Performance of tests of significance based on stratification by a multivariate confounder score or by a propensity score

Abstract

Stratification by a multivariate confounder score or by a propensity score has been proposed for the multi-confounder situations that are commonly encountered in epidemiologic evaluations of the effects of a treatment or a triage decision. However, the use of these scores in clinical research has been limited, perhaps in part because of the concern that the stated level of statistical significance may be exaggerated when there is a high degree of correlation between the exposure and the set of confounders. We present a specific example and computer simulations to suggest that exaggeration of statistical significance occurs only under unusual circumstances when the correlation between the exposure and the confounders is extreme. Our simulations also suggest that an analysis based on stratification by a propensity score is less affected by high correlation between the exposure and the confounders than is an analysis based on a multivariate confounder score.