Jon Wakefield

"Ecological Inference in Epidemiology"

Presented to the Center for Statistics and the Social Sciences, University of Washington, 12 March 2002.

Ecological regression studies in which an aggregate level response is regressed upon an aggregate level predictor, are a dangerous enterprise since relationships at the level of the group only agree with relationships at the level of the individual under very strict circumstances. This can lead to the possibility of the so-called ecological fallacy which occurs because of within-group variability in exposures and confounders. Usual biases in individual observational studies are far more complicated in an ecological setting, and there are additional biases that are unique to ecological studies. In this talk I will describe the various types of bias and characterize the direction and magnitude in a number of situations. Examples from spatial epidemiology will be presented, in particular the association between cancer incidence and socio-economic variables.