Fast uncertainty quantification in EZ cognitive models

Classical approaches to uncertainty quantification in cognitive modeling rely on computationally expensive Monte Carlo methods or resampling of raw trial data, which creates a barrier for real-time analysis and large-scale studies. We present a computationally efficient bootstrap method that operates directly on summary statistics, exploiting the synthetic likelihood structure of a small class of cognitive models that includes the simple diffusion model and the circular diffusion model in addition to signal detection and multinomial processing trees. The method relies on a numerical transformation-of-variables technique in which known sampling distributions of summary statistics of behavior are parametrically bootstrapped, after which the resampled statistics are transformed to parameter estimates with a known analytical system. This approach does not require additional assumptions beyond those already made by the models themselves, but achieves over 1000-fold speed improvements over already efficient fully Bayesian methods. The proposed method makes real-time uncertainty quantification accessible and enables new applications in adaptive testing, meta-analyses, and exploratory data analysis.