OBJECTIVES: Investigate how selection criteria and statistical model specifications affect comparisons of alternative medications using retrospective database analyses. METHODS: Data from the Medi-Cal Program were used to conduct a series of head-to-head comparisons of alternative antipsychotics to test the sensitivity of results to sample selection and model specification. OLS models were estimated for duration of therapy defined based on breaks in therapy > 15 days. Five models were compared: (1) a baseline model with only demographic independent variables and selection criteria limited to age (18-100) and prior eligibility > 6 months; (2) screening for one year of post-treatment data; (3) screening for schizophrenia; (4) addition of prior diagnoses and utilization; and (5) addition of independent variables for episode type. RESULTS: The baseline sample consisted of 263,206 episodes. Average unadjusted days of therapy for typical antipsychotics, olanzapine, risperidone and quetiapine were 63, 138, 143 and 131, respectively. Model 1 found longer duration for all atypical antipsychotics relative to conventional drugs (63-67 days). Risperidone exhibited longer duration relative to olanzapine (+6 days, p<0.0001) while olanzapine duration exceeded quetiapine by 8 days (p<0.0001). Duration for typicals increased from 63 days to 112 days in the schizophrenia analysis (N=70,650), reducing the estimated differences favoring atypicals to 29-32 days (p<0.0001). Adding independent variables for prior utilization and diagnostic mix reversed risperidone's advantage over olanzapine from +6 to – 4 days (p<0.0001). Results favoring quetiapine over olanzapine in duration of therapy on all antipsychotics was reversed from +4 to +16 days to – 3 days when covariates for episode type were included in the model. CONCLUSIONS: Differences in duration of antipsychotic therapy exist across diagnostic group and episode type. Differences also exist in the diagnostic and episode mix across drugs. Therefore, disaggregated patient samples and expanded model specifications provide more accurate estimates of differences in treatment duration.