New methods for modelling EQ-5D-5L value sets: an application to English data

Background: The EQ-5D is a widely used questionnaire that describes and values health related quality of life. Recently, a five level version was developed. Updated methods to estimate values for all health states are required. Data: 996 respondents representative of the English general population completed Time Trade-Off (TTO) and Discrete Choice Experiment (DCE) tasks. Methods: We estimate models, with and without interactions, using DCE data only; TTO data only; and TTO/DCE data combined. TTO data are interpreted as both left and right censored. Heteroskedasticity and preference heterogeneity between individuals is accounted for. We use maximum likelihood estimation in combination with Bayesian methods. The final model is chosen using the deviance information criterion (DIC). Results: Censoring and taking account of heteroskedasticity has important effects on parameter estimation. Regarding DCE, models with different dimension parameters and similar level parameters are best. Considering models for both TTO and DCE/TTO combined, models with parameters for all dimensions and levels perform best, as judged by the DIC. Taking account of heterogeneity improves fit, and a three latent group multinomial model has the lowest DIC. Conclusion: Studies to elicit values for the EQ-5D-5L need new approaches to estimate the underlying value function. This paper presents approaches which suit the characteristics of these data and recognise preference heterogeneity.