Accounting for Unobserved Confounding to Reduce False Discoveries in Microbiome Research

Oregon Convention Center 

Microbiome research often conducts differential abundance analysis (DA) to identify microbial features associated with covariates of interest. Recently, concerns with false discoveries from DA have increased, and related statistical research usually attributes this to compositionality (microbial abundances are relative). In this work, we examine another potential cause: unobserved, microbiome-wide confounding (e.g., population structures, unmeasured technical effects). Such effects, often ignored during DA, have been noted to inflate false discovery rates (FDR) in molecular epidemiology, where research shows low-dimensional factor structures of the data can act as surrogates for confounding and be adjusted for to control FDR. We demonstrate systemic, real-data-based evidence that unobserved confounding consistently inflates FDR in microbiome DA. However, existing factor-based correction methods with simple modifications can effectively address this. We implement such methods with open-source software, to be conveniently integrated with existing DA. Our work is one of the first efforts to evaluate and correct for unobserved confounding to control FDR in microbiome DA.

False discovery rate

Unobserved confounding

Microbiome

Differential abundance

Latent factor models 

Section on Statistics in Genomics and Genetics