Strauss D
Department of Statistics, University of California, Riverside 92521, USA.
J Clin Epidemiol. 1998 Mar;51(3):233-6. doi: 10.1016/s0895-4356(97)00283-7.
Consider a regression model for the effect of a treatment on an outcome variable in the presence of potential confounders. It is common to test for main effects and interactions of the treatment variable; if the confounding variables are discrete, one might then compare the treatment groups within strata formed by suitable covariate patterns. An alternative, due to Miettinen, is to stratify according to a "multivariate confounder score" and test for treatment effects within strata. This test has been shown to be flawed, and the method appears largely to have fallen into disuse. Here we show that such stratification nevertheless provides a sound and often useful graphical comparison of treatment and control groups.
