nucleation

LIT-563: HDDM: Hierarchical Bayesian Estimation of the Drift-Diffusion Model in Python

3 min read

Properties5
StatusActive — read in full 2026-10-03 (NOTE-466); worth reading as the methods paper for HDDM, the Python package that fits the drift diffusion model by hierarchical Bayesian estimation (PyMC, MCMC), with subject parameters drawn from group distributions and trial-by-trial regressions of parameters on covariates such as EEG power. Its simulations show hierarchical Bayes recovering parameters with less error than per-subject χ²-quantile, maximum-likelihood or non-hierarchical Bayesian fits, most of all with few trials, and detecting drift-rate differences and covariate effects more often. The comparisons are the authors' own, on data simulated from the model being fitted, and its hypothesis tests on the group posterior differ from the t-tests used for the other methods.
Tagsprobabilistic-modeling, cognition, neuroscience, behavioral-integration, psychometrics
ExtendsLIT-576 — The Diffusion Decision Model: Theory and Data for Two-Choice Decision Tasks
Cited as “paper” byNOTE-466 — HDDM: Hierarchical Bayesian Estimation of the Drift-Diffusion Model in Python
Cited as “source” byTHEORY-475 — In simulations from the drift-diffusion model itself, with 12 subjects and 20 to 150 trials each, hierarchical Bayesian estimation recovers every DDM parameter with less error than per-subject maximum likelihood, χ²-quantile fitting or non-hierarchical Bayes, most of all at few trials; whether the gain survives misspecification, and how much of the power gain is the hierarchy's, is not shown

Graph View

Thomas V. Wiecki, Imri Sofer and Michael J. Frank (2013), Frontiers in Neuroinformatics 7:14 — DOI-10.3389/fninf.2013.00014

Key takeaways

  • The drift diffusion model can be fitted hierarchically: each participant’s drift, threshold, starting point and non-decision time are drawn from group distributions estimated at the same time, so participants who resemble each other constrain each other’s estimates, and every parameter comes with a posterior instead of a point estimate. The likelihood is the closed-form first-passage density (Navarro & Fuss approximation), with across-trial variability in drift integrated analytically and in starting point and non-decision time numerically.
  • Parameters can be regressed on measurements trial by trial. In the tutorial, frontal theta power predicts a higher decision threshold on high-conflict trials, and that relation reverses when the subthalamic nucleus is stimulated (data of Cavanagh et al. 2011).
  • In the authors’ simulations hierarchical Bayes has the lowest recovery error and the highest detection rate for drift differences and covariate effects, against non-hierarchical Bayes, maximum likelihood and the χ²-quantile method, with the largest gains at small trial counts. Applying a t-test to the hierarchical estimates also raised detection, so the gain is credited to the pooling, not to the Bayesian test.

Standing in the record

Filed on 2026-10-03 at the owner’s request, as the estimation tool the evidence-accumulation readings are applied with. It is a software and methods paper for cognitive modelling. The boundary was checked: the model it estimates is a model of human decisions, its use is to measure cognitive processes and relate them to brain activity, and it carries no instruction for machine-learning practice. So it belongs here rather than in the anthology. Its primary topic is probabilistic-modeling (hierarchical Bayesian inference and MCMC outside deep learning). The package is at https://github.com/hddm-devs/hddm.

It extends Ratcliff and McKoon (LIT-576): what HDDM estimates is that paper’s full diffusion model, drift v, threshold a, starting point z and non-decision time t with across-trial variabilities sv, sz and st, and HDDM cites it, with Ratcliff and Rouder 1998, as the model’s statement. The method it replaces, Ratcliff and Tuerlinckx’s χ²-quantile fit to group-averaged data, is the one Ratcliff and McKoon use. Ratcliff and Tuerlinckx (2002) is the method HDDM was compared against, but it is not in the record, so no compared_against is declared.

It is the bridge from the models to the parameters the other filings interpret. Bogacz et al. (LIT-569) read threshold as caution set by reward and starting point as prior odds; HDDM is how such parameters are estimated from individual people’s data, and its regression feature is how they are linked to a neural signal trial by trial.

Backlinks