Smoke and Mirrors in Causal Downstream Tasks

Machine Learning and AI have the potential to transform data-driven scientific discovery, enabling accurate predictions for several scientific phenomena. As many scientific questions are inherently causal, this paper looks at the causal inference task of treatment effect estimation, where the outcome of interest is recorded in high-dimensional observations in a Randomized Controlled Trial (RCT). Despite being the simplest possible causal setting and a perfect fit for deep learning, we theoretically find that many common choices in the literature may lead to biased estimates. To test the practical impact of these considerations, we recorded ISTAnt, the first real-world benchmark for causal inference downstream tasks on high-dimensional observations as an RCT studying how garden ants (Lasius neglectus) respond to microparticles applied onto their colony members by hygienic grooming. Comparing 6 480 models fine-tuned from state-of-the-art visual backbones, we find that the sampling and modeling choices significantly affect the accuracy of the causal estimate, and that classification accuracy is not a proxy thereof. We further validated the analysis, repeating it on a synthetically generated visual data set controlling the causal model. Our results suggest that future benchmarks should carefully consider real downstream scientific questions, especially causal ones. Further, we highlight guidelines for representation learning methods to help answer causal questions in the sciences.

Paper

References (60)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Zfn96/10 · confidence 3/52024-06-20

Summary

the paper makes a deep dive into the various types of biases that can arise in RCTs that would invalidate causal estimants

Strengths

- Interesting paper that would greatly benefit the discussions of the community - Identifies crucial sources of bias, offers interesting potential solutions to them - Experimentally shows the biases in a large number of models

Weaknesses

- The literature review only focuses on the last 2 years and ignores prior works that are still relevant - there is little theoretical argumentation in the proposed solutions - there is little experimental evidence on ways to overcome the identified problems

Questions

- not much to be honest, the identification and codification of the RCT biases and how these effect causal estimates makes the paper pass the bar for me

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

- adequately addressed

Reviewer WAFK7/10 · confidence 1/52024-07-10

Summary

This paper theoretically reveals that many common choices in the literature may lead to biased estimates. To test the practical implications of these considerations, this paper recorded the first real-world benchmark for causal inference downstream tasks on high-dimensional observations through an RCT studying how garden ants (Lasius neglectus) respond to microparticles applied to their colony members by hygienic grooming. By comparing 6,480 models fine-tuned from state-of-the-art visual backbones, they found that the sampling and modeling choices significantly affect the accuracy of the causal estimates, and that classification accuracy is not a reliable proxy for this accuracy.

Strengths

1. To facilitate future research on representation learning for causal downstream tasks, the authors have formulated representation desiderata to obtain accurate estimates for causal queries, along with best practices. 2. The authors have designed and collected a new dataset, conducted extensive experiments, and proposed a new synthetic benchmark. 3. This paper is well-structured, with thorough theoretical derivations and experimental discussions.

Weaknesses

1. The link to the data set provided by the author cannot be accessed. 2. The author's presentation of the provided data set is a little bit limited, especially in terms of visualizations. It would be helpful to include some actual images of the physical data.

Questions

1. Could authors provide more details on the dataset collection process and explain why ants were used as an example? 2. Could the method proposed in this paper be generalized to other fields, beyond the ecological example mentioned?

Rating

7

Confidence

1

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors thoroughly discussed the limitations of this paper.

Reviewer Erid8/10 · confidence 3/52024-07-11

Summary

This paper considers the task of causal effect estimation $P(Y|do(T))$, where the treatment is mediated through a (potentially high-dimensional) observation $X$. Additionally, a semi-supervised setting is studied, where labels $Y$ are only available in a subset of the data. A set of possible biases affecting the treatment effect estimation that can arise in this setting are discussed together with possible mitigation strategies. Two datasets are introduced as new benchmarks: (i) An Ant Dataset, where groups of ants are filmed under the treatment of putting microparticles on the body surface. Grooming behaviour is observed through video recordings. (ii) CausalMNIST: A dataset that adds colour coding to MNIST data and adds a downstream causal task. The biases discussed above are empirically evaluated on a large battery of models that are fine-tuned in few- and many-shot learning.

Strengths

I think this paper has many great contributions and approaches the questions it tries to answer systematically. - The biggest contribution IMO is the ant video dataset. Causal representation suffers from overly simplistic settings and restrictive assumptions. Such datasets have the potential to move this field closer to practically useful applications and out of the ivory tower of identifiability under very restrictive assumptions. - I love Sec. 3 and the fact that it is written as a sort of self-help book on causal effect estimation. This was refreshing to read. - The writing is superb. - The systematic evaluation of a large number of different models is great.

Weaknesses

The biggest weakness of this paper is that it tries to do a bit too much. Given the NeurIPS page limit, this leads to things being condensed to a point that makes it difficult to follow at times. For example, Theorem 3.1 tries to convey an idea without properly defining most of its elements, and IMHO it fails to convey much at all. Without looking up the formal definition in the appendix, it's difficult to understand.

Questions

- Sec. 5.1: What exactly are $X$ and $Y$ for this dataset?

Rating

8

Confidence

3

Soundness

3

Presentation

3

Contribution

4

Limitations

- "Model bias from the encoder choice" (L146): This point is a bit wishy-washy and, as it is, it has limited use to the practitioner. It is too generic and doesn't offer much practical guidance on how to e.g. "attempt to quantify its biases". Maybe this space would be better spent on expanding on other parts of the paper.

Reviewer ZnDR6/10 · confidence 2/52024-07-14

Summary

The paper explores the challenges associated with using machine learning, particularly deep learning, to estimate causal treatment effects from high-dimensional data, such as images, in Randomized Controlled Trials (RCTs). The authors point out that standard practices in machine learning, such as selecting models based on accuracy and discretizing predictions, can lead to biases in causal estimates. To substantiate their claims, they provide both theoretical analyses and empirical results. They introduce a novel real-world benchmark dataset that involves the grooming behavior of ants, which they assert is the first benchmark designed for causal inference downstream tasks. The authors stress the necessity of incorporating causal downstream tasks into benchmark design and offer guidelines to enhance causal inference in scientific applications through machine learning.

Strengths

1. The paper addresses an important and under-explored area of causal inference, which will facilitate further research. 2. The authors provide some theoretical analysis to identify potential sources of bias in treatment effect estimation using deep learning models. 3. The paper is well-written and well-organized. 4. The experiments involve fine-tuning thousands of machine learning models with various encoders, enhancing the reliability of the conclusions.

Weaknesses

I am not very familiar with this topic, so I may adjust my rating based on further discussions with the authors and reviewers. 1. The main theoretical analysis focuses on binary classification, which may not generalize to many real-world settings. On line 251, the authors claim that Theorem 1 is validated on their dataset. Could you provide more details on how Theorem 1 is validated? I found it difficult to follow. 2. The setting of this paper differs from many causal representation learning algorithms, making it challenging to directly evaluate those algorithms on this benchmark. This may limit the applicability of the proposed benchmarks. Could you apply some existing causal representation learning methods to the proposed datasets and compare the conclusions? The experiments only test different pre-trained models, but it would be more compelling to demonstrate that this new benchmark can help answer causal questions with training in a causal manner. 3. The authors frequently mention causal inference downstream tasks. Can you further elaborate on why the proposed dataset is superior to other benchmarks for answering causal questions? The experiments primarily show that bias can arise, but I am not fully convinced that the proposed dataset is better than the existing ones. 4. While the authors acknowledge limitations such as keeping the backbones of models frozen and not incorporating semi-supervised learning, these constraints might affect the generalizability of their findings.

Questions

Please see the weaknesses below. I may reconsider my score after further discussion with the authors and reviewers. My main concern is the significance of the proposed benchmark and its utility in evaluating different causal representation algorithms. Why is this dataset better suited for answering causal questions and reducing bias compared to existing datasets?

Rating

6

Confidence

2

Soundness

2

Presentation

3

Contribution

2

Limitations

Yes.

Reviewer Erid2024-08-08

Thank you for the clarifications. On the limitation: I'm happy with either solution (if you expect me to weigh in on which option to choose).

Authorsrebuttal2024-08-08

Thank you for the feedback. We are also happy either way and will decide based on the amount of space we have available as we prepare the next revision.

Reviewer WAFK2024-08-10

Thanks for the author's rebuttal. I will keep my score.

Reviewer Zfn92024-08-12

Acknowledgement

I acknowledge that I have read the authors rebuttal, and I maintain my score of acceptance

Reviewer ZnDR2024-08-12

Sorry for the late reply. Thank you so much for the detailed rebuttal. Most of my concerns are addressed. Regarding my first question, I meant Theorem 3.1—sorry for the typo. In lines 301-302, I saw that Theorem 3.1 is empirically validated, but adding more analysis could be beneficial. After reading other reviews, I cannot find significant issues, and I am willing to increase my score to 6 while keeping the confidence level at 2.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC