Hester van Eeren

| Chapter 4 4 | 72 Generalized PS: case study The generalized PS was applied according to the protocol described by Spreeuwenberg and colleagues (2010) The generalized PS was estimated as a combination variable of treatment duration and severity of problems: short-term treatment for patients having mild problems (reference category, N=268), short-term treatment for patients having severe problems (N=34), long-term treatment for patients having mild problems (N=338) and long-term treatment for patients having severe problems (N=87). The same list of covariates selected in the univariate PS was included in the generalized PS estimation. Here, we followed the simulation results of scenario 2. The PS was estimated by multinomial regression analysis, with the combination variable of treatment duration and severity of problems as dependent variable and not including interaction terms between the covariates. However, the number of cases in 2 groups was small and validity of the model fit was therefore uncertain (Austin, 2009; Shah et al., 2005; Stürmer et al., 2006; Weitzen et al., 2004). Because the four estimated generalized PSs add up to 1 and are complementary, only 3 of 4 were used in further analyses. As required when using the PS, the ranges of the estimated PS scores showed overlap (Figure 2). In the final regression model on the outcome GSI, 3 generalized PSs and 3 dummies indicating group membership were included [Eq. (6)]: β β β β β β β = + + + + + + 0 1 1 2 2 3 3 4 1 5 2 6 3 OUTCOME PS PS PS LongMild ShortSevere LongSevere These results indicated that long- term treatment was more favorable for patients having mild problems. For patients having severe problems, both treatment options were almost equally effective, just as was presented when the univariate PS was applied while taking the interaction effect into account (Table 3). To compare the relative effects of this model to the results of using the univariate PS, we used the standardized coefficients. These coefficients can be interpreted independent of the intercept and PS scores added in each model. The coefficient for patients having severe problems in short-term treatment was 0.240 using the univariate versus 0.129 using the generalized PS. The coefficient of patients having mild problems in long-term treatment was almost equal: 0.092 using the univariate PS versus 0.099 using the generalized PS. For patients having severe problems in long-term treatment we combined the standardized coefficients of the model in which the univariate PS was applied and compared it to the corresponding coefficient in the model of the generalized PS. For these patients, the combined coefficient was 0.248 using the univariate versus 0.139 using the generalized PS (Table 3).