A closer look at decision and analyst error by including nonlinearities in discrete choice models: implications on willingness-to-pay estimates derived from discrete choice data in healthcare.

ABSTRACT

BACKGROUND: Most researchers in health economics cite random utility theory (RUT) when analysing discrete choice experiments (DCEs). Under RUT, the error term is associated with the analyst's inability to properly capture the true choice processes of the respondent as well as the inconsistency or mistakes arising from the respondent themselves. Under such assumptions, it stands to reason that analysts should explore more complex nonlinear indirect utility functions, than currently used in healthcare, to strive for better estimates of preferences in healthcare. OBJECTIVE: To test whether complex indirect utility functions decrease error variance for models that either implicitly (i.e. the multinomial logit (MNL) model) or explicitly (i.e. entropy multinomial logit (EMNL) model) account for error variance in health(care)-related DCEs; and to determine the impact of complex indirect utility functions on willingness-to-pay (WTP) measures. METHODS: Using data from DCEs aimed at healthcare-related decisions, we empirically compared (1) complex and simple indirect utility specifications in terms of goodness of fit, (2) their impact on WTP measures, including confidence intervals (CIs) based on the Delta method, the Krinsky and Robb-procedure, and Bootstrapping, and (3) MNL and EMNL model results. RESULTS: Complex indirect utility functions had a better model fit than simple specifications (p < 0.05). WTP estimates were quite similar across alternative specifications. The Delta method produced the most narrow CIs. The EMNL model showed that respondents apply simplifying strategies when answering DCE questions. CONCLUSION: Complex indirect utility functions reduce error arisen from researchers, which can have important implications for measures in healthcare such as the WTP, whereas EMNL provides insights into the behaviour of respondents when answering DCEs. Understanding how respondents answer DCE questions may allow researchers to construct DCEs that minimise scale differences, so that the decision error made across respondents is more homogeneous and therefore taken out as additional noise in the data. Hence, better estimates of preferences in healthcare can be provided.