Study Design (psci-study-design)

Psychological Science expects studies that are adequately powered, transparently planned, and robust to researcher degrees of freedom. Authors must justify sample size (a formal power analysis where appropriate). This skill hardens the design before data collection.

When to trigger

• Planning a study or a multi-study package
• Writing a preregistration / pre-analysis plan or a Registered Report Stage 1
• A reviewer questioned power, design, confounds, or analytic flexibility
• Justifying sample size and stopping rules

Design standards

1. Sample-size justification. Provide an explicit basis for N — a power analysis for the smallest effect of interest, a precision/AIPE rationale, or (for sequential/Bayesian designs) the decision rule. State the assumed effect size and where it came from.
2. Preregister the confirmatory core. Specify hypotheses, design, conditions, measures, exclusion rules, and the analysis plan in advance (OSF/AsPredicted, or a Registered Report Stage 1). This is what converts a claim from exploratory to confirmatory.
3. Control researcher degrees of freedom. Decide in advance: conditions, the full set of measures, exclusion criteria, covariates, and how stopping is determined. Undisclosed flexibility inflates false positives.
4. Confounds and validity. Address random assignment, manipulation/attention checks, order effects, demand characteristics; argue construct and external validity for the population claimed.
5. Multi-study logic. If using several studies, say what each adds (generalization, mechanism, boundary condition) — not just repetition.

Registered Reports (strongest design path)

• Stage 1 reviews the theory + design + analysis plan before data; in-principle acceptance commits the journal regardless of outcome if you execute the plan. Ideal for confirmatory and replication work, and it neutralizes publication bias. For prior-collected data, use RR with Existing Data and declare provenance.

Sample-size justification — worked example (illustrative)

For the two-study attention package, justify N before collecting, tied to the smallest effect of interest (SESOI), not a round number per cell.

Smallest effect of interest: d = 0.30 (below this, the premise is not
            practically load-bearing for downstream clinical models).
Study 1 (between-subjects, two groups):
            target 80% power, two-sided alpha .05 → N ≈ 278; we collect 240
            and report honestly that we have ~80% power for d = 0.36, i.e.
            the design is calibrated to a slightly larger effect — stated, not hidden.
Study 2 (direct replication + moderation):
            increase to N = 300 for the interaction term; precision goal is a
            half-width ≤ 0.25 on the replication d.
Stopping rule: fixed-N; no optional stopping. (For sequential designs, state
            the decision boundary and alpha-spending in advance.)

State the assumed effect size and its source (prior meta-analytic estimate, a pilot, or a SESOI argument). A power analysis anchored to an inflated published effect is a known failure mode here.

Pre-data lockdown checklist

Design-stage reviewer pushback and the venue fix

• "50 per cell, no justification" → replace with a SESOI-anchored power or precision argument.
• "Manipulation may not have worked" → preregister and report a manipulation/attention check; if it fails, the confound objection lands hard at this venue.
• "Looks like flexible exclusions" → preregister exclusion rules and report the estimate with and without them (handoff to psci-data-analysis).
• "Three near-identical studies" → make each study add inference (generalization, mechanism, boundary).

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. Psychological Science is short-format experimental psychology with strong open-science norms; preregister, run randomization inference, and report effect sizes with family-wise corrections.

• 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

• "We collected 50 per cell" with no power/precision justification
• Optional stopping or undisclosed exclusion rules
• Flexible measure/condition selection revealed only after results
• Underpowered single studies chasing a surprising effect
• A multi-study paper where studies are near-duplicates with no added inference

Output format

【Sample size】N + justification (power for smallest effect of interest / precision / decision rule)
【Preregistration】confirmatory core preregistered? where?
【Degrees of freedom】conditions, measures, exclusions, covariates fixed in advance? [Y/N]
【Validity】confounds / checks / population addressed
【Design path】Research Article vs Registered Report (S1)
【Next】psci-data-analysis

Supplementary resources

• ../../resources/external_tools.md — G*Power, simr, Superpower, preregistration templates
• ../../resources/official-source-map.md — sample-size-justification and preregistration policy