Testing and improving the robustness of amortized bayesian inference for cognitive models

Contaminant observations often cause problems when estimating the parameters of cognitive models. In this study, we tested and improved the robustness of parameter estimation using amortized Bayesian inference. We conducted systematic analyses in two settings: a toy example (i.e., a normal distribution with an unknown mean) and a popular cognitive model, the drift diffusion model. First, we studied the stylized sensitivity curve and the breakdown point of the estimators. Next, we proposed a simple data augmentation approach that incorporated a contamination distribution into the data-generating process during training to train robust estimators. We examined several robust estimators with different contamination distributions, and evaluated their performance and cost in terms of accuracy and efficiency loss relative to a standard estimator. Introducing contaminants from a Cauchy distribution during training significantly increases the robustness of the neural density estimator, as measured by bounded sensitivity functions and a substantially higher breakdown point. Overall, the proposed method is straightforward and practical to implement, with broad applicability in fields where outlier detection or removal is challenging. (PsycInfo Database Record (c) 2026 APA, all rights reserved).