Kernel Deconvolution Regression Estimator for Compositional Data

Compositional data arise in many scientific fields and require specialized statistical methods due to the constant-sum constraint. We propose a nonparametric regression approach for compositional responses with either real-valued or compositional predictors subject to measure-ment error. Specifically, a unified treatment is achieved through a kernel-based approach using a suitable class of kernels on the simplex.A real data application is presented, analyzing the relationship between moss layer composition and O-horizon layer compositionin soil.