On the Propensity Score Can Result in Biased Estimation of Common Measures of Treatment Effect:

In medical research logistic regression and Cox proportional hazards regression analysis, in which all the confounders are included as covariates, are often used to estimate an adjusted treatment effect in observational studies. In the last decade the method of propensity scores has been developed as an alternative adjustment method and many examples of applications can be found in the literature. Frequently this analysis is used as a comparison for the results found by the logistic regression or Cox proportional hazards regression analysis, but researchers are insufficiently aware of the different types of treatment effects that are estimated by these analyses. This is emphasized by a recent simulation study by Austin et al. in which the main objective was to investigate the ability of propensity score methods to estimate conditional treatment effects as estimated by logistic regression analysis. Propensity score methods are in general incapable of estimating conditional effects, because their aim is to estimate marginal effects like in randomized studies. Although the conclusion of the authors is correct, it can be easily misinterpreted. We argue that in treatment effect studies most researchers are interested in the marginal treatment effect and the many possible conditional effects in logistic regression analysis can be a serious overestimation of this marginal effect. For studies in which the outcome variable is dichotomous we conclude that the treatment effect estimate from propensity scores is in general closer to the treatment effect that is of most interest in treatment effect studies.