Error control for microbiome mediation analysis: benchmarking and remedy

Abstract

Identifying microbial taxa that mediate the effects of exposures on health outcomes can elucidate causal pathways and suggest therapeutic targets. Microbiome sequencing typically measures relative abundances (RA) of taxa, while absolute abundances (AA)—often the relevant scale for mechanistic interpretation—are not directly observed. Although compositionality-induced false positives have been documented in other microbiome analyses (e.g., differential abundance testing), systematic benchmarking of microbiome mediation methods remains scarce. We benchmarked nine mediation methods using extensive simulations and a real data application. We anchor our simulations to an experimentally derived AA template dataset, generating AA profiles that better preserve empirical structure than commonly used simulators. Across simulation settings, many methods exhibit inflated error rates, with inflation worsening as compositional effects intensify. In the real data application, methods exhibiting severe inflation in simulations also returned substantially larger mediator sets. As a remedy, we introduce CAMRA, which infers and tests AA-level mediation effects from standard RA data. CAMRA improves FDR calibration and power for taxon-level mediator discovery while maintaining favorable runtime. Reliable microbial mediator discovery requires well-calibrated inference. Our benchmark highlights systematic challenges for RA-level mediation testing. CAMRA offers a remedy by targeting AA-level mediation effects.