Alcohol Drinking and Cutaneous Melanoma Risk
Alcohol Drinking and Cutaneous Melanoma Risk
We carried out a systematic literature search in Medline, using PubMed, for all epidemiological studies published as original articles in English up to 30 April 2012, investigating the association between alcohol drinking and CM. We followed the Meta-Analysis of Observational Studies in Epidemiology guidelines. For the literature search, we used the following search string: [(ethanol OR alcohol drinking) AND (skin neoplasms OR melanoma) OR (alcohol OR alcoholic beverages OR ethanol OR alcohol drinking) AND melanoma], which comprises the following medical subject heading terms: 'Ethanol', 'Alcohol Drinking', 'Alcoholic Beverages' and 'Melanoma' or 'Skin Neoplasms'. The process for article selection is shown in Figure 1. No studies were excluded a priori for design weakness or low-quality data. Three investigators (M.R., E.P. and L.S.) independently screened each retrieved study for inclusion in the meta-analysis. In case of doubts or disagreement, a fourth investigator (V.B.) was consulted, and consensus was reached. We retrieved a total of 1044 published papers, of which 997 were excluded as they were not relevant to the topic of our meta-analysis. From a detailed review of the reference lists of the remaining 47 potentially relevant articles, we identified two further publications of interest. From a total of 49 articles, 33 were excluded because they did not satisfy the inclusion criteria: (i) studies investigating nonmelanocytic skin cancer only; (ii) studies reporting neither relative risks (RRs) nor odds ratios (ORs) and the corresponding 95% confidence intervals (CIs), or sufficient information to calculate them; (iii) studies conducted on special populations (e.g. alcoholics or cancer survivors); and (iv) studies reporting only the result for specific alcoholic beverages (e.g. beer, wine or liquor/spirit). The latter studies were not included in the analyses as nondrinkers of a specific alcoholic beverage may drink other beverages, leading to a likely underestimation of the association.
(Enlarge Image)
Figure 1.
Flowchart of study selection.
Finally, 16 studies were included in this meta-analysis: 14 case–control and two prospective cohort studies. For each study, we extracted the following information: study design, location, number of subjects (cases, controls or cohort size), sex, type of controls (hospital or population based) and period of enrolment for case–control studies, duration of follow-up for cohort studies, RR estimates for categories of alcohol consumption along with the corresponding 95% CIs, and the variables that were adjusted for/matched in the analysis.
We used ORs, RRs and hazard ratios as comparable estimates of the RR. We extracted multivariate-adjusted RR estimates whenever available. If RRs were not reported, we computed crude RRs using the frequency distributions presented in the original reports.
As different units were used to express the amount of alcohol consumed, we converted all measures into grams of ethanol per day as a standard measurement unit, defining 1 drink as 12·5 g of ethanol if not otherwise specified in the original report, 1 mL as 0·80 g, and 1 ounce as 28·35 g of ethanol. The dose associated with each RR estimate was computed as the midpoint of each exposure category, and for the open-ended upper category, as 1·2 times its lower boundary. When possible, we chose nondrinkers as the reference category, but in some studies occasional drinkers were also included. In the Million Women Study, we derived the floated variances – which describe the uncertainty in RRs without reference to a predefined category – from the 95% floated CI provided by the authors, in order to derive RRs and corresponding 95% CIs for different categories of alcohol consumption compared with nondrinkers.
We defined the daily amount of alcohol consumption as light (≤ 1 drink, or ≤ 12·5 g of ethanol) or moderate to heavy (> 1 drink, or > 12·5 g of ethanol per day). As only two studies investigated high daily amounts of alcohol, we could not examine the effect of heavy drinking (> 50 g of ethanol per day) on the risk of CM. When more than one category of alcohol consumption fell in the same level, we combined the corresponding estimates using the method proposed by Hamling et al., which takes into account the correlation between estimates. This method uses the dose-specific covariate-adjusted risk estimates and the numbers of cases and noncases for each category of exposure to derive a set of pseudonumbers of cases and noncases consistent with both of the adjusted estimates. The pseudonumbers of two or more categories of exposure can then be combined to provide adjusted risk estimates for light or moderate-to-heavy alcohol drinking.
All of the meta-analytical estimates were obtained using random effects models. Between-study heterogeneity was assessed using the χ-test, and inconsistency was measured using Higgins I statistics, which gives the proportion of total variation contributed by between-study variance.
We conducted sensitivity analyses by excluding one study at a time from the meta-analysis, in order to evaluate each study's impact on the final pooled estimate. In order to investigate possible sources of between-study heterogeneity, we conducted stratified analyses according to potentially relevant factors (study design, source of controls for case–control studies, geographical area, sex and adjustment for sun exposure).
We assessed the dose–response relationship between alcohol intake and CM using flexible nonlinear meta-regression models. In this analysis, we considered only studies reporting RRs estimates for at least three exposure categories, including the referent category.
The presence of publication bias was assessed by examination of the contour-enhanced funnel plot, and also by applying the Egger's test for funnel-plot asymmetry.
Statistical analyses were performed using SAS (version 9.1.3; SAS Institute Inc., Cary, NC, U.S.A.) and STATA (version 11; StataCorp, College Station, TX, U.S.A.).
Materials and Methods
Identification of Studies and Data Collection
We carried out a systematic literature search in Medline, using PubMed, for all epidemiological studies published as original articles in English up to 30 April 2012, investigating the association between alcohol drinking and CM. We followed the Meta-Analysis of Observational Studies in Epidemiology guidelines. For the literature search, we used the following search string: [(ethanol OR alcohol drinking) AND (skin neoplasms OR melanoma) OR (alcohol OR alcoholic beverages OR ethanol OR alcohol drinking) AND melanoma], which comprises the following medical subject heading terms: 'Ethanol', 'Alcohol Drinking', 'Alcoholic Beverages' and 'Melanoma' or 'Skin Neoplasms'. The process for article selection is shown in Figure 1. No studies were excluded a priori for design weakness or low-quality data. Three investigators (M.R., E.P. and L.S.) independently screened each retrieved study for inclusion in the meta-analysis. In case of doubts or disagreement, a fourth investigator (V.B.) was consulted, and consensus was reached. We retrieved a total of 1044 published papers, of which 997 were excluded as they were not relevant to the topic of our meta-analysis. From a detailed review of the reference lists of the remaining 47 potentially relevant articles, we identified two further publications of interest. From a total of 49 articles, 33 were excluded because they did not satisfy the inclusion criteria: (i) studies investigating nonmelanocytic skin cancer only; (ii) studies reporting neither relative risks (RRs) nor odds ratios (ORs) and the corresponding 95% confidence intervals (CIs), or sufficient information to calculate them; (iii) studies conducted on special populations (e.g. alcoholics or cancer survivors); and (iv) studies reporting only the result for specific alcoholic beverages (e.g. beer, wine or liquor/spirit). The latter studies were not included in the analyses as nondrinkers of a specific alcoholic beverage may drink other beverages, leading to a likely underestimation of the association.
(Enlarge Image)
Figure 1.
Flowchart of study selection.
Finally, 16 studies were included in this meta-analysis: 14 case–control and two prospective cohort studies. For each study, we extracted the following information: study design, location, number of subjects (cases, controls or cohort size), sex, type of controls (hospital or population based) and period of enrolment for case–control studies, duration of follow-up for cohort studies, RR estimates for categories of alcohol consumption along with the corresponding 95% CIs, and the variables that were adjusted for/matched in the analysis.
Statistical Analyses
We used ORs, RRs and hazard ratios as comparable estimates of the RR. We extracted multivariate-adjusted RR estimates whenever available. If RRs were not reported, we computed crude RRs using the frequency distributions presented in the original reports.
As different units were used to express the amount of alcohol consumed, we converted all measures into grams of ethanol per day as a standard measurement unit, defining 1 drink as 12·5 g of ethanol if not otherwise specified in the original report, 1 mL as 0·80 g, and 1 ounce as 28·35 g of ethanol. The dose associated with each RR estimate was computed as the midpoint of each exposure category, and for the open-ended upper category, as 1·2 times its lower boundary. When possible, we chose nondrinkers as the reference category, but in some studies occasional drinkers were also included. In the Million Women Study, we derived the floated variances – which describe the uncertainty in RRs without reference to a predefined category – from the 95% floated CI provided by the authors, in order to derive RRs and corresponding 95% CIs for different categories of alcohol consumption compared with nondrinkers.
We defined the daily amount of alcohol consumption as light (≤ 1 drink, or ≤ 12·5 g of ethanol) or moderate to heavy (> 1 drink, or > 12·5 g of ethanol per day). As only two studies investigated high daily amounts of alcohol, we could not examine the effect of heavy drinking (> 50 g of ethanol per day) on the risk of CM. When more than one category of alcohol consumption fell in the same level, we combined the corresponding estimates using the method proposed by Hamling et al., which takes into account the correlation between estimates. This method uses the dose-specific covariate-adjusted risk estimates and the numbers of cases and noncases for each category of exposure to derive a set of pseudonumbers of cases and noncases consistent with both of the adjusted estimates. The pseudonumbers of two or more categories of exposure can then be combined to provide adjusted risk estimates for light or moderate-to-heavy alcohol drinking.
All of the meta-analytical estimates were obtained using random effects models. Between-study heterogeneity was assessed using the χ-test, and inconsistency was measured using Higgins I statistics, which gives the proportion of total variation contributed by between-study variance.
We conducted sensitivity analyses by excluding one study at a time from the meta-analysis, in order to evaluate each study's impact on the final pooled estimate. In order to investigate possible sources of between-study heterogeneity, we conducted stratified analyses according to potentially relevant factors (study design, source of controls for case–control studies, geographical area, sex and adjustment for sun exposure).
We assessed the dose–response relationship between alcohol intake and CM using flexible nonlinear meta-regression models. In this analysis, we considered only studies reporting RRs estimates for at least three exposure categories, including the referent category.
The presence of publication bias was assessed by examination of the contour-enhanced funnel plot, and also by applying the Egger's test for funnel-plot asymmetry.
Statistical analyses were performed using SAS (version 9.1.3; SAS Institute Inc., Cary, NC, U.S.A.) and STATA (version 11; StataCorp, College Station, TX, U.S.A.).