How do Design Characteristics Affect Respondent Engagement? Assessing Attribute Non-attendance in Discrete Choice...

Discrete choice experiments (DCEs) are increasingly applied to develop value sets for health-related quality-of-life instruments, but respondents may adopt various simplifying heuristics that affect the resulting health state values. Attribute level overlap can make these DCE tasks easier and thereby increase respondent engagement. This study uses choice tasks involving EQ-5D-5L health states to compare designs with and without overlap, constructed using different methods (generator-developed design, Ngene, SAS, and Bayesian D-efficient design) to assess respondent non-attendance to attributes.

A multi-arm DCE using the EQ-5D-5L was conducted in the Australian general population. The performance of designs with various properties was compared using the level of respondent engagement. Respondent engagement was quantified through the inferred attribute non-attendance (ANA) estimated by the equality constrained latent class model. Utility decrements derived using all respondents (i.e., including non-attendees) were compared with estimates obtained only from those who attended to all EQ-5D-5L attributes.

The inclusion of overlap improved full attendance rates from 22.3-28.4% to 28.2-54.2%. Within designs with overlap, modified Fedorov designs (constructed using either Ngene or SAS macros) had higher full attendance rates than other designs. The relative attribute importance of the EQ-5D-5L also differed significantly before and after data exclusion using ANA analysis, but there was no clear pattern in the differences.

This study found evidence to support the use of modified Fedorov designs (constructed using Ngene or SAS) with attribute overlap to reduce ANA and improve respondent engagement in DCE studies. It highlights the potential value of ANA analysis as a quality-control tool for the inclusion and exclusion of respondents in future health valuation work for the EQ-5D-5L.