Gütemaße und Kriterien bei der Anwendung von Propensity Scores

Abstract

Propensity scores (PS) have been established as a valid alternative to conventional regression models when evaluating non-randomized treatment studies. The PS describes the probability for an individual to receive a treatment, conditional on a set of observed covariates. PS analyses are performed in two steps. In the first step, the PS is generally estimated via logistic regression. In the second step, the actual treatment effect is estimated. The quality of a PS analysis depends on whether it is possible to achieve a sufficient balance of the patient characteristics in the treatment groups in the first step. This is the only way to ensure that these patient characteristics do not bias the estimate of the treatment effect. Various measures have been proposed to measure this balance, e.g. the standardized difference or the z-difference. Closely related to the balance of patient characteristics (and thus also a measure of the quality for a PS model) is the overlap, i.e. the similarity in the distribution of the estimated propensity scores in the two treatment groups. A valid comparison of the treatments is only possible in regions of sufficient overlap. In this article, the two concepts of balance and overlap are presented and discussed using an example from cardiac surgery.