorrhagic stroke subtypes suggest that different mechanisms underlie associations with the different stroke types. Alternatively, the slightly different results may be due to fewer haemorrhagic stroke events compared with ischemic stroke events, and thus the statistical power could be too low for reliable risk estimates for haemorrhagic stroke alone. Heavy alcohol intake may be associated with an increased risk of haemorrhagic events because heavy alcohol intake potentially results in high blood pressure, reduced platelet aggregation, and enhanced fibrinolysis [12,13]. Further, alcohol consumption is associated with increased high density lipoprotein cholesterol levels, decrease platelet aggregation, increase fibrinolysis, and decrease plasma fibrinogen levels and this might help explain the lower risk of ischemic stroke [12,13].

From analysis using ADH genotypes as instrument for alcohol consumption, we did not obtain results that supported causality of the observational finding. The strength of using genetic instruments for lifestyle exposures is that confounding from associated risk factors and reverse causation are unlikely, and causal interpretation of results may be more straight forward than causal interpretation of observational findings [17]. Further, selection bias due to non-participation is less likely to affect results for genetic instruments [42] whereas as for results obtained for self-reported alcohol intake, bias cannot be ruled out: if alcohol intake was associated with non-participation and with risk of stroke, results would be influenced [43]. Hence, heavy drinkers who participate may be at better health than heavy drinkers in the underlying population, whereas light and moderate drinkers who volunteer may be more representative of moderate and non-drinkers drinkers generally. Ultimately, such a mechanism would lead to more accurate results among the light to moderate drinkers and underestimated results among the heavy drinkers.

A limitation is that the ADH1B and ADH1C genotypes are not strong instruments, meaning that the influence on the amount of alcohol consumption is limited; 10% of the variation in alcohol intake was explained by the ADH1B and ADH1C genotypes in this population. Hence, the lack of a positive finding does not preclude that alcohol is causally associated with stroke. Further, using these genotypes as instruments implies a linear model between genetic alcohol consumption and the risk of stroke; if the association between alcohol and risk of stroke for instance is indeed threshold-shaped, then the genetic instrument has little power to pick it up.

Further, another limitation is lack of power in the MR analysis. We computed the lowest HR that can be picked up with the present set-up, for the ADH1B/ADH1C combination with the highest alcohol intake (i.e., the 1/1;1/2 group, please refer to Table 2), with a 90% power and a significance level of 0.05: this HR is 0.65, meaning that we do not have the adequate power to assess a HR of 0.84.

The reason that ADH genotypes is associated with alcohol consumption is most likely that—for a given level of intake—individuals with fast alcohol degradation have higher levels of acetaldehyde leading to unpleasant symptoms such as nausea and flushing compared with individuals with slow degradation, causing them to drink less. Thus, peak levels of acetaldehyde, known to be carcinogenic and playing important roles in the development of alcohol-related diseases [44], will be higher in fast compared to slow metabolisers, meaning that by using ADH1B and ADH1C genotypes as instruments the difference in amount of alcohol cannot be studied independently of difference in acetaldehyde levels, which constitute a final limitation of this approach.

A candidate for another instrument of alcohol intake to be used in future MR studies is the recently identified single nucleotide polymorphism (SNP) in Klothoβ (KLB) gene (encoding β-Klotho), demonstrated to have a role in the regulation of alcohol drinking, observed in independent samples [45]. While the mechanism linking the KLB gene with alcohol intake is still unknown, making the potential for using it as an instrument for alcohol intake unsure it is a promising candidate to be used alone or in combination with ADH1B and ADH1C genotypes forming stronger instruments. Last, it was not possible to separate former drinkers from abstainers and hence, some drinker misclassification might have occurred.

Study strengths include the large number of individuals and stroke endpoints, and that the hospital register information on the stroke diagnosis was highly sensitive and specific [36,37]. Further, the National Danish Patient Register covers all hospitalizations in Denmark including outpatients and emergency wards since 1995, and only if participants were treated in another country or by a general practitioner solely, the information would be lacking. Hence, loss to endpoint findings is considered to be negligible.