Hester van Eeren

| Chapter 4 4 | 76 matching for example can result in very small comparison groups (Spreeuwenberg et al., 2010). To improve precision of the effect estimates, those methods can also be used in combination with regression analysis (Hirano & Imbens, 2001; Imbens, 2004). Thirdly, when using the PS, researchers should assess carefully which method is most appropriate for their specific research question. For example, when the treatment effect itself is more important, using the univariate PS and adjusting for extra covariates additionally reduces the bias of the treatment effect estimation (Rubin & Thomas, 2000; Stürmer et al., 2006). However, incorporating effect modification reduces the direct interpretability of the main treatment effect (Liem et al., 2010; Sturmer et al., 2006). Furthermore, if the distributions of the effect modifying variables vary highly, the overall estimates may differ across PS quantiles (Lunt et al., 2009). Different adjustment methods can thus result in divergent results, which all may be correct, but strongly depends on the research question and the population in which the estimation is most suitable (Kurth et al., 2006; Liem et al., 2010). Our study has several limitations. First, only a basic simulation study was designed. The characteristics of the simulated data were known in advance to the analyzer, which could have influenced the analysis and method chosen. Testing the methods using new simulated data that could be based on a real dataset, is recommended to further investigate the performance of the methods. Furthermore, the number of simulated datasets was rather small, which could have caused the small differences and inconsistencies in the simulation results, due to Monte Carlo error. Secondly, the overlap for the generalized PSs in the case study appeared to be less than optimal (Figure 2). This could have caused the difference in the estimated coefficients when the PS methods were compared. A distance score defined by Cochran and Rubin (1973) can be used to precisely test and define the overlap. A third limitation deals with the selection of variables into the PS. We left out the subgroup variable in the univariate PS, while Rubin and Thomas (2000) state that no prognostic variable should be left out. Although the results of our simulations and the case study only slightly changed when we added the subgroup variable to the univariate PS, we recommend investigating its influence in more detail. Fourth, although we controlled for observed pre-treatment variables, hidden bias due to unobserved confounders could not be controlled for in the case study. As we did not include hidden bias in the simulated datasets either, we do not know the effect of hidden bias in using the PSs in subgroup analysis. This study supports the idea that the generalized PS can be used in estimating the treatment effect when this is modified by a subgroup variable. As patient-tailored treatment becomes more and more important in outcomes research (Norcross & Wampold, 2011), this study contributes to the literature on how to handle effect estimation in non-randomized outcomes studies of patient subgroups using the PS.