OBJECTIVES: The potential for antipsychotic-induced diabetes is an important issue. Retrospective studies using large patient databases have had conflicting findings regarding diabetes risks associated with different antipsychotics. METHODS: Claims data for thousands of psychosis patients treated or untreated with antipsychotics were analyzed. Screening for preexisting diabetes, identification of diabetes with prescription claims only, and requirement of antipsychotic monotherapy provide better control for confounding influences and represent a stronger study design. Diabetes odds ratios for risperidone, olanzapine, quetiapine, or conventional antipsychotics versus non-treatment were estimated for all patients and for patients stratified by dose levels. Logistic regression controlled for age, sex, type of psychosis, length of observation/treatment, preexisting excess weight, and use of other drugs with diabetogenic effects. RESULTS: Under a weaker study design, all of the antipsychotics were associated with significantly higher odds of diabetes relative to non-treatment. Odds ratios (95% confidence intervals [CI]) were: risperidone 1.388 (1.276 - 1.509), olanzapine 1.331 (1.224 - 1.446), quetiapine 1.394 (1.247 - 1.559), and conventional antipsychotics 1.365 (1.238 -1.503). Under a stronger study design, relative odds for risperidone and quetiapine declined, becoming statistically insignificant, whereas odds for olanzapine and conventional antipsychotics increased. Odds ratios (95% CI) were: risperidone 1.224 (0.962 - 1.562), olanzapine 1.858 (1.549 - 2.238), quetiapine 1.087 (0.742 - 1.612), and conventional antipsychotics 1.755 (1.381 - 2.221). With quetiapine, odds of diabetes were not significantly increased at any dose level relative to non-treatment. Odds were significantly increased at all dose levels with conventional antipsychotics, at medium and high doses with, olanzapine, and at high doses with risperidone. CONCLUSIONS: In database studies, estimated risks of diabetes among antipsychotics are affected by study design. When a more reliable design was used, risks associated with quetiapine and risperidone were lower than those of olanzapine and conventional antipsychotics and not significantly different from those in untreated patients.