How many trials are required for parameter estimation in diffusion modeling? A comparison of different optimization...

Area of Science:

• Cognitive psychology
• Computational neuroscience
• Psychometrics

Background:

• Diffusion models offer detailed insights into cognitive processes like information accumulation and response bias in binary decision tasks.
• These models leverage response time distributions for both correct and error responses, enabling robust parameter estimation.

Purpose of the Study:

• To compare the efficiency and robustness of parameter recovery in diffusion models with varying complexity and trial numbers.
• To evaluate different optimization criteria (maximum likelihood, Kolmogorov-Smirnov, chi-square) within the fast-dm software.
• To compare fast-dm performance against the EZ approach and a Bayesian implementation.

Main Methods:

• Simulation studies were conducted using the fast-dm software (Voss et al., 2015).
• Model complexity ranged from few to many free parameters, with trial numbers from 24 to 5,000.
• Three optimization criteria were assessed: maximum likelihood, Kolmogorov-Smirnov, and chi-square.

Main Results:

• Maximum likelihood estimation demonstrated superiority with uncontaminated data.
• Kolmogorov-Smirnov outperformed other methods when fast response time contaminants were present.
• Chi-square estimations generally yielded less precise results compared to maximum likelihood and Kolmogorov-Smirnov.
• The fast-dm software's performance was competitive with EZ and Bayesian approaches.
• Robust parameter estimation is achievable even with fewer than 100 trials under specific conditions.

Conclusions:

• The choice of optimization criterion in diffusion modeling is crucial and depends on data quality, particularly the presence of contaminants.
• The fast-dm software provides a robust platform for diffusion model analysis.
• Recommendations for optimal trial numbers are provided based on model complexity, suggesting that sufficient data for reliable parameter estimation can be obtained even with relatively small sample sizes.