OBJECTIVES: Cox proportional hazards models comparing patient outcomes achieved across alternative drug therapies may be sensitivity to alternative specifications for censoring of the data prior to an event. Three criteria are used in drug studies: duration of therapy; time to a change in therapy; and end of available data. METHODS: California Medicaid data for 1994-2003 were used to identify 165,013 episodes of antipsychotic drug therapy initiated during CY2000-2002 by patients with schizophrenia. Cox models were estimated separately for acute or a psychiatric hospital admissions controlling for patient demographics, diagnostic profile, prescription drug profile and prior health care use. Four types of treatment episodes were defined based on discontinuation of therapy: patients restarting or switching drugs after a break in therapy > 15 days and patients switching or augmenting a pre-existing antipsychotic regimen with no break in treatment. Clozapine or ziprasidone episodes were excluded due to small sample size. Patients initiating treatment using a typical antipsychotic (TAP) were combined into a single comparison group. RESULTS: Patients using atypical antipsychotics achieved longer duration of therapy than TAP patients for all episode types, creating a longer period of risk exposure for an event in Cox models that censor data based on duration of therapy. The Cox models for all episode types combined did find that the hazard ratios for all three atypical antipsychotics were reduced relative to TAP when duration of therapy was dropped as a censoring criterion. However, changes in the estimated hazard ratios were small and the results from separate analyses by episode type were not always consistent with the aggregate model results. Patients switching therapies were consistently found to be at higher risk for hospitalization than patients restarting a drug used in their previous treatment attempt. CONCLUSION: Cox hazard ratios are sensitive to censoring specifications based on duration of therapy.