Hierarchical Bayesian measurement models for continuous reproduction of visual features from working memory

Area of Science:

• Cognitive Psychology
• Computational Neuroscience
• Psychometrics

Background:

• Visual working memory (VWM) research often uses continuous reproduction tasks to measure response distributions on circular feature dimensions.
• Existing models, like mixture models with von-Mises distributions, face challenges with estimation bias, particularly in low-performance conditions.
• There is a need for robust statistical frameworks to accurately model VWM response distributions and parameterize memory representations.

Purpose of the Study:

• To introduce and evaluate Bayesian hierarchical modeling frameworks for two distinct measurement models of visual working memory.
• To assess the performance of these models using parameter recovery simulations, focusing on precision and bias.
• To apply the developed models to experimental data to investigate memory set size effects and interference mechanisms.

Main Methods:

• Development of two Bayesian hierarchical measurement models for VWM: a mixture model and a novel interference-based model.
• Parameter recovery simulations were conducted to evaluate model performance under varying trial and subject numbers.
• Application of the models to two experimental datasets, analyzing effects of set size and feature interference.

Main Results:

• The Bayesian hierarchical framework yields precise parameter estimates, particularly when a large number of subjects compensates for a small number of trials.
• The mixture model, within the hierarchical framework, successfully minimizes previously observed estimation bias for memory precision in low-performance scenarios.
• The interference model provides unbiased and reasonably precise estimates, although certain parameters require substantial data for accurate measurement.

Conclusions:

• Bayesian hierarchical modeling provides a robust approach for analyzing visual working memory data, offering improved precision and reduced bias.
• The interference model offers a viable alternative for understanding feature-based interference in VWM, supported by experimental data.
• These modeling frameworks advance the quantitative understanding of visual working memory representations and their limitations.