RM3 - IS PATIENT CHOICE PREDICTABLE? THE IMPACT OF DISCRETE CHOICE EXPERIMENT DESIGNS AND MODELS

Abstract

Increased use of discrete choice experiments (DCEs) in healthcare requires establishing whether stated preferences are predictive of observed healthcare utilization. This study aimed to determine whether the number of alternatives in a DCE choice task should reflect the actual decision context, and how complex the choice model needs to be to predict real-world choices correctly at an aggregate and individual level in healthcare. Two randomized controlled trials (RCTs) involving choices for influenza vaccination and colorectal cancer screening were used. Each RCT had three study conditions: DCE choice tasks with (i) two alternatives, (ii) three alternatives, or (iii) both. Two samples of 1,200 respondents each were randomly assigned to one of the conditions. Each respondent answered 16 DCE choice tasks (for the derivation of the decision model) plus a choice task mimicking the real-world choice (to keep the decision context the same). The data was analysed in a systematic way using random-utility-maximization (RUM) and random-regret-minimization (RRM) choice processes with scale and/or preference heterogeneity. Irrespective of the number of alternatives per choice task, the choice to opt for influenza vaccination and colorectal cancer screening was correctly predicted by DCE at an aggregate level, if scale and preference heterogeneity were taken into account. At an individual level, three alternatives per choice task and using heteroscedastic model plus preference heterogeneity seemed to be most promising, correctly predicting the real-world choice in 81.7% and 87.9% of the cases for influenza vaccination and colorectal cancer screening respectively. No evidence was found that RRM outperformed RUM. Our study shows that DCEs hold the potential of being externally valid at an aggregate level if at least scale and preference heterogeneity are taken into account. Further research is needed to determine if this result remains in other contexts, and to optimise choice prediction at an individual level.