Reducing Uncertainty in EQ-5D Value Sets: The Role of Spatial Correlation

Abstract

Objective. Scoring algorithms for the EQ-5D-3L are constructed subject to a large degree of uncertainty (a credible interval width of 0.152, which is significant in comparison to the reported minimal important differences). The purpose of this work is to explore modeling techniques that will reduce the extent of this uncertainty. Methods. We used the US valuation study data. A Bayesian approach was used to calculate predicted utilities and credible intervals. A spatial Gaussian correlation structure was used to model correlation among health states (HS), thus allowing directly valued HS to contribute to the predicted utility of nearby unvalued HS. Leave-one-out cross-validation was used to compare model performances. Results. The average posterior standard deviation was 0.039 for the unvalued health states and 0.011 for the valued health states. Using cross-validation, the US D1 model had 31% coverage probability. Models with independent and Gaussian correlation had coverage probabilities of 95% and 93%, respectively. Moreover, the Gaussian correlation structure resulted in a 25.6% reduction in mean squared error (SE) and 13.2% reduction in mean absolute error (AE) compared to the independent correlation structure (mean SE of 0.00131 v. 0.00176 and mean AE of 0.02818 v. 0.03248). Conclusion. Uncertainty was substantially lower for the directly valued HS compared to unvalued HS, suggesting direct valuation of as many health states as possible. Incorporation of a spatial correlation significantly reduced uncertainty. Hence, we suggest incorporating this when constructing scoring algorithms.