Identifying Heterogeneity in Regression Compositional Data Integration with Many Categories
share
In compositional data, detecting which part of the whole delineates heterogeneity is important. The aim is to propose a procedure to quantify this term in the multivariate regression context without abandoning the data's natural restriction. A single probabilistic model with a hierarchical structure was built for multiple compositional data. An objective criterion based on skewness and kurtosis metrics provides support to characterize each component's performance as well as to assist in choosing one component as a reference avoiding model identifiability issues. The inference procedure was done under the Bayesian approach using the Hamiltonian Monte Carlo (HMC) method to obtain the posterior distribution of interest. The Kullback-Leibler divergence (KLD) from information theory and the Aitchison distance metrics are calculated to compute the similarity between compositions to compare scenarios in the model validation process. The proposal was motivated by a composition structure with high uncertainty in the Abrolhos Reefs of Brazil as a consequence of a dam rupture. The results support an understanding of patterns in the studied process recognizing local effects on each component as well as quantifying the precision parameter. These highlights contribute to characterizing the marine life community in areas that were affected by anthropogenic damage.
READ FULL TEXT