Model-based experimental manipulation of probabilistic behavior in interpretable behavioral latent variable models.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Behavioral Economics

Background:

• Interpretable latent variable models link behavior to cognitive processes.
• These models offer insights into cognition but are underutilized for prediction.
• Predictive power can guide experimental design and adaptive testing.

Purpose of the Study:

• To demonstrate the use of latent variable models for generating precise behavioral predictions.
• To apply this framework to the cognitive process of delay discounting (DD).
• To design adaptive experiments that elicit specific behavioral outcomes.

Main Methods:

• Inferred delay discounting (DD) models from behavioral data.
• Leveraged inferred models to design a second, adaptive DD task.
• Experimental trials were designed to elicit nine graded discounting probabilities.
• Validated models by assessing out-of-sample prediction error against alternatives.

Main Results:

• The proposed framework successfully induced nine distinct levels of discounting probabilities.
• The applied model showed high validity with low prediction error compared to alternatives.
• Evidence for inter-individual differences in the most suitable underlying models was found.

Conclusions:

• Latent variable models can generate falsifiable behavioral predictions for adaptive experimentation.
• This framework enhances the resolution of cognitive constructs and their neural substrates.
• The method is adaptable for studying other cognitive processes like reinforcement learning.