Tree-aggregated compositional regression under measurement error
Compositional data arise in many applications, including microbiome studies, where taxonomic or phylogenetic trees describe relationships among taxa. These trees allow taxa with a common regression coefficient to be modeled as a group. An association may be specific to an individual genus or shared across a broader taxonomic group. Measurement error can make an association shared across a group appear specific to smaller groups, even when coefficient signs are preserved. We develop Tree-Aggregated Regression with Correction for Observation Error, or TARCO, to estimate latent coefficients and identify groups along the tree. We establish finite-sample conditions for exact recovery of the true subtree groups and derive bounds on prediction and coefficient estimation errors. The guarantees extend to settings where the measurement error covariance is estimated from auxiliary replicates. Simulations show improvements in prediction and coefficient estimation over tree aggregation without measurement error correction, and illustrate how the choice of tree penalty affects group recovery. In a gut microbiome analysis, TARCO identifies shared associations with body mass index and shows how correction changes their relative strengths across taxonomic groups.