Residuals and diagnostics in Dirichlet regression
Aug 01, 2006
DOI:
Published in: Proceedings of the Joint Statistical Meetings 2006, American Statistical Association
Publisher: American Statistical Association
Compositional data are rarely analyzed with the usual multivariate statistical methods. One approach to model such data is Dirichlet regression. We present various diagnostic methods for Dirichlet regression models. We discuss the use of quantile residuals to check the distributional assumptions. Measures of total variability and goodness of fit are proposed to assess the adequacy of the suggested models. An R-square measure based on Aitchison’s distance is introduced. The likelihood distance is employed to identify the influential compositions. Finally, an example with real data is presented and discussed.