psci-data-analysis

Data Analysis (psci-data-analysis)

Psychological Science holds analyses to high credibility standards: effect sizes with confidence intervals for major results, full disclosure of how the data were handled, and a clean confirmatory vs. exploratory separation. Analysis scripts and data are shared and can be checked.

When to trigger

• Running and reporting the main and supporting analyses
• A reviewer asked for effect sizes, intervals, robustness, or disclosure
• Reconciling preregistered analyses with exploratory follow-ups
• Preparing analysis scripts and a data dictionary for deposit

Reporting norms Psychological Science expects

1. Effect sizes + uncertainty. Report a standardized or unstandardized effect size and a measure of uncertainty (e.g., confidence intervals) for major results — not just p-values and stars.
2. Full disclosure (the "21-word-solution" spirit). Report how sample size was determined, all data exclusions (and reasons), all manipulations/conditions, and all measures. Total excluded observations must be stated.
3. Confirmatory vs. exploratory. Label preregistered confirmatory analyses separately from exploratory ones; do not present exploratory results as predicted (no HARKing).
4. Appropriate inference. Justify the model; report assumptions/diagnostics; correct for multiple comparisons when testing many outcomes; consider robust or Bayesian alternatives where apt.
5. Replicability of the analysis. Provide analysis scripts and a data dictionary; results should regenerate from the shared data in a fresh session (see psci-open-science-and-transparency).

Robustness

• Show the result survives reasonable alternative specifications and exclusion choices; report sensitivity rather than a single fragile model. For small samples, be candid about uncertainty.

Worked micro-example (illustrative numbers)

A preregistered two-study package on selective attention. Study 1 (N = 240, between-subjects) tests whether a brief mindfulness induction reduces attentional capture by emotional distractors. The confirmatory analysis is a single preregistered contrast on reaction-time cost.

Confirmatory (preregistered) — Study 1
  Effect: induction vs. control on capture cost (ms)
  d = 0.34, 95% CI [0.08, 0.59], t(238) = 2.66, p = .008
  Sensitivity: holds with/without the 6 preregistered RT-outlier exclusions
               (d shifts 0.34 → 0.31), and under log-RT (d = 0.33)
Exploratory (labeled) — Study 1
  Trait-anxiety × condition interaction surfaced post hoc; reported as
  exploratory, flagged for confirmation in Study 2's preregistration
Confirmatory (preregistered) — Study 2 (N = 300, direct + extension)
  Replicates direct effect (d = 0.29, 95% CI [0.06, 0.51]) and
  preregisters the anxiety moderation that was exploratory in Study 1

Why this passes Psychological Science scrutiny: every confirmatory number carries an effect size and a CI; the anxiety interaction is honestly demoted to exploratory and then promoted to confirmatory only after preregistration in Study 2; the sensitivity line pre-empts the "fragile-to-exclusions" reviewer.

Analysis-stage reviewer pushback and the venue fix

Reviewer pushback	What it signals here	Psychological Science fix
"p = .048 — too close to the line, and the CI nearly spans zero"	post-credibility-revolution distrust of just-significant single tests	report the CI prominently, add the Study 2 replication, lead with the pooled estimate
"Which exclusions were preregistered?"	suspicion of undisclosed forking paths	give the disclosure table: rule, count, preregistered vs. post hoc, and the estimate with vs. without
"Means hide the distribution"	bar-of-means aesthetic distrusted	recompute and show effect size + CI; route exhibit to psci-tables-figures
"Is this confirmatory?"	HARKing concern	point to the preregistration timestamp; relabel anything generated after data as exploratory
"Reviewer 2 could not rerun your code"	reproducibility gate	ship a fresh-session run log; see psci-open-science-and-transparency

Calibration anchors

• One adequately powered effect with a tight CI beats three stars on an underpowered model — the journal's cautionary history is flashy-but-fragile single studies.
• Prefer estimation language ("the induction reduced capture cost by ~0.3 SD, 95% CI [...]") to dichotomous "significant/not." Bare p-value sentences read as pre-reform here.
• When N is modest, state the smallest effect the design could detect rather than implying precision you do not have; hedge magnitude claims to what the interval supports.

Anti-patterns

• p-values and stars with no effect size or confidence interval
• Selectively reporting conditions, measures, or exclusions (undisclosed flexibility)
• HARKing exploratory findings into confirmatory hypotheses
• Optional-stopping / garden-of-forking-paths analyses presented as planned
• Analysis code that does not reproduce the reported numbers

Output format

【Main result】effect size + confidence interval + meaning
【Disclosure】N-determination + all exclusions + all conditions + all measures reported? [Y/N]
【Confirmatory vs exploratory】clearly separated? [Y/N]
【Inference】assumptions/diagnostics, MHT handled?
【Reproducible】scripts + data dictionary + fresh-session check? [Y/N]
【Next】psci-tables-figures

Supplementary resources

• ../../resources/external_tools.md — effectsize, emmeans, metafor, JASP/jamovi, reproducible-report tooling
• ../../resources/official-source-map.md — statistical and disclosure requirements