Inclusion criteria

Included studies were original English-language research articles published in the peer-reviewed literature that quantified the relationship between all-cause mortality and alcohol consumption among human populations in cohort studies. All genders, age groups, and subjects from any racial, ethnic, cultural, or religious groups were eligible for inclusion, regardless of geographic region. Studies were excluded if all-cause mortality outcomes could not be separated from morbidity outcomes. Studies were also excluded if the sample was defined in terms of pre-existing illness or poor health status. When more than one publication of the same study was available, the most recent or comprehensive in its treatment of potential bias and confounding was selected.

Data sources

We identified all potentially relevant English-language articles published up to December 31, 2014, by searching PubMed (last searched February 25, 2015) and the Web of Science and through reference list cross-checking of previous meta-analyses.

Search strategy

We used the following key words and subject headings to identify relevant articles in electronic databases: [mortality OR death OR coronary heart disease OR coronary artery disease OR ischemic heart disease OR atherosclerotic heart disease] AND [alcohol OR consumption OR ethanol OR alcohol drinking] AND [cohort OR prospective OR longitudinal].

Study selection

Two trained reviewers read the titles of all the citations retrieved from the electronic database searches and removed those clearly unrelated to the relationship between mortality and alcohol consumption. At the next stage of study selection, abstracts were reviewed to further exclude studies that clearly did not meet inclusion criteria. At the third stage, the full articles were checked for eligibility with cross-checking by senior investigators. Reasons for exclusion were documented at each stage. The reference lists from two previous meta-analyses—Fillmore et al. (2006) and Ronksley et al. (2011)—also were searched for additional eligible studies that would not otherwise have been included.

Data extraction

Two reviewers extracted and coded data from all studies fulfilling the inclusion criteria, and any disagreements were resolved by discussion with the investigators. The original Fillmore et al. (2006) codebook was refined to provide more detailed classifications of the type of reference groups used and methods of quantifying alcohol consumption and study characteristics. The coding of all variables in the analysis presented here was double-checked by the first two authors (T.S. and J.Z.).

Data items

Quality assessment

To identify potential study-level covariates to be controlled in multivariable meta-regression analyses (Greenland, 1998; Normand, 1999), each study was coded for publication year, sample size, population characteristics (age, gender, country), and whether covariates (e.g., smoking status, previous illness) were controlled for in individual studies. Covariates available for all selected studies were median age of study participants at first assessment, sex, country in which a study was conducted, date a study was conducted, number of years of follow-up, whether persons with previous illnesses were excluded, and quality of the measure used for typical daily alcohol intake.

On theoretical grounds, it was expected that a long follow-up period, inclusion of individuals with previous illnesses, and an earlier age at intake would be study characteristics that reduce selection biases (Bergmann et al., 2013; Stockwell & Chikritzhs, 2013). Studies were classified according to the presence or absence of two key types of potential bias: (a) including former drinkers and/or (b) including occasional drinkers in the abstainer reference category.

Following Fillmore et al. (2006), lifetime abstention was strictly defined as zero consumption and did not include studies with any level of occasional lifetime or past-year drinking (e.g., less than 12 drinks or “rarely” or “hardly ever” drinking). Such self-reported infrequent drinkers have been shown to greatly underreport their personal consumption (Stockwell et al., 2014; Ye et al., 2013). When studies assessed usual or typical drinking patterns over a month or a week, it was assumed that individuals classified as abstainers by this method would include occasional drinkers (e.g., abstaining in a typical week is still consistent with drinking less than once a week). We coded a drinking measure as “adequate” for the purpose of estimating average daily alcohol intake if both quantity and frequency of drinking were assessed for a period of at least 1 week. Given that simple quantity–frequency measures of drinking typically result in substantial underreporting (e.g., Stockwell et al., 2014), we recognize these are minimal criteria for adequacy of measurement, necessitated by the poor overall quality of drinking measures in this literature.

Analyses

Visual inspection of the data suggested the presence of extreme outliers among estimates of the risk of all-cause mortality from drinking. Estimates of RR were classified as “extreme” when they were outside of the interval of the sample mean of natural log RR ± 2 times the standard deviation of estimates within each drinking category (Acuna & Rodrigues, 2014; Pagano & Gauvreau, 2000; Woodward, 2000). This procedure identified 11 risk estimates significantly below the mean (RR range: 0.1–0.46) and 18 risk estimates markedly higher (OR range: 1.89–4.57). Compared with other available methods (Cook & Weisberg, 1982; Viechtbauer & Cheung, 2010), this is a conservative approach excluding relatively few risk estimates. Removal of outliers made no substantive difference to the results; therefore, models are presented without any outlier estimates excluded.

Publication bias was assessed first through visual inspection of the funnel plot of log-RR of all-cause mortality due to alcohol consumption against the inverse standard error of log-RR (see Figure C1 in Appendix C) (Woodward, 2000), and also by Egger’s linear regression method (Egger et al., 1997). We also assessed between-study heterogeneity of RRs using Cochran’s Q (Cochran, 1954) and the I² statistic (Higgins & Thompson, 2002). When significant heterogeneity was detected, mixed-effects models were used to obtain the summarized RR estimates.

Mixed regression analyses were performed in which drinking groups and control variables were treated as fixed effects with a random-intercept study effect (Normand, 1999). The dependent variable was the natural log of the RR estimated using the rate ratio or hazard ratio of each drinking group in relation to the abstainer category. Analyses were weighted by the inverse of the estimated variance of the natural log RR. Variance was estimated from reported standard errors or confidence intervals (CIs). The weights for each individual study were created using the inverse variance weight method used in mixed regression analysis in order to get maximum precision for the main results of the meta-analysis (Woodward, 2000). Studies with large or small estimates and/or variance can be highly influential. Sensitivity analyses also were run after excluding such studies to detect influential cases (Woodward, 2000).

Drinking levels were examined in terms of predefined specific consumption levels. Drinking categories were defined as (a) lifetime abstainers; (b) former drinkers now completely abstaining; (c) current occasional drinkers, up to one drink per week (<1.30 g per day); (d) low-volume drinkers, up to two drinks or 1.30–24.9 g per day; (e) medium-volume drinkers, up to four drinks or 25–44.9 g per day; (f) high-volume drinkers, up to six drinks or 45–64.9 g per day; and (g) higher volume drinkers, more than six drinks or 65 g per day. The adequacy of methods used by studies also was assessed against minimal quality criteria as discussed above.

Median cohort age, sex, country, quality of drinking measure, and abstainer bias variables were included as covariates in adjusted models. The country in which a study was conducted was dichotomized into those with mainly Caucasian populations versus without variable to reflect evidence that health protection from moderate drinking was more likely to be observed among Caucasians (Kerr et al., 2011).

Other covariates were selected for inclusion on empirical grounds based on p values of bivariable tests of the natural log-RR and each covariate, and significant correlations with other variables. Based on bivariable analysis of the data set, any variable producing a bivariable test result with p < .20 was considered a candidate for the multivariable regression analyses of the natural log-RR of all-cause mortality and alcohol consumption (Hosmer & Lemeshow, 2000). Independent variables with particularly high intercorrelations (>.30) were identified, less precise measures were excluded (Stokes et al., 2000), and the variable pool was reduced to avoid synonymous variables and collinearity (Pagano & Gauvreau, 2000).

Median age of a study cohort was treated as a continuous variable, whereas other variables with fixed multiple response options were all reduced to two or three category variables (see Table 1 and Table D1 in online Appendix D) to remove options with few or no values and to make the models more efficient. Median age of the study population at intake, gender, or being a mainly Caucasian versus non-Caucasian population were tested as possible effect modifiers of the relationship between alcohol consumption and mortality. No significant interactions were observed; therefore, pooled meta-analyses of all studies are presented.

Table 1.

Characteristics of all studies included in meta-analyses on alcohol use and all-cause mortality

Open in a separate window

	Sample size
	Studies (N = 87)		Risk estimates (N = 523)
Study characteristics	n	%	n	%
Gender
Male	61	52.14	276	52.77
Female	37	31.62	178	34.03
Both	19	16.24	69	13.19
Age (Mdn = 54.5 years, SD = 10.5)
19–49	32	36.78	189	36.14
50–59	29	33.33	194	39.09
60–78	26	29.89	140	26.77
Mainly Caucasian vs not
North America, Europe, Australia	77	88.51	460	87.95
Japan, China, India	10	11.47	63	12.05
Years of follow-up (M = 13.4 years, SD = 6.6)
3.7–9	37	42.53	191	36.52
10–15	22	25.29	178	34.03
16–40	28	32.18	154	29.45
Disease
No exclusion	53	60.92	320	61.19
Unhealthy excluded	34	39.08	203	38.81
Daily alcohol intake measure
Adequate	64	73.56	394	75.33
Not adequate	23	26.44	129	24.67
Studies with abstainer biasesa
Both former and occasional drinker biases	41	47.13	180	34.42
Former drinker bias only	24	27.59	140	26.77
Occasional drinker bias only	9	10.34	76	14.53
No bias	13	14.94	127	24.28

Open in a separate window

Notes: Mdn = median.

^aBoth = studies in which reference group included both former and occasional drinkers; former = studies in which reference group included former drinkers; occasional = studies in which reference group included occasional/low-volume but not former drinkers, and no bias = studies in which the reference group only included lifetime abstainers.

All significance tests assumed two-tailed p values or 95% CIs. All statistical analyses were performed using SAS Version 9.3 (SAS Institute Inc., Cary, NC), and the SAS PROC MIXED procedure was used to model the log-transformed RR.

Study characteristics including controls for potential bias

The 87 selected studies included a total of 523 estimates of the risk relationships between levels of alcohol consumption and all-cause mortality. Among these, 30 studies reported separate estimates for men and women, 31 for men only, 7 for women only, and 19 for both combined. Only 13 of these studies (127 risk estimates) were coded as free of abstainer biases because they strictly defined lifetime abstainers as the reference group. Table 1 summarizes these and other study characteristics. A summary of potentially confounding variables controlled for or not in each study is provided in online Appendix D (Table D1).

Results of individual studies

Two forest plots illustrate the range of RR estimates for (a) any level of drinking (Figure C2 in online Appendix C) and (b) low-volume drinking (Figure 2) across individual studies (Woodward, 2000). Consistent with most previous meta-analyses, these indicate (a) a wide range of estimates across different studies and (b) mean estimates for low-volume drinking significantly below unity, indicating health protection in comparison with abstainers. When all drinking outcomes are considered collectively in each study and compared against those for abstainers, no significant overall difference is observed, although again there is great variation across studies (see also Figure C2 in online Appendix C).

Open in a separate window

Figure 2.

Estimates of the relative risk of all-cause mortality associated with low-volume drinking in 81 studies. CI = confidence interval.

Limitations

A number of limitations and caveats around our findings need to be acknowledged. A major limitation involves imperfect measurement of alcohol consumption in most included studies. Self-reported alcohol consumption is mostly under-reported (Stockwell et al., 2014), and even the classification of drinkers as lifetime abstainers can be unreliable (Kerr et al., 2002). The number of available studies in some stratified analyses was small, and therefore there may be limited power to control for potential study-level confounders. However, the required number of estimates per variable for linear regression can be much smaller than in logistic regression, and a minimum of at least two estimates per variable is recommended for linear regression analysis (Austin & Steyerberg, 2015), suggesting that the sample sizes were adequate in all models presented.

It has been demonstrated that a pattern of heavy episodic (i.e., “binge”) drinking is not associated with the appearance of reduced health risks even when average daily volume is low (Roerecke & Rehm, 2010). Too few studies adequately controlled for this variable to investigate its effect on different outcomes across studies. Finally, our findings only apply to the net effect of alcohol at different doses on all-cause mortality, and different risk relationships likely apply for specific disease categories.

We acknowledge with deep gratitude the contributions of our late colleague and principal investigator Kaye Middleton Fillmore, Ph.D., to our thinking and the analyses reported in this article.