OBJECTIVES: An algorithm was recently developed to predict median US population preferences for EQ-5D health states. The primary objective of this study was to provide evidence regarding the estimation efficiency of the median-based EQ-5D index relative to the existing mean-based index. A secondary objective was to evaluate the sensitivity of findings to the application of robust statistical procedures. METHODS: Data were taken from the 2002-2003 Medical Expenditure Panel Survey. Adult survey participants completed the EQ-5D and 12-item Short-Form Health Survey (SF-12); rated their physical and mental health; and reported the presence of chronic conditions. Associations of mean- and median-based EQ-5D index scores with health status measures and chronic conditions were estimated using least squares (LS) and rank (R) regression. Associations of changes in index scores with changes in health status measures were similarly estimated. For each set of analyses, a relative efficiency (RE) statistic was derived as the ratio of model F statistics (median-based index/mean-based index). RESULTS: RE statistics (LS/R) for associations of index scores with health status measures and chronic conditions were as follows: SF-12 mental summary, 0.50/1.35; SF-12 physical summary, 0.40/0.75; mental health rating, 0.63/1.10; physical health rating, 0.54/0.99; angina, 0.55/0.89; arthritis, 0.50/0.88; asthma, 0.77/0.98; coronary heart disease, 0.54/0.91; diabetes, 0.60/0.95; emphysema, 0.65/1.03; hearing problems, 0.50/0.98; myocardial infarction, 0.51/0.97; stroke, 0.66/1.00; and visual problems, 0.65/1.03. RE statistics (LS/R) for associations of index score changes with changes in SF-12 physical and mental summaries were 0.63/0.89 and 0.59/1.17, respectively. CONCLUSIONS: Parametric analyses exhibited reduced power when applied to the median-based index scores. However, robust procedures suffered from minimal efficiency losses and exhibited efficiency gains when used to analyze associations of index scores with mental health measures. These findings reflect the difference in skewness between the mean- and median-based index scores, which can be attributed to the difference in score ranges.