Please join the Statistical Computing Workshop as they host Sameer Deshpande is an Assistant Professor in Statistics at the University of Wisconsin–Madison.
Description: The Bayesian Additive Regression Trees (BART) model works by estimating unknown functions using a sum of several binary regression trees (i.e., piecewise-constant step functions). BART typically delivers extremely accurate predictions and well-calibrated uncertainty intervals "off-the-shelf": users do not need to pre-specify the function's parametric form in advance and do not have to perform extensive hand-tuning. It's generally excellent predictive performance and ease-of-use have made BART an increasingly popular tool for estimating heterogeneous causal effects.
I will introduce an extension of BART, which is known as VCBART, that fits varying coefficient models in which an outcome is linearly related to some predictors but those relationship (i.e., the corresponding slopes in the regression model) change as a function of other variables known as effect modifiers. I will show how varying coefficient models arise naturally in several common causal estimation settings including those with binary or nominal treatments and in difference-in-difference designs. I'll provide an overview of on-going work to extend the basic VCBART framework to estimate simultaneously the heterogeneous effects of multiple treatments on multiple outcomes and to fit fully nonparametric multi-level models with complex grouping structures. I’ll conclude by introducing a new implementation of BART & VCBART that allows users to fit even more flexible and expressive BART-based models.
This talk is based primarily on Deshpande et al. (2024), Deshpande (2025), and Kokandakar et al. (2023), but it builds on earlier works on BART & causal inference by Hill (2011) and Hahn et al. (2024).
Bio: Sameer Deshpande is an Assistant Professor in Statistics at the University of Wisconsin–Madison. He earned his Ph.D. from the Wharton School at the University of Pennsylvania. His research interests include Bayesian hierarchical modeling, Bayesian trees regression, variable selection, and causal inference. Much of his methodological research, which focuses on understanding how the effects of multiple exposures on multiple, inter-dependent outcomes vary over time and across the population, is motivated by a long-running collaboration that seeks to understand the long-term health risks and benefits of playing sports in adolescence. His applied sports analytics work includes computing how individual NBA players contribute to their team's chances of winning; quantifying the value of pitch framing, plate discipline, and the third time through the order penalty in Major League Baseball; and estimating the upper limits of human athletic performance using data from elite decathletes. He was named a finalist in the 2019 NFL Big Data Bowl and received the 2021 Significant Contributor award from the Sports in Statistics Section of the American Statistical Association. He is a long-suffering fan of Dallas professional sports teams and can often be spotted wearing a giant Texas belt buckle.
The Statistical Computing Workshop (formerly R Workshop) is a year-long series that meets three times per quarter during the academic year. The purpose of the workshop is to learn, practice, and update cutting edge statistical programming skills as they apply to quantitative and computational social science.
Workshop meetings will feature internal or external speakers introducing a new tool, method, or research project involving statistical computing in the broadest sense. All meetings are hybrid or fully virtual.
Ariel Sowers
(847) 491-7454
Email