The WIG (weighted individual and group) shrinkage estimator

Area of Science:

• Behavioral science
• Statistical modeling
• Machine learning

Background:

• Regularization techniques constrain parameter estimates to improve model prediction.
• Despite their utility in other fields, regularization methods are underutilized in behavioral journals like JEAB.
• Potential reasons for this lack of adoption include complexity and specific objections to existing schemes.

Purpose of the Study:

• To review common regularization methods and explore their limited use in behavioral research.
• To propose a novel shrinkage estimator, the weighted individual and group (WIG) data method.
• To evaluate the performance of WIG against traditional methods in a specific modeling context.

Main Methods:

• A review of existing regularization techniques (e.g., parameter shrinkage towards a group mean or zero).
• Development of the weighted individual and group (WIG) data mixing estimator.
• A simulation study comparing WIG with ordinary least squares (OLS) regression for model selection.

Main Results:

• The proposed WIG estimator offers an alternative approach by mixing data rather than constraining parameters.
• In a simulation study focused on selecting matching-law-based punishment models, WIG demonstrated superior performance.
• On average, WIG outperformed OLS regression in the context of this specific model selection task.

Conclusions:

• The WIG method provides a potentially valuable tool for behavioral researchers seeking improved predictive accuracy.
• This approach addresses some objections associated with traditional regularization techniques.
• Further application and validation of WIG in diverse behavioral modeling scenarios are warranted.