OBJECTIVES: The objective was to identify the optimal statistical method for regression analysis of the EQ-5D index. Specifically, we compared the performance of alternative regression methods in estimating incremental preference-based health related quality of life scores from the EQ-5D index.  Importance of this work is high as preference-based scores derived from the EQ-5D index are used to calculate quality-adjusted life years (QALYs), the most common measure of health outcomes used in cost-effectiveness analysis.  Many health utility variables are censored from the top by construction, i.e. full health at unity. This is a utility measurement dilemma as 46% of US respondents report a perfect EQ-5D score.  Also, EQ-5D is treated as a continuous variable; however due to its derivation algorithm there is a gap nearly equivalent to one standard deviation in the US preference-based scores of the EQ-5D index (no values between 0.8603 and 1). METHODS: Simulation analyses were implemented to compare the performance of OLS, median regression, Tobit, and robust extensions of Tobit models. First, pooled 2000-2003 Medical Expenditure Panel Survey data was randomly divided into independent derivation and validation sets to estimate the relationship between the EQ-5D index and SF-12 physical summary scores. Second, the performance of the same estimation methods was compared in a Monte Carlo simulation under ten non-normal distributions. RESULTS:  Median regression outperformed all other methods in the first simulation analysis followed by the re-censored Tobit method. Median regression also resulted in the smallest mean squared prediction errors in the Monte Carlo simulations, followed by the Tobit method with logistic distribution. CONCLUSIONS: Median regression appears to be the most robust method to use in regression analysis of the EQ-5D index. If normality and homoscedasticity assumptions are not met, then logistic-Tobit regression can be used as a robust extension of the classical Tobit method.