The question before the experiment

A useful question names the system or population, the conditions being compared, the outcome, and the observation window. “Does recovery occur?” leaves these choices open. “Does locomotor activity return within a defined tolerance of a time-matched control before a later challenge?” makes the measurement and decision explicit.

The target quantity, sometimes called the estimand, might be a difference in average duration, a change in choice probability, or a difference in response to a challenge. Those are different questions. Decide which one matters before selecting a convenient summary.

A feasibility study has a different target again: can the assay detect the relevant range, can the comparison be made, and is the expected precision sufficient? It can answer those questions without establishing the proposed biological mechanism.

The unit receiving an independently assigned treatment may differ from the unit being measured. If a treatment is assigned to a culture well, cells from that well are subsamples of that treatment unit. Repeated images improve characterization of the well; they do not create new independently treated wells. Lazic and colleagues make this distinction explicit.

Technical repeats characterize the measurement of a sample; biological replication samples distinct biological material. Those labels do not replace the design: samples sharing one treatment assignment still do not provide independent treatment replicates.

For a behavioral study, record the assignment, repeated trials, individual identity, and any shared batch or enclosure. The analysis must reflect that structure. The number of rows in a file is not, by itself, the experimental sample size.

Source: Lazic, S. E., Clarke-Williams, C. J., & Munafò, M. R.

Write down the expected direction and time course under at least two plausible accounts. A learning explanation might predict selective updating after informative outcomes; a performance explanation might predict slower responses and more omissions regardless of outcome. These predictions should determine which observations are collected, before a model gives them a name.

Our planning standard is to record the primary outcome, comparison, exclusions, and analysis before inspecting the relevant group contrast. For an existing dataset, the record must also say what the analyst has already seen. A plan written after exploration cannot erase that knowledge.

Exploration can identify a valuable pattern. Its value is preserved by labeling the question as newly generated and defining how it could be assessed in independent or appropriately reserved data. Reusing the same observations to discover and confirm a pattern changes the evidential role of the second analysis.

Specify the meaningful effect and report its estimate with uncertainty. A small estimated effect with a narrow interval, a large effect with a wide interval, and a near-zero estimate with a wide interval support different next decisions. A significance label compresses those distinctions.

Ask which remaining uncertainty is limiting the conclusion. Unclear source records call for provenance work. A detector that misses relevant movement calls for measurement validation. A poorly estimated contrast calls for more independent information or a revised design.

Designing the next study means stating the effect worth resolving, plausible variability, independent replication, and the control that separates the strongest alternative. More measurements within one sample cannot automatically substitute for additional treatment units.

The adaptive-behavior workstream is currently in study development. Its next useful output is a comparison that can distinguish meaningful alternatives, together with a defensible analysis plan and a decision rule for proceeding.