Frank DiTraglia: Bayesian Double Machine Learning for Causal Inference

Seminar
  • Date and time: Wednesday 26 February 2025, 1pm to 2pm
  • Location: In-person only
    A/EW105 Alcuin East Wing
  • Audience: Open to staff, students
  • Admission: Free admission, booking not required

Event details

Author: Francis DiTraglia (Oxford)

Abstract: This paper proposes a simple, novel, and fully-Bayesian approach to causal inference in a partially linear model with treatment D and control variables X. Because the dimension of X can often be large in practice, researchers increasingly turn to off-the-shelf machine learning methods to regularize their coefficient estimates. However, this can lead to unreliable estimates in causal problems, a phenomenon known as regularization-induced confounding. To address this problem, we propose a Bayesian Double Machine Learning (BDML) method, which modifies a standard Bayesian multivariate regression model so that the causal effect of interest is absorbed by the reduced-form covariance matrix and can then be recovered through a simple transformation of the posterior. Our approach is related to the burgeoning frequentist literature on DML while addressing its limitations in finite-sample inference. Moreover, we offer the first Bayesian approach based on a fully generative probability model in the DML context, adhering to the likelihood principle. In simulations, we show that our method achieves lower RMSE, better frequentist coverage, and shorter confidence interval width than alternatives from the literature, both Bayesian and frequentist.

Host: Francesco Bravo (York)