OBJECTIVES: Health-related quality-of-life data has distributional properties that raise numerous statistical challenges. Our objective was to compare the performance of different predictive models in handling health utility data.

METHODS: We simulated three utility datasets from a randomized controlled trial using the Dutch EQ-5D-5L scores, one of the most common instruments to estimate utility scores. Each dataset resembled a situation of a different known issue around utility analysis, in particular, multimodality and ceiling effect (right and left-bounded). We compared the performance of ordinary least squares (OLS), linear mixed effects, generalized linear model (GLM), tobit, generalized estimating equations (GEE), repeated measured mixed effects (RMME) and adjusted limited dependent variable mixture (ALDVM) models. The types of regression models considered were based on a targeted literature review, which found that these are the most frequently used for analyzing utility data. Each regression model was tested across the simulated datasets and its performance was assessed in terms of mean absolute error (MAE), root mean square error (RMSE) and visual inspection.

RESULTS: The RMME model achieved the best performance according to the three criteria in the multimodality (MAE=0.042; RMSE=0.067) and right-bounded (MAE=0.048; RMSE=0.073) datasets. In the left-bounded dataset, the linear mixed model came best, with MAE=0.093 and RMSE=0.136. ALDVM model successfully predicted the multimodality, although the MAE and RMSE were high, 0.203 and 0.261, respectively. ALDVM model performed less well in the right-bounded and left-bounded datasets. Across all three datasets, the OLS, GLM, GEE and tobit models had the poorest performance according to all criteria.

CONCLUSIONS: Common rejection reason in health technology appraisals is related to inappropriate utility data handling. Simulations are a pragmatic way to explore statistical issues surrounding utility data. This analysis identified the strengths and weaknesses of different regression approaches, which can reduce the uncertainty of cost-utility analyses.