Linkage failures in ecological studies.

Abstract

Ecological studies require a methodological theory distinct from that used in individual-level epidemiological studies. This article discusses the special problems that need to be considered when planning ecological studies or using the results of such studies. Ecological studies are much more sensitive to bias from model mis-specification than are results from individual-level studies. For example, deviations from linearity in the underlying individual-level regressions can lead to inability to control for confounding in ecological studies, even if no misclassification is present. Conditions for confounding differ in individual-level and ecological analyses. For ecological analyses of means, for example, a covariate will not be a confounder if its mean value in a study region is not associated with either (i) the mean exposure level across regions, or (ii) the mean outcome (disease rate) across regions. On the other hand, effect modification across areas can induce ecological bias even when the number of areas is very large and there is no confounding. In contrast to individual-level studies, independent and nondifferential misclassification of a dichotomous exposure usually leads to bias away from the null hypothesis in aggregate data studies. Failure to standardize disease, exposure and covariate data for other confounders (not included in the regression model) can lead to bias. It should be borne in mind that there is no method available to identify or measure ecological bias. While this conclusion may sound like a general criticism of ecological studies, it is not. It does, however, serve as a reminder of the problems that need to be considered when one designs, analyses, or critically evaluates ecological studies.