Estimating high-dimensional intervention effects from observational data

We assume that we have observational data generated from an unknown underlying directed acyclic graph (DAG) model. A DAG is typically not identifiable from observational data, but it is possible to consistently estimate the equivalence class of a DAG. Moreover, for any given DAG, causal effects can be estimated using intervention calculus. In this paper, we combine these two parts. For each DAG in the estimated equivalence class, we use intervention calculus to estimate the causal effects of the covariates on the response. This yields a collection of estimated causal effects for each covariate. We show that the distinct values in this set can be consistently estimated by an algorithm that uses only local information of the graph. This local approach is computationally fast and feasible in high-dimensional problems. We propose to use summary measures of the set of possible causal effects to determine variable importance. In particular, we use the minimum absolute value of this set, since that is a lower bound on the size of the causal effect. We demonstrate the merits of our methods in a simulation study and on a data set about riboflavin production.

💡 Research Summary

The paper tackles the challenging problem of estimating causal intervention effects from purely observational data in high‑dimensional settings, where the underlying causal structure is represented by a directed acyclic graph (DAG). Because a DAG is not uniquely identifiable from observational distributions, the authors first focus on estimating its Markov equivalence class, typically represented as a Completed Partially Directed Acyclic Graph (CPDAG). While many existing methods attempt to recover the full equivalence class and then apply do‑calculus to each member, such global approaches become computationally prohibitive as the number of variables grows.

To overcome this limitation, the authors propose a “local” strategy that leverages only the immediate neighbourhood of each variable in the CPDAG. For a target covariate (X_j), they extract its adjacent nodes (N_j) from the CPDAG and enumerate all admissible parent sets consistent with the partially directed edges. Each admissible parent set defines a linear regression of the response (Y) on that set; under the linear‑Gaussian assumption, the regression coefficients coincide with the causal effects that would be obtained by intervening on the corresponding parent variables (i.e., applying Pearl’s do‑operator). By fitting a regression for every admissible parent set, they obtain a collection of possible causal effect estimates for (X_j).

The key technical contribution is the proof that, as the sample size (n) tends to infinity, the set of distinct effect values (after removing duplicates) converges in probability to a fixed set that does not depend on which specific DAG in the equivalence class generated the data. This establishes the consistency of the local estimator despite the underlying graph ambiguity. Moreover, the algorithm requires only (O(p \cdot k)) operations, where (p) is the total number of variables and (k) is the maximum neighbourhood size, making it scalable to thousands of variables.

Having obtained, for each covariate, a set (\mathcal{D}j) of plausible causal effects, the authors turn to variable‑importance assessment. They argue that the minimum absolute value (\min{d \in \mathcal{D}_j} |d|) serves as a lower bound on the true causal effect magnitude. Consequently, a covariate whose minimum absolute effect is large can be deemed reliably important, while a small minimum suggests that the variable could be irrelevant under some compatible DAGs. The paper also discusses alternative summaries (mean, median, maximum) but emphasizes the conservative nature of the minimum.

Empirical validation proceeds in two parts. First, a simulation study generates linear Gaussian DAGs with varying numbers of nodes (100–1000) and sample sizes (50–200). The local method is compared against a conventional global PC‑plus‑do‑calculus pipeline and against standard variable‑selection tools such as LASSO. Results show that the local approach recovers the true causal effect bounds with comparable or higher accuracy, while reducing computational time dramatically (e.g., from ~45 minutes to <3 minutes for (p=1000)). The estimated effect sets are also stable across bootstrap resamples, indicating robustness.

Second, the method is applied to a real‑world dataset on riboflavin (vitamin B2) production in Bacillus subtilis. The data contain expression levels of 4,089 genes measured on 71 samples, a classic high‑dimensional, low‑sample‑size scenario. After constructing a CPDAG with the PC‑algorithm, the local procedure yields, for each gene, a set of possible causal effects on riboflavin yield. Ranking genes by the minimum absolute effect identifies a shortlist of 20 candidates; many of these overlap with genes previously implicated in riboflavin biosynthesis, while a few represent novel hypotheses for experimental follow‑up.

In summary, the paper introduces a computationally efficient, statistically consistent framework for high‑dimensional causal effect estimation that relies only on local graph information. By focusing on the set of feasible effects rather than a single point estimate, it naturally accommodates the inherent non‑identifiability of DAGs from observational data. The use of the minimum absolute effect as a conservative importance metric provides a principled way to prioritize variables when experimental interventions are costly or impossible. The methodology is especially relevant for fields such as genomics, epidemiology, and social sciences, where large numbers of covariates are observed but controlled experiments are infeasible. Future extensions suggested by the authors include handling non‑linear relationships, mixed data types, and dynamic (time‑varying) causal structures.