Factor selection in observational studies - An application of nonlinear factor selection to propensity scores

Abstract

Observational studies have become a major research methodology in scientific disciplines where experiments are hard to perform. These most notably include the health sciences, social sciences and (macro-)economics. It is difficult to estimate treatment effects due to the non-randomised character of these studies. Propensity scores solve this problem to some extent by incorporating the control variables into one measure (Rosenbaum and Rubin 1983). An estimation of these propensity scores can be performed by any method of categorical regression (logit etc.). Nevertheless, estimated effects are very sensitive to the propensity score (Drake 1993; Heckman et al. 1998; Brookhart et al. 2006) and for this reason, the propensity score should be estimated with great care. In order to avoid a poor generalisation performance of the propensity score estimation, it is important to choose factors appropriately. Unfortunately, it is not possible to select factors during the estimation of the propensity score and independently from treatment calculation. In this paper, an integrated factor selection method is proposed, which considers the treatment effect estimation during the propensity score estimation. © 2010 Springer-Verlag Berlin Heidelberg.