Meta-Analysis Methods (psychbull-meta-analysis-methods)

This is the quantitative core of a Psychological Bulletin meta-analysis: turning coded study statistics into effect sizes, pooling them under a defensible model, and characterizing heterogeneity honestly. Psychological Bulletin expects MARS-compliant methods. This skill covers estimation; moderators and publication-bias diagnostics live in psychbull-moderators-and-bias.

When to trigger

• Computing effect sizes from coded statistics
• Choosing fixed-effect vs. random-effects (vs. multilevel) models
• Handling multiple effect sizes per study (dependency)
• Quantifying and interpreting heterogeneity

Effect sizes

• Pick a metric that matches the designs and is comparable across studies: standardized mean difference (Hedges' g, small-sample corrected), correlation r → Fisher's z, log odds ratio / risk ratio. Convert disparate metrics to a common scale and document conversions.
• Compute with a transparent tool (e.g., metafor::escalc); record the formula and the inputs used.
• Track the direction/sign so effects align with the substantive hypothesis.

Model choice

1. Random-effects (or mixed-effects) by default. Psychological literatures vary across populations, measures, and procedures, so a common true effect is implausible; estimate τ² and a summary effect with a random-effects model. Justify any fixed-effect use explicitly.
2. Dependent effect sizes (multiple per study/sample) violate independence. Use robust variance estimation (RVE) (robumeta/clubSandwich) or a multilevel/three-level model (metafor::rma.mv); do not naively treat all effects as independent.
3. Weighting by inverse variance; report the estimator for τ² (e.g., REML).

Heterogeneity

• Report Q (test), I² (proportion of variance from heterogeneity), and τ²/τ (absolute between-study SD), plus a prediction interval for the range of true effects.
• Interpret heterogeneity substantively — it motivates the moderator analysis, it is not a nuisance to hide. High heterogeneity with a tiny CI on the mean can mislead without the prediction interval.

Execution bridge (StatsPAI / Stata MCP)

Meta-analysis itself is largely outside this causal-inference toolchain — use dedicated tools (e.g. metafor) for pooled effects and meta-regression. Full map (for primary-study reanalysis): execution-with-mcp. Psychological Bulletin is a meta-analytic review venue.

• Moderator / meta-regression tests: apply the multiple-testing haircut (romano_wolf / benjamini_hochberg) — many moderators inflate false positives.
• Reanalyzing a primary dataset: the design→fit→audit chain applies (detect_design → recommend → fit → audit_result).
• Exhibits: etable / plot_from_result for any reanalysis tables/figures.

Be explicit about what is meta-analytic (dedicated tools) vs primary reanalysis (this chain). JF execution walkthrough.

Anti-patterns

• A fixed-effect model imposed on an obviously heterogeneous literature
• Treating multiple effects per study as independent (understated SEs)
• Reporting only the pooled point estimate and CI, with no I²/τ²/prediction interval
• Mixing incomparable effect-size metrics without conversion
• Letting reported numbers diverge from the analysis script (the database is deposited and checkable)

Model-choice expectations at the review flagship

Psychological Bulletin, the APA's flagship review journal, expects state-of-the-art meta-analytic modeling — the model is where methods reviewers concentrate. The decision table they apply:

Worked vignette — fitting the pooled model

Illustrative numbers only — not real data. The self-affirmation synthesis codes 51 effects from k = 42 studies (9 studies contribute 2–4 effects each). Under this skill's rules:

• Effect size: Hedges' g with small-sample correction; 6 effects originally reported as r or t are converted to g and the conversions are documented.
• Model: a random-effects estimate is implausible to treat as a single true effect across varied populations, so REML random-effects is the base; clustered effects are handled with RVE (robumeta/clubSandwich) so the nine multi-effect studies do not understate the SEs.
• Pooled result: g = 0.34, 95% CI [0.24, 0.44].
• Heterogeneity: Q(41) significant, I² = 68%, τ² = 0.041 (τ = 0.20), and a 95% prediction interval of roughly [−0.10, 0.78] — far wider than the CI, which is the honest signal that true effects vary and motivates the moderator analysis rather than a single-number headline.

Referee pushback → venue-specific fix

• "You report a CI but no prediction interval." → Add the PI; with I² = 68% the CI alone misleads about the spread of true effects across settings.
• "Multiple effects per study were treated as independent." → Refit with RVE or a three-level model and report how the SEs and τ² change.
• "A fixed-effect model is indefensible here." → Switch to random-effects, justify in text, and name the τ² estimator.

Output format

【Effect-size metric】g / z(r) / logOR + conversions noted
【Model】random-effects / multilevel / RVE (+ τ² estimator)
【Dependency】handled via RVE / multilevel? [Y/N]
【Pooled effect】estimate + 95% CI
【Heterogeneity】Q, I², τ², prediction interval
【Next】psychbull-moderators-and-bias

Supplementary resources

• ../../resources/external_tools.md — metafor, robumeta/clubSandwich, Stata meta, CMA
• ../../resources/official-source-map.md — MARS reporting of model, effect sizes, heterogeneity