A1278
Title: Regression by composition
Authors: Simon Schoenbuchner - Cardiff University (United Kingdom) [presenting]
Rhian Daniel - Cardiff University (United Kingdom)
Daniel Farewell - Cardiff University (United Kingdom)
Abstract: Regression by composition is a modular framework for parametric statistical modelling. Parametric models are decomposed into covariate-dependent linear predictors of transformations (flows) that act on a reference distribution to model the conditional distribution of an outcome. These flows function as theoretical and computational building blocks of standard regression models, including generalised linear models, but they can also be composed in new ways to yield bespoke models. The likelihood is constructed by recursively combining (folding) a list of flows, with the reference distribution as an initial value. The framework provides a generic tool for specifying a wide variety of parametric models. The distributional transformations that constitute a regression by composition can be visualised using a sequence of parallel-coordinate plots, allowing investigation of counterfactual contrasts and understanding of how different covariate values give rise to different outcome distributions.
