TESTING THE ACCURACY OF COMPOSITE HEALTH STATE UTILITY ESTIMATORS – AN ANALYSIS OF MEDICAL CONDITION PAIRS FROM MEPS 2001 AND 2003

Purpose: Numerous estimators have been suggested for estimating disutility for patients with multiple conditions for which single condition values are known. The objective of this study was to test the accuracy of these estimators on empirical patient data while controlling for several methodological weaknesses in previous studies – in particular (contaminated) estimates of single condition disutility based on respondents with multiple other conditions.

Methods: We used pooled data from the 2001 and 2003 Medical Expenditure Survey Panel (MEPS), with EQ-5D (index and self ratings on the Visual Analogue Scale, VAS) and disease data for 40,846 non-institutionalized respondents. Using estimated values for condition categories with at least 100 respondents having only that condition, we compared estimated and observed mean scores for respondents with pairs of the same conditions. We used both US EQ-5D index values, and VAS ratings adjusted to a 0-1 scale where 1 was set using the mean VAS value from respondents in the US EQ-5D valuation study reporting no health problems (state 11111).

Results: Additive and multiplicative paired condition disutility estimates (index and VAS) were not statistically different, while both were significantly larger than the disutility of the worst condition in the pairs. Respondents in the pairs had on average 4.2 medical conditions outside the pairs. Excluding respondents with extra-pair conditions resulted in too few observations to conduct analyses. Without adjusting for extra-pair conditions, all disutility estimates were too small. Adjustig for extra-pair conditions by adding the mean difference between contaminated and uncontaminated single-condition estimates, both the additive and multiplicative methods resulted in statistically insignificant overestimates  for both VAS (difference: additive=.015, SD .036, multiplicative .009, SD .037) and EQ-5D index data (additive .034, SD .033, multiplicative .025, SD .035). Using the mean of the worst condition in the pair resulted in statistically significant underestimates.

Conclusions: When controlling for contamination from conditions outside the scope of the single-to-pair analysis, both the additive and the multiplicative estimators performed well. Single condition estimates based on respondents with multiple other conditions resulted in overestimates using these methods in a previous study (Fu & Kattan, 2008). These results are promising in terms of the ability to estimate values for composite health states, but fail to indicate which of the two estimators is most appropriate.