select search filters
roundups & rapid reactions
before the headlines
Fiona fox's blog

health effects of alcohol across different age groups

A paper published in the BMJ has examined the effect of alcohol on the health of various age groups. They analysed mortality and report that previously beneficial effects of alcohol may be strongest in women of over 65. Roundup comments accompanied this analysis.


Title, Date of Publication & Journal

All cause mortality and the case for age specific alcohol consumption guidelines: pooled analyses of up to 10 population based cohorts

10 Feb 2015



Study’s main claims – and are they supported by the data

– The paper suggests that previous studies indicating low alcohol consumption is associated with a protective effect against cardiovascular disease and all-cause mortality have been due to improper reference comparison and failing to account for important factors.

– The two key points of the paper: comparisons to a highly variable reference group, and failing to adjust analyses for important factors, are valid points.

– However, and importantly, the paper does not provide strong evidence against a protective effect of low alcohol consumption.




– A great strength of the dataset is the larger age range and more population representative sample (although, not actually population representative, issues of non-response to the survey are discussed but due to only recently being able to re-weight the study to account for this, the raw data are used in this analysis) compared to many studies.

– The authors correctly state that many studies pre-screen individuals with prior medical conditions and co-morbidities at the start of follow-up, which may lead to under estimating the impact of alcohol on all-cause mortality. However, within their own analysis, there are no covariates to reflect co-morbidities and thus these are not adjusted for.



– The title of the paper, and some sentences, are slightly mis-leading as to the scale and strength of the underlying dataset. First, although there are strictly 10 cohorts, they are all part of the same longitudinal study. This is not a meta-analysis combining individual patient data from multiple studies. Second, it is slightly unclear from the paper whether the same individuals are present in multiple cohorts. It is implied that even if they are, this fact is not linked between cohorts in the authors’ dataset.

– Although the paper refers to alcohol-related mortality several times, the data and analysis are only about all-cause mortality. Categorising deaths into alcohol-related or not is a complex analysis with numerous sources of error and mis-classification. From this analysis, it is not possible to answer whether low alcohol consumption increases risk of alcohol-related mortality but decreases risk of non-alcohol-related mortality.

– A significant limitation of the study is the inclusion of many covariates (factors such as social class, ethnicity, etc.) as an attempt to account for other reasons for a difference in all-cause mortality. However, as seen in Figure 1, the cost of including these covariates is attrition of the sample. In Figure 1, we see that for the weekly data, “Insufficient covariates available” results in the loss of 24,729 individuals – approximately a third of the original sample are dropped due to missing some/all measures of BMI, social class, ethnicity, etc.

– The authors themselves note that self-reported alcohol consumption has been shown to be quite unreliable, and the possibility of self-selection biases in the covariates reported cannot be ignored.

– Unlike previous studies, this paper splits individuals who have never drank and former drinkers into two groups. This resulted in the previously seen protective effect of low alcohol consumption being not statistically significant in this dataset. Splitting the reference group has reduced the power of the study to detect a true difference.

– Unlike previous studies that have not adjusted for covariates, the authors have limited their sample size and reduced the power to detect effects (which we expect to be small, and so a loss of statistical power to detect them may be the reason for the null result – the null result being there is no evidence against the hazard of all-cause mortality being the same regardless of alcohol consumption).

– Previous research has suggested a J-shaped dose-response to alcohol consumption, i.e. low levels provide a protective effect and high levels increase risk of mortality. Tables 3 and 4 illustrate the authors’ hypothesis that after changing the reference group and adjusting for covariates none of the low alcohol consumption groups have a reduced hazard ratio (as none of the p-values are significant). However, note that the high alcohol groups show no increased risk (hazard ratio greater than one).

– Although not statistically significant, the pattern of hazard ratios still shows a decrease at low-consumption. However, possibly due in part to the sample size, many of the results are not statistically significant, but the presented tables do not completely rule out the possibility of a protective effect.

– An important unanswered question is that of interpreting the possible correlations between alcohol consumption and the adjusted covariates. The hazard ratios of the adjusted covariates are not reported in the tables, but we cannot causally disentangle the two if they are highly correlated. Certain groups (as defined by covariates) may tend towards specific alcohol consumption levels. So although the protective effect seems to be attenuated, if we consider that some of the adjusting covariates may be merely proxy measures of alcohol consumption, then there could still be a protective effect.


In summary, the article presents an interesting analysis of a complex issue and attempts to account for two issues with previous studies (adjusting for covariates and composite reference group). The final analysis does not provide strong evidence against a protective effect of low alcohol consumption.


Before The Headlines is a service provided to the SMC by volunteer statisticians: members of the Royal Statistical Society (RSS), Statisticians in the Pharmaceutical Industry (PSI) and experienced statisticians in academia and research.  A list of contributors, including affiliations, is available here.

in this section

filter Headlines by year

search by tag