Coffee Consumption and Mortality
Coffee Consumption and Mortality
We performed a literature search in the PubMed database for articles published from January 1966 through December 2013, using the terms "(prospective or cohort) and (fatal or death or mortality) and (hot beverages or coffee or caffeine)." The search was limited to studies carried out in humans. We followed the Meta-Analysis of Observational Studies in Epidemiology (MOOSE) guidelines for conducting meta-analyses and reporting results. Two authors (A.C., A.D.) separately retrieved the studies reporting data on the association between coffee consumption and all-cause mortality, as well as associations for CVD and all-cancer mortality. Discrepancies were discussed and resolved.
Studies were eligible for inclusion in the meta-analysis if they met the following criteria: 1) the study had a prospective design; 2) the exposure of interest was coffee consumption; 3) the outcome was all-cause mortality, CVD mortality, and/or all-cancer mortality; 4) the investigators reported relative risks with 95% confidence intervals for 3 or more quantitative categories of coffee consumption; and 5) the reported relative risks had been adjusted at least for smoking status.
The following information was extracted from each study: first author's surname, publication year, study location, study period, duration of follow-up (years), sex, number of subjects (total number of deaths and total cohort size or total number of deaths and person-years of follow-up), mortality outcomes, coffee consumption categories, type of coffee, covariates adjusted for in the multivariable analysis, and relative risks (with their 95% confidence intervals) for all categories of coffee consumption. We extracted the relative risks that reflected the greatest degree of adjustment for potentially confounding variables. If investigators reported the adjusted relative risks but not the corresponding confidence intervals, we calculated the confidence intervals for the crude relative risks and related them to the adjusted relative risks. For studies that presented data separately on both coronary heart disease and stroke, we combined the results as indicated by Hamling et al.
For each study, the median or mean coffee consumption within each exposure interval was assigned the corresponding relative risk. When median or mean consumption per category was not reported, we assigned the midpoint of the upper and lower boundaries for each category as the average consumption. If the upper bound for the highest category was not provided, we assumed that the category had the same amplitude as the adjacent one.
We performed a 2-stage random-effects dose-response meta-analysis to examine a potential nonlinear relationship between coffee consumption and 3 different outcomes: all-cause mortality, CVD mortality, and cancer mortality. This was done by modeling coffee consumption using restricted cubic splines with 3 knots at fixed percentiles (25%, 50%, and 75%) of the distribution. In the first stage, a restricted cubic spline model with 2 spline transformations (3 knots minus 1) was fitted taking into account the correlation within each set of published relative risks. In the second stage, we combined the 2 regression coefficients and the variance/covariance matrices that had been estimated within each study, using the multivariate extension of the method of moments in a multivariate random-effects meta-analysis. We calculated an overall P value by testing that the 2 regression coefficients were simultaneously equal to zero. We calculated a P value for nonlinearity by testing that the coefficient of the second spline was equal to zero.
We excluded from the main analysis those studies that did not report the number of subjects (total number of deaths and total cohort size or total number of deaths and person-years of follow-up) in order to avoid biases in the estimates for the variances. We considered the excluded studies in a sensitivity analysis.
We performed stratified analysis by study location, sex, type of smoking adjustment (smoking status, categories of cigarette smoking, or number of cigarettes smoked per day (continuous variable)), and alcohol adjustment. Statistical heterogeneity among studies was assessed using the χ test and was defined as a P value less than 0.10. Statistical heterogeneity was further quantified through the multivariate generalization of the I statistic. Low heterogeneity is defined by I values less than 25%, while values greater than 75% are indicative of high heterogeneity. Publication bias was assessed with Egger's regression test. All statistical analyses were conducted with the dosresmeta and metafor packages in R (R Foundation for Statistical Computing, Vienna, Austria).P values less than 0.05 were considered statistically significant.
Methods
Literature Search and Selection
We performed a literature search in the PubMed database for articles published from January 1966 through December 2013, using the terms "(prospective or cohort) and (fatal or death or mortality) and (hot beverages or coffee or caffeine)." The search was limited to studies carried out in humans. We followed the Meta-Analysis of Observational Studies in Epidemiology (MOOSE) guidelines for conducting meta-analyses and reporting results. Two authors (A.C., A.D.) separately retrieved the studies reporting data on the association between coffee consumption and all-cause mortality, as well as associations for CVD and all-cancer mortality. Discrepancies were discussed and resolved.
Studies were eligible for inclusion in the meta-analysis if they met the following criteria: 1) the study had a prospective design; 2) the exposure of interest was coffee consumption; 3) the outcome was all-cause mortality, CVD mortality, and/or all-cancer mortality; 4) the investigators reported relative risks with 95% confidence intervals for 3 or more quantitative categories of coffee consumption; and 5) the reported relative risks had been adjusted at least for smoking status.
Data Extraction
The following information was extracted from each study: first author's surname, publication year, study location, study period, duration of follow-up (years), sex, number of subjects (total number of deaths and total cohort size or total number of deaths and person-years of follow-up), mortality outcomes, coffee consumption categories, type of coffee, covariates adjusted for in the multivariable analysis, and relative risks (with their 95% confidence intervals) for all categories of coffee consumption. We extracted the relative risks that reflected the greatest degree of adjustment for potentially confounding variables. If investigators reported the adjusted relative risks but not the corresponding confidence intervals, we calculated the confidence intervals for the crude relative risks and related them to the adjusted relative risks. For studies that presented data separately on both coronary heart disease and stroke, we combined the results as indicated by Hamling et al.
For each study, the median or mean coffee consumption within each exposure interval was assigned the corresponding relative risk. When median or mean consumption per category was not reported, we assigned the midpoint of the upper and lower boundaries for each category as the average consumption. If the upper bound for the highest category was not provided, we assumed that the category had the same amplitude as the adjacent one.
Statistical Analysis
We performed a 2-stage random-effects dose-response meta-analysis to examine a potential nonlinear relationship between coffee consumption and 3 different outcomes: all-cause mortality, CVD mortality, and cancer mortality. This was done by modeling coffee consumption using restricted cubic splines with 3 knots at fixed percentiles (25%, 50%, and 75%) of the distribution. In the first stage, a restricted cubic spline model with 2 spline transformations (3 knots minus 1) was fitted taking into account the correlation within each set of published relative risks. In the second stage, we combined the 2 regression coefficients and the variance/covariance matrices that had been estimated within each study, using the multivariate extension of the method of moments in a multivariate random-effects meta-analysis. We calculated an overall P value by testing that the 2 regression coefficients were simultaneously equal to zero. We calculated a P value for nonlinearity by testing that the coefficient of the second spline was equal to zero.
We excluded from the main analysis those studies that did not report the number of subjects (total number of deaths and total cohort size or total number of deaths and person-years of follow-up) in order to avoid biases in the estimates for the variances. We considered the excluded studies in a sensitivity analysis.
We performed stratified analysis by study location, sex, type of smoking adjustment (smoking status, categories of cigarette smoking, or number of cigarettes smoked per day (continuous variable)), and alcohol adjustment. Statistical heterogeneity among studies was assessed using the χ test and was defined as a P value less than 0.10. Statistical heterogeneity was further quantified through the multivariate generalization of the I statistic. Low heterogeneity is defined by I values less than 25%, while values greater than 75% are indicative of high heterogeneity. Publication bias was assessed with Egger's regression test. All statistical analyses were conducted with the dosresmeta and metafor packages in R (R Foundation for Statistical Computing, Vienna, Austria).P values less than 0.05 were considered statistically significant.
Source...