Assumption-Lean (Causal) Modeling - Prof. Stijn Vansteelandt

Assumption-Lean (Causal) Modeling
Description:
Traditional inference in (semi-)parametric models, such as generalized linear models, assumes that models are correctly specified and pre-determined. However, this approach is increasingly inadequate because models are often adaptively selected based on the data, introducing unacknowledged uncertainty. Furthermore, since models rarely represent a true underlying mechanism, standard inference is prone to bias from model misspecification; this is especially a concern in causal modeling, where even small degrees of misspecification in the range of the observed data can give rise to large biases. Recent advances in debiased machine learning and targeted learning have addressed these issues by reducing reliance on correct model specification. However, their model-free nature can limit their applicability and the insight they can deliver in complex settings.

Assumption-lean modeling rethinks the trade-off between model correctness, parsimony, and interpretability. It begins with data-adaptive outcome predictions, which are then projected onto specific model parameters. This projection is designed to ensure that the parameters remain interpretable or meaningful, even under model misspecification. By incorporating debiased machine learning techniques, assumption-lean modeling minimizes bias, maximizes interpretability, and provides valid confidence intervals that account for both model uncertainty and model misspecification.

In this talk, I will introduce the core principles of assumption-lean modeling, focusing on its application to generalized linear models for accessibility. The presentation will draw on the work of Vansteelandt and Dukes (2022) that was presented in a discussion paper for the Journal of the Royal Statistical Society: Series B. I will also discuss extensions to causal inference and time-to-event data, showcasing recent advancements aimed at balancing efficiency with interpretability.

Biography:
Stijn Vansteelandt graduated as Master in Mathematics at Ghent University in 1998, and obtained a PhD in Mathematics (Statistics) in 2002 at the same university. After postdoctoral research at the Department of Biostatistics of the Harvard School of Public Health, he returned to Ghent University in 2004, where he is now Full Professor in the Department of Applied Mathematics, Computer Science and Statistics. From 2017-2021, he had joint appointment in the Department of Medical Statistics at the London School of Hygiene and Tropical Medicine as Professor of Statistical Methodology.

Stijn Vansteelandt is an expert in causal inference. He has authored over 200 peer-reviewed publications in international journals on a variety of topics in biostatistics, epidemiology and medicine, such as the analysis of longitudinal and clustered data, missing data, mediation and moderation/interaction, instrumental variables, time-to-event analysis, family-based genetic association studies, analysis of outcome-dependent samples and phylogenetic inference. He previously served as (Co-)Editor of Biometrics, the leading flagship journal of the International Biometrics Society, and as Associate Editor for the journals Biometrics, Biostatistics, Epidemiology, Epidemiologic Methods, the Journal of Causal Inference, and the Journal of the Royal Statistical Society- Series B. He is holder of an Advanced ERC Grant (2024-2029).