**Seminars will be held In Person unless otherwise noted**
Seminar One
September 8, 2026
Griffin Floyd Hall, Room 100, 4pm to 5pm
Omiros Papaspiliopoulos
Università Bocconi
Title:
Variational inference for bilinear mixed models
Abstract:
The talk provides an overview of a book I am currently writing together with Max Goplerud (Department of Government at the University of Texas at Austin). The underlying modeling framework provides a unification of foundational model structures used throughout applied statistics, including GLMMs, GAMs, factor models/item response theory, spatial random effects, probabilistic matrix factorization, etc. I show three motivating applications: deep interactions for survey data, ideal point models for political ideology, and high-resolution demography. The situations we are interested in, involve large scale inference, where both the amount of training data and the number of model parameters are large, in the thousands or larger. We develop variational inference methods that are partly generic and partly tailored to these model structures, and some theory that establishes their properties in the large scale regime, in terms of convergence, uncertainty quantification and model selection. We have an associated modular, publicly availabe software in R, which is in synergy with lme4, mgcv, brms, and inla.
A couple of relevant recent publications:
Goplerud, M., Papaspiliopoulos, O. and Zanella, G. (2025) Partially factorized variational inference for high-dimensional mixed models, Biometrika, 112:2 Pandolfi, A., Papaspiliopoulos, O. and Zanella, G. (2026) Conjugate gradient methods for high-dimensional GLMMs, J. Amer. Statist. Assoc., 121(554), 1103–1115
Seminar Two
September 17, 2026
Griffin-Floyd Hall, Room 100, 4pm to 5pm
Kuan Liu
University of Toronto
Title:
Bayesian Causal Inference with Latent Variables for Cognitive Aging
Abstract:
Modifiable behavioural and psychosocial risk factors, including depression and loneliness, and environmental exposures such as heat and air pollution may contribute to accelerated cognitive aging and dementia risk. Motivated by the Canadian Longitudinal Study on Aging, this talk examines Bayesian causal inference when scientifically relevant variables are not directly observed. The first part presents a Bayesian nonparametric latent trajectory model for multivariate longitudinal cognitive outcomes, combined with posterior predictive standardization to estimate exposure effects on cognitive trajectory membership. The second part considers Bayesian sensitivity analysis for time-varying unmeasured confounding and its connection to ongoing work in environmental health, where individual-level exposure is latent. These examples showcase how Bayesian causal approaches can provide a flexible and principled framework for incorporating latent structure and uncertainty in challenging applied health research.
Seminar Three
October 1, 2026
Griffin-Floyd Hall, Room 100, 4pm to 5pm
Andrew Thomas
University of Iowa
Title:
Nested peak inference for hotspots in noisy images with cubical persistent homology
Abstract:
In this talk, we will use topological data analysis to develop a method to test intensity peaks in images and iteratively estimate corresponding hotspots—assessing the statistical significance of each subsequent noise hypothesis beyond the union of the iteratively identified regions of interest. Our ability to coherently estimate hotspots will follow from the use of cubical persistent homology, which respects the connectivity structure of our image. The nested nature of the sequence of random hypotheses that we test ensures control of the familywise error rate at the desired level. Our method employs permutation testing for assessing significance, which can be computationally intensive, so we use anytime-valid permutation testing to minimize the number of permuted images needed. Finally, we need to account for the fact that our hypotheses are selected based on the image data we observe. We do so by detailing scenarios in which our framework yields similar conclusions to that of selective inference. Our development not only provides rigorous theoretical guarantees and estimated hotspots with meaningful topology but also yields performance as good and in many cases better than existing methods in simulation experiments. Finally, when applied to two real-world datasets in vibrothermography and neuroimaging, we will see that our methodology provides excellent performance with interpretable results.
October 15
Griffin-Floyd Hall, Room 100, 4pm to 5pm
Sumit Mukherjee
Columbia University
Title:
TBD
Abstract:
TBD
Seminar Five-Challis Lectures
October 29-30, 2026
Orange and Blue Room, O’Connell Center, 4pm to 5pm
Andrea Rotnitzky
University of Washington
Title:
TBD
Abstract:
Seminar Six
November 12, 2026
Griffin-Floyd Hall, Room 100, 4pm to 5pm
Hiya Bannerjee
Eli Lilly
Title:
TBD
Abstract:
TBD