OBJECTIVE: Medical records provide a potential wealth of information about treatment effects; however, differences in pretreatment patient or other characteristics may influence treatment assignment. This, in turn, could lead to biased estimates of treatment effects in nonrandomized studies. We developed a statistical model using propensity scores to reduce treatment selection bias in analyses based on retrospective data. METHODS: As part of a study described elsewhere, we abstracted retrospective data from the medical records of 327 patients treated for schizophrenia or schizoaffective disorder with risperidone, olanzapine, or quetiapine at 3 acute inpatient mental health facilities. Data were collected on patients from the inpatient hospitalization through 60 days following initiation of study drug. Using a multinomial logistic regression analysis of pretreatment patient and other characteristics, we developed a predictive model of treatment assignment to risperidone, olanzapine, or quetiapine. RESULTS: The following variables were significantly predictive of treatment assignment: age at admission, gender, race, smoker at admission, history of substance abuse, prior use of clozapine, and facility. The following variables were among those not significantly predictive of treatment assignment: prior use of atypical antipsychotics other than clozapine, body mass index at admission, age at first hospitalization for mental illness, and history of suicide attempts, violence, glucose abnormalities, or seizures. CONCLUSION: The propensity score model offered a means to adjust for treatment selection bias in a nonrandomized study comparing treatment effects of risperidone, olanzapine, and quetiapine in an inpatient setting. In addition, the propensity score methodology can be used by researchers responsible for designing non-randomized studies of healthcare interventions and decision-makers who are responsible for evaluating and interpreting the results in this disease area.