Using Process Data to Improve Choice Predictions

Empirical studies of consumer choice increasingly make use of cognitive process models that incorporate behavioral insights while maintaining quantitative rigor. At the same time, important questions remain regarding their ability to deliver insights that are at once accurate, generalizable, and managerially relevant. We take a step toward addressing these issues by conducting a series of consumer choice experiments and comparing the accuracy and generalizability of the canonical multinomial logit model (MNL) against the drift diffusion model (DDM), a cognitive model that incorporates response times in addition to choice data. We find that the DDM robustly outperforms MNL in providing accurate forecasts on several managerially relevant measures, and that these improvements generalize to out-of-sample scenarios involving new consumers and new choice environments. Perhaps most impressively, relative performance improvements when generalizing across individuals are greater in magnitude than when using holdout choices within the same individuals. Additional analyses further show that these improvements derive from the ability of the cognitive model to capture heterogeneity in the tradeoff between time spent on deliberation and the probability of mistakenly choosing a lower-valued option. We conclude with a discussion of the implications of these findings for theoretical and applied work in consumer choice modeling.