Abstract
The estimation of causal treatment effects from observational data is a fundamental problem in causal inference. To avoid bias, the effect estimator must control for all confounders. Hence practitioners often collect data for as many covariates as possible to raise the chances of including the relevant confounders. While this addresses the bias, this has the side effect of significantly increasing the number of data samples required to accurately estimate the effect due to the increased dimensionality. In this work, we consider the setting where out of a large number of covariates $X$ that satisfy strong ignorability, an unknown sparse subset $S$ is sufficient to include to achieve zero bias, i.e. $c$-equivalent to $X$. We propose a common objective function involving outcomes across treatment cohorts with nonconvex joint sparsity regularization that is guaranteed to recover $S$ with high probability under a linear outcome model for $Y$ and subgaussian covariates for each of the treatment cohort. This improves the effect estimation sample complexity so that it scales with the cardinality of the sparse subset $S$ and $\log |X|$, as opposed to the cardinality of the full set $X$. We validate our approach with experiments on treatment effect estimation.
Introduction
In causal inference, the goal is to determine the effect of a treatment ($T$) on an outcome ($Y$). Using observational data requires controlling for "confounders"—variables that influence both the treatment and the outcome. To avoid missing any critical confounders, researchers often collect a high-dimensional set of covariates ($X$). However, this "kitchen sink" approach triggers the curse of dimensionality: as the number of features ($p$) grows, the number of samples ($n$) required to estimate the effect reliably increases dramatically.
This paper addresses the challenge of identifying the minimal sufficient subset of features for unbiased estimation. While it is known that only a subset of features is necessary (a property called c-equivalence), finding this subset in high-dimensional settings is computationally difficult. The authors propose a method that focuses on identifying variables that drive the outcome, which leads to better sample efficiency and lower variance compared to methods focusing solely on treatment drivers.
Methodology: Jointly Sparse Variable Selection
The proposed method partitions covariates into three categories based on their causal relationships:
- $X_1$: Variables affecting only the treatment.
- $X_2$: Confounders affecting both treatment and outcome.
- $X_3$: Predictors affecting only the outcome.
While adjusting for $X_1 \cup X_2$ removes bias, the authors show that adjusting for $S = X_2 \cup X_3$ is optimal. Including outcome predictors ($X_3$) reduces the variance of the final estimate, whereas including treatment drivers ($X_1$) can actually increase it.
To identify $S$, the researchers use Jointly Sparse Variable Selection. By solving a single optimization problem with a non-convex regularizer (such as MCP or SCAD) across all treatment cohorts, the model identifies a consistent set of features. This "joint" approach pools information across groups, making feature recovery significantly more robust than independent selection.
Experimental Results
The approach was evaluated using synthetic data and real-world benchmarks including the IHDP and Cattaneo2 datasets.
Support Recovery and Scaling
Synthetic tests confirmed that the joint sparse algorithm correctly identifies the true subset $S$ with far fewer samples than independent lasso-style methods. As shown in the figure below, the recovery rate for $S$ reaches 1.0 efficiently, with sample complexity scaling with the logarithm of the total features ($\log p$). This allows the method to handle datasets with a high number of irrelevant variables.
Estimation Accuracy
On the IHDP dataset, the "Sparse Doubly Robust (DR)" estimator achieved an Average Treatment Effect (ATE) of 4.49, significantly closer to the true value (4.36) than the non-sparse version (4.76). In the Cattaneo2 maternal smoking dataset, the sparse method produced an estimate (-239.3g) consistent with historical empirical intervals, whereas non-sparse methods often underestimated the effect.
Discussion and Limitations
The method provides a powerful tool for high-dimensional causal inference, though some limitations exist:
- Linearity: Theoretical guarantees currently assume a linear outcome model.
- Treatment Type: The framework is designed for discrete treatments; continuous treatment regimes remain a topic for future research.
- Admissibility: The method assumes no hidden confounders exist within the original covariate set $X$.
By leveraging joint sparsity and prioritizing outcome predictors, this work offers a path toward more sample-efficient and precise causal effect estimation in complex, high-dimensional data environments.
Summary updated