Finite sample-optimal adjustment sets in linear Gaussian causal models

TL;DR

This paper introduces the concept of finite sample-optimal adjustment sets in linear Gaussian causal models, focusing on minimizing mean squared error in causal effect estimation for limited data scenarios.

Contribution

It proposes a method to identify MSE-optimal adjustment sets considering sample size and joint distribution, differing from traditional asymptotic approaches.

Findings

MSE-optimal adjustment sets can outperform asymptotic sets in finite samples

Graphical criteria help efficiently identify optimal adjustment sets

Simulation results demonstrate practical benefits in limited data contexts

Abstract

Traditional covariate selection methods for causal inference focus on achieving unbiasedness and asymptotic efficiency. In many practical scenarios, researchers must estimate causal effects from observational data with limited sample sizes or in cases where covariates are difficult or costly to measure. Their needs might be better met by selecting adjustment sets that are finite sample-optimal in terms of mean squared error. In this paper, we aim to find the adjustment set that minimizes the mean squared error of the causal effect estimator, taking into account the joint distribution of the variables and the sample size. We call this finite sample-optimal set the MSE-optimal adjustment set and present examples in which the MSE-optimal adjustment set differs from the asymptotically optimal adjustment set. To identify the MSE-optimal adjustment set, we then introduce a sample size criterion for comparing adjustment sets in linear Gaussian models. We also develop graphical criteria to reduce the search space for this adjustment set based on the causal graph. In experiments with simulated data, we show that the MSE-optimal adjustment set can outperform the asymptotically optimal adjustment set in finite sample size settings, making causal inference more practical in such scenarios.