cogpsych-study-design

Experiment Design (cogpsych-study-design)

Cognitive Psychology expects tightly controlled cognitive experiments whose design is engineered to discriminate models, organized as a multi-experiment program in which each experiment adds inference. The craft is in stimulus construction, counterbalancing, confound control, and powering the critical contrast — not just the main effect. Co-design the experiments with the model (cogpsych-theory-and-hypotheses).

When to trigger

• Designing an experiment or a multi-experiment series
• Constructing stimuli, item sets, and counterbalancing schemes
• A reviewer questioned confounds, power, design logic, or whether the design discriminates the models
• Justifying sample size for the critical contrast (often an interaction)

Design standards

1. Design for discrimination. Build the design so the data produce the signature that separates the models (e.g., a manipulation that the rival accounts predict to diverge). A design that both models predict equally well wastes the experiment.
2. Control researcher and stimulus degrees of freedom. Counterbalance condition/item assignment; control low-level confounds (frequency, length, familiarity, response mapping); randomize order; use attention/manipulation checks. Document the full stimulus pool, not a curated subset.
3. Power the critical contrast. Justify N (and trials per cell) for the discriminating effect — often an interaction or a model parameter — not the easy main effect. State the assumed effect size and its source. Trials-per-participant is part of power for within-subjects designs.
4. Multi-experiment logic. Say what each experiment adds: rules out a confound, extends scope, replicates the critical pattern, or tests a further model prediction. Avoid near-duplicate runs.
5. Validity. Argue construct validity (does the task measure the process the model is about) and the generality of the claim across the stimulus space and population.

Powering the critical contrast — worked example (illustrative)

For the recognition-memory program, power the z-ROC shape contrast, not just overall accuracy.

Critical contrast: the diagnostic difference in z-ROC curvature between
            UVSD and DPSD predictions.
Within-subjects: trials per participant drive ROC precision — target enough
            old/new trials per confidence bin to estimate the slope reliably
            (state the per-bin minimum, not just N).
Sample size: justified by simulation under each model (generate data from
            UVSD and DPSD at plausible parameters; find N + trials at which
            the model-recovery rate exceeds the target).
Across experiments: Exp 1 establishes the pattern; Exp 2 rules out a list-
            composition confound; Exp 3 tests a further divergent prediction.
Stopping rule: fixed N + fixed trials; no optional stopping.

Justify sample size by model/parameter recovery simulation where the contrast is a model parameter, not only by a textbook power formula for a mean difference — this is the venue-appropriate move.

Pre-data lockdown checklist

Design-stage reviewer pushback and the venue fix

• "Both models predict this design equally" → redesign so a manipulation makes the model predictions diverge; the signature must be diagnostic.
• "Possible stimulus confound (frequency/length)" → control or counterbalance it; report the matched pools; this objection lands hard here.
• "Underpowered for the interaction / too few trials" → power the critical contrast via simulation; report trials per cell, not only N.
• "Three near-identical experiments" → make each add inference (confound control, scope, further prediction).

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. Cognitive Psychology is experimental — within-subject designs and mixed models dominate; report the model, the effect size, and multiple-comparison control.

• detect_design → recommend → fit with as_handle=true → audit_result.
• Observational causal claims: staggered DiD (callaway_santanna / sun_abraham + bacon_decomposition + honest_did_from_result); IV (effective_f_test + anderson_rubin_ci); RDD (rdrobust + mccrary_test).
• Experiments: randomization-based inference, romano_wolf for many-outcome family-wise control, and mediate for mediation (not naive controlling-away).
• Sensitivity: oster_delta / sensemakr for observational claims.

Report the effect size in interpretable units; route the full battery to the appendix/supplement. A run end-to-end (synthetic data, real returns) is in the JF execution walkthrough.

Anti-patterns

• A design that both rival models predict equally well (non-diagnostic)
• Uncontrolled low-level stimulus confounds or a curated stimulus subset
• Powering the easy main effect while the critical interaction/parameter is underpowered
• Too few trials per condition to estimate the model quantities reliably
• A multi-experiment paper of near-duplicate runs with no added inference

Output format

【Discrimination】does the design produce the model-separating signature? [Y/N]
【Confound control】counterbalancing + low-level controls + checks? [Y/N]
【Power】N + trials/cell justified for the critical contrast (simulation)? [Y/N]
【Degrees of freedom】stimuli, models, exclusions, stopping fixed in advance? [Y/N]
【Multi-experiment logic】what each experiment adds
【Next】cogpsych-data-analysis

Supplementary resources

• ../../resources/external_tools.md — stimulus tools, power/recovery simulation, preregistration templates
• ../../resources/official-source-map.md — design and reporting expectations