Shrinkage priors for Bayesian Substitute Confounders

A new statistical framework proposes using Bayesian shrinkage priors to learn sparse substitute confounders from multi-cause observational data, addressing a key limitation in causal inference when unmeasured variables are present, according to research submitted on 16 June 2026 [1]. The work builds on the deconfounder method introduced by Wang and Blei in 2019, which exploits dependence among multiple causes to infer a lower-dimensional substitute for an unobserved confounder [1]. That approach preserves shared assignment variation needed for stable causal adjustment, but flexible assignment models can over-encode the treatment vector, collapse overlap, or capture single-cause variation rather than genuine confounding structure [1]. The new Bayesian factor assignment framework uses shrinkage priors to favour low-dimensional factors supported by multiple causes while discouraging effectively single-cause factors [1]. The priors induce an ordering of latent factors through progressive shrinkage, providing a natural tool for latent structural learning [1]. The theoretical guarantees are stated at the level of posterior concentration, factor score contraction, and overlap-preserving assignment geometry, meaning the results do not depend on a particular choice of shrinkage prior [1]. Under the relevant latent variable identification assumptions, the regression-adjusted estimators are consistent for mean potential outcomes [1]. This formalization connects to broader statistical principles: estimation of covariance matrices from observational data plays a foundational role in principal component analysis and factor analysis, where sample estimates are used to study inter-relationships among variables and for model checking [5]. Synthetic experiments in the paper illustrate the roles of signal strength, outcome validity, and geometry-aware regularization [1]. The framework also includes an adaptive shrinkage study linking the learned geometry to regularized causal estimation [2]. In an applied case study using baseline data from the Alzheimer's Disease Neuroimaging Initiative, the sparse substitute scores recovered much of the adjustment obtained by directly conditioning on invasive cerebrospinal-fluid biomarkers [1]. Collapse diagnostics were used to identify when fitted factors reduced to individual observed measurements [1]. The approach relates to a broader class of shrinkage-based methods for confounding adjustment. Prior work has introduced spike-and-slab priors on regression coefficients to prioritize variables associated with the treatment, using lasso to fit the exposure model and setting the prior inclusion probability higher for covariates with non-zero coefficients in that model [4]. Such methods aim to shrink coefficients for instrumental variables or noise variables toward zero while protecting important confounders from excessive shrinkage [4]. The new framework extends this logic to the latent factor space, where the goal is learning a substitute score rather than selecting among observed covariates [1].