Generalized Balancing Weights via Deep Neural Networks

Estimating causal effects from observational data is a central problem in many domains. A general approach is to balance covariates with weights such that the distribution of the data mimics randomization. We present generalized balancing weights, Neural Balancing Weights (NBW), to estimate the causal effects of an arbitrary mixture of discrete and continuous interventions. The weights were obtained through direct estimation of the density ratio between the source and balanced distributions by optimizing the variational representation of $f$-divergence. For this, we selected $α$-divergence as it presents efficient optimization because it has an estimator whose sample complexity is independent of its ground truth value and unbiased mini-batch gradients; moreover, it is advantageous for the vanishing-gradient problem. In addition, we provide the following two methods for estimating the balancing weights: improving the generalization performance of the balancing weights and checking the balance of the distribution changed by the weights. Finally, we discuss the sample size requirements for the weights as a general problem of a curse of dimensionality when balancing multidimensional data. Our study provides a basic approach for estimating the balancing weights of multidimensional data using variational $f$-divergences.

Paper

References (39)

Scroll for more · 27 remaining

Similar papers

Reviewer bV7F4/10 · confidence 2/52023-06-29

Summary

This paper proposes a method for estimating the balancing weights in causal effect estimation. The balancing weights are given by the density ratio of the counterfactual distribution and the observed distribution, and the density ratio can be estimated by solving the optimization of the variational expression of the f-divergence between the two distributions. The authors claim that using the $\alpha$-divergence is useful because it addresses the vanishing-gradient problem when we use a neural network for the model. The paper presents some theoretical results.

Strengths

- Expressing $\phi$ as $\exp(T)$ in Eq. (14) is interesting since the domain of the optimization variable becomes simpler. - The discussion about vanishing gradient is interesting, and the suggested way of setting $\alpha$ may be useful to address this issue.

Weaknesses

- There are several technical parts that might not be precise. See the Questions section below. - The writing could be improved. The paper rather looks like a collection of definitions of results. Connections between sections are not smooth, and the message of the paper as a whole is not clear. - The abstract says, "we selected $\alpha$-divergence as it presents efficient optimization because it has an estimator whose sample complexity is independent of its ground truth value and unbiased mini-batch gradients; moreover, it is advantageous for the vanishing-gradient problem." I am not sure what makes the $\alpha$-divergence special in terms of these aspects. Also, the explanation about the vanishing-gradient issue in Section 5.1 is not convincing enough to make me believe that the authors' suggestion really addresses the issue. - The paper relies on references (including one without open access) for important definitions. The paper could be more self-contained. - Overall, the paper is mainly about estimating density ratios by the variational expression of the f-divergence, which is not novel. - There is no experiment in the main part of the paper.

Questions

- Is Definition 3.1 really the same as the standard definition from Pearl (2009)? Which part of the book are the authors referring to? - I believe there is something wrong with Eq. (4) and Eq. (7). The right hand side depends on $P(\boldsymbol{X}_1), ..., P(\boldsymbol{X}_n)$ while the left hand side does not. - In lines 207-211, what is $\overset{\to}{\infty}$? - "We note that the selecting of α does not cause instability in numerical calculations for cases where $E[\nabla_{\theta} \mathcal{L}_{\alpha}(θ)] \to \infty - \infty$." Is there any theoretical result to support this claim? - Theorem 6.1 looks strange because Eq. (19) concerns $\tilde{P}$ although Eq. (18) has nothing concerning $\tilde{P}$. - What does it mean by using test data for validation? Is this validation performed only for performance evaluation or also for hyper-parameter tuning? - As far as I know, people just perform as many epochs as they can afford while keeping the model with the best validation performance. So, fixing the early stopping time is not very useful in practice unless. Am I missing something? - The $\alpha$-divergence "has an estimator whose sample complexity is independent of its ground truth value": is this true? Eq. (13) seems to depends on both $P$ and $Q$. - "unbiased mini-batch gradients": isn't this true for any $f$-divergence? ----- Edit on Aug 2: There is another thing that I would like the authors to address if it is not too late: Is Eq. (10) correct? Is $f'$ the derivative of $f$ and are we optimizing the objective with respect to $\phi$? I thought that the authors use the result from [15], but I cannot exactly match the equation with the one in the paper. [15] XuanLong Nguyen, Martin J Wainwright, and Michael Jordan. Estimating divergence functionals and the likelihood ratio by penalized convex risk minimization. Advances in neural information processing systems, 20, 2007.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

2 fair

Presentation

1 poor

Contribution

3 good

Limitations

The authors discuss limitations of the work.

Reviewer bV7F2023-08-21

Remaining questions

If the authors still have time to respond, I encourage them to comment on these points: > The $\alpha$-divergence "has an estimator whose sample complexity is independent of its ground truth value": is this true? Eq. (13) seems to depends on both $P$ and $Q$ (I should have looked at Eq. (105) instead of Eq. (13)) > "unbiased mini-batch gradients": isn't this true for any $f$-divergence?

Authorsrebuttal2023-08-22

We apologize for not answering the reviewer's question of concern. > The $\alpha$-divergence "has an estimator whose sample complexity is independent of its ground truth value": is this true? Eq. (13) seems to depends on both The term ‘sample complexity’ would have been ambiguous. We use this to mean ‘the upper bound of the sample sizes to guarantee that the variance of estimators is below a constant’, or, more strictly, ‘the upper bound of the sample sizes to guarantee that the ratios of the variance of estimators to the truth values are below a constant’. > "unbiased mini-batch gradients": isn't this true for any $f$-divergence? Some previous studies have addressed the problem of biased mini-batch gradients of $f$-divergences ([4], Section 3.2. in [B3]). [4] Bellemare, M. G., Danihelka, I., Dabney, W., Mohamed, S., Lakshminarayanan, B., Hoyer, S., & Munos, R. (2017). The cramer distance as a solution to biased wasserstein gradients. arXiv preprint arXiv:1705.10743. [B3] Belghazi, M. I., Baratin, A., Rajeswar, S., Ozair, S., Bengio, Y., Courville, A., & Hjelm, R. D. (2018). Mine: mutual information neural estimation. arXiv preprint arXiv:1801.04062.

Reviewer bV7F2023-08-22

Unbiasedness

Thank you for sharing the references. I think the paper [B3] addresses the biased mini-batch updates for the "Donsker-Varadhan representation" not the $f$-divergence. In fact, they mention in Footnote 6 that > From the optimization point of view, the $f$-divergence formulation has the advantage of making the use of SGD with unbiased gradients straightforward I agree with them because $f$-divergence has the expectation operation in the very outer side of the expression and taking the empirical approximation of this expectation will give us an unbiased estimate no matter what $f$-divergence we use. The point I am making here is that the unbiasedness comes as a general property of $f$-divergences, but as a special property of the $\alpha$-divergence. It might make sense for the paper to mention this to justify the choice of $f$-divergence over other divergences, but I don't think the unbiasedness makes the $\alpha$-divergence special compared to other $f$-divergences.

Reviewer LPz44/10 · confidence 4/52023-07-02

Summary

This paper proposes an approach for estimating balancing weights. The ultimate target is to use these weights for the estimation of the casual effects of interventions. The search of balancing weights are formulated as density ratio estimation. To tackle this problem, the authors employ a variational representation of the $f$-divergence. . The density ratio is modeled by a neural network. Specifically, this paper advocates the use of $\alpha$-divergence, and provide various ways to improve the practical performances of the proposed weights.

Strengths

- This paper successfully makes use of the density ratio learning tools based on $f$ -divergence for causal inference applications. - This paper summarizes various related existing results on density ratio learning via variational form of $f$ -divergence. - This paper provides detailed discussion on various practical improvement of the proposed algorithm, which would benefit practitioners.

Weaknesses

- This paper lacks theoretical results on the causal effect estimation (not the density ratio learning), which is regarded as a key target of this work. - It is unclear whether the proposed estimator obtain optimal convergence rates and (semi-parametric) efficiency, when degenerated to standard settings like average treatment estimation, and conditional average treatment estimation. - This work is not impressive in terms of novelty, as many significant components follow more directly from existing work. - The variational form of $f$ -divergence and the corresponding estimation of density ratio has been explored by references [15] and [16]. - The main theoretical result (Theorem 6.1) seems to be a fairly standard dual result (e.g., Theorem 4.4 of https://arxiv.org/pdf/1003.5457.pdf) - The theoretical results are not stated clearly. Many results are not stated in rigorous statements with careful listing of assumptions. Some seem to be exclude certain important cases like mixture distributions, as claimed to be a major contribution of this proposed work. (see my question below.)

Questions

- When the target causal effect is a scalar, does the proposed weighted estimator achieve $\sqrt{N}$ rate? is the corresponding weighted estimator efficient? - For conditional treatment estimation, the target quantity is a function, and could be nonparametric in general. Would the weighted estimator converge? Does it converge with optimal rate? - How to initialize the algorithm? - How to choose $C$ in (23) in practice? - Line 277, can the author formalize their claim that $W_1(Q, \hat{Q}_{K_0}^{(N)})$ (i.e., with clear conditions) and provide a proof? Since the estimator is parametrized by neural network, it leads to a non-convex optimization in general. The loss landscape in terms of the neural network parameters seem complicated, and the performance of the estimator resulting from early stopping seems to be related to initialization. (In the supplement (Section C.3), it seems that the authors analyze a different algorithm (instead of Algorithm 1) where the GD is not applied to the neural network parameters.) - While it is great that the authors provide a detailed discussion of various ways to improve the performance of the proposed estimator in Section 7.1, these strategies seems relatively standard. - The authors emphasize that their estimator can handle arbitrary mixture of discrete and continuous interventions, but I don't see much discussion/innovation related to this challenge throughout the paper. It seems to me the only trick that is advocated to deal with this challenge is to view the weights as Radon-Nikodym derivatives (which seems to be pretty well known) and then use NN to model. Am I missing anything? - Also, in the case of mixture interventions, can the authors confirm and discuss the applicability of the theoretical results in Section 7.1 and 8? It seems that the analysis of the early stopping does not apply to the mixture settings (Assumption E4).

Rating

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

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Section 8 discusses limitations on sample size requirements. The authors claim that the sample size required for controlling the error of $\hat{Q}_{K_0}^{(N)}$ has an exponential dependence on the dimension. Although I have questions regarding this theoretical results (as stated above), I would not be surprised about the curse of dimensionality. However, since the ultimate target is the causal effect, I am more interested in a direct error or sample complexity analysis of the causal effect. The lack of such analysis is a major limitation of this work.

Reviewer L7FM4/10 · confidence 2/52023-07-07

Summary

This paper derives a novel method of reweighting covariates for the purpose of causal inference. This is done through learning optimal change of measure weights modeled via neural networks and trained using alpha-divergence measures between the true joint distribution and a mutually independent one. Further tweaks are introduced to help enable this procedure to be more practical in its implementation. A brief disclaimer: I am not very familiar with causal inference (to be honest I am not quite sure how I received this paper to review); however, I am quite familiar with variational inference and related techniques. I will focus primarily on the latter during my review, but please keep all of this in mind while reading my comments.

Strengths

All of the work presented is very precise and detailed, with novel contributions that I can see with regards to the learning of the weights via the alpha-divergence objective. I found the result of the main algorithm to be particularly concise and intuitive for such an involved derivation (e.g., almost trivial to describe but definitely not to prove). As far as I could tell, all of the decisions made throughout the proposal were well justified and deliberate.

Weaknesses

The precision presented definitely came at the cost of readability in my opinion. The paper is very notation heavy and introduces many concepts rapidly and with great precision that lead me to having to re-read sections many times over to understand the message. It feels like the paper could benefit from summarizing some of the technical details in the main paper and reproduce the more exact version in the appendix. Additionally, I understand the work faced space limitations; however, I believe it should definitely have the numerical experiments presented in the main paper. If pressing for pages, I could see including the experiments in the main paper and moving section 7.1 (either partially or completely) to the appendix. I found no weaknesses in the technical information itself from what I could understand.

Questions

The derivations showcased that optimal values of alpha range from 0 to 1, in your experience was there any intuition over what value in this range alpha should be? Or put differently, is there much difference between different values so long as they are all in this range? I am primarily interested both in terms of the resulting divergence value, but also in terms of the resulting downstream effect that such weights would have on applications.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

4 excellent

Presentation

2 fair

Contribution

4 excellent

Limitations

The authors very adequately discussed limitations.

Reviewer vQ5H4/10 · confidence 1/52023-07-07

Summary

This work presents the way to estimate the balance weight, that represents the causal effects of arbitrary mixture of intervention, by using the neural network. Specifically, authors recognize this balance weight as the density ratio between the source and balanced distributions, and estimates this ratio by optimizing the variational representation of $\alpha$-divergence with $\alpha \in (0,1)$. In this procedure, authors justify why the $\alpha$-divergence should be used in terms of the property of the estimator, such as the sample complexity and unbiasedness, and the vanishing gradient issue for training.

Strengths

* This work seems to have a solid theoretical explanation for the balance weight estimation through neural network.

Weaknesses

First of all, I confess that I am not expert in this field, and thus ask for understanding if my feedback does not make sense in this field. * The proposed has not been compared with other baselines for balance weights. > Authors claim that the proposed approach is a general method that can estimate balancing weight even when the datasets are generated from arbitrary mixture of discrete and continuous interventions. This means that the proposed method could be comparable with other baselines if either discrete interventions or continuous interventions exists on datasets. I believe that this comparison seems necessary because it can validate whether the proposed method can be regarded as the general method or not by showing that the proposed method is competitive with the existing methods. However, since the proposed method was validated only for mixed variable interventions, its effectiveness on single intervention (either discrete or continuous) is not clear. * The advantages of the proposed approach seems less clear as compared to the existing baseline. > In appendix, table 3 shows that the baseline approach (Entropy Balancing) outperforms the proposed method in most cases. Therefore, I am skeptical about why the proposed approach is meaningful. Does the proposed method have any other advantages over the baseline approach, that are not shared at current draft ?

Questions

See above Weaknesses.

Rating

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

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

See above Weaknesses.

Reviewer LPz42023-08-14

Thank you for your reply, and the additional analysis for the nonparametric setting. - To your first response: Thank you for attempting to answer my question. I didn't intend to focus on the specific setting with independence between " $Y$ " and " $X$ and $Z$ ". I was referring to settings such that the estimand (the target causal effect) is a scalar, like average treatment effect (ATE). For instance, many weighted estimators for ATE in the literature are $\sqrt{N}$ -consistent, as well as semi-parametric efficient. - To your second response: Thank you very much for extending your analysis to answer my question. In your statement, you didn't specify $\mathbf{x}$ . The convergence seems to be interpreted on a single **fixed** (but artibrary) $\mathbf{x}$ , and the convergence is pointwise. I would note that common guarantees are usually in $L_2$ , empirical $L_2$ , or uniform sense. Currently, I still think the paper has several issues that would benefit from a more substantial revision and so I intend to keep my original rating. I encourage the authors to keep the momentum and look deeper into the theory of the weighted estimator, which is regarded as the target of the current work, in the future.

Authorsrebuttal2023-08-19

We apologize for the delay in responding to your comment. > To your first response: Thank you for attempting to answer my question. I didn't intend to focus on the specific setting with independence between "$Y$" and "$X$" and "$Z$". I was referring to settings such that the estimand (the target causal effect) is a scalar, like average treatment effect (ATE). For instance, many weighted estimators for ATE in the literature are $\sqrt{N}$-consistent, as well as semi-parametric efficient. Thank you for explaning the settings you refered to. We consider that even in semi-parametric settings, the weights would only guarantee the same rates as the nonparametric minimax optimal rates of ATE. > To your second response: Thank you very much for extending your analysis to answer my question. In your statement, you didn't specify $\mathbf{x}$. The convergence seems to be interpreted on a single fixed (but artibrary) $\mathbf{x}$, and the convergence is pointwise. I would note that common guarantees are usually in $L_2$, empirical $L_2$, or uniform sense. Thank you for reviewing the proposition in detail and noting the preferable manners to guarantee convergence rates for ATE. In the case of $L_2$ or uniform sense, the proposition may require additional assumptions. For example, the boundedness of $dQ/dP$ would be sufficient. However, we recognize that these discussions should be done carefully and in detail, which would benefit other researchers. We appreciate this detailed review and constructive comments.

Reviewer bV7F2023-08-15

Re: Rebuttal by Authors

Thank you for the answers. Unfortunately, some of the answers are not clear enough to address the concerns. > We refered to Equation (3.14) in section 3.2.3. Note that in our definition, a probability distribution containing all observed and unobserved variables is considered first, and the causal effect is defined by marginalizing out variables that are not used. The Equation (3.14) in Section 3.2.3 is very different from the definition in Eq. (1) of the paper. Could you write down both equations in the same notation and explain the correspondence precisely? > I believe there is something wrong with Eq. (4) and Eq. (7). The right hand side depends on while the left hand side does not. Let me try to further precise the question: in Eq. (4), the left-hand side represent a probability distribution of $Y$, but the right-hand side represent a probability distribution of $(\boldsymbol{X}, Y)$. So, there must be something wrong. The same goes for Eq. (7). > Eq. (5) in [15] corresponds to Eq. (10) in the present study. > > Note that > in [15] corresponds to in Eq. (10), and in [15] to in Eq. (10), here denotes the sub-differential of (not the partial differential of ), and that in [15] is identical to in Eq. (10). I still see a difference between them: Eq. (10) applies $f'(\cdot)$ to $\phi$, which is not present in Eq. (5) in [15]. Where does this difference come from?

Authorsrebuttal2023-08-19

We thank the reviewer for the detailed questions regarding this study. We apologize for the delay in responding to your comment. > The Equation (3.14) in Section 3.2.3 is very different from the definition in Eq. (1) of the paper. Could you write down both equations in the same notation and explain the correspondence precisely? Herein, to avoid confusion, we use only the notations used in [15] for the following discussions, except for using the bold font type to write sets of variables and their values. In particular, $D\_{A}$ denotes the domain of $A$, which is the same as $\mathcal{X}_A$ in us. First, please note that (3.14) in [15] defines the causal effects determined by the following $do$ operation: $$ \begin{aligned} &P(x_1,\ldots,x_n|\hat{\mathbf{s}}) \\\\ &= P(X_1=x_1,\ldots,X_n=x_n|do(\mathbf{S}=\mathbf{s})). \end{aligned} $$ Let $\mathbf{PA}\_{\mathbf{S}}= \bigcup\_{X_i \in \mathbf{S}} PA_i$. From Therom 3.2.2 in [15], we have $$ \begin{aligned} & P(x_1,\ldots,x_n|\hat{\mathbf{s}})\\\\ &= \sum\_{\mathbf{pa}\_{\mathbf{s}} \in D\_{\mathbf{PA}\_{\mathbf{S}}}} P(x_1,\ldots,x_n|\mathbf{s}, \mathbf{pa}\_{\mathbf{s}}) \cdot P(\mathbf{pa}\_{\mathbf{s}})\\\\ &= \sum\_{\mathbf{pa}\_{\mathbf{s}} \in D\_{\mathbf{PA}\_{\mathbf{S}}}} \frac{P(x_1,\ldots,x_n, \mathbf{s}, \mathbf{pa}\_{\mathbf{s}})}{P(\mathbf{s} |\mathbf{pa}\_{\mathbf{s}})}. \end{aligned} $$ Here, the last equality is obtained using the following equation: $$ \begin{aligned} &P(x_1,\ldots,x_n|\mathbf{s}, \mathbf{pa}\_{\mathbf{s}}) \cdot P(\mathbf{pa}\_{\mathbf{s}}) =\frac{P(x_1,\ldots,x_n,\mathbf{s}, \mathbf{pa}\_{\mathbf{s}})}{P(\mathbf{s} |\mathbf{pa}\_{\mathbf{s}})}. \end{aligned} $$ Now, we note that our definition considers the general case in which other variables from $\mathbf{X}={X_1, X_2, \ldots, X_n}$ exist. Then, we use the notation $\mathbf{V}$ for all the observed random variables: $\mathbf{X} \subset \mathbf{V}$, and write $\mathbf{V}'$ for $\mathbf{V} \setminus (\mathbf{X} \cup \mathbf{PA}\_{\mathbf{S}} \cup \mathbf{S})$. Subsequently, $$ \begin{aligned} &P(X_1=x_1,\ldots,X_n=x_n, \mathbf{V'}=\mathbf{v}'|do(\mathbf{S}=\mathbf{s}))\\\\ &=P(x_1,\ldots,x_n, \mathbf{V'}=\mathbf{v}'|\hat{\mathbf{s}}) \\\\ &=\sum\_{\mathbf{pa}\_{\mathbf{s}} \in D\_{\mathbf{PA}\_{\mathbf{S}}}} \frac{P(x_1,\ldots,x_n, \mathbf{V'}=\mathbf{v}'|\mathbf{s}, \mathbf{pa}\_{\mathbf{s}})}{P(\mathbf{s} |\mathbf{pa}\_{\mathbf{s}})}. \end{aligned} $$ By marginalizing $\mathbf{V}'$, we obtain $$ \begin{aligned} &P(X_1=x_1,\ldots,X_n=x_n|do(\mathbf{S}=\mathbf{s}))\\\\ &=\sum\_{\mathbf{v'} \in D\_{\mathbf{V}}} P(X_1=x_1,\ldots,X_n=x_n, \mathbf{V'}=\mathbf{v}'|do(\mathbf{S}=\mathbf{s}))\\\\ &=\sum\_{\mathbf{v'} \in D\_{\mathbf{V}}} \sum\_{\mathbf{pa}\_{\mathbf{s}} \in D\_{\mathbf{PA}\_{\mathbf{S}}}} \frac{P(x_1,\ldots,x_n, \mathbf{V'}=\mathbf{v}'|\mathbf{s}, \mathbf{pa}\_{\mathbf{s}})}{P(\mathbf{s} |\mathbf{pa}\_{\mathbf{s}})}. \end{aligned} $$ From this, we see that (1) in the present study coincides with (3.14) in [15]. ($\{X_1, X_2, \ldots, X_n\}$ in the above equality corresponds to $\mathbf{Y}$ in (1), and $\mathbf{S}$ to $\mathbf{X}$.)

Reviewer bV7F2023-08-20

Definition 3.1

Thank you for the detailed answer to my question about Definition 3.1. It addressed my concern. I think it is important to mention that the paper adopt Theorem 3.2.2 of Pearl (2009) as the definition.

Authorsrebuttal2023-08-19

We thank the reviewer for the detailed questions regarding this study. We apologize for the delay in responding to your comment. > Let me try to further precise the question: in Eq. (4), the left-hand side represent a probability distribution of $\mathbf{Y}$, but the right-hand side represent a probability distribution of $(\mathbf{X}, Y)$. So, there must be something wrong. The same goes for Eq. (7). We provided the definition of $\overline{do}$ as a probability distribution of $(\mathbf{X}, Y)$. Thus, the left-hand side represents a probability distribution of $(\mathbf{X}, Y)$ not $\mathbf{Y}$. Indeed, $P(\mathbf{Y}| do(\mathbf{X}))$ operation defines a (conditional) probability distribution of $\mathbf{Y}$. However, we believe that this manner has the disadvantages described in line 115-121: $do(\mathbf{X})$ operation cannot represent the intervention of the variables that preserve functional relationships. To consider operations that preserve the functional relationship, we believe that the definition must define a probability distribution for all variables rather than a conditional probability on variables of intervention.

Reviewer bV7F2023-08-20

Definition of $P(Y | do(X))$

If the authors intend to define $P(Y | \overline{do}(\boldsymbol{X}))$ as a probability distribution of $(\boldsymbol{X}, Y)$, there seems there is no problem. However, it is hard for a reader to understand that Deinition 3.2 is overriding even the notation $P(Y | \cdot)$ without any explanation. Also, it is very confusing to use the notation $P(Y | \cdot)$ for a joint distribution. I strongly recommend changing the notation to something clearer.

Authorsrebuttal2023-08-19

We thank the reviewer for the detailed questions regarding this study. We apologize for the delay in responding to your comment. > I still see a difference between them: Eq. (10) applies $f'$ to $\phi$, which is not present in Eq. (5) in [15]. Where does this difference come from? Eq. (5) in [15] considers a general function class for $f$ that contains non-differentiable functions. Therefore, the sub-differential $\partial f$ was used instead of $f'$. As shown in line 175, we assume that $f$ is twice differentiable. We can then consider $f'$ because $\partial f=f'$ holds for any differentiable $f$. We would like to note that $f$ is assumed to be twice differentiable for the strong monotonicity of $f'$, which guarantees the uniqueness of $\phi = dQ/dP$. In addition, the method proposed in [15] considers an optimization of KL divergence that satisfies this assumption. Therefore, we believe that the assumption is reasonable.

Reviewer bV7F2023-08-20

$f'$ in Eq. (10)

I was not asking about the difference between $\partial f$ and $f'$. I simply did not see anything corresponding to $\partial f$ or $f'$ in the Eq. (5) in [15]. However, I finally managed to understand how the authors got Eq. (10). The authors use the fact from Lemma 1: "Equality holds if and only if the subdifferential $\partial \phi(q_0/p_0)$ contains an element of $\mathcal{F}$." I really think the paper should not have omitted this detail.

Reviewer bV7F2023-08-20

My current evaluation of the paper

Overall, the authors follow-up response has addressed my major concerns about the definitions. I would like to raise my score from 3 to 4. Looking at the supplementary material (although I have not read all of it), the authors may have very interesting theoretical results. However, I think the presentation of the paper (especially in the main manuscript) should be improved a lot before it can be published. To me, the main contributions seem to be proving some nice properties of the alpha divergence which are useful for general density ratio estimation. I honestly don't know why the authors decided to limit the scope of the paper to the causal effect estimation, which is just one of many applications of density ratio estimation. I think the causal effect estimation part is overly complicating the notation and making the main messages of the paper hard to understand. I would recommend reorganizing the paper in a way that the reader can understand the main contributions before going into the specific application.

Authorsrebuttal2023-08-20

We thank the reviewer for the detailed and constructive feedback. We appreciate the reviewer's time to understand our study. We will incorporate the reviewer's suggestion regarding the scope of the study.

Reviewer L7FM2023-08-19

Replay to Authors

My apologies for taking so long to reply to your initial response; I was waiting to read the discussion amongst the other reviewers since I am not the most familiar with causal inference. After having read all of the other reviews and corresponding responses, I defer to what I believe the common thread amongst them which is that there needs to be some major revisions in order to better link the theory developed and the setting of interest, as well as more adequately analyze the performance and success of the method actually being used for causal inference. I do think that the theory developed in this work is promising and should continue to be investigated; however, the current iteration seems to require too major of revisions for this specific submission to be viable.

Authorsrebuttal2023-08-19

We thank the reviewer for the detailed and thoughtful review and appreciate the constructive feedback. We will incorporate the reviewer's suggestion regarding the presentation of this study.

Program Chairsdecision2023-09-21

Decision

Reject

© 2026 NYSGPT2525 LLC