Rethinking Model Selection in Choice Behavior

Mathematical models of choice behavior have traditionally relied on a small set of canonical forms leading to recurrent debates about which formulation is “correct.” Yet the primary aim of behavior analysis is not to identify a universally true model, but to predict and influence choice by manipulating behavior–environment relations. This paper argues that model evaluation should therefore focus on how different models capture distinct, manipulable components of the environment, and on how these components combine to support prediction and control. Using both simulated data and a reanalysis of two classic concurrent VI datasets, this paper demonstrates that the adequacy of a single parametric form depends on the structure of the choice context. Static ratio-based models (e.g., generalized matching law) perform well when contingencies are stable and closely match their assumptions, whereas in more weakly structured environments, multiple models—logit, probit, and reinforcement-history formulations—each capture different aspects of behavior. Ensemble methods, specifically Bayesian model averaging (BMA) and stacking, provide a principled way to integrate these complementary structures. The divergence between BMA and stacking weights reflects distinct theoretical questions: whether one prioritizes structural fidelity or predictive generalization. Across all analyses, the results support reframing model evaluation not in terms of universal superiority but in terms of fit, scope, and relevance to manipulable variables. Ensemble methods offer a practical and theoretically coherent framework for constructing flexible hypothesis spaces, aligning model-based inference more closely with the behavior-analytic goals of description, prediction, and control. This approach also facilitates integration with dynamic, history-sensitive formulations and offers a pathway toward more context-sensitive theories of choice.