gamlss2: The Next Generation of Distributional Regression

Nikolaus Umlauf, University of Innsbruck

Co-authors: Mikis Stasinopoulos, University of Greenwich; Fernanda De Bastiani, University of Pernambuco; Gillian Heller, University of Sydney; Thomas Kneib, University of Goettingen; Andreas Mayr, University of Marburg; Robert Rigby, University of Greenwich; Reto Stauffer, University of Innsbruck; Achim Zeileis, University of Innsbruck

Abstract: Few methodological developments are as closely aligned to the Statistical Modelling Society and its workshop as Generalized Additive Models for Location, Scale, and Shape (GAMLSS; Rigby and Stasinopoulos, 2005), with the possible exception of P-splines. To advance the success of this phenomenally flexible modeling approach to the next generation, we have developed a new computational framework for distributional regression. Given the increasing importance of distributional regression in modern statistical analysis – allowing for the modeling of entire distributions rather than just the mean – such a framework is crucial in many fields where quantiles, probabilities, and other distributional characteristics are of interest. Building on the well-established gamlss framework, we introduce gamlss2, an advanced and user-friendly R package for distributional regression. It provides a modular infrastructure for defining custom distributions and model terms, along with tools for efficient computation, diagnostics, and visualization. We introduce the core features of gamlss2 and demonstrate its capabilities through various real-world applications (see https://gamlss-dev.github.io/gamlss2/).