Study Design — The Multi-Study Package (jpsp-study-design)

This is the skill that most distinguishes JPSP from short-report journals. A JPSP paper is a coherent set of related studies built to test a theory, not a single experiment. The package must converge: each study should add something the previous one could not establish, and the set should withstand the question "could one study break the whole story?"

When to trigger

• Planning the sequence and roles of studies in the package
• Powering each study and the package as a whole
• Choosing designs (experiment, survey, longitudinal, dyadic/APIM, intensive-longitudinal, archival)
• Deciding what to preregister and what is exploratory

Designing the package

1. Give every study a job. A common arc: establish the effect → test the mechanism (mediation/process) → probe boundary conditions / moderators → demonstrate generalization (population, context, method). Avoid a pile of near-identical replications.
2. Triangulate methods. Combine, e.g., an experiment (causal) with a field/longitudinal study (external validity) so the package is robust to any single design's weaknesses.
3. Mind the section's study budget. IRGP caps the main text at 5 studies; additional studies go to supplemental materials with results summarized briefly. Prioritize the studies that carry the argument. (Section-specific; verify — 待核实.)
4. Power each study explicitly. Plan against the smallest effect size of interest, not a pilot estimate. Use simulation for multilevel/dyadic/within-subject designs; justify N per study.
5. Preregister. Register hypotheses, design, sampling/stopping rule, and analysis for confirmatory studies; mark exploratory studies as such. JPSP is TOP Level 2 and asks you to state preregistration status (see jpsp-open-science-and-transparency).
6. Design for the internal meta-analysis. Use comparable measures/effect metrics across studies so effects can be pooled later (jpsp-data-analysis); plan it now, not after the fact.
7. Build in alternative-explanation tests. At least one study should rule out the most salient alternative account (construct, confound, alternative mediator).

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. JPSP is predominantly experimental social/personality psychology; randomization inference, mediation done right (mediate, not naive controlling-away), power, and family-wise corrections are decisive.

• 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 "package" that is one study run several times with cosmetic changes
• Underpowered studies whose null results are then waved away
• Adding studies reactively in revision instead of designing the set up front
• Designing studies with incomparable measures, making internal meta-analysis impossible
• Treating preregistration as paperwork rather than a constraint on later analysis

Post-credibility-revolution power calibration

Since the open-science reforms, JPSP reviewers treat underpowering as a central limitation. The anchors below are illustrative, not mandated thresholds — confirm any quantitative expectation against the journal's current submission guidelines, since JPSP publishes no fixed N or power floor.

Plan against the smallest effect of interest, never a noisy pilot d: a pilot d = 0.6 "powering" a study at N = 30 per cell is the classic way JPSP packages collapse on replication. For interactions and partner paths, simulate (simr, DeclareDesign) rather than a closed-form G*Power main-effect calculation.

Worked vignette: powering a three-study ASC package

Illustrative numbers — invented to show design logic, not real findings.

Claim: incidental gratitude broadens construal level (an ASC social-cognition effect).

• S1 (establish). Gratitude vs. neutral recall; DV = construal. Smallest effect d = 0.35 → N = 260 (~85% power); preregistered. Result d = 0.34, 95% CI [0.10, 0.58].
• S2 (mechanism). Adds self-transcendence mediator; bootstrap indirect path needs more N than the total effect, so N = 320. Result ab = 0.12, 95% CI [0.04, 0.21].
• S3 (boundary). 2×2 gratitude × time-pressure, community sample; interaction at f = 0.10 → N ≈ 520. Interaction d = 0.28, 95% CI [0.05, 0.51].

A referee checks for a comparable construal metric across all three (so they pool into one internal meta-analysis), S3 ruling out a mood-valence confound, and a non-student sample answering "is this just undergraduates?"

Output format

【Study set】S1 (establish) · S2 (mechanism) · S3 (boundary) · S4 (generalize) …
【Designs】experiment / longitudinal / dyadic-APIM / archival per study
【Power】N per study + smallest effect size of interest + method (sim?)
【Study budget】≤ section cap? (IRGP ≤5 in main text) extras → supplement
【Preregistration】what is confirmatory vs exploratory
【Meta-analysis ready】comparable effect metrics across studies? [Y/N]
【Next】jpsp-data-analysis

Supplementary resources

• ../../resources/external_tools.md — power (G*Power, simr, Superpower), DeclareDesign, dyadic/multilevel tools
• ../../resources/official-source-map.md — section study caps and length rules