Marginal Causal Flows for Validation and Inference

Investigating the marginal causal effect of an intervention on an outcome from complex data remains challenging due to the inflexibility of employed models and the lack of complexity in causal benchmark datasets, which often fail to reproduce intricate real-world data patterns. In this paper we introduce Frugal Flows, a novel likelihood-based machine learning model that uses normalising flows to flexibly learn the data-generating process, while also directly inferring the marginal causal quantities from observational data. We propose that these models are exceptionally well suited for generating synthetic data to validate causal methods. They can create synthetic datasets that closely resemble the empirical dataset, while automatically and exactly satisfying a user-defined average treatment effect. To our knowledge, Frugal Flows are the first generative model to both learn flexible data representations and also exactly parameterise quantities such as the average treatment effect and the degree of unobserved confounding. We demonstrate the above with experiments on both simulated and real-world datasets.

Paper

References (62)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer J42v7/10 · confidence 3/52024-07-03

Summary

This paper introduces _Frugal Flows_ a method that learns the data distribution of data for causal effect estimation; namely outcome $Y$, binary treatment $X$ and pretreatment covariates $\mathbf{Z}$. Through a combination of frugal parametrisation, normalizing flows and copulas, separate components for the marginal causal effect $p_{Y| do(X)}$, the probability integral transforms of $\mathbf{Z}$ and the propensity score are leaned. (The components of) the learned model, can be used for (i) estimating the marginal effect and (ii) to generate synthetic data with a fixed marginal effect for benchmarking other causal inference methods. In the second application the component for the marginal effect is switched out for another with desired properties. (i) is demonstrated on small synthetic datasets. (ii) is demonstrated by fitting FFs to two real-world datasets and generating synthetic data with adjusted properties.

Strengths

- The paper is well-written. - It tackles an important problem in causality research. Since, randomized data is hard and expensive to get, many causal methods are only evaluated on synthetic data and generating realistic/semi-synthetic data is hard. This paper makes a great contribution towards improving synthetic data generation. If the code for the method is provided in a user-friendly manner, I could see this having a big impact on the causality community.

Weaknesses

- Normalizing Flows have been used in the causal modelling context before (see [1, 2]). While prior works solve different problems (the inferred latents correspond to exogenous variables of an SCM, not directly applied to causal effect estimation), I think it would still be valuable to contrast this work to what has been done before for future reference in the literature. - L59: The basic causal assumptions aren't explicitly stated. What are the causal assumptions on $X$, $Y$ and $\mathbf{Z}$? It seems like the method wouldn't hold if $\mathbf{Z}$ was a mediator (I suppose the equation after L60 wouldn't hold). A reference to a 500+ page book is given for the assumptions, which feels like a slap in the face for the reader. - The notation for interventional distributions is confusing: what's the difference between using an asterisk and explicitly using the do-notation? In the equation after L60, the LHS seems to be an interventional quantitiy (asterisk, but no do-notation), whereas Equation (1) has the do-notation, but no asterisk. Do the two notation elements mean different things? - I think this paper would greatly benefit from a visual abstract showing how the different flows and distributions come together. Maybe this is something that could be added for the camera-ready. Minor: - L201: typo [1] Javaloy et al. "Causal normalizing flows: from theory to practice." NeurIPS 2023 [2] Wendong et al. "Causal component analysis" NeurIPS 2023

Questions

- Fig. 1: What's the meaning of the red undirected edges. Does it mean they could go either way and/or they could be confounded? Please specify this somewhere in the paper or appepndix. - Def. 1, App. A: I struggle to understand this definition. The cartesian product on the RHS makes sense to me, but what does the LHS of the equation mean? What is the "x" operation between functions? Do the two functions map to the same space? - Fig. 2: Why is the first line needed? If I understand correctly, this just makes all pretreatment variables uniform. Couldn't you put $\mathbf{Z}$ directly in the second line? - You show synthetic data generation based on two datasets in Sec. 4.2. Why couldn't you use the same datasets to test causal effect estimation in Sec. 4.1? - How many datapoints are in the Lalonde dataset? - In training, how did you check whether the training has succeeded? I suppose you minimize the log-likelihood, how did you define "good enough"? I'm asking because the training seems pretty fast (App. D2.4). What's the total number of parameters for each of the datasets?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

- The biggest limitation of the work seems to be the requirements for the dataset size, needed to train normalizing flows. The authors mention this in Sec. 4.1. There could be some applications that have enough data for using Frugal Flows for causal effect estimation (e.g. online businesses with many customers, or well-curated medical datasets like the ones from healthcare providers in Isreal). However, for most applications the data won't be enough.

Reviewer Cwuj7/10 · confidence 4/52024-07-11

Summary

The paper introduces a generative modeling approach called Frugal Flows, designed to learn the data generation process with an explicit parametrization for the marginal causal effect of treatment on outcomes. Inspired by the frugal parametrization of marginal structural models, this approach models the marginal intervention distribution $p(Y|do(X))$ directly, rather than the joint distribution $p(Y|Z, do(X))$. This helps in preserving any constraints on the average treatment effect while flexibly modeling the data generation process. Frugal Flows employs copula flows to parameterize the model, accommodating constraints on the average causal effect and handling unobserved confounding during data generation. The authors validate the proposed method through experiments on both synthetic and real-world datasets, demonstrating its ability to generate realistic datasets with user-specified constraints.

Strengths

* The paper's approach to validating causal models using simulated datasets is indeed impactful and relevant. It addresses a significant gap by allowing for general constraints on quantities of interest, such as average causal effect and unobserved confounding, during data generation. This capability is crucial because many prior generative modeling approaches for causal datasets either do not offer such flexibility or cannot ensure the preservation of these constraints, thus making this work a notable advancement in the field. * The paper is well-written, with clear explanations in the background sections on frugal parametrization and flows, which help the reader grasp the proposed approach. The details of the approach are well-articulated, and the experimental results are presented effectively. * The proposed approach is indeed novel. While it builds on established concepts like frugal parametrization, the specific application of normalizing flows for parametrization and its focus on average causal effect estimation represent a significant and innovative contribution.

Weaknesses

My main concern with the work is the limited empirical validation of the proposed approach. Given that the primary contribution is the learning methodology rather than theoretical analysis, I would expect a more extensive set of experiments to validate its effectiveness. For example, prior research on generative modeling for causal inference, such as the work by [1], includes comprehensive experiments with various statistical tests to assess the realism of generated samples and a broader benchmarking of causal estimators. This paper would benefit from similar depth in its empirical evaluation. It would nice if the authors can conduct similar experiments to asses whether learned generative model generates realistic samples and evaluate it on more datasets. Also, the authors should compare with the prior works [1, 2] as baselines to establish which approach is the best at capturing the underlying data generation process, and empirically validate their claim (Section 2.6) that the proposed approach would be better than prior works at capturing used-specified constraints on the average causal effect. References [1] Neal, Brady, Chin-Wei Huang, and Sunand Raghupathi. "Realcause: Realistic causal inference benchmarking." arXiv preprint arXiv:2011.15007 (2020). [2] Harsh Parikh, Carlos Varjao, Louise Xu, and Eric Tchetgen Tchetgen. Validating causal inference methods. In International conference on machine learning, pages 17346–17358. PMLR, 2022.

Questions

* A suggestion for the notation is that the authors could use $T$ instead of $X$ to denote the treatment variables in the paper. This way the notation would be less confusing, as $X$ represents a general random variable in Section 2.5 and Section 2.6 * Maybe there is a typo in Figure 2? The top row should be $\mathcal{F_{Z_i}}^{-1}(.)$ as we transforming the covariate $\{ Z_i \}$ to the correlated uniform variables $\{ V_{i} \}$ * I don't understand the Section 3.1.1 on copula flow for $X$ on $Z$. Why don't we directly model $p(X|Z)$ using a normalizing flow and why do we need to parametrize using copula flow as $p(X|Z)= p(X).c(X|Z)$?

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

Yes, the authors have adequately addressed the limitations of their work.

Reviewer SZi86/10 · confidence 3/52024-07-12

Summary

This paper proposed a generative model called Frugal Flows making use of copula flows to infer about marginal causal effects by simulating the data generating process.

Strengths

- The problem of inferencing marginal causal effects is an interesting and important problem - The idea of using generative models to estimate the marginal effects in the paper is interesting

Weaknesses

See questions.

Questions

I have two question -- 1. A highly related (and I suspect might be viewed as a "dual" appproach to your Frugal Flow) is the statistical matching (e.g., bipartite matching) to estimate the average treatment effect. It would be very informative to compare this as one of the baselines in your benchmarking and validation, as this also shed lights on how these two different schools of causal inference may (or may not) converge on ATE estimation. 2. I think it is good to use real data for benchmarking/validation (but perhaps benchmarking is a bit strong here since only two datasets were used), but usually it is unclear how to interpret the results since the ground truth is unknown. Can you design and run some controlled synthetic experiment to verify the model?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

See questions.

Reviewer PATR7/10 · confidence 3/52024-07-12

Summary

This work proposes to leverage existing neural density estimators (specifically, normalizing flows) to exploit a newly-proposed "frugal parametrization" that can capture the causal marginal distribution of an underlying causal model. Under this parametrization, the authors show how to specify and train each component of the model, and thus train the proposed Frugal-Flows to match the observational distribution as closely as possible, while being able to tune the marginal causal effect present in the generated data. This way, frugal flows can be used to generate synthetic causal benchmarks that closely represent the _observational_ data while having more difficult-to-estimate causal effects, putting existing approaches to the test.

Strengths

- **S1.** The proposed frugal flows provide a way of generating new datasets that can be challenging from a causal-inference point of view, which I believe _important_ to test new and existing methods. - **S2.** The construction of the proposed architecture is quite rich in details. - **S3.** I find the frugal parametrization conceptually quite interesting. - **S4.** The authors motivate different scenarios for frugal flows in Sec. 3.2, as well as empirically show positive results on some synthetic and real-world scenarios.

Weaknesses

- **W1.** I find the frugal parametrization to be extremely under-explained, relying too much on the reader having full knowledge of the referenced work. Similarly, there is little to no explanation/intuition on why the frugal parametrization would properly capture the marginal causal distributions. - **W2.** The lack of explanations also applies to other concepts, e.g., "conditional ignorability" (line 39) "variation independence" (line 82, and I know the definition is later in App. A), or why copula-based flows would target conditional causal effects instead of marginal causal ones (line 182). (similar with lines 221 and 229) - **W3.** There are no mention to related works that propose similar ways of constructing causal benchmarks. From a 1-min search in google scholar, I already found some likely relevant works: [Work 1](https://arxiv.org/abs/2406.08311), [Work 2](https://arxiv.org/abs/2011.15007). - **W4.** I find the experiments a bit underwhelming, specially those from Section 4.1. The authors should at least show how is the fitting of the observational likelihood, and if they want to show the capabilities of frugal flows for causal inference (and not only causal-benchmark generation), they should compare with other methods like [Causal Normalizing Flows](https://arxiv.org/abs/2306.05415).

Questions

- **Q1.** I am not sure that I understand what does the dotted red line represent in the boxplots. - **Q2.** Doesn't the statement in lines 291-294 directly contradict what you say later in lines 296-297? - **Q3.** Why is Figure 1 placed there?

Rating

7

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

I think limitations are properly discussed.

Reviewer PATR2024-08-11

I thank the authors for their super detailed responses and the additional experiments. I do think they add quite some value, and if the authors use the feedback to improve the readability of the manuscript (and thus how welcoming it is for newer audiences), I think it has potential to be quite a good paper. With respect to the new results, I am a bit surprised by the results of CNF on model 1. Just to make sure, I'd encourage the authors to check that they used a model different from MAF for the base model, as this is not a universal density approximator. In any case, I am happy with the response from the authors and I will update my score.

Reviewer Cwuj2024-08-11

Thanks for the detailed response! My questions have been addressed and I have increased my rating accordingly!

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC