OBJECTIVES: Policymakers need cost utility estimates when evaluating new biologic therapies in RA. Most clinical trials of these therapies measure physical function, then estimate utility from physical function. Adalimumab (D2E7, Abbott) trials measured both physical function and utility, providing an opportunity to evaluate estimates of utility from physical function. METHODS: A total of 2070 patients with active RA participated in 4 clinical trials (ARMADA, DE011, DE019 and STAR). The Health Assessment Questionnaire Disability Index (HAQ) measured physical function for all patients and the Health Utilities Index Mark-3 (HUI3) measured utility for 2000 patients at baseline and months 3 and 6. HAQ scores range from 0 (good physical function) to 3. HUI3 was regressed on HAQ using both repeated measure mixed models and cross-sectional models controlling for age, sex, disease duration, clinical trial and fatigue (measured by the Functional Assessment of Chronic Illness Therapy-Fatigue (FACIT-F)). The cross-sectional models used end-of-study data with Last Observation Carried Forward (LOCF) imputation. These analyses include the 1990 patients with non-missing values for these variables. RESULTS: The mixed models and the cross-sectional models provided almost identical coefficient estimates. Disease duration and trial were not significant and were deleted from the model. Age (in years) was statistically significant but trivial (?=0.00087). The basic estimated cross-sectional model was: HUI3=0.76-0.28*HAQ+0.05*FEMALE, (p<0.0001 for each regressor, Adj. R2=0.49). However, the relationship between HUI3 and HAQ appears to be nonlinear: coefficients for HAQ-squared and HAQ-cubed were significant (p=0.013 and p=0.003, respectively) when added to the regression. Adding FACIT-F to the basic cross-sectional model substantially improved model fit (Adj. R2=0.63). CONCLUSIONS: The basic algorithm developed in this study is consistent with published predictions of utility from HAQ (Kobelt et al., Arthritis and Rheumatism, 1999). However, these algorithms' predictions are limited and should only be used when direct utility scores are not available.