OBJECTIVES: Many cost-effectiveness analyses in rheumatoid arthritis (RA) rely on statistical models relating Health Assessment Questionnaire Disability Index (HAQ) scores to health utilities. Linear models can produce out-of-bound estimates of Health Utilities Index Mark 3 (HUI3) scores. We estimated bounded, nonlinear relationships between HAQ and HUI3 based on patient-level data. METHODS: Bayesian generalized linear models (GLMs) were developed to predict baseline HUI3 conditional on baseline HAQ using patient-level data from the PREMIER (2-year controlled study in early RA) and ARMADA (24-week controlled study in longstanding RA) trials. HUI3 was rescaled to the interval [0,1] and modeled using a beta distribution and logistic link function. Normal-linear models were also estimated. Alternative specifications included age, sex, and HAQ-squared as additional predictors. Model parameters were estimated using WinBUGS 1.4.3. Models were compared using the deviance information criterion (DIC); lesser values imply better fit. Predicted values from beta-logistic models were linearly retransformed to the original HUI3 scale. RESULTS: Results were similar in early and late RA. Based on DIC, the beta-logistic models were more likely to generate the observed data than were the normal-linear models (PREMIER: –673.0 vs. –614.4; ARMADA: –226.1 vs. –215.8). Qualitatively, predictions from the beta-logistic models differed modestly from the normal-linear model. At low disability (HAQ=0.0), predicted HUI3 utilities were 0.75 vs. 0.81 (PREMIER) and 0.74 vs. 0.79 (ARMADA) for the beta-logistic and normal-linear models, respectively. At high disability (HAQ=3.0), predicted HUI3 utilities were –0.03 vs. –0.05 (PREMIER) and 0.01 vs. 0.01 (ARMADA). Age, sex, and HAQ-squared did not improve DIC.CONCLUSION: There is a strong negative relationship between HAQ and HUI3. Although the overall relationship is nonlinear, the linear approximation seems close across the relevant range of HAQ scores (0–3). Considering the complexity of the GLM approach, normal-linear regression may be adequate for cost-effectiveness analyses.