How the brain accumulates noisy evidence over time to reach a binary decision — and why speed and accuracy trade off.
The Drift Diffusion Model (DDM), formalised by Roger Ratcliff (1978), is the dominant computational account of two-choice decision-making. It treats a decision as a noisy evidence-accumulation process: a particle drifts from a starting point toward one of two boundaries while being buffeted by random noise. The first boundary reached determines the response; the time taken determines the reaction time. The DDM simultaneously explains choice accuracy and the full shape of reaction-time distributions in a single coherent framework.
Drift rate (v): The average speed and direction of evidence accumulation. High |v| means strong, reliable evidence — fast correct responses and few errors. Low |v|, near zero, means ambiguous evidence — slow responses with many errors. Negative v biases accumulation toward the wrong boundary.
Boundary separation (a): The total distance between the two decision thresholds. Wide boundaries (high a) require more evidence before committing — slower but more accurate. Narrow boundaries (low a) are fast but error-prone. This parameter captures the speed–accuracy trade-off directly.
Starting point (z): Where the evidence accumulator begins relative to the two boundaries. A bias toward one boundary (z ≠ a/2) produces faster responses for that option — modelling prior expectation or response bias.
Non-decision time (t₀): Time attributable to peripheral processes — sensory encoding and motor execution — that do not involve the decision itself. It shifts the whole RT distribution without affecting accuracy.
Why people can be fast or accurate, but rarely both: Lowering the boundary separation (a) speeds up every decision because less evidence is needed, but the accumulator is more likely to be caught by noise on the wrong side — increasing errors. Raising a has the opposite effect. Crucially, the same change in task difficulty (changing v) affects both speed and accuracy simultaneously, which is the hallmark prediction that separates DDM from simpler threshold models.
Urgency signals: In deadline paradigms, decision boundaries often collapse over time, trading accuracy for speed as time pressure increases — a dynamic extension of the standard model called the collapsing-boundary DDM.
Cortical ramping activity: Neurons in areas such as the lateral intraparietal cortex (LIP) and dorsomedial frontal cortex show firing rates that rise gradually toward a fixed threshold before a saccade decision — a neural signature of evidence accumulation that maps directly onto the DDM accumulator.
Basal ganglia and boundaries: The striatum and subthalamic nucleus influence the boundary separation, providing a biological substrate for speed–accuracy adjustments. Parkinson's disease, which damages these circuits, typically narrows boundaries — faster but more error-prone decisions.
BOLD signals and DDM parameters: fMRI studies correlate trial-by-trial drift-rate estimates with BOLD activity in sensory cortex, and boundary-separation estimates with prefrontal and ACC activity — linking model parameters to distinct neural systems.
Random Dot Motion: Varying dot coherence changes drift rate while boundary and non-decision time stay constant — the cleanest laboratory manipulation of v.
Flanker Task: Incongruent flankers reduce drift rate (conflicting evidence) and may widen boundaries (increased caution) — decomposing the congruency effect into distinct cognitive mechanisms.
Stroop Task: Colour–word conflict reduces drift rate toward the correct colour response, and elevated non-decision time reflects additional processing overhead from word reading.
Lexical Decision: High-frequency words produce higher drift rates than low-frequency words or non-words, capturing lexical access speed within the DDM framework.