Psychbull Moderators And Bias

---
name: psychbull-moderators-and-bias
description: Use when explaining heterogeneity and probing robustness in a Psychological Bulletin meta-analysis — moderator/subgroup analysis, meta-regression, and publication-bias diagnostics (funnel, Egger, trim-and-fill, PET-PEESE, p-curve, selection models) plus sensitivity analyses. Extends the core model; estimation lives in psychbull-meta-analysis-methods.
---

# Moderators & Publication Bias (psychbull-moderators-and-bias)

Once a pooled effect and its heterogeneity exist, two questions decide the paper's credibility: **what
explains the variation** (moderators), and **is the effect an artifact of selective reporting**
(publication bias). Psychological Bulletin reviewers scrutinize both, and **MARS** requires reporting
bias assessment. This skill extends the core model in `psychbull-meta-analysis-methods`.

## When to trigger

- Testing pre-specified moderators / meta-regression to explain heterogeneity
- Running publication-bias diagnostics
- A reviewer asks for sensitivity / robustness analyses
- Reconciling conflicting signals across bias tests

## Moderators & meta-regression

- **Pre-specify** moderators in the protocol; treat unplanned ones as **exploratory** and label them.
- Use **mixed-effects meta-regression** (categorical subgroups and continuous moderators); report the
  moderator coefficient, its CI, **residual heterogeneity**, and **R² analog** (variance explained).
- Beware **ecological/aggregation** bias (study-level moderators ≠ individual-level), **multiple
  testing** across many moderators, and **confounded** moderators; interpret cautiously.

## Publication-bias diagnostics (run several, not one)

1. **Funnel plot** (with contour enhancement) — visual asymmetry; not proof on its own.
2. **Egger's regression** / rank tests — small-study effects, with the usual caveats under high
   heterogeneity.
3. **Trim-and-fill** — imputes "missing" studies; treat as sensitivity, not truth.
4. **PET-PEESE** — regression-based bias-adjusted estimate.
5. **p-curve / p-uniform** — evidential value and right-skew vs. p-hacking signatures.
6. **Three-parameter selection models** (`weightr`) — model the selection process directly.

No single test is decisive; **converging evidence** across methods is the standard, and all are weak
under strong heterogeneity — say so.

## Sensitivity & robustness

- **Leave-one-out** and influence/outlier diagnostics; refit without high-leverage studies.
- Sensitivity to **effect-size metric**, **model** (RVE vs. multilevel), and **inclusion borderline**.
- Subset by **study quality / risk of bias**; published vs. grey literature.

## Anti-patterns

- Mining dozens of moderators and theorizing the one that hits (HARKing); no multiple-testing caution
- A single bias test reported as if it settled the question
- Trim-and-fill or PET-PEESE reported as the "true" effect rather than a sensitivity bound
- Ignoring that bias diagnostics behave poorly under high heterogeneity
- Subgroup claims from tiny k (few studies per cell)

## Output format

```
【Moderators】pre-specified vs exploratory; meta-regression coef + CI + R²
【Residual heterogeneity】after moderators
【Bias diagnostics】funnel / Egger / trim-fill / PET-PEESE / p-curve / selection — converge? 
【Sensitivity】leave-one-out, metric, model, quality subsets
【Bottom line】is the effect robust? [statement]
【Next】psychbull-theory-integration
```

## Supplementary resources

- [`../../resources/external_tools.md`](../../resources/external_tools.md) — `metafor`, `dmetar` (PET-PEESE), `weightr`, `puniform`, p-curve
- [`../../resources/official-source-map.md`](../../resources/official-source-map.md) — MARS bias-assessment reporting