On June 17th, 2026, Mark Steel(Statistics, University of Warwick) will deliver a short course on Model Uncertainty and Bayesian Model Averaging. The course is intended for academic staff and postgraduate students.
On the following day, June 18th, Professor Steel will also present a research seminar as part of the Research Seminars in Applied Mathematics for Economics and Management, organised by CEMAPRE.
Room: Novo Banco Lecture Hall 4th floor, Quelhas 6 building
Time: 10.00-12.00 and 14.00-16.00
Attendance is free, subject to prior registration.
Mark Steel is Professor of Statistics at the University of Warwick, where he was Head of the Department of Statistics from 2014 until 2018. Mark is interested in theoretical and applied Bayesian Statistics and Econometrics. He works on a variety of topics, for example, currently his main focus is on Bayesian model averaging, causal modelling and inference in models with latent variables. So far, he has published over 100 papers in international journals and leading machine learning proceedings, which have collected over 14,600 citations (Google Scholar).
Mark Steel recently served as the Editor-in-Chief of the journal Bayesian Analysis and has previously been part of editorial boards of other leading journals in Statistics and Econometrics, such as the Journal of the Royal Statistical Society, Series B, Journal of Business and Economic Statistics and the Journal of Econometrics. He has performed a variety of roles in the International Society for Bayesian Analysis (ISBA) and in the Royal Statistical Society.
Course Abstract: “Model Uncertainty and Bayesian Model Averaging”
In this short course, we introduce the basic ideas of model uncertainty and how to address it, through traditional model selection methods and throughBayesian Model Averaging (our main focus).
We briefly cover the basics of Bayesian inference and MCMC methods. The general framework of Bayesian Model Averaging is explained, and we discuss in some detail its implementation in a normal linear regression model with uncertainty about the inclusion of covariates. The main numerical issue in the Gaussian linear regression case is to efficiently explore a huge model space. A simple Metropolis sampler (MC^3) often works very well for problems with up to 2^100 models or so.
An application to cross-country growth data illustrates the main ideas and showcases what the analysis can provide.
In an extension of the standard regression model, we also consider the situation (quite common in social science applications) where one or more of the regressors of a Gaussian linear regression model can be endogenous (or affected by unobserved confounding) which invalidates the usual statistical properties. We investigate a solution through instrumental variables, while allowing for model uncertainty concerning which regressors and instruments to include. Applications to country growth and returns to schooling are briefly discussed.
Finally, we discuss some publicly available software and resources that can be used to implement BMA on your favourite dataset.