OBJECTIVES: Patient-level utility scores derived using regression models may exhibit suboptimal performance when confronted with the ceiling effect. This study aims to identify the most suitable model for addressing the ceiling effect in EQ-5D scores from the MYR301 clinical trial.

METHODS: EQ-5D health state profile at the patient level was transformed into an index score (utility) by applying UK general population preference weights to each level in every dimension. Various regression models were employed to analyze EQ-5D scores and generate coefficients linked to covariates. The models' predictive accuracy/goodness-of-fit was evaluated using pseudo-R-squared; a higher value indicates better prediction.

RESULTS: The analysis of EQ-5D index scores from MYR 301 indicated a noticeable leftward skew in the distribution, with approximately 45-50% subjects achieving the highest score of 1. This observation strongly suggests the presence of ceiling effect in the data, characterized by a substantial proportion of subjects attaining maximum scores on the observed variables. Among different regression models, Tobit around median exhibited the most favorable performance on EQ-5D data with a pronounced ceiling effect, boasting predictive accuracy metric pseudo-R² of 0.23. Tobit model around the mean (pseudo-R² = 0.18) and Log transformed two-part model (TPM) with median (pseudo-R² = 0.15), and with mean (pseudo-R² = 0.14) ranking next in performance. Beta regression and its extensions did not perform satisfactorily in this context.

CONCLUSIONS: Tobit around the median exhibited the highest performance, followed by Tobit around the mean and log transformed TPM with Tobit, effectively addressing heavy ceiling effects, surpassing all other approaches. In the presence of a ceiling effect in the EQ-5D index scores, Tobit analysis provides more accurate and robust results by accounting for censoring in the regression models. From the perspective of health technology assessment (HTA) submissions, detailed assessment of regression models should be considered in order to derive the most appropriate utility scores.