Optimal DCE design for modelling nonlinear time preferences in EQ-5D-5L valuation studies: exploration of data from Denmark and Peru

Background: Discrete choice experiment (DCE) methods are an increasingly popular valuation method, particularly for the EQ-5D-5L. While EQ-5D-5L value sets developed using DCE have traditionally assumed linear time preferences, this assumption has been challenged. This has led to the development of DCE modelling methods that allow for nonlinear time preferences. The aim of this study was to explore the impact of a model that accounts for nonlinear time preferences with DCE choice set formats and design construction methods for EQ-5D-5L value sets. Methods: This study used a four-arm (2 × 2) between-subjects design to investigate the impact of two commonly used DCE choice set formats (i.e. a third option of either immediate death or full health) and two commonly used DCE design construction methods (i.e. generator-developed and efficient designs) on EQ-5D-5L value sets. Mixed logit models that used exponential discounting to account for nonlinear time preferences were estimated in OpenBUGS. This was tested in a sample of respondents from Peru (n = 942) and Denmark (n = 988). Results: Across all arms and for both countries, discounting was found to be present when modelling explicitly for nonlinear time preferences. Although estimated discount rates varied widely from 1 to 117%, both type of choice set format and type of design construction method influenced the utilities for more severe health states. Choice sets with full health tended to produce a wider range of utility weights, while choice sets with immediate death tended to produce higher estimated discount rates. Generator-developed designs tended to produce the highest and lowest utility weights for health states compared to the efficient designs. Conclusions: This study provides a comparison of DCE choice set format and design construction method when nonlinear time preferences were explicitly modelled. Limitations to this study are discussed including data quality issues with the Peruvian dataset and small sample sizes. Further investigation is needed to confirm the suitability of models that account for nonlinear time preferences in EQ-5D-5L valuation studies.