acceptodds
Under review as a conference paper at ICLR 2027

End-to-End Mediation Analysis at Scale: From High-Dimensional Causal Discovery to Path-Level Effects, with a Human Phenotype Application

Abstract

Causal mediation analysis reveals how an exposure affects an outcome, but requires the mediator and the variables to adjust for to be specified in advance; with thousands of variables, both must be inferred from a causal structure learned from the data. Learned structures are uncertain and may disagree on what to adjust for, and existing pipelines face three obstacles at this scale: structure learning grows quadratically with the number of variables, averaging estimates across structures targets no single adjustment, and estimators must commit to an unknown functional form. We propose Channel-Additive Robust Mediation (Charm), an end-to-end framework that estimates mediation effects from incomplete, high-dimensional data. Charm searches for structure only among each variable's strongest partners and registered role pairs, so its cost grows linearly; fixes, before estimation, one adjustment set valid under every structure supporting a path, and withholds the path when none exists; and estimates every path with one flexible model that learns how the mediator acts on the outcome, rather than fixing its form in advance. In simulations, screened discovery matches the accuracy of the full search at a fraction of its cost on data with of values missing; under a complete public tier order, Charm recovers the target path and estimates its effects accurately up to ten thousand variables, never issues an invalid adjustment set where the structures disagree, and stays accurate across linear and nonlinear mechanisms. On the Human Phenotype Project, it recovers mediation chains that survive multiple-testing correction and reproduce reported biology, such as Prevotella acting on blood glucose through branched-chain amino acids.

Primary area:
causal reasoning
Keywords:
high-dimensional causal discovery, causal mediation analysis, missing data, real-data application, adjustment-set selection.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_DJsNOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes an end-to-end framework for high-dimensional causal mediation analysis that automatically discovers candidate mediation paths, selects adjustment sets, and estimates natural direct and indirect effects from observational data. The proposed method has three stages. First, it performs causal discovery on strongly corrected pairs of variables. Then, it identifies the common adjustment set of all sampled possible causal graphs. The final step is to estimate the mediation effects using a channel-additive model. Experiments demonstrate scalability to thousands of variables and promising estimation accuracy, including an application to the Human Phenotype Project.

Critical strengths
  1. Scalable automated mediation analysis with high-dimensional observational data is an important problem in practice, especially for large biological datasets. The proposed method scales linearly with the number of variables and can estimate the mediation effect end-to-end.

  2. Extensive experiments validate the soundness of the method. The paper separately evaluates causal discovery, effect estimation, and the complete end-to-end pipeline. The experiments cover up to 10,000 variables, compare against several established methods, investigate missing data and adjustment-set disagreements, and include a real-data application to the Human Phenotype Project.

Critical weaknesses
  1. The framework generates candidate DAGs by randomly sampling consistent extensions of a learned partially directed graph and subsequently applying heuristic score-based refinements. However, the resulting distribution over DAGs is neither a posterior distribution nor a statistically calibrated confidence set. It is totally possible that one adjustment set is valid across all sampled DAGs but still fails in the true causal graph. I wonder if you observe such failure in your experiments.

  2. The end-to-end experiments assume a complete five-tier ordering of variables and restrict the adjustment sets to at most two variables. Furthermore, the synthetic generator is designed with a valid two-variable adjustment set, matching the algorithm's imposed restriction K=2K=2. Could the authors evaluate the framework on DGPs requiring larger adjustment sets (e.g., three or more variables), while varying (K) accordingly? It would also be interesting to investigate how performance changes when the complete tier ordering is unavailable or only partially specified.

Questions for authors

Please see the weakness section.

Reviewer_L5r3OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Channel-Additive Robust Mediation (CHARM), which searches for structure only among each variable’s strongest partners and registered role pairs; fixes one adjustment set valid under every structure supporting a path, and withholds the path when none exists; and estimates every path with one flexible model that learns how the mediator acts on the outcome.

Critical strengths
  1. The research problem is important and interesting.
Critical weaknesses
  1. The methodological novelty seems limited: It mostly combines existing components.
  2. To my understanding, the estimator has no valid statistical inference guarantees.
  3. Real data analysis quietly switches to a different estimand and estimator, making it confusing.
  4. The presentation needs improvement. In Figure 1 and Tables 1--3, the font size is too small.
  5. The main text does not discuss limitations.
Questions for authors
  1. If there is a DAG between candidate mediators, does that affect the proposed procedure?
  2. Does the method have any statistical guarantees about false discovery or coverage?
Reviewer_ApAxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes the CHARM pipeline for causal mediation analysis in high-dimensional spaces. The pipeline consists of three stages: a score-based causal discovery algorithm based on FGES, the selection of an adjustment set from the sampled graphs, and the estimation of natural direct and indirect effects using g-computation. Under these heuristics and assumptions, the paper shows empirically that the framework can be used to estimate effects on large synthetic graphs, along with an application to the Human Phenotype Project.

Critical strengths
  • The paper proposes an approach that can make causal inference more practical, even though several assumptions and heuristics are required.
  • The paper conducts experiments not only on synthetic data but also on a real-world biological dataset, showcasing the scalability of the results.
  • In general, the paper is well-structured and easy to follow.
Critical weaknesses
  1. I think the paper could make it clearer that the improvements in computational efficiency are not free — they are actually obtained through approximations based on heuristics and assumptions. For example, in FGES, only the 32 highest-correlated neighbors are considered; when selecting the adjustment set, it only considers a candidate pool of 12 variables and restricts the set to at most 2 variables. It also assumes that the mediation effects are additive. These assumptions sacrifice accuracy in theory and can be hard to justify.

  2. Each stage in the framework is not particularly novel — they are mostly engineering modifications of existing methods. Therefore, the paper's contribution lies in the end-to-end assembly and the application part. The paper could be more upfront about which pieces are new and which are repurposed.

Minor comments:

  1. Line 116: it is difficult to understand a definition that will only be introduced in later sections.
  2. Lines 237–243: this part is quite confusing. I think it would be better to illustrate it with an example.
Questions for authors
  1. In general, I believe the missing data problem in causal inference requires special caution. How do you ensure that K-MICE does not introduce additional bias?
  2. How sensitive are the values of Q and K in the adjustment set generation?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.