useR! 2024

Compositional data analysis is growing rapidly in modern application areas, e.g. microbiome analysis, time-use studies and archaeometry to name a few. The dominant statistical methods use work of J Aitchison from the 1980s, where log-ratio transformation of data (compositional measurements) was developed as the key to use of standard multivariate methods with composition data. I outline some difficulties with log-ratios, and show that the assumptions underpinning log-ratios actually lead to a simple variance-covariance function that is readily used to build appropriate GLMs with logit link. The theory is in [1], but emphasis here will be on practice, and working in R. A key focus will be examples from in-development R package _compos_, which implements the new approach in the style of standard `lm` or `glm` methods. The new `colm` model class in _compos_ not only mimics R's `lm` and `glm` classes, it also neatly uses `lm` internally in a robust fitting algorithm (similar to the "Poisson trick", used for fitting multinomials with `glm`). The package is under open-source development [2] for CRAN submission by June 2024. [1] arxiv.org/abs/2312.10548 [2] github.com/DavidFirth/compos