A Causal Framework for Decomposing Spurious Variations

One of the fundamental challenges found throughout the data sciences is to explain why things happen in specific ways, or through which mechanisms a certain variable $X$ exerts influences over another variable $Y$. In statistics and machine learning, significant efforts have been put into developing machinery to estimate correlations across variables efficiently. In causal inference, a large body of literature is concerned with the decomposition of causal effects under the rubric of mediation analysis. However, many variations are spurious in nature, including different phenomena throughout the applied sciences. Despite the statistical power to estimate correlations and the identification power to decompose causal effects, there is still little understanding of the properties of spurious associations and how they can be decomposed in terms of the underlying causal mechanisms. In this manuscript, we develop formal tools for decomposing spurious variations in both Markovian and Semi-Markovian models. We prove the first results that allow a non-parametric decomposition of spurious effects and provide sufficient conditions for the identification of such decompositions. The described approach has several applications, ranging from explainable and fair AI to questions in epidemiology and medicine, and we empirically demonstrate its use on a real-world dataset.

Paper

References (19)

Scroll for more · 7 remaining

Similar papers

Peer review

Reviewer EguC8/10 · confidence 3/52023-06-27

Summary

The paper proposes a novel theoretical framework for decomposing spurious causal effects. For this purpose, the authors introduce the concept of partial abduction and prediction. Here, only a subset of the latent variables in a given structural causal model (SCM) is updated when conditioning on observed evidence. The authors use this concept to derive nonparametric decompositions of spurious effects and provide identifiability results both for Markovian and semi-markovian SCMs. The paper ends with an empirical example using the COMPAS dataset.

Strengths

- The paper addresses a relevant problem: decomposing the spurious effect in the presence of multiple confounders. This has potential applications in e.g., causal algorithmic fairness. - Novelty: I am not aware of any other work done in this regard. The authors develop a novel theory for defining such decompositions based on Pearl's SCM framework. - The paper is technically solid and rigorous. Definitions are clear and precise, all assumptions are stated transparently, and proof and illustrative examples for theoretical results are provided.

Weaknesses

- My only main concerns are regarding the applicability of the proposed framework in practice. To my understanding, the full (semi-markovian) causal graph needs to be known (including the spurious arrows between the observed confounders) in order to apply the decomposition result from Theorem 3. This may be the case for simple examples like the COMPAS dataset. However, in many applications, confounders are potentially high-dimensional and the causal graph is unknown. - I think the paper would benefit from expanding the experiment section, i.e., by including a second example dataset. Minor points: - Eq. (7): I suggest using the notation $p(Y_x = y)$ so that $y$ on the right-hand side of the equation is well-defined. - Typo in Definition 7: path causal path - Proposition 2: I suggest moving the definitions of TV and TE from Appendix A1 to the main paper.

Questions

- It seems like the decomposition in Theorem 1 depends on the chosen topological ordering of the observed confounders $Z_i$. Consequently, would the "spurious contribution" of confounder $Z_i$ change when selecting a different ordering? This would seem counterintuitive for a framework that aims at decomposing the overall spurious effect into individual "spurious contributions".

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

I suggest including a section on limitations. Here, the authors should particularly stress the assumptions necessary for the identification of spurious decompositions from observational data. Furthermore, a discussion on these assumptions in the context of the empirical example in Section 5 should be included.

Reviewer 1quh7/10 · confidence 2/52023-07-08

Summary

This paper considers the problem of decomposing spurious variations in both Markovian and Semi-Markovian models, which are potentially useful to many real applications. A non-parametric decomposition of spurious effects is given given and identification issue is also discussed under suitable conditions.

Strengths

- decomposing spurious variations can be potentially useful to many real applications - Presentation is very good with well-structured preliminary and examples - Proof is complete and concise

Weaknesses

It seems that there is no discussion on limitation of the developed theory.

Questions

The paper is very clear in both motivation and technical part. As the first work on this problem with a general setting, the content of the present version is sufficient for an acceptance.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

It seems that there is no discussion on limitations. Developing more identification theories is worth future investigation, as also mentioned by the authors.

Reviewer G3e38/10 · confidence 3/52023-07-08

Summary

The authors tackle the problem of decomposing the source of variation in causal analysis. In many real-world problems, there are many factors that can introduce spurious correlation between a treatment and outcome, and breaking down the contribution of each variable to that spurious correlation can be useful in many fields where an explanation is important. The authors propose a methodology for both Markovian and semi-Markovian models, based on the 'abduction-action-prediction' method. It allows for the application of evidence to only a subset of exogenous variables, which then allows us to separate the variation from each exogenous variable.

Strengths

This paper is well-written and well-motivated. The problem it seeks to address is important, and the introduction does a good job at setting the theoretical foundation and describing the problem. For the most part, the terminology is introduced at a good pace, making the overall narrative easier to follow. I appreciate the authors interspersing examples with figures, since without them, the paper would likely bog down with all the math. Overall, I found this a compelling approach to an important problem.

Weaknesses

I think, for such an equation-heavy paper, the notation could be made clearer/more explicit in sections. For example, Proposition 1 is the first time that we see the $e^{U_1}$ notation, which it seems like means 'this evidence is applied to everything except for $U_1$'. That's a fairly counter-intuitive notation for such a thing, since generally a subscript or superscript implies that we're using those variables in someway, rather than explicitly excluding them (I realize you need to use them to exclude them, but conceptually, the notation is non-obvious for me). If I'm not misunderstanding what this notation means, then a more explicit definition before Proposition 1 would be helpful, given how central it is to the rest of the paper. As another notation point, Theorem 2 uses the notation $z_{-[i]}$, but until now, the $Z_{[i]}$ notation has meant, as defined in Definition 2, "all $Z_i$'s, from 1 to i", and it's not obvious how this translates to a negative. The COMPAS demonstration is great, but since it clearly lacks ground truth, additional synthetic experiments would be helpful to help demonstrate both the correctness of the theory and how this method can be used in practice on other types of data.

Questions

Is there a list of all of assumptions required for your method? In Section 4.1, you say that the decomposition needs to follow a topological order of the variables U_i, but are there others?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

In terms of social impact, this method is more likely to have a positive social impact, since causal inference is often applied in domains where spurious variation is present but where the decisions made can have significant effects on people's lives (e.g., the COMPAS example)

Reviewer VtmC7/10 · confidence 3/52023-07-20

Summary

This paper provides a new framework for decomposing spurious effects and its identification conditions. In addition, they proposed a nonparametric method for decomposing the spurious effect under sufficient conditions.

Strengths

- In general, this paper is well-motivated. - It provides a procedure to decompose spurious effects. - The theoretical results in this paper are sound.

Weaknesses

- The paper could clarify the difference between "spurious effect" and "spurious variations", or state if they are used interchangeably. More explanation of how these two terms are defined would help the reader understand if they refer to the same concept. - More details could be provided on how the procedure for Semi-Markovian spurious decomposition is implemented. In particular, explaining the steps for removing the effects of all exogenous variables would make the approach clearer. - Comparing the proposed method against existing methods on estimating causal effects could highlight an advantage. Since the proposed method can estimate redundant effects, the resulting causal effect estimates may be more accurate compared to current approaches. This comparison could emphasize the improved performance enabled by accounting for spurious variations.

Questions

See the weaknesses above.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

See the weaknesses above.

Reviewer dHJ75/10 · confidence 2/52023-07-25

Summary

The manuscript studies the decomposition of spurious variations. Specifically, a tool is developed for the non-parametric decompositions in both Markovian and Semi-Markovian models.

Strengths

1. The considered problem is practical. As stated in the introduction, the related tools are almost entirely missing in the literature (though I'm not totally sure about it). 2. The theoretical framework is rather complete.

Weaknesses

1. The empirical evaluation seems to be not strong enough. The proposed tool has only been applied in one specific case. The generality and stableness of the method could be better illustrated by more types of evaluation, perhaps including some simulations. 2. Since causal inference in the presence of latent confounders is a well-studied problem, it will be helpful if more discussion on the considered task and the other related ones can be conducted. This can provide additional insight and broaden the understanding of the subject matter.

Questions

1. What could be the limitations of the proposed tool?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

I didn't find a discussion on the limitations.

Reviewer EguC2023-08-10

Thank you for your clarifications

I appreciate the detailed response and additional experiments. I still have some concerns regarding practical applicability, but as pointed out by the authors, strong assumptions are to be expected and are in line with previous literature. Overall the paper is a valuable contribution to the causal inference literature. I raise my score accordingly and recommend acceptance.

Authorsrebuttal2023-08-11

We thank the reviewer for acknowledging our response, and providing swift feedback. We also acknowledge the strength of assumptions needed for using the tool, but remark that we are in particular optimistic about the usage of cluster diagrams in this context. In particular, we are grateful that the reviewer sees our framework as a valuable contribution to the causal inference literature, which is quite encouraging. Also, thank you for adjusting the score!

Reviewer VtmC2023-08-13

Thank you for the clarifications

Thank you for the clarifications. I’m happy to increase my evaluation score.

Authorsrebuttal2023-08-14

We thank the reviewer for a constructive review process and acknowledging our clarifications. We are quite encouraged by the grade increase from the reviewer, thanks!

Reviewer dHJ72023-08-18

Thank the authors for the clarification. I will keep my score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC