Small-Study-Effect and Asymmetry Methods in Anatomical Prevalence Meta-Analysis: A Simulation-Informed Roadmap

Anatomical prevalence meta-analyses often combine studies that differ in population, modality, specimen source, variant definition, and observational unit.Funnel-plot asymmetry is frequently interpreted as publication bias, although it may reflect transformation, event yield, heterogeneity, or prevalence-studysize associations.We evaluated six two-stage procedure families and a one-stage binomial-normal mixedeffects framework (BN-GLMM) across rare/low (0.05-5%) and broader (5-50%) prevalence domains, each with a controlled heterogeneity anchor and stochastic module.The 80,000 meta-analyses yielded 240,000 transformation-specific analyses under ordinary arcsine-square-root (PAS), Freeman-Tukey double-arcsine (PFT), and continuity-corrected two-stage logit (PLO), plus 240,000 BN-GLMM fits.Among two-stage analyses, PAS and PFT were most stable.Under PLO, heterogeneity alone generated Egger, Begg, and LFK signals even at 5-50% prevalence, whereas study-size-based regressions were more stable.BN-GLMM directly modeled counts and was highly evaluable.In rare data, all-zero evidence bases were non-estimable, and moderator estimates were unstable when only one or two studies contained events.Wald-z BN inference was well calibrated without heterogeneity but mildly anti-conservative at K = 10 with heterogeneity.T-reference sensitivity improved calibration and coverage with modest power loss but did not consistently restore nominal performance.Direct BN signaling was 8.8% for selection alone versus 61.8% for a generated size-prevalence association alone in rare data, and 9.4% versus 82.1% at 5-50% prevalence.Correct-direction detection declined as heterogeneity increased, estimated I² did not map consistently to true logit-scale heterogeneity, and selection could shift observed prevalence without detectable asymmetry.Within the evaluated assumptions, the provisional roadmap supports count-based modeling and transformation-aware triangulation.Formal inference should require K ≥ 10, sufficient event information, and meaningful study-size variation.No asymmetry or moderator signal establishes publication bias.