Modelling the impact of alcohol consumption on cardiovascular disease mortality for comparative risk assessments: an overview

The calculation of the Alcohol-Attributable Fraction (AAF) requires the combination of data on (i) the prevalence of different drinking statuses (current drinker, former drinker, lifetime abstainer), (ii) average consumption among current drinkers, (iii) the prevalence of people who are heavy episodic drinkers, and (iv) per capita consumption of alcohol. Exposure estimates for drinking status by sex, age (age groups: 15 to 34 years of age, 35 to 64 years of age, and 65 plus years of age; as usual for comparative risk assessments, no consumption was assumed for people under 15 years of age) and country, as well as data on the average drinking level among drinkers and the prevalence of regular and irregular heavy drinking occasions, were obtained from the Global Information System on Alcohol and Health (GISAH) (for a brief description of data sources, see [27]; http://​apps.​who.​int/​gho/​data/​node.​main.​GISAH). As alcohol is a legal substance, there are sales and/or production, export and import data for every country. However, as the prevalence of sick quitters (i.e. those who stopped consuming alcohol because of health reasons) among former drinkers may be different based on the country (e.g., in Thailand, where lifetime abstention rates are high, the prevalence of sick quitters among former drinkers is likely lower than that of European countries such as Russia), modifications were made to the prevalence of former drinkers when modelling the burden of IHD and IS attributable to alcohol consumption (see below). Total per capita consumption of alcohol data were obtained from the GISAH (these estimates were also the basis for exposure estimates used in the WHO’s 2014 GSRAH [19]).

Data on mortality by sex, age, cause of death and country for the year 2012 were obtained from the WHO Global Health Estimates (http://​www.​who.​int/​healthinfo/​global_​burden_​disease/​en/​).

In the WHO 2014 GSRAH, the AAFs for most causes of death were calculated for each cardiovascular cause of death separately by sex, age (age groups: 15 to 34 years of age, 35 to 64 years of age, and 65 plus years of age), and country. To calculate the AAF, the Relative Risks (RR) for CVD mortality for different groups in comparison to lifetime abstainers were multiplied with the prevalence of these groups. The groups used were former and current drinkers, where the latter were modelled continuously based on their average drinking per day level. This modelling strategy has become standard in comparative risk assessments for alcohol, both for the Global Burden of Disease Studies [12] and the WHO GSRAH.

Alcohol attributable deaths by sex, age, and CVD cause of death categories were obtained for each country by multiplying the respective AAFs with the absolute numbers of death in the respective categories as defined by the various categories listed above. For instance, if there had been 1000 ischemic heart disease deaths for women ages 35–64 in a country, and if the AAF for this sex, age and cause of death category was 14 % for this country, the absolute number of deaths was 140.

As with all comparative risk analyses for alcohol, uncertainty was estimated with Monte Carlo methodology, described in more detail in Additional file 1: Web Appendix 1.

We re-conducted the analyses from the WHO’s 2014 GSRAH using data on risk relations (i.e., RR functions) from new meta-analyses for IHD (for an overview of the changes, see Table 1), and for IS (where the only changes were in modelling the “sick quitter effect” [28, 29] and age-specific alcohol RR functions).

The following information was used to model the relationship between alcohol consumption and both IHD and IS:

The methodology described above was used for all countries except for Russia and surrounding countries, in which case we used country-specific RRs based on a large epidemiological study from Zaridze and colleagues [35]. The RRs provided by Zaridze and colleagues have been recently corroborated by a large cohort study [34], and are in line with natural experiments such as the Gorbachev reform [43], where a marked reduction of alcohol consumption was associated with a marked reduction of cardiovascular mortality [44]. This shows that for Russia and surrounding countries, the detrimental effects of regular and irregular heavy drinking occasions by far outweigh the beneficial effects. Additionally, for these countries, the risk relations were based on [34] (see Additional file 1: Web Appendix 4 for details).

The modelling strategy for IHD compared to that used in the 2014 GSRAH is summarized in Table 1. For IS mortality, we used the alcohol RRs from the meta-analyses of Patra and colleagues (for risk curves see [8]); for people with heavy drinking occasions and an average volume of alcohol consumption of up to 60 g of pure alcohol per day, we set the RR to 1 (for the rationale, see [14] and above).

This modelling approach is in line with biological pathways, where regular light to moderate drinking is associated with more favourable blood lipid profiles, fibrinogen levels, inhibition of platelet activation, and anti-inflammatory effects [3, 17, 45]; irregular binge episodes, which are characteristic of the Russian style of drinking, lead to opposite effects, and an increase in hypertension and arrhythmias [3, 46‐48]).

The resulting alcohol-attributable burden estimated using data from the most recent and comprehensive evidence were compared to the results from the last WHO comparative risk assessment for alcohol consumption for the year 2012 presented in the 2014 GSRAH [19].

The study was approved by the Research Ethics Board of the Centre for Addiction and Mental Health, Toronto, Canada, as part of the Comparative Risk Analyses for alcohol. All the data sources used can be obtained by the authors upon request.

Figures 1 and 2 present the new risk relations for IHD for men and women, respectively. These graphs take into account the various RRs by range of underlying exposures found in meta-analyses (see Table 1 for details).

Table 2 provides an overview of the recalculated burden for alcohol-attributable ischaemic diseases by region and globally. Russia and surrounding countries are presented as a separate category, as the burdens for these countries were calculated based on country-specific risks (see [35], and above). Overall, the new risk relations resulted in a lower estimated burden, which was driven mainly by a lower estimated burden for women. Proportionally, the effect was substantial, and about the same for IHD and IS; the newly estimated burden for female mortality due to alcohol-attributable ischaemic diseases is approximately one-half of what had been previously estimated (Table 3). The specific contribution of Russia and surrounding countries to the global burden is apparent: for men and the risk of IHD, and for women and the risk of IS, the global alcohol-attributable burden of ischaemic diseases would be negative (i.e., exhibiting a beneficial effect) if results from Russia and surrounding countries were not included.

Table 3 outlines the estimated burden of CVDs attributable to alcohol using the old and new methodologies. Using the new methodology, alcohol-attributable CVDs accounted for 1.4 % of all deaths (1.5 % for men and 1.3 % for women) and for 26.6 % of all alcohol-attributable deaths (20.1 % for men and 45.7 % for women). The percentages obtained using the new methodology differ from the 2014 GSRAH estimates [19] by−0.6 % with respect to all deaths (−0.2 % for men and−1.1 % for women) and by−10.6 % with respect to all alcohol-attributable deaths (−2.6 % for men and−28.2 % for women).

We used the most recent meta-analyses to estimate the global mortality burden for 2012, and found marked differences in the burden for both IHD and IS, with the new results being 348,000 deaths lower globally (primarily in women) compared to current WHO GSRAH estimates. However, the overall impact of alcohol consumption on CVD is still detrimental, with approximately 780,000 deaths due to CVD attributable to alcohol consumption globally. While this article is restricted to CVD outcomes, it should be stated, however, that the overall detrimental effect of alcohol is mainly via cancer, alcohol use disorders, liver cirrhosis, and unintentional and intentional injury [19].

In our revision, we used the most recent meta-analyses [21‐23] and restricted ourselves to the data range of these analyses to avoid problems based on just extrapolating trends which is especially problematic for exponential trends. A further strength of this strategy is that all of the studies to date were included in a systematic fashion. However, the results of all systematic reviews and meta-analyses can only be as good as the underlying data, and in the field of alcohol epidemiology, there is still a scarcity of well conducted studies which separate the effects of different patterns of drinking (for listing of relevant studies see [22, 46]; for the general question see [49]). Thus, while the proposed methodology is based on the best available evidence systematically collected, there are still many holes in the literature on the impact of different drinking pattern on CVD outcomes.

Second, we set the RR (compared to lifetime abstainers) to at least 1 if people had irregular or regular heavy drinking occasions, which is justified by the underlying recent meta-analyses. Finally, we used disease-specific RRs for former drinkers (i.e., sex-specific RRs for IHD and IS from [36, 37]), and capped the prevalence of former drinkers in regions where sick quitters are less likely (e.g., in regions where lifetime abstention is the norm and where many former drinkers likely drank only for a short period of time). The capping of the prevalence of former drinkers was determined to be the most important factor contributing to the differences between the 2014 GSRAH estimates and the estimates provided in this paper. Clearly, more research is necessary on former drinkers in low- and middle income countries, which will hopefully inform further iterations of the As a result, it is likely that the estimates presented in this paper are a more accurate representation of the burden of alcohol consumption than are the 2014 GSRAH estimates.

The exposure data from the 2014 GSRAH have been validated by experts in the respective countries (for detailed description of the procedures used: [27]; see also [50]), and are based on a triangulation of per capita consumption with survey data [51, 52]. While this procedure cannot eliminate all biases, especially in countries with high unrecorded consumption [53, 54], it does correct for errors and biases stemming from survey-based data (for reasoning, see [51, 52]; for estimating uncertainty, see [55]). In other words, the underestimation of consumption via surveys, where in some countries like Canada the survey responses covered less than 35 % of the per capita consumption which is mostly based on taxation ([56]; general discussion in [51, 52]), could be avoided. However, as indicated above, unrecorded consumption may contain considerable measurement bias [53, 54].

Mortality data are mainly based on verbal autopsies, especially in low-income countries, and while they certainly contain errors, they are the best available current data [57]. Lastly, the new RR estimates used in this study were based on a larger pool of studies, and thus constitute an improvement. In addition, we were able to base alcohol burden estimates on disease- and sex-specific meta-analyses, and to use age-specific relative risk functions.

As indicated above, this study is the only one to date which incorporated the latest evidence from systematic reviews and meta-analyses. Thus, it constitutes the best effort to date to model the impact of alcohol on CVDs. All prior efforts [11, 18, 58] used the same principal methodology, but employed less comprehensive meta-analyses on risk relations, and fewer systematic searches for exposure data.

The best possible estimate of the impact of alcohol on CVD mortality revealed an important global detrimental effect. A relatively large proportion of these deaths stem from Russia and surrounding countries, with their specific combination of overall high volume of alcohol consumption coupled with prolonged binges [34, 35, 59‐61]. Accordingly, policy makers should act to reduce overall alcohol consumption as well as heavy drinking occasions in order to reduce the alcohol-attributable burden of disease in general and the alcohol-attributable non-communicable disease burden in particular [62].

There is an ongoing debate as to whether the beneficial effects of alcohol consumption in terms of a protective effect on ischaemic diseases are real. The newest contribution to this discussion comes from Mendelian randomization studies [63, 64], where Holmes and colleagues [63] recently investigated the role of alleles of the ADH1B gene on IHD risk. The rs1229984 A-allele has been found to lead to an increased metabolization of alcohol into acetaldehyde [65] (acetaldehyde is associated with negative effects, including a flushing response). Holmes and colleagues showed that carrying the rs1229984 A-allele was associated with less consumption in terms of both average alcohol consumption and binge drinking, and with a lower IHD risk. While Holmes and colleagues concluded that there is no beneficial effect of alcohol consumption on IHD risk, their findings do not rule out a beneficial effect of low average alcohol consumption without binge drinking. However, the findings highlight the need for careful interpretation of the evidence. Meta-analyses of high quality observational studies, supported by short-term experimental studies on biomarkers for IHD risk, identified a curvilinear relationship between average alcohol consumption and IHD risk, and a modifying effect from episodic heavy drinking (see also [66]). We would expect on average low or moderate alcohol consumption coupled with no binge drinking to show less IHD risk compared to other drinkers. If this rather complex relationship is not reflected in the analysis, interpretation becomes increasingly difficult. Moreover, certain assumptions must be met in order to assess the validity of the findings from a Mendelian randomization analysis, and the conclusions of such an analysis depend on these assumptions. First, it was assumed that all of the effect of the allele was mediated through alcohol consumption. However, only a small percentage of the variation of alcohol consumption is explained by ADH1B genotypes, or by other genes related to alcohol consumption, such as ALDH2. Furthermore, it is assumed that the risk for IHD is identical for carriers and non-carriers, except for the effect of the allele. Additionally, the effect of the allele, other than a reduction in alcohol exposure, must be independent from IHD risk. However, Holmes and colleagues demonstrated that the rs1229984 A-allele was associated with several IHD risk factors. Another problem lies in the statistical power required to examine such a complex association. Only a small percentage of the population studied by Holmes and colleagues carried the rs1229984 A-allele and, thus, investigations in populations where carriers are more common might remedy this problem. While Mendelian randomization has theoretical advantages over observational studies, it is unknown whether such an investigation of the ADH1B gene meets the assumptions necessary to yield unbiased results, particularly in more complex analyses. More research on the influence of several different genes on alcohol consumption and their influence on IHD risk is necessary to validate and correctly interpret the findings from this and other Mendelian randomization studies.

The above discussion illustrates that the relationship between alcohol and CVDs is still very much under examination. However, the new data presented in this article reflect the current state of knowledge and are based on a more robust evidence base than any prior estimates. Thus, unless the knowledge base changes substantially, these new estimates can be considered the best estimates for modelling the impact of alcohol consumption on CVDs.

Secondary analyses of published data. The study was approved by the Research Ethics Board of the Centre for Addiction and Mental Health, Toronto, Canada, as part of the Comparative Risk Analyses for alcohol.

Not applicable.

We used only publicly available data as specified in the manuscript. The R programs to program the AAFs used can be obtained from the first author (jtrehm@gmail.com).