OBJECTIVES: To assess the preference-based health status, the EQ-5D index is preferable and is a widely used instrument that could serve as the basis for summarizing and comparing health outcomes. Various regression models have been used to generate the utility estimate for the economic models from patient-level utility scores. However, some of the regression models do not perform well in the case of the ceiling effect. The research aimed to compare the different regression models in case of a heavy ceiling effect in the EQ-5D scores.

METHODS: The data for EQ-5D scores and covariates were simulated in R (v4.0.2) with an assumption that 75% of patients had the EQ-5D scores of 1. Different regression models such as Ordinary Least Square (OLS), Tobit/Censored (around mean/median), Beta regression and its extensions (bias correction and bias reduction), Two-part model (TPM) with OLS, TPM with Tobit, Log-transformed TPM with OLS and Log-transformed TPM with Tobit were fitted on EQ-5D scores to generate the coefficient associated with covariates. The predictive accuracy/goodness-of-fit of all the models was compared using absolute and squared errors, that are, R¹ and R².

RESULTS: Log-transformed TPM with OLS performed best on EQ-5D data with a heavy ceiling effect, having predictive accuracy metric R¹ (-24.39) and R² (-444.36) followed by Log-transformed TPM with Tobit (R¹=-22.98, R²=-395.88). The performance of the two-part models (with OLS/Tobit) (R²=-73.36) was superior to the Tobit model around mean (R²=-5.71) and median (R²=-3.13), followed by OLS (R²=0.18), whereas the beta regression and its extensions did not perform well.

CONCLUSIONS: The log-transformed TPM model with OLS was superior followed by log-transformed TPM with Tobit to all the other approaches in handling data with heavy ceiling effects.