Liesel: A Python Framework for Graph-Based Bayesian Modeling and Customizable MCMC with Support for Generalized Additive Models

Abstract

Liesel is a Python framework for Bayesian model building and posterior computation with dedicated support for generalized additive regression models that is designed to reduce friction in methodological work. The framework consists of three components. The first component, Liesel-Model, represents models as directed acyclic graphs and supports interactive model construction, modification, visualization, prediction, and prior or posterior predictive simulation. The second component, Liesel-Goose, provides a modular MCMC framework based on reusable kernels and supports blocked componentwise sampling as well as user-defined Gibbs and Metropolis--Hastings updates, while leveraging JAX for automatic differentiation, just-in-time compilation, and hardware acceleration. The third component, Liesel-GAM, supplies high-level building blocks for generalized additive regression models and provides functionality for formula-based model specification, summaries, diagnostics, and effect visualization. Together, these parts make Liesel an effective tool for the rapid development, testing, and application of Bayesian models and MCMC algorithms. Liesel's modular architecture allows users to extend the software to models and inference algorithms beyond generalized additive models and MCMC, offering considerable flexibility for a wide spectrum of statistical research.