Purpose: QALY weights for the EQ-5D have typically been obtained through time trade-off (TTO) surveys using a sub-set (n=17 or n=43) of the 243 health states. Discrete choice experiments (DCEs) are an alternative, potentially more flexible approach. We explore the development of EQ-5D algorithms based on both approaches and the impact of selection of health states in each.
Method: In phase one, respondents recruited via an on-line panel (n=228) were randomly allocated to three DCE designs: A (health states from the original UK EQ-5D valuation set); B (excluded health states which combined Mobility Level 3 was with Usual Activities Level 1 or Self Care Level 1); and C (no restrictions on the selection of health states). A pooled model and separate models for each design were estimated. In phase two an on-line panel sample (n=1000) completed the DCE for the preferred design. Each choice set presented two health profiles (EQ-5D state and survival duration), and death. In phase three, a population sample (n=417) completed a computer based TTO task and the DCE. The TTO incorporated 198 of the 243 health states (the remainder excluded as implausible). Respondents were randomly assigned to value 11 health states and the worst health state. Separate models were estimated for the DCE and the TTO.
Result: Comparison of the models suggest that the different choice experiment designs lead to differences in parameter estimates as well as differences in variance (scale). Response rates were higher for Design A, but there is a considerable loss of statistical efficiency arising from the reduced set of health states included in the experimental design. For the TTO data, models were estimated based on the functional form for the published algorithm for the EQ-5D, and allowing for interaction terms. While relatively few interaction terms are significant, a likelihood ratio test demonstrates that inclusion of interaction terms improves the fit of the model.
Conclusion: While the results for both approaches are broadly consistent with the previously published EQ-5D algorithms, there are some important differences. The DCE approach allows us to explore interactions between health states and duration that cannot be estimated with the TTO approach. In both approaches a design that allows for more extensive coverage of the EQ-5D space is appropriate.
Candidate for the Lee B. Lusted Student Prize Competition