Alcohol Consumption Over Time and Risk of Death: A Systematic Review and Meta-analysis

Harindra Jayasekara; Dallas R. English; Robin Room; Robert J. MacInnis


Am J Epidemiol. 2014;179(9):1049-1059. 

In This Article


Literature Search and Selection

We systematically searched the following electronic databases for potentially relevant original papers published through August 2012: Medline, Web of Science (science citation index expanded, social sciences citation index, and arts and humanities citation index), Cumulative Index to Nursing and Allied Health Literature (CINAHL) Plus, and Scopus. We used broad search criteria because alcohol consumption over time has not been measured uniformly. We used the following keywords and subject headings to identify relevant articles in electronic databases: (alcohol* OR ethanol) AND (lifetime drinking OR lifetime consumption OR lifetime intake OR cumulative drinking OR cumulative consumption OR cumulative intake OR drinking over time OR consumption over time OR intake over time OR change* in drinking OR change* in consumption OR change* in intake OR drinking pattern) AND (mortality OR death* OR coronary heart disease OR coronary artery disease OR coronary disease OR ischemic heart disease OR ischemic heart disease OR cardiovascular disease OR myocardial infarction OR sudden cardiac death OR angina pectoris OR coronary death) AND (case OR retrospective OR cohort OR prospective OR longitudinal OR follow OR ratio* OR risk*). No language restrictions were imposed. We did not include informally published written material, such as reports, in our search. When an article was not available electronically or otherwise, we contacted the authors to obtain a copy. Standard criteria for analysis and reporting the results were followed.[11]

Eligible articles included original publications (excluding letters, editorials, conference abstracts, reviews, and commentaries) of cohort studies reporting hazard ratios, relative risks, or odds ratios (referred to herein using the general term, "relative risk") and their 95% confidence intervals or information allowing us to compute the standard error of the relative risk of the association between alcohol consumption over time (measured as an individual's alcohol consumption history for different periods of life or as repeated assessments of an individual's alcohol consumption over time) and the risk of death. We excluded studies that characterized alcohol exposure qualitatively using such terms as "problem drinkers." If multiple publications from the same study cohort were available, the 1 with the most comprehensive data on alcohol consumption was included. One author (H.J.) performed the search and excluded studies at the first exclusion pass on the basis of titles and abstracts. The studies for our meta-analysis were identified from the remaining articles that reported any assessment of alcohol consumption over time and mortality risk.

Case Definition

We defined death from all causes as the outcome of interest and accepted outcomes based on death indices, registry data, medical records, and reports.

Data Extraction

Information from the identified studies was extracted by H.J. with assistance from R.J.M. We abstracted the following information from each study included in the analysis by using a standard pro forma: title; authors; year of publication; study name; study design; country; region; ethnicity; age; sex; sample size; percent lost to follow-up; exposure and follow-up times; exposure assessment and the comparability of reference categories; end points; measures of association; steps taken to minimize bias; and covariates included in the multivariable analysis. We extracted the maximally adjusted relative risks with corresponding 95% confidence intervals for each category of alcohol consumption. If results were reported for 2 multivariable models, we extracted relative risks from the models that did not adjust for possible intermediaries in the causal pathway (e.g., hypercholesterolemia). For studies in which nondrinkers were not the reference category, we recalculated relative risks and confidence intervals for categories of drinkers, making nondrinkers the reference category. We contacted the authors whenever additional information or clarifications were necessary.

For the dose-response meta-analysis, the median alcohol consumption in grams per day for each category of average intake was assigned to each corresponding relative risk. When studies defined alcohol intake over time by comparing intakes at baseline and follow-up (e.g., abstainer-to-moderate), we derived an average median intake based on the intake categories at the 2 time points. To calculate the median consumption for each intake category, we first converted the upper and lower boundaries into grams per day of alcohol from milliliters or standard drinks per day by considering the type of alcohol and the size of a standard drink in the study's country of origin.[12] Because the shapes of the alcohol intake distributions are similar across countries,[13] we derived the median age- and sex-specific consumption values for each category from Australian National Health Survey data.[14] For pooled analyses of comparable categories of alcohol consumption over time, we extracted relative risks under 5 broadly defined categories: nondrinkers (average intake over time 0 g/day) and intakes of less than 1 g/day, 1–29 g/day, 30–59 g/day, and 60 or more g/day. We mapped the intake categories of individual studies into these categories using their calculated median age- and sex-specific consumption values.

Statistical Analysis

A 2-stage random-effects meta-analysis was used to examine a potential nonlinear relationship between alcohol and all-cause mortality risk.[15] Alcohol consumption was modeled using restricted cubic splines with 3 knots at fixed percentiles (10%, 50%, and 90%) of the distribution.[16] Restricted cubic spline models were initially computed for each study, taking into account the within-study correlation; then, a random-effects meta-analysis was performed using the regression coefficients and the variance-covariance matrix from each individual study.[17,18] Nonlinearity of the dose-response curve was assessed by testing the null hypothesis that the coefficient of the second spline was equal to 0. In a separate analysis, we calculated pooled relative risks for comparable categories of alcohol intake over time using DerSimonian-Laird random effects models.[19] We used nondrinking as the reference category for both analyses. The inconsistencies across studies and their impact on the analysis were quantified by the I2 statistic.[20] Publication bias was assessed through visual inspection of funnel plots[21] and by using Egger's regression test.[22] All statistical analyses were performed using Stata, version 12.1, software (StataCorp LP, College Station, Texas). P values of less than 0.05 were considered statistically significant.