Drift Diffusion Model
A 600-millisecond button press can contain a theory of mind. Roger Ratcliff's 1978 drift diffusion model did not just ask which choice a person made; it asked why the wrong answers were fast, why the right answers were slow, and why both formed a distribution instead of a point. The model's core trick is plain: treat a decision as noisy evidence wandering until it hits a wall.
How it works
The drift diffusion model describes a two-choice decision as one variable, x, moving between two boundaries. The upper boundary means one answer. The lower boundary means the other. The movement is partly signal, partly noise.
dx = v dt + s dW
v is drift rate, the average strength of evidence. s is noise scale. dW is a Wiener process step, the mathematical object behind Brownian motion. Raise the boundary and the person waits longer before acting. Raise the drift rate and the person gets faster and more accurate.
Ratcliff's full version adds starting-point bias, non-decision time, and across-trial variation. That matters because a reaction time includes the retina seeing dots, the cortex forming a judgment, and the finger moving toward a key.
Why reaction time matters
Accuracy alone hides the mechanism. A subject who gets 80% correct might be cautious with weak evidence, biased toward one answer, or fast with a high motor delay. Reaction time separates those stories.
| Parameter | Meaning | Behavioral signature |
|---|---|---|
v |
Evidence quality | Faster correct responses |
a |
Boundary height | Slower responses, fewer errors |
z |
Starting bias | More fast choices on one side |
t0 |
Perception plus movement time | Whole distribution shifts later |
The sharp line is this: the error is not the residue after accuracy. The error is data.
Where it shows up
The model became a workhorse for two-choice tasks: memory recognition, lexical decision, visual motion discrimination, and speed-accuracy tradeoff experiments. In monkey motion tasks, neurons in lateral intraparietal cortex ramp before a saccade, which made the accumulator story feel biologically plausible. Gold and Shadlen's 2007 review is the canonical bridge between the behavioral model and neural decision-making.
The statistical ancestor is Abraham Wald's 1947 sequential analysis. Wald asked when a factory inspector should stop sampling and make a call. Drift diffusion gives the same skeleton to a brain: sample, accumulate, stop.
What's contested
The model fits reaction-time distributions, but a fit is not a circuit diagram. A brain is not one clean accumulator with Gaussian noise; it is recurrent, parallel, value-sensitive tissue with motor preparation mixed into perception.
The live question is how literal the accumulator is. Ramping activity in LIP looks like evidence accumulation in some motion tasks, but other decisions involve urgency signals, collapsing boundaries, multiple options, and changing rewards. Drift diffusion may be the right compression, not the right anatomy.
Why this has to do with other realms
Drift diffusion sits beside concept bayes theorem because each evidence sample updates the odds of one answer over another. It sits beside concept information theory because stronger signals reduce uncertainty faster. It also sharpens concept signal detection theory: signal detection asks how separable two signals are; drift diffusion asks how long a noisy system should wait before paying the cost of delay.
That makes the model useful outside neuroscience. A radiologist reading a scan, a trader watching a price tape, and a founder choosing whether to ship a flawed build all face the same shape of problem: stop too early and error rises, wait too long and the world moves.
An open question
What is the smallest decision model that keeps the drift diffusion model's discipline, but handles three choices, changing rewards, and a deadline that moves while the mind is still sampling?
Key Sources
- Ratcliff, R. 1978. "A theory of memory retrieval." Psychological Review 85(2):59-108. The canonical paper joining choice and reaction time in one model.
- Gold, J. I. and Shadlen, M. N. 2007. "The neural basis of decision making." Annual Review of Neuroscience 30:535-574. The main bridge from accumulator models to neural activity.
- Bogacz, R., Brown, E., Moehlis, J., Holmes, P., and Cohen, J. D. 2006. "The physics of optimal decision making." Psychological Review 113(4):700-765. Formal comparison of diffusion models and optimal stopping.
- Ratcliff, R. and McKoon, G. 2008. "The diffusion decision model: theory and data for two-choice decision tasks." Neural Computation 20(4):873-922. Later account of parameters, fits, and model behavior.
Further Reading
- Wald, A. 1947. Sequential Analysis. The statistical root of deciding when enough evidence is enough.
- Laming, D. R. J. 1968. Information Theory of Choice-Reaction Times. Early reaction-time theory before diffusion models became the default tool.
- concept prediction error - another small signal that explains behavior across time.
- concept reinforcement learning - where evidence accumulation meets reward, policy, and action.
See Also
- concept bayes theorem
- concept information theory
- concept signal detection theory
- concept prediction error
- concept reinforcement learning
- concept queueing theory
Abhishek's take
What grabs me is not that the model is elegant; it is that it refuses to throw away the timing data. Most operational mistakes look like bad choices from a distance, but the timing often tells the real story: weak signal, low caution, biased start, or a deadline that punished waiting. I like models that make a decision leave fingerprints.
Tags: #decision-making #reaction-time #evidence-accumulation #cognitive-models