PURPOSE: Medical cost-effectiveness analyses (CEA) traditionally use utility values that average across the whole population of persons who might be appropriate for the treatment. However, most treatments are used only when the patient accepts them, suggesting that the preferences of patients who would choose a treatment if it were offered are the relevant ones for CEA. This study uses the example of intensive therapy for diabetes to study the effect of accounting for patient choice in incorporating patient preferences into CEA. METHODS: We interviewed 394 adults with type 2 diabetes to determine their utility values for health states relating to common diabetic complications and treatments and their choice of therapy. We then used a simulation model developed previously by the CDC to assess the cost-effectiveness of intensive therapy for the population as a whole and for the set of patients who chose intensive therapy. RESULTS: The mean age of subjects was 62 years. 58% were women. 43% were African-American, 30% White, and 27% Latino. 31% of patients reported using intensive therapy. The mean utility for intensive glucose control was 0.67 (SD 0.33) and the mean value for conventional control was 0.75 (0.30). Utilities for complications were similar to prior estimates, and were similar between patients on intensive and conventional therapies. However, patients on intensive therapy had higher utilities for intensive therapy than did patients on conventional therapy. When patient utilities were used in the cost-effectiveness model, intensive therapy for diabetes was harmful on average for the full population of patients with diabetes (mean DELTA costs $8007, mean DELTA QALYs -0.35) but beneficial and cost-effective for the patients who select it (mean DELTA costs $7777, mean DELTA QALYs 0.18, ICER $43K/QALY). CONCLUSIONS: Cost-effectiveness results for diabetes care are dramatically altered by accounting for preference heterogeneity and patient selection of treatment choice. Cost-effectiveness models that do not account for patient selection may be seriously misleading for treatments whose cost-effectiveness is sensitive to variations in patient preferences.