GWAM: Ghost-Weighted Aggregate Meta-Analysis for registry-based publication-bias sensitivity

Abstract

Background. Publication bias threatens the validity of meta-analysis, yet routine detection tools (funnel-plot asymmetry, trim-and-fill) rest on assumptions that are hard to verify. Trial registries offer an external anchor for how much of the evidence base is actually visible.

Methods. GWAM (Ghost-Weighted Aggregate Meta-Analysis) is a registry-anchored sensitivity method. It classifies ClinicalTrials.gov records as ghost (completed, with no publication and no posted results), results-posted, or published (PMID-linked); computes an enrollment-weighted integrity ratio λ (the published-participant fraction); and scales the published pooled effect by λ. A Monte-Carlo layer and a hierarchical variance-propagation extension carry the uncertainty about unobserved trials.

Results. For escitalopram in depression (18 trials) λ was 0.096; the GWAM-adjusted log-odds-ratio was 0.025 (95% SI −0.041 to 0.087; SI denotes a simulation interval, not a confidence interval) versus a published random-effects log-odds-ratio of 0.255 (OR 1.29). Across a benchmark of 7,467 meta-analyses from 398 Cochrane reviews the median λ was 0.871, and 7.9% of meta-analyses with a random-effects estimate of magnitude at least 0.10 were attenuated below 0.10. For pregabalin (35 trials, log-risk-ratio) λ was 0.014, flagged as likely incomplete PMID linkage. In simulation (2,000 replications) bias was minimal at μ=0 but grew negative for μ>0 (−0.162 at μ=0.2) and raw coverage fell to 7.1%; the hierarchical extension restored near-nominal coverage.

Conclusions. GWAM is a transparent, reproducible registry-anchored sensitivity analysis, best interpreted as a conservative framework that complements existing publication-bias methods; its performance depends on the quality of ClinicalTrials.gov publication linkage.