What do diffusion models reveal about multisensory benefits that response times and accuracy miss?

Why can two groups look identical on accuracy yet differ inside?

Accuracy and reaction time are summary numbers. Accuracy collapses a whole session into a percentage correct, and reaction time is usually an average across trials. A diffusion model instead fits the full spread of responses — both correct and incorrect, fast and slow — and estimates the parameters that could have produced that spread [5]. That is why it can detect a difference that the summaries wash out.

A concrete case: in one study, older and younger adults showed no significant difference in choice accuracy between multisensory and unisensory conditions, yet modeling revealed that older adults still gained a multisensory benefit in how efficiently evidence accumulated, especially on easier, lower-difficulty trials [2]. The accuracy score said "no difference"; the model said the underlying evidence-gathering was still being helped.

The same logic applies to numerosity judgments. In auditory-visual number discrimination, hierarchical drift diffusion modeling uncovered a multisensory advantage that the classic accuracy and reaction time measures had obscured [1]. The benefit was real but lived in the decision process, not in the headline scores.

What does the model actually pull apart?

Think of a decision as a race: evidence for each option accumulates until one crosses a threshold. The drift rate is how fast and reliably that evidence builds — a measure of processing efficiency. The starting point captures a prior bias toward one answer before any evidence arrives. Non-decision time covers the sensing and motor steps around the choice itself.

This separation matters because different mechanisms can produce the same outward behavior. In a study of autistic and non-autistic adults judging whether audiovisual speech was simultaneous, response times and accuracy were fitted to a diffusion model so that drift rate and starting point could be estimated separately. There were no between-group differences in either drift rate or starting point [3]. Without the model, a difference in response bias could have been mistaken for a difference in perception — the model lets the two be told apart.

The model also flags when a benefit is not about the decision at all. In work on auditory accessory stimuli and visual processing, combining electrophysiology with a diffusion model indicated that response times and accuracy did not fully capture the effects; components such as motor activation and non-decision processing were influenced instead [4]. In other words, some multisensory gains are about getting the response out, not about deciding faster.

When is the hidden benefit most likely to appear?

Difficulty is one lever. The age-related accumulation benefit was clearest on lower-difficulty trials [2], which suggests the model is most informative when the task is not already near ceiling or floor — exactly the conditions where summary scores tend to flatten.

Uncertainty is another. When visual uncertainty was manipulated in a multisensory decision task, drift-diffusion modeling of the behavioral data, alongside electrophysiological recordings, pointed to adaptive adjustments in how the senses are combined that standard reaction time and accuracy measures did not capture [5]. The takeaway for a reader is practical: if you only compare average speed and percent correct across conditions, you may conclude "no multisensory effect" when the effect has simply moved into a parameter you did not measure.

What the modeling approach does not settle on its own

A model-based benefit is an inference about a latent process, not a direct observation of it. The evidence here is consistent with multisensory input improving evidence accumulation or non-decision stages, but it does not by itself prove a specific neural mechanism. The accessory-stimulus work, for instance, points to motor and non-decision components as candidates [4] rather than pinning down a single cause.

The approach also depends on how constructs are defined. In the simultaneity-judgment study, exploratory analysis found that operationalizing the temporal binding window in different ways changed whether a group difference appeared at all [3]. That is a caution about measurement choices, not about whether diffusion models are useful: the same data can yield different conclusions depending on the definition you bring.

Finally, the payoff is diagnostic rather than automatic. Diffusion modeling tells you where in the decision a multisensory benefit sits — efficiency, bias, or response execution — and that is precisely the information accuracy and reaction time leave out [1][2][5]. It does not, by itself, tell you why the senses are being combined that way.

Sources used in this answer

Related research questions