OBJECTIVES: Antipsychotic polypharmacy is frequently used in schizophrenia although the incremental effect for relapse prevention is uncertain. The purpose of this study is to compare risks of relapse between polypharmacy and monotherapy based on an observational study. Propensity score matching (PSM) and a latent variable model (LVM) were used to correct for selection bias. METHODS: Analyses were based on data from a multinational prospective cohort study that enrolled 1208 patients, followed over 2 years (EuroSC). The effect of treatment (combination therapy vs. monotherapy) on the risk of relapse was estimated using Cox models. PSM was used to create two comparable groups of patients. Propensity scores were derived from 11 variables measuring disease severity, quality of life, depression and functioning at baseline. A LVM was used to account for bias related to unmeasured confounding factors. It consisted of two parts estimated simultaneously: a logistic model predicting treatment choice and a Cox model on time to relapse. The model assumed that an unmeasured factor affecting risk of relapse was taken into account in the choice of treatment. All analyses were performed in a Bayesian framework, using WinBugs. RESULTS: Matched groups each included 344 patients, with relapse rates of 51% for monotherapy and 59% for polypharmacy. The hazard ratio for relapse with monotherapy versus polypharmacy was 0.77 (95% CI: 0.63, 0.94) based on the simple Cox model and 1.34 (0.87, 2.39) based on LVM. The LVM was better according to the Deviance Information Criterion. CONCLUSIONS: Results of the LVM suggest that polypharmacy is associated with a reduced risk of relapse, and contradict results based on PSM ignoring bias related to unmeasured confounding factors. The LVM appears to be a useful method to detect unmeasured confounding. Further research on the specification of latent variables is recommended to accurately quantify treatment effects.