Computational Science
The Slowest Responses May Tell a Different Story Than the Mean
2026-09-10
A higher mean reaction time can reflect a broad shift or a small number of very slow trials. Those explanations imply different modeling decisions. Computational reanalysis can examine the distribution before a team describes the change as slower processing, reduced attention or a product benefit in reverse.
The computational starting point
Look for an existing task dataset containing trial-level response times, accuracy, stimulus conditions and participant identifiers. OpenNeuro studies sometimes include behavioral records, but availability must be checked study by study. Preserve anticipatory responses, missing responses and exclusion reasons rather than silently deleting inconvenient observations before inspecting their structure.
Sources: OpenNeuro documentation.
Where intuition enters
An intuitive impression of intermittent interruption becomes a mixture hypothesis: most responses remain similar while a smaller tail changes. The alternative is a distribution-wide shift. Neither model licenses a psychological label on its own. Both need a declared relationship between the observed data and the proposed mechanism.
A test that can disagree
Compare participant-level quantiles and accuracy before fitting a distributional model. Fit a shift model and a constrained mixture model on training blocks, using an estimation method appropriate to the chosen likelihood. Numerical fitting tools such as SciPy do not decide whether a mixture component has psychological meaning.
Evaluate held-out tail probabilities, not only average error. Repeat with predefined exclusion thresholds and a model for censored or missing responses where appropriate. Include simulations in which only a few logging errors create extreme times. If the mixture interpretation depends on one participant or an arbitrary cutoff, retain that sensitivity prominently.
Sources: SciPy nonlinear least squares.
An illustrative decision
Suppose median response time barely changes, but the upper tail grows. That supports investigating a distributional change, not automatically diagnosing lapses of attention. If the tail follows browser timing interruptions in the metadata, a recording explanation becomes more plausible than a new cognitive mechanism. The original intuition must yield to that evidence.
What the research would deliver
The resulting brief would show distributional contrasts, alternative explanations and the minimum defensible claim. It could help a behavioral-software team choose an endpoint for further computational validation. It is not an assessment of a person's mental health, and it should not turn an aggregate benchmark into an individual diagnosis.
Questions this raises
Why not simply remove the slowest trials?
Because they are central to the question. Exclusion rules should be justified and frozen, with their consequences visible.
Can this show that an intervention works?
Only to the extent permitted by the original design. Reanalysis cannot manufacture randomization or remove every source of confounding.
Sources and their limits
- OpenNeuro documentation. Archive access, not confirmation that a particular study fits this question.
- SciPy nonlinear least squares. Numerical fitting, not physical identification.
Prepared with AI assistance. The linked sources support the specified technical points; they do not validate applied psionics as a whole or guarantee a result for a client.
Read the editorial and evidence standard.
Continue reading
Explore Scientific Oracle consultingfor a scoped review of an existing-data research decision. Start with a non-confidential outline of the question, available evidence and the decision it needs to inform.