Improving Parameter Recovery in Computational Models: Employing Outlier-Insensitive Loss Functions.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Decision Science

Background:

• Parameter estimation reliability is vital for computational models in decision-making.
• Log-likelihood loss functions are standard but sensitive to outlier data.
• Outliers can bias parameter estimation and subsequent analyses.

Purpose of the Study:

• To compare the performance of outlier-insensitive loss functions against the standard log-likelihood function.
• To evaluate parameter recovery in computational models under varying proportions of outlier data.
• To assess the impact of outliers on statistical power in cognitive modeling.

Main Methods:

• Compared three outlier-insensitive loss functions with log-likelihood.
• Tested performance in a reinforcement learning model and an intertemporal choice model.
• Systematically varied outlier presence from 0% to 25% of data.

Main Results:

• Log-likelihood shows substantial impairment in parameter identification with even small outlier proportions.
• Outlier-insensitive loss functions markedly improve computational model parameter recovery.
• As few as 5% outlier trials can undermine statistical power to detect condition differences.

Conclusions:

• Outlier-insensitive loss functions are recommended for non-hierarchical model estimation.
• These functions enhance the reliability of cognitive models as measurement tools.
• Accounting for outliers is crucial for accurate parameter estimation and valid conclusions.