Rethinking Model Selection in Choice Behavior

Technical Summary: Rethinking Model Selection in Choice Behavior

Problem Statement
Traditional research in the experimental analysis of choice behavior has relied heavily on a "theoretical monism," where a single canonical model (most often the Generalized Matching Law or specific logit/probit formulations) is selected to describe choice across diverse contexts. This approach frames model comparison as a search for the single "correct" formulation. However, the author argues that this perspective is fundamentally misaligned with the goals of behavior analysis, which prioritize the prediction and control of behavior through the manipulation of environmental variables (e.g., reinforcement rates, magnitudes, changeover delays).

The paper posits that no single parametric form adequately describes choice across all contexts. Static ratio-based models may perform well under stable contingencies but fail to capture history-dependent dynamics or weakly structured environments. Furthermore, the reliance on a single model often obscures the fact that different models capture distinct, manipulable components of the environment. The current methodological framework, which seeks a universal "best" model, limits the ability to construct flexible hypothesis spaces necessary for prospective control.

Methodology
The paper employs a unified Bayesian framework to evaluate and integrate multiple models of choice behavior. The methodology consists of three primary components:

1. Ensemble Learning Techniques: Instead of selecting a single "winner," the study utilizes two ensemble methods:

   • Bayesian Model Averaging (BMA): Calculates posterior model probabilities, assuming one of the candidate models is the true data-generating mechanism (an "M-closed" perspective).
   • Stacking: Determines optimal weights by maximizing the leave-one-out (LOO) predictive log density of a mixture distribution, assuming no single candidate is fully correct (an "M-open" perspective). Stacking emphasizes predictive complementarity over structural fidelity.

2. Simulation Studies:

   • Simulation 1 (Static): Virtual choice data ($N=2000$) were generated using a logistic function. Three static Generalized Linear Models (GLMs) were tested: a Log-ratio model (Generalized Matching Law), a Logit model, and a Probit model.
   • Simulation 2 (Dynamic): Virtual data ($N=3000$) were generated using a history-dependent process (Value Difference model). Four models were tested: a Simple model (previous consequence only), an Interaction model (choice $\times$ consequence), a Value Difference model (dynamic value updating), and a Dynamic GML model (ratio-based dynamic updating).

3. Empirical Reanalysis:

   • Two classic datasets on concurrent variable-interval (VI) schedules (Alsop & Elliffe, 1988; Elliffe & Alsop, 1996) were reanalyzed.
   • Because the data were aggregated at the condition level, Beta regression was used with subject-specific random effects.
   • Five models were compared: three static models (GML, Logit-difference, Probit-difference) and two pseudo-dynamic models incorporating exponentially weighted moving average (EWMA) reinforcement histories.
   • Sensitivity to the EWMA decay parameter ($\lambda$) and predictive accuracy (ELPD) were assessed.

Key Contributions

• Reframing Model Selection: The paper argues that model selection should not be viewed as a contest for a universal law but as an evaluation of "reasonable approximations" to specific theoretical questions. The adequacy of a model is conditional on the purpose (explanation vs. prediction), the mechanism, and the context.
• Distinction Between BMA and Stacking: The study demonstrates that BMA and Stacking weights often diverge. BMA weights reflect structural fidelity to the assumed generative process, while Stacking weights reflect predictive complementarity. This divergence highlights that different models may capture distinct, non-overlapping aspects of behavioral organization (e.g., steady-state sensitivity vs. transient history effects).
• Hybrid Model Design: The paper proposes "hybrid" formulations that integrate static ratio-based structures with dynamic history-dependent terms. This allows for the systematic testing of how different temporal mechanisms (e.g., short-term repetition vs. long-term accumulation) jointly contribute to choice controllability.
• Context-Dependent Hypothesis Spaces: The analysis shows that the "best" hypothesis space is not invariant; it reorganizes based on the structure of the arranged contingencies. In some contexts, a single model dominates; in others, the space is diffuse, requiring multiple models to explain the data.

Results

• Simulation 1: When the true generative process was a Logit model, BMA correctly assigned the highest weight to the Logit model (0.72). However, Stacking assigned the highest weight to the Probit model (0.70), with the Logit model receiving 0.30. This occurred because the Probit model's error structure was slightly more complementary to the Logit in the mixture, despite the Logit being the true generator. The Log-ratio model performed poorly in both metrics due to a structural mismatch (ratio vs. difference).
• Simulation 2: When the true process was history-dependent (Value Difference model), the true model received the highest BMA weight (0.97). However, Stacking distributed weights across the Interaction model (0.23) and Dynamic GML (0.31), indicating that these models captured complementary temporal patterns (short-term interactions vs. long-term trends) that improved overall predictive accuracy.
• Empirical Reanalysis (1988 Dataset): The hypothesis space was "sharply concentrated" around the Generalized Matching Law (GML). Static ratio-based models dominated both BMA and Stacking weights, and dynamic terms added little predictive value. This suggests the contingencies in this dataset produced stable, ratio-based control.
• Empirical Reanalysis (1996 Dataset): The hypothesis space was "diffuse." No single model achieved consistent dominance. Model weights were distributed across GML, Logit, and Probit formulations, and dynamic models showed unstable directional changes. This suggests the contingencies were weakly structured, and the mapping between reinforcement and choice was not invariant.
• Robustness: The sensitivity of the reinforcement-history coefficient ($\beta_1$) and predictive accuracy (ELPD) were robust across a wide range of EWMA decay parameters ($\lambda$), indicating that the conclusions regarding history effects were not artifacts of specific parameter choices.

Significance and Claims
The paper claims that the traditional search for a single "correct" model of choice is conceptually misguided given the context-dependent nature of behavioral control. By adopting an ensemble learning framework, researchers can:

1. Construct Flexible Hypothesis Spaces: Move beyond "M-closed" assumptions to acknowledge that multiple models may be partially valid and capture different components of the behavioral process.
2. Enhance Prospective Control: Ensemble methods provide more robust predictive distributions for forecasting behavior under new or non-stationary contingencies (e.g., changes in reinforcement rates or schedules) than any single model.
3. Align with Behavior-Analytic Goals: This approach aligns model-based inference with the core aims of behavior analysis—description, prediction, and control—by treating models as tools to identify manipulable variables (sensitivity, bias, history kernels) rather than as fixed laws of nature.

The author, Naoki Kamiya, who was affiliated with The Institute of Statistical Mathematics when the work reported in the manuscript was conducted and who now conducts research through Relations & Behavior Lab, Tokyo, Japan, concludes that theoretical progress in behavior analysis depends less on defending a specific parametric form and more on designing richer hypothesis spaces that reflect the manipulable dimensions of the environment and the multiple processes through which reinforcement exerts control.