A classifier returns a label. A person returns a label and a latency.

That extra number is not administrative dust. Reaction time is a measurement of the computation. It says how much evidence the system needed, how noisy the world felt, how cautious the person was, and sometimes which mistake was already waiting in the hand.

The drift-diffusion model is one of the cleanest little bridges between psychology and statistics. It says:

evidence arrives noisily
the decision variable accumulates it
a choice is made when the variable hits a boundary

The model is not a metaphor for indecision. It is a stochastic process with a stopping rule. That is why it keeps reappearing in memory, perception, neuroscience, psychometrics, human-computer interaction, and sequential testing. It gives choice and time the same explanation.

A Choice Has a Stopwatch Inside

For a two-alternative task, write the latent evidence as \(X_t\). Positive values favor the upper response. Negative values favor the lower response. The simplest continuous-time model is:

\[dX_t = v\,dt + \sigma\,dW_t.\]

Here \(v\) is the drift rate, \(\sigma\) is the noise scale, and \(W_t\) is Brownian motion. The process starts at \(X_0=z\) and stops at:

\[\tau = \inf\{t: X_t \ge a \text{ or } X_t \le -a\}.\]

If \(X_\tau = a\), the upper response is chosen. If \(X_\tau=-a\), the lower response is chosen. The observed reaction time is usually modeled as:

\[\mathrm{RT} = \tau + t_0,\]

where \(t_0\) is nondecision time: encoding, motor execution, display latency, and other time not spent accumulating evidence.

Each parameter has a behavioral interpretation:

  • drift rate \(v\): quality and direction of evidence;
  • boundary height \(a\): caution or required certainty;
  • start point \(z\): prior bias toward one response;
  • noise \(\sigma\): internal or stimulus uncertainty;
  • nondecision time \(t_0\): everything around the decision, not inside it.

Ratcliff’s 1978 model of memory retrieval used this diffusion machinery to explain not only accuracy but also the shape of reaction-time distributions.1 Ratcliff and McKoon’s review is still a good entry point because it keeps the core promise visible: a serious model of two-choice behavior must explain the joint distribution of choices and response times, not merely percent correct.2

A Race You Can Slow Down

The lab below simulates a positive-drift task where the upper boundary is the correct response. The paths panel shows individual evidence trajectories. The histogram shows reaction times by outcome. The sweep changes only the boundary, so you can see how caution alters accuracy, speed, and reward rate.

Try three moves:

  1. raise Boundary height and watch decisions slow down while accuracy improves;
  2. lower Evidence drift and notice that no threshold setting can fully save a weak signal;
  3. push Start bias negative and see fast errors appear before the evidence has had a fair hearing.
Correct Error Timeout Start / current Reward rate Best boundary

Deterministic synthetic experiment. Evidence evolves by dX_t = v dt + sigma dW_t with fixed symmetric boundaries. Reward rate uses correct choices minus error-cost-weighted mistakes, divided by reaction time plus a 250 ms inter-trial interval. The sweep reuses the same parameter values while changing only the boundary height.

At the default setting, the lower boundary is not crossed often because drift is positive. The errors that remain are mostly noise paths that wandered downward early. Raise the boundary and you give those paths more time to recover, but you also make every ordinary correct decision wait longer.

This is the central trade:

low boundary: quick, confident-looking, mistake-prone
high boundary: accurate, patient, expensive in time

The word “threshold” is sometimes used as if it were a mere implementation detail. In this model the boundary is the policy. It encodes how much delay is worth buying for a marginal reduction in mistakes.

Speed-Accuracy Has Prices

The speed-accuracy tradeoff is often described qualitatively: be careful and you get slower; hurry and you get noisier. The diffusion model makes it quantitative.

With positive drift and symmetric boundaries, increasing \(a\) usually increases accuracy because a random walk with upward drift has more room to reveal its direction before it is forced into a response. But the mean decision time rises because the process must travel farther before stopping.

The right boundary depends on the task environment. If errors are cheap and new trials arrive quickly, a lower boundary can maximize reward per unit time. If errors are costly, irreversible, or embarrassing, caution earns its keep.

Bogacz, Brown, Moehlis, Holmes, and Cohen analyzed two-alternative forced-choice tasks from this optimal-decision perspective, connecting diffusion-style models to reward-rate maximization and classical sequential decision rules.3 That framing is useful because it changes the question from:

what threshold seems reasonable?

to:

what boundary is optimal under this payoff schedule and time budget?

This is why the lab’s best boundary moves when you raise Error cost. The same psychophysical stimulus can justify different caution policies in a game, a medical screen, a fraud system, or a trading strategy.

The Prior Starts Moving Before Evidence

The start point \(z\) is the part of the model that feels psychological because it turns a prior into a physical head start.

If a participant expects the upper response, evidence can start closer to the upper boundary. This increases the probability of an upper choice and often speeds that choice. If the prior is wrong, it creates fast errors: the system is not merely biased at the final comparison; it begins the race already leaning.

This is one reason reaction times are diagnostic. Two datasets can have similar accuracy and different mechanisms:

  • stronger drift means the evidence itself improved;
  • lower boundary means the observer became less cautious;
  • positive start bias means the observer expected one answer;
  • lower nondecision time means the surrounding pipeline got faster.

Accuracy alone confounds these. Reaction-time distributions help separate them. The separation is not automatic, and real fits need care, but the goal is exactly right: infer the latent computation from both what was chosen and when.

The Delay Around the Decision

Nondecision time \(t_0\) is easy to dismiss because it is not the glamorous part of the model. That is a mistake.

A person may see the stimulus slowly, map evidence to keys awkwardly, or press a button through a laggy device. A machine may have feature extraction, queueing, network, or logging delay around its classifier. Those delays stretch reaction times without changing evidence quality.

If you ignore \(t_0\), you may explain peripheral latency as cognitive caution. If you overfit \(t_0\), you may hide real decision-time changes in a nuisance parameter. The small symbol is doing real hygiene work.

The same warning applies in machine-learning systems. End-to-end latency is not the same as model uncertainty. A retry, cache miss, tool call, or moderation pass can look like “harder reasoning” unless the pipeline decomposes time.

A Sequential Test With Fingers

The diffusion model also has a statistical cousin: sequential hypothesis testing. A system observes evidence over time and stops when the log-likelihood ratio is large enough to choose one hypothesis over another. Wald’s sequential probability ratio test formalized this stop-when-evidence-is-sufficient logic in statistics.4

The diffusion model is not identical to every sequential probability ratio test, but the family resemblance matters. Both replace fixed-sample thinking with a stopping rule. The sample size is not chosen in advance; it is a random variable produced by the evidence path.

This is a healthier way to read a reaction time:

not "how slow was the subject?"
but "how much evidence did this stopping policy require on this trial?"

Gold and Shadlen’s review traces the neuroscience version of the same idea: perceptual decisions can be described as evidence accumulation toward a bound, with neural activity in sensorimotor circuits reflecting the evolving decision variable.5 The exact neural implementation is richer than the toy model, but the computational grammar is recognizable.

The Toy Is Cleaner Than the Data

The toy lab is honest about the mechanism and dishonest about the mess.

Real data bring across-trial variation in drift, starting point, and nondecision time. Some responses are contaminant lapses. Boundaries may collapse with time. Participants may learn, fatigue, switch strategies, or trade speed for accuracy within a block. Multi-alternative choices are not just two-boundary diffusion with extra labels.

There is also a practical identifiability problem. A change in observed reaction times can be explained by several parameter combinations. Better stimuli can raise drift. Different instructions can lower boundaries. A response-mapping change can alter nondecision time. A base-rate cue can move start point.

So the model should not be used as a storytelling machine:

fit parameters
choose the most narratively pleasing one
publish the adjective

It should be used as a constrained generative model. Fit it, simulate from it, check whether it reproduces accuracy, conditional reaction-time quantiles, errors, correct responses, and the effect of task manipulations. If the posterior predictive checks fail, the parameters are not yet explanations.

Evidence-Accumulation Checklist

Before I believe a diffusion-model claim, I want to know:

  • whether the model predicts both choices and reaction-time distributions;
  • how outlier fast responses and long lapses were handled;
  • whether nondecision time is identifiable from the design;
  • whether drift, boundary, and starting point were manipulated independently;
  • whether hierarchical pooling was used when individual subjects are sparse;
  • whether the same parameters predict held-out conditions;
  • whether the payoff structure justifies the claimed optimal boundary;
  • whether conclusions survive plausible contaminant processes;
  • whether the task is truly two-alternative or only forced into that shape;
  • whether posterior predictive checks are shown, not merely parameter tables.

The diffusion model earns its place because it refuses to throw away time. A response is not just a class label. It is a path that stopped.

That is a useful lesson far outside psychology. Any system that waits for more evidence has a boundary somewhere: an abstention rule, a retry budget, a human review trigger, a stop-loss, a diagnostic threshold, a beam-search termination rule, a confidence gate. The question is not whether there is a threshold. The question is what cost function taught it to stand there.

Source Notes

  1. Roger Ratcliff, “A theory of memory retrieval”, Psychological Review, 1978. ↩

  2. Roger Ratcliff and Gail McKoon, “The Diffusion Decision Model: Theory and Data for Two-Choice Decision Tasks”, Neural Computation, 2008. An open full-text copy is available through PMC. ↩

  3. Rafal Bogacz, Eric Brown, Jeff Moehlis, Philip Holmes, and Jonathan D. Cohen, “The physics of optimal decision making: A formal analysis of models of performance in two-alternative forced-choice tasks”, Psychological Review, 2006. ↩

  4. Abraham Wald, Sequential Analysis, John Wiley & Sons, 1947. For historical context, see Wald’s earlier sequential-test paper in The Annals of Mathematical Statistics: “Sequential Tests of Statistical Hypotheses”, 1945. ↩

  5. Joshua I. Gold and Michael N. Shadlen, “The Neural Basis of Decision Making”, Annual Review of Neuroscience, 2007. ↩