Which Discounting Model is Best?

A research paper by Exponent's Jordan Bailey, Ph.D., and University of Kentucky's Mark J. Rzeszutek, Ph.D., and Mikhail Koffarnus, Ph.D., examines methods for selecting discounting models — behavioral models that offer explanations for how a commodity's subjective value changes as a result of delay to its receipt. 

The study compares three discounting model selection techniques — Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and Leave-One-Out Cross-Validation (LOOCV) — to evaluate their effectiveness in identifying the best-fitting models and their ability to generalize unseen data. 

The authors' findings emphasize the importance of balancing model fit, parsimony, and generalization in discounting research. Combining AIC/BIC with cross-validation provides a more comprehensive evaluation of model performance, particularly for predicting unseen data. Future research should explore cross-validation methods further and extend these approaches to other areas of operant behavioral economics, such as demand modeling. 

This study highlights the need for researchers to consider both explanatory and predictive qualities when evaluating models. The authors recommend supplementing information criteria with cross-validation to ensure robust model selection, particularly in low-density data sets where incorrect model selection is more likely. The research provides valuable insights for advancing discounting model selection and improving the generalization of findings in behavioral economics. 

From the publication: "Given the results observed with MLM and the two forms of information criteria, we believe that it is important for researchers to assess model fits with information criteria but to supplement with cross-validation, at least when it is important to assess generalization."