Abstract

Author: Yufei Wu, Stefan Radev, Francis Tuerlinckx This paper addresses the challenge of achieving robust and reliable parameter estimation in cognitive models when contaminations and outliers are present. Studies have demonstrated that amortized Bayesian inference (ABI) has the potential to facilitate inference of arbitrarily complex cognitive models via simulations and neural networks. However, it remains unclear whether ABI is robust to outliers and how it can be robustified. Our findings show that by simulating a more realistic data-generation process that incorporates contaminants and using the contaminated data to train the neural networks, the inference with ABI achieves a high degree of robustness while maintaining efficiency. These results are significant because they provide a powerful yet practical solution to handle outliers in empirical research with ABI .