BACKGROUND: The benefits of some second-generation antipsychotics (SGAs) must be weighed against the increased risk for diabetes mellitus. This study examines whether the association between SGAs and diabetes differs by dose. METHODS: Patients were >/=18 years of age from three US healthcare systems and exposed to an SGA for >/=45 days between November 1, 2002 and March 31, 2005. Patients had no evidence of diabetes before index date and no previous antipsychotic prescription filled within 3 months before index date.49,946 patients were exposed to SGAs during the study period. Person-time exposed to antipsychotic dose (categorized by tertiles for each drug) was calculated. Newly treated diabetes was identified using pharmacy data to determine patients exposed to anti-diabetic therapies. Adjusted hazard ratios for diabetes across dose tertiles of SGA were calculated using the lowest dose tertile as reference. RESULTS: Olanzapine exhibited a dose-dependent relationship for risk for diabetes, with elevated and progressive risk across intermediate (diabetes rate per 100 person-years = 1.9; adjusted Hazard Ratio (HR), 1.7, 95% confidence interval (CI), 1.0-3.1) and top tertile doses (diabetes rate per 100 person-years = 2.7; adjusted HR, 2.5, 95% CI, 1.4-4.5). Quetiapine and risperidone exhibited elevated risk at top dose tertile with no evidence of increased risk at intermediate dose tertile. Unlike olanzapine, quetiapine, and risperidone, neither aripiprazole nor ziprasidone were associated with risk of diabetes at any dose tertile. CONCLUSIONS: In this large multi-site epidemiologic study, within each drug-specific stratum, the risk of diabetes for persons exposed to olanzapine, risperidone, and quetiapine was dose-dependent and elevated at therapeutic doses. In contrast, in aripiprazole-specific and ziprasidone-specific stratum, these newer agents were not associated with an increased risk of diabetes and dose-dependent relationships were not apparent. Although, these estimates should be interpreted with caution as they are imprecise due to small numbers.