How much data do we need to estimate computational models of decision-making? The COMPASS toolbox

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Psychology

Background:

• Estimating computational model parameters requires sufficient data.
• Current goodness-of-recovery studies may yield suboptimal sample sizes.
• A generalized concept of statistical power is needed for data requirement determination.

Purpose of the Study:

• To propose a novel approach for determining data needs in computational modeling.
• To introduce a generalized concept of statistical power for parameter estimation.
• To provide a practical tool for sample size calculation in computational modeling.

Main Methods:

• Formulated a generalized statistical power concept.
• Developed a Python-based toolbox (COMPASS) for sample size determination.
• Utilized simulations to evaluate data requirements for the Rescorla-Wagner model.

Main Results:

• The proposed method offers an alternative to standard goodness-of-recovery studies.
• COMPASS facilitates sample size calculations for specific computational models.
• Simulations indicated a high number of trials per person is essential for power.

Conclusions:

• The novel approach and COMPASS toolbox aid in determining appropriate sample sizes.
• High trial counts per participant are critical for robust parameter recovery in learning models.
• This work advances the methodology for data-driven computational modeling research.