A method for discriminating between models describing compositional data

Abstract

Until-recently, only the Dirichlet class of distributions was available to describe compositional data, but now there is an alternative more flexible class of logistic normal distributions. To choose between these two classes requires methods for testing separate families of distributions, but application of the methods described in the literature involves extremely tedious calculation. In this paper, a conceptually straightforward and computationally simpler method is introduced and illustrated.