Selection bias in ecological studies

An ecological study is an observational epidemiological study which uses aggregated data instead of individual data, where the aggregation is usually done at some geographical level such as a town, a region, or a country.1 They are typically carried out when an exposure of interest is fully defined at the group level.2 A classical example is the association between air pollution and a specific health outcome. While it is difficult (or even meaningless) to measure individual exposure to air pollution, it is easier to measure air quality in a region. In order to investigate the association between air pollution and the occurrence of some disease, one may correlate the available measures of air quality with the prevalence (or incidence) of the disease over a set of regions. This approach faces several limitations; in the literature about ecological studies, one can find numerous warnings against the ‘ecological fallacy’, which states that a correlation at the group level is not of the same magnitude (and not always with the same sign) than a correlation at the individual level,3 and about confounding.4 The present communication aims to be a reminder about another serious limitation of ecological studies which is rarely mentioned in the literature: selection bias. Yet, as illustrated below, a deception due to a selection bias might be still more spectacular than any confounding issue.