Evidence Accumulation & Response Times

In previous work, we introduced a probabilistic formulation of the EZ drift diffusion model (EZ-DDM) that enables hyper-efficient Bayesian inference of the drift rate, boundary separation, and non-decision time parameters from binary choice and response time (RT) data. By deriving sampling distributions for the summary statistics underlying the EZ estimators, we constructed a “proxy” likelihood that can be implemented in JAGS, Stan, and PyMC. This formulation supports hierarchical Bayesian models with excellent sampling properties, allowing researchers to decompose variability in model parameters across individuals, experimental conditions, and stimuli. We now present a robust extension of the hierarchical Bayesian EZ-DDM that addresses the sensitivity of standard EZ-based estimators to contaminant RTs. The robust variant replaces the mean and variance of RTs with quartile-based statistics, while preserving the generative model structure. Simulation studies varying sample size, trial count, effect size, and contamination level show that the robust implementation matches the accuracy of the standard hierarchical EZ-DDM on clean data and remains stable under contamination, without sacrificing computational efficiency. Beyond simulations, we illustrate the potential of our robust hierarchical Bayesian EZ-DDM with two experimental datasets: a large lexical decision study and a recognition memory study with symbolic stimuli. In both cases, we model the effect of stimulus-level properties on model parameters indexing information processing, response caution, and non-decision time components, and explore how these effects vary across participants and conditions. Together, these examples demonstrate how the robust hierarchical EZ-DDM supports scalable, experiment-based inferences about cognitive processes under realistic data conditions.

This is an in-person presentation on July 18, 2026 (09:00 ~ 09:20 EDT).

An emerging approach addressing Cronbach’s (1957) experimental-correlational-psychology divide combines two components, a structural-equation model (SEM) and evidence-accumulation models (EAMs, e.g., Schmiedek et al., 2007, Ratcliff et al., 2010, Schubert et al., 2016, 2019a, 2019b, Lerche et al., 2020, Löffler et al., 2024). The aim is to use hierarchical latent structures to understand relationships among behavior in different decision tasks performed by the same individual. The first layer consists of EAMs that have been refined and validated in the experimental-psychology literature over many years based on data from simple choice tasks. The EAMs distil behavioural measures (e.g., choices and associated response-time distributions) into a small set of psychologically meaningful parameters. Further layers decompose correlations among these parameters, both within and between tasks, into individual differences in higher-order abilities and preferences. Recently, Rey-Mermet et al. (2025) applied such structural-equation evidence-accumulation models (SEEAMs) to a wide variety of tests of executive-function based on standard EAMs developed for simple-choice tasks. We focused their approach on a subset of these tests, decision-conflict (e.g., Stroop and Simon) tasks, and extended it in two ways. First, we used EAMs tailored to provide a more accurate fine-grained description of decision conflict (Heathcote et al., 2026). Second, we used seamless simultaneous fitting of EAM and SEM components to address overconfidence and attenuation in previous SEEAMs caused by two-stage estimation (i.e., first fitting EAMs then feeding the resulting parameter estimates into a SEM). We applied these innovations to simple choice and conflict task data from Eisenberg et al. (2019) and report the degree to which a range of latent constructs—including simple decision ability, conflict-resolution ability, and response-caution preferences—are task general.

This is an in-person presentation on July 18, 2026 (09:20 ~ 09:40 EDT).

Parametric models of human behavior (e.g., response time models) are usually estimated using advanced computational techniques that vary in sophistication. In turn, the use of such models requires a level of expertise that presents a barrier of entry to beginners in the field. Additionally, many of these methods rely on iterative techniques that in some cases do not converge, or worse, provide estimates that miss the mark by converging at a point that is locally optimal, but not globally optimal. In contrast, closed-form estimators allow a user to get parameter estimates by applying relatively simple formulas to a set of summary statistics from the observed response times. In this talk, we present and evaluate two closed-form techniques for estimating ex-Gaussian parameters from summaries of response time data. The first uses the method of moments to transform observed summary measures of mean, standard deviation, and skewness directly into ex-Gaussian parameters. The second is a novel, robust approach using order statistics to obtain closed-form approximations to the ex-Gaussian parameters via a quadrature-based decomposition of interquartile half-spreads into orthogonal symmetric and skew components. Initial parameter recovery work indicates that both methods provide fast, reasonably accurate computation of ex-Gaussian parameters, though the new robust method may be preferred when contaminant trials are present. Because of their balance of accuracy and ease of use, these methods may be an excellent way for new users to "go beyond the mean" when analyzing response time data.

This is an in-person presentation on July 18, 2026 (09:40 ~ 10:00 EDT).

Amortized Bayesian model comparison reframes evidence estimation as a supervised learning problem: simulate from candidate cognitive models, train a classifier, and read off Bayes factors (or approximate evidence) at test time. It’s fast, scalable, and increasingly popular, but it can also be confidently wrong. Over the past ten years, the field has explored many new concepts, some of which have seen limited uptake in cognitive modeling. This talk presents a guide to the dos and don’ts of Bayesian model comparison as classification for cognitive models. It compares training objectives and loss functions, shows why accuracy is a poor proxy for evidence quality, and highlights diagnostics based on calibration, coverage, and self-consistency. I then survey newer ideas—such as ensembling, joint learning, and explainable ML, discussing when they fix problems versus when they merely expose or hide them. The takeaway is a set of concrete recommendations for building, validating, and reporting amortized evidence estimates that remain trustworthy.

This is an in-person presentation on July 18, 2026 (10:00 ~ 10:20 EDT).