Hester van Eeren

Subgroup analysis using the propensity score | 4 67 | The primary outcome measure was psychiatric symptomatology and was measured with the Global Severity Index (GSI), which is the mean score of the 53 items of the Brief Symptom Inventory (Arrindell & Ettema, 2003; Derogatis, 1986). The GSI ranges from 0 to 4, with higher scores indicating more problems. Three treatment institutes conducted their follow-up measures on the GSI at 12, 24, 36 and 60 months after baseline. The 3 remaining treatment institutions conducted their follow-up measures at the end of treatment, 6 and 12 months after end of treatment, and again at 36 and 60 months after baseline. As in an earlier study by Spreeuwenberg and colleagues (2010) we used the mean GSI score of all follow-up measures as a primary outcome measure (range 0.01 - 3.17) (American Psychiatric Association, 2000). We excluded 114 cases that had missing values on one of the potential confounders, leaving 727 patients in the final sample. The excluded cases were not significantly different on the outcome GSI. The potential confounders were assessed at baseline, that is, age, gender, civil status, living situation, care of children, employment, level of education, duration of psychological complaints, treatment history, alcohol and drug abuse, motivation, treatment preferences, level of psychiatric symptomatology, level of personality pathology, interpersonal functioning, social role functioning, quality of life, number of Diagnostic and Statistical Manual (DSM)-IV Axis II cluster A personality disorders, number of DSM-IV Axis II cluster B personality disorders, number of DSM-IV Axis II cluster C personality disorders, and psychological capacities. For specific details of this study, we refer the reader to the literature (Bartak, Andrea, Spreeuwenberg, Thunnissen, et al., 2011; Bartak, Andrea, Spreeuwenberg, Ziegler, et al., 2011; Bartak et al., 2010). Computation The analyses were performed with IBM SPSS for Windows, version 20 (SPSS Inc., Chicago, IL). All simulations were performed in R programming language, version 2.13.0 (R Development Core Team, 2010). Results Monte Carlo simulation results We evaluated the bias, MSE and standard error of the relevant effects in the simulation study (see Table I, Supplemental Material for the bias, MSE and SE in scenario 1; for scenario 2, see Table II, Supplemental Material). Because taking an average over 3 estimated bias values related to the 3 relevant coefficients per PS method can average out positive and negative bias values, the MSE was used to find which PS estimations was most efficient (Table 2). In almost all simulated datasets within scenario 1, when the subgroup was not related to the treatment assignment, the MSE was closest to zero if the variables related to the outcome only were included in the univariate PS and the generalized PS (Table 2). If the subgroup was related to the treatment assignment, as in scenario 2, the MSE was closest to zero when the variables related to the outcome were included in the PS model in all simulated datasets (Table 2). In both the scenarios, including the subgroup variable in the univariate, PS estimation gave larger MSE values.