The Institute for Statistics and Mathematics is pleased to invite you to the next research seminar, taking place on campus:

Gregor Zens (International Institute for Applied Systems Analysis (IIASA))
Advances in Bayesian Model Averaging for Latent Variable Regression
Wednesday, March 25, 2026, 17:15, in Building D4, Room D4.0.127.

Abstract:
Latent variable models are widely used for Bayesian regression analysis with non-Gaussian outcomes. Bayesian variable selection and model averaging techniques offer principled frameworks for addressing model uncertainty in these settings. However, practical implementation is often hindered by intractable marginal likelihoods, exponentially growing model spaces, and computational bottlenecks within individual models. This talk discusses recent advances that help overcome these challenges. First, we establish a model averaging framework for a broad class of latent Gaussian models. We develop theoretical foundations - covering posterior existence and model selection consistency - and present a robust Markov chain Monte Carlo algorithm for exact probabilistic inference. Second, to make these methods scalable to large datasets, we introduce a novel approximation to marginal likelihoods based on variational inference, outlining its theoretical properties and practical implementation. Finally, we examine a computational shortcut that approximately preserves model selection behavior while substantially reducing runtimes in large-scale settings, often with negligible loss of accuracy. Throughout, we illustrate the practical benefits of the methods through simulations and a range of real-world applications.

This talk is based on joint work with Mark F. J. Steel (University of Warwick).
 

For further information and the seminar schedule, please see:
www.wu.ac.at/en/statmath/research/resseminar