Roger Ratcliff and Gail McKoon (2008), Neural Computation 20(4):873–922 — DOI-10.1162/neco.2008.12-06-420
Key takeaways
- The diffusion model separates three things that RT and accuracy confound: the quality of the evidence (drift rate), how much evidence is required (boundary separation), and everything outside the decision (non-decision time). With across-trial variability in drift and starting point it also predicts when errors are slower than correct responses and when they are faster.
- Its constraints come from distribution shape. A drift change moves the .9 RT quantile about four times as much as the .1 quantile; a boundary change moves them in roughly a 2:1 ratio. Across conditions only location and spread change, not shape. In the three experiments (14 to 17 participants each) difficulty was fitted by drift alone, speed and accuracy instructions by boundaries alone, and a 75:25 stimulus proportion mainly by starting point, about a third of the way toward the likelier boundary.
- Fitted to individuals, the parameters dissociate. Across 18 data sets accuracy correlated with drift and mean RT with boundary separation. Older adults’ slowing was mostly wider boundaries, not poorer evidence, and aphasic patients’ slowing was mostly boundaries and longer non-decision time.
Standing in the record
Filed on 2026-10-03 at the owner’s request, as the record’s statement of the single-process diffusion model, beside its competing-accumulator rival (Usher and McClelland, LIT-565), the analysis that relates the two and ties both to the sequential probability ratio test (Bogacz et al., LIT-569), and the Bayesian estimation package built on it (HDDM, LIT-563). It is chosen over Ratcliff’s original 1978 paper because it is a lawfully readable, current statement of the full model with its variability parameters and its uses. The 1978 paper is not filed.
It is a model of a cognitive process in people and monkeys, and it gives no
instruction for machine-learning practice, so it belongs here rather than in
the anthology. It is tagged behavioral-integration for the vocabulary’s
“evidence accumulation to a decision”, and probabilistic-modeling because it
is used, above all, as a statistical measurement model that turns RT
distributions into parameters.
Its relation to the record’s other readings is mainly through LIT-569 and LIT-563. It does not bear on the record’s emotion-regulation theory (THEORY-056) or its free-will theories.
No relation is declared. The paper runs no comparison against another filed model: its §6 judges, from other work, that recent accumulator models “may be as successful as” the diffusion model, and its §7.1 objects that Bogacz et al.’s account of criterion setting cannot explain calibration from verbal instruction or without feedback. That is a limitation it alleges, not a correction of a claim.