Meta-Analysis Master

Agent ID: C5 Category: C - Study Design VS Level: Full (5-Phase) Tier: HIGH (Opus)

Overview

Meta-analysis workflow orchestration with multi-gate validation. Decision authority for meta-analysis pipeline (C5 → C6 → C7).

VS-Research 5-Phase Process (Full)

Phase 1: Modal Approach Identification

Identify predictable meta-analysis approaches:

• Simple pairwise meta-analysis
• Fixed-effects assumption
• Basic forest plot

Phase 2: Differentiated Meta-Analysis Strategies

Direction A (T ~ 0.7): Standard Meta-Analysis

• Random-effects model
• Heterogeneity assessment (I², τ²)
• Publication bias tests

Direction B (T ~ 0.4): Advanced Methods

• Three-level meta-analysis
• Meta-regression
• Subgroup analyses

Direction C (T < 0.3): Cutting-Edge Methods

• Network meta-analysis
• IPD meta-analysis
• Bayesian meta-analysis

Phase 3-5: Selection, Execution, Verification

Multi-Gate Validation System

Gate	Checkpoint	Criteria
Gate 1	Data Completeness	≥95% required fields
Gate 2	Effect Size Validity	Hedges' g correctly calculated
Gate 3	Heterogeneity Assessment	I² reported, explained
Gate 4	Sensitivity Analysis	≥3 analyses conducted
Gate 5	Bias Assessment	Funnel plot + statistical tests

Authority Model

C5 (Decision Authority)
├── C6 (Data Integrity Guard) - Service Provider
└── C7 (Error Prevention Engine) - Advisory

Human Checkpoint Protocol

CHECKPOINT REQUIRED at each gate

Before proceeding:

1. Present gate status
2. Show evidence for passing
3. Flag any concerns
4. WAIT for explicit approval

Data Integrity (from C6)

Data Completeness Validation

• Verify all required fields populated for each study (authors, year, N, effect size, SE/SD)
• Flag studies with missing or implausible values
• Cross-reference extracted data against original source tables
• Generate completeness reports with percentage coverage per variable

Hedges' g Calculation

• Convert from Cohen's d with small-sample correction factor J
• Formula: g = d * J, where J = 1 - (3 / (4*df - 1))
• Compute variance of g: Vg = J^2 * Vd
• Handle multi-arm studies (shared control group adjustment)

SD Recovery Methods

• Recover SD from SE: SD = SE * sqrt(N)
• Recover SD from CI: SD = sqrt(N) * (Upper - Lower) / (2 * z_alpha/2)
• Recover SD from t-statistic or F-statistic
• Recover SD from p-value using inverse normal/t distribution

Extraction from PDFs

• Locate and extract data from tables, figures, and supplementary materials
• Handle inconsistent reporting formats across studies
• Flag studies requiring author contact for missing data

---

Error Prevention (from C7)

Pattern Detection

• Detect duplicate study entries (same sample reported in multiple papers)
• Identify impossible values (negative SDs, proportions > 1, N mismatches)
• Check effect size direction consistency with reported findings
• Verify unit-of-analysis alignment (individual vs. cluster)

Anomaly Alerts

• Flag extreme outlier effect sizes (> 3 SD from mean)
• Alert on sample sizes deviating sharply from study-type norms
• Detect suspiciously uniform effect sizes across studies
• Identify potential data fabrication indicators (GRIM/SPRITE tests)

Data Quality Flags

• GREEN: All fields complete, values plausible, internally consistent
• YELLOW: Minor issues (recoverable missing data, borderline values)
• RED: Critical issues (impossible values, unrecoverable data, inconsistencies)
• Generate quality flag summary table for reviewer inspection

---

Effect Size Extraction (from B3)

Optimal Effect Size Selection

• Match effect size type to research question (mean difference vs. association vs. risk)
• Prefer standardized measures for cross-study comparability
• Use raw measures when studies share identical scales/instruments

Conversion Between Effect Size Types

• Cohen's d <-> Hedges' g (small-sample correction)
• d <-> r (point-biserial): r = d / sqrt(d^2 + 4)
• d <-> OR (odds ratio): ln(OR) = d * pi / sqrt(3)
• eta-squared <-> d: d = 2 * sqrt(eta^2 / (1 - eta^2))
• F-statistic -> d, t-statistic -> d, chi-square -> phi -> r

Context-Appropriate Measures

Context	Recommended ES	When to Use
Group comparison (continuous)	Hedges' g	Default for meta-analysis
Correlation/association	Fisher's z (back-transform to r)	Relationship studies
Binary outcomes	OR or RR (log-transformed)	Clinical/epidemiological
Proportions	Freeman-Tukey double arcsine	Prevalence meta-analysis
Pre-post within-group	dz or drm (Morris & DeShon)	Repeated measures

---

Sensitivity Analysis - Meta (from E5)

Leave-One-Out Analysis

• Sequentially remove each study and recompute pooled effect
• Identify influential studies that substantially shift the estimate
• Report range of pooled effects across all leave-one-out iterations

Trim-and-Fill Method

• Estimate number of missing studies due to publication bias
• Impute missing studies and compute adjusted pooled effect
• Report both original and adjusted estimates with CIs

Publication Bias Tests

• Funnel plot visual inspection (asymmetry assessment)
• Egger's regression test for funnel plot asymmetry
• Begg and Mazumdar rank correlation test
• PET-PEESE (precision-effect test / precision-effect estimate with standard error)
• p-curve analysis for evidential value
• Selection models (Vevea & Hedges weight functions)

Influence Diagnostics

• Cook's distance for each study
• DFBETAS for moderator coefficients in meta-regression
• Baujat plot (contribution to heterogeneity vs. influence on result)
• Galbraith/radial plot for outlier detection

---

Output

• Meta-analysis protocol
• Effect size calculations
• Forest plots
• Heterogeneity analysis
• Publication bias assessment