Data-Driven Confounder Selection via Markov and Bayesian Networks

TL;DR

This paper introduces a data-driven method for selecting confounders in causal inference using probabilistic graphical models, effectively identifying relevant covariate subsets to improve causal effect estimation.

Contribution

It proposes a novel approach combining Markov and Bayesian networks to select confounders without prior causal structure knowledge, validated through simulation.

Findings

Outperforms random forests and LASSO in confounder selection accuracy

Achieves lower mean squared error in causal effect estimation

Effective in high-dimensional data settings

Abstract

To unbiasedly estimate a causal effect on an outcome unconfoundedness is often assumed. If there is sufficient knowledge on the underlying causal structure then existing confounder selection criteria can be used to select subsets of the observed pretreatment covariates, $X$, sufficient for unconfoundedness, if such subsets exist. Here, estimation of these target subsets is considered when the underlying causal structure is unknown. The proposed method is to model the causal structure by a probabilistic graphical model, e.g., a Markov or Bayesian network, estimate this graph from observed data and select the target subsets given the estimated graph. The approach is evaluated by simulation both in a high-dimensional setting where unconfoundedness holds given $X$ and in a setting where unconfoundedness only holds given subsets of $X$. Several common target subsets are investigated and the selected subsets are compared with respect to accuracy in estimating the average causal effect. The proposed method is implemented with existing software that can easily handle high-dimensional data, in terms of large samples and large number of covariates. The results from the simulation study show that, if unconfoundedness holds given $X$, this approach is very successful in selecting the target subsets, outperforming alternative approaches based on random forests and LASSO, and that the subset estimating the target subset containing all causes of outcome yields smallest MSE in the average causal effect estimation.