Light Alcohol Drinking and Cancer

A Meta-Analysis

V. Bagnardi; M. Rota; E. Botteri; I. Tramacere; F. Islami; V. Fedirko; L. Scotti; M. Jenab; F. Turati; E. Pasquali; C. Pelucchi; R. Bellocco; E. Negri; G. Corrao; J. Rehm; P. Boffetta; C. La Vecchia


Ann Oncol. 2013;24(2):301-308. 

In This Article

Materials and Methods

Search Rationale

We restricted the investigation on light drinking to those tumor sites for which there is evidence of an increased risk associated with alcohol consumption in general. For this purpose, we started from the indications drawn up by the International Agency for Research on Cancer,[3] which listed all the tumor sites for which there is sufficient evidence for carcinogenicity (oral cavity and pharynx, larynx, esophagus, liver, colorectum and female breast). For the esophagus, we decided to restrict our analyses to esophageal squamous cell carcinoma (SCC), since the alcohol–cancer association was mostly appreciable in that histological subtype.[7,8]

Search Strategy

We carried out a literature search in Medline, ISI Web of Science (Science Citation Index Expanded) and EMBASE for epidemiological studies published before December 2010. For the sake of completeness, we also reviewed references from all relevant studies, reviews and meta-analyses published on the alcohol–cancer association to identify additional studies. The key words used for the literature search are reported in supplementary material S1, available at Annals of Oncology online. We limited our search to studies published in English.

Inclusion Criteria

Articles were included in the meta-analysis only if they satisfied the following criteria:

  • Case–control or cohort studies published as original articles (abstracts, letters, reviews and meta-analyses were excluded).

  • Studies that reported findings expressed as odds ratio (OR), relative risk (RR) or hazard ratio (or reporting sufficient data to compute them) for light drinkers (≤12.5 g ethanol; ≤1 drink) versus non-drinkers.

  • Studies that reported standard errors or confidence intervals (CIs) of the risk estimates, or provided sufficient data to calculate them.

We excluded studies reporting on a specific type of alcoholic beverage only (e.g. beer only) because in those studies the non-drinkers of a specific beverage were possibly drinkers of other types of alcoholic beverages.

Data Abstraction

The reports available for each cancer site were independently reviewed by one of the authors to determine the eligibility of each article for inclusion in the meta-analysis. Doubts or disagreement were resolved by consensus among all the investigators. When the results of the same study were published in more than one paper, only the most recent and/or complete article was included in the analysis.

For each included study, we extracted details on study design, outcome, country, gender, RR estimates and 95% CIs, adjustment variables, and, when available, the number of cases and controls (case–control studies) or number of events and subjects at risk/person-years (cohort studies) for light drinkers and non-drinkers. We also recorded whether the reference category of non-drinkers included occasional drinkers or not. Where possible, separate estimates were extracted for males and females.

Since different studies used different units of measurement to express alcohol consumption (grams, milliliters, ounces or drinks consumed every day, week, month or year) we used grams per day (g/day) as a standard measure of ethanol intake using the following equivalencies: 0.8 g/ml = 28 g/ounce = 12.5 g/drink. Moreover, since the included studies usually reported alcohol exposure in intervals, we decided to consider as light every interval whose midpoint was ≤12.5 g per day (or one drink per day) of alcohol. Also, some studies reported two or more adjusted risk estimates for light drinking (e.g. 6 g/day and 12 g/day). In that case, we combined them into a single estimate using the method for pooling non-independent estimates described by Hamling et al.[9] This method uses the number of exposed to different levels of alcohol and non-exposed subjects and the associated reported risk estimates to derive a set of pseudo-numbers of cases and controls/subjects at risk, by taking into account the correlation between the original estimates due to the common reference group. These pseudo-numbers can then be used to calculate a single pooled adjusted risk estimate and CI.

Statistical Methods

Because cancer is a relatively rare outcome, we assumed that ORs, risk ratios and rate ratios were all comparable estimates of the RR.[10] Measures of association and the corresponding CIs were translated into log(RR)s and the corresponding variances.

We computed a pooled RR of site-specific cancer for light drinkers versus non-drinkers, using random-effects models. We used random-effects models to estimate pooled RRs in order to take into account the heterogeneity, albeit small, of the risk estimates. Each study log(RR) was weighted by the inverse of its variance. Weights were taken equal to the inverse of the reported within-study variance plus the between-study variance component τ2. The moment estimator of the between-study variance was used.[11]

We evaluated the statistical heterogeneity among the studies using I2, which is the proportion of total variation contributed by the between-study variance.[12] We examined the publication bias through the funnel plots and the Begg's rank correlation method.[13]

We carried out subgroup analyses and meta-regression models to investigate potential sources of between-study heterogeneity. We used a chi-square statistic to test for differences of summary estimates among the subgroups.[10]

We estimated the proportion and number of cancer deaths attributable to light alcohol drinking and to alcohol drinking at any dose using the methods described in Gmel et al.[14] For each cancer site and World Health Organization (WHO) subregion, we obtained the age-specific and dose-specific distribution of drinkers among adults for 2004 from the Global Burden of Disease project,[15] along with number of deaths. For light drinking (up to 1 drink/day), we considered the pooled RRs estimated in the present meta-analysis. For drinking at any dose, we used the dose-specific RRs estimated in the meta-analysis of Bagnardi et al.[4] The RRs were specific for sex, but not for age or WHO subregion.

We carried out all analyses using SAS software, version 9.1 (SAS Institute Inc., Cary, North Carolina). All P values were two-sided.