THE IMPACT OF HEALTH STATE “FEASIBILITY” ON THE ESTIMATION OF VALUES FOR EQ-5D

Purpose: Estimating values for all 243 EQ-5D health states is generally achieved through an estimation model based on the direct valuation of a subset of states. Typically such models explain around 40% of the variance. Model performance is affected by data quality that is itself determined by respondent perceptions. Infeasible health state descriptions are likely to elicit less reliable valuation data. This study reports on the effect of health state feasibility on estimation performance.

Methods: This study is based on a Korean valuation study in which 101 EQ-5D health states were valued using Time Trade-Off (TTO) collected from a national representative sample of 1,307 adults. The performance of random effects models based on the full valuation dataset was compared with identical models in which less feasible health states have been removed. The means absolute difference (MAD) between observed and estimated values was computed in all cases. Three alternative exclusion criteria were constructed to identify less feasible health states. Criterion A excludes states where level 3 (extreme problems) on mobility is combined with level 1 (no problems) on self-care or usual activities dimensions; criterion B excludes infrequently occurring EQ-5D states reported in the national Health Survey for England (HSE); criterion C excludes states with a standard deviation > 0.4.

Results: Criterion A excluded 9 states (6% observed data). Corresponding values for criteria B and C are 49 states (51%) and 13 states (18%). The MAD based on the full valuation dataset is 0.032 (SD 0.025). MAD for both criteria A and C are virtually identical 0.030 (0.024). Under criterion B a model could not be computed as there was insufficient data on severe EQ-5D health states.

Conclusions: Feasibility appears not to be an issue in terms of improving model performance based on censored data. This has significant implications in the design of valuation studies for EQ-5D (and probably other generic health status classifications). Nonetheless, if health states are limited to those observed in general population studies (criterion B), then an estimation model may be impossible as more severe health states are excluded. Even though MAD may be a useful index when alternative models are applied to a single dataset, it is of doubtful value when used to compare model results based on nested datasets.