Causal Contrastive Learning for Counterfactual Regression Over Time

Estimating treatment effects over time holds significance in various domains, including precision medicine, epidemiology, economy, and marketing. This paper introduces a unique approach to counterfactual regression over time, emphasizing long-term predictions. Distinguishing itself from existing models like Causal Transformer, our approach highlights the efficacy of employing RNNs for long-term forecasting, complemented by Contrastive Predictive Coding (CPC) and Information Maximization (InfoMax). Emphasizing efficiency, we avoid the need for computationally expensive transformers. Leveraging CPC, our method captures long-term dependencies in the presence of time-varying confounders. Notably, recent models have disregarded the importance of invertible representation, compromising identification assumptions. To remedy this, we employ the InfoMax principle, maximizing a lower bound of mutual information between sequence data and its representation. Our method achieves state-of-the-art counterfactual estimation results using both synthetic and real-world data, marking the pioneering incorporation of Contrastive Predictive Encoding in causal inference.

Paper

References (84)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Nj8s6/10 · confidence 3/52024-07-01

Summary

the paper proposes a method based on CPC to estimate the causal effect of treatments over a period of time

Strengths

- Interesting paper - well written , although some things could have been cleaner - good experimentation and ablation study - interesting use of cpc

Weaknesses

- Not clear from an intuition perspective if the contrastive approach incentivises the model to learn what the authors want - The complexity of the method is significant , raising questions about applicability and reproducibility - The point about invertible processes is not a great one as invertible processes severely limit the function space available without offering much in return. They also increase computational complexity significantly - not the first use of a contrastive objective in causal learning , id suggest that this contribution is toned down a bit

Questions

not many, mostly how susceptible is the method to violations of the assumptions the authors have set out ?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

adequately discussed

Reviewer DSyv6/10 · confidence 4/52024-07-07

Summary

This paper leverages Contrastive Predictive Coding (CPC) with RNN for counterfactual regression over time to provide a compelling alternative to transformer-based approaches (which are challenging to interpret). By leveraging CPC to capture long-term dependencies and InfoMax for "reconstructable" representation, the method achieves state-of-the-art results in counterfactual estimation.

Strengths

- The method seems sound. - The paper is well-written. - Previous work leveraging contrastive learning for causal inference applies only to the static setting with no theoretical grounding. This paper frames the representation balancing problem from an information-theoretic perspective. The authors show that the suggested adversarial game yields theoretically balanced representations using MI's Contrastive Log-ratio Upper Bound (CLUB), computed efficiently. - Achieved good empirical performance with fewer patient data. Good ablation studies.

Weaknesses

- Experiments: performance in random trajectories for the pharmacokinetic-pharmacodynamic model of tumor growth is missing. - While good results are achieved with fewer patient data, it is important to check if the results are consistent as the patient number increases. - Line 96: "*...the role of invertible representation in improving counterfactual regression. Here, we introduce an InfoMax regularization term to make our encoder easier to invert.*" The authors provide some intuitions (lines 204-213) in favor of invertibility (beyond reconstructable) but they are not sufficient. A formal proof is needed. - For a “reconstructable” representation of the process history $H_{t}$, While a lower bound of InfoMax objective ( regularization) is optimized during pre-training of the encoder, there is no guarantee that finetuning of pre-training will retain the desired “reconstructable” property of the representation. Hence, the role of invertible (or reconstructable) representation in improving counterfactual regression is not well understood even though ablation studies show improvement when $\mathcal{L}^{Infomax}$ is introduced.

Questions

- Line 178-179: *$H_{t}$ is a sequence of high-dimensional covariates and the computation of such loss is computationally demanding.* - Line 165-167: *...where at each horizon the discriminator parameter matrix $\Gamma_{j}$ is of a small number of dimensions since being a map between two lower-dimensional representations*. Lower bounds are motivated by the problem of high-dimensionality of $H_{t}$. What is the dimension of $H_{t}, Z_{t}$ and $C_{t}$ in the performed experiments?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

See Weaknesses.

Reviewer 6PYm6/10 · confidence 3/52024-07-10

Summary

The paper introduces a novel algorithm for long-term counterfactual forecasting over time through representation learning of historical information and balanced representation learning that predicts the outcome given balanced treatments. The representation learning process combines Contrastive Predictive Coding and Information Maximization, while the training of balanced representation learning involves an adversarial game in the actual outcome and treatment networks.

Strengths

1. The paper introduces a novel algorithm for predicting counterfactual responses over time which outperforms the baselines in both a synthetic and a semi-synthetic dataset. 2. The proposed algorithm is robust. 3. The paper is well-structured and comprehensive, covering theoretical analysis, experiments, ablation studies, and falsifiability tests. 4. The experiment settings are described in detail. 5. Most of the notation is clear and easy to understand.

Weaknesses

1. Is it possible to conduct experiments on real datasets? 2. It appears there is only one synthetic dataset and one semi-synthetic dataset. In the synthetic experiment, is the standard deviation of the errors calculated from experiments with 5 different seeds instead of random settings of the data generating process? Would it be feasible to conduct experiments on multiple synthetic datasets with random settings to obtain error bars, rather than relying on just one synthetic dataset? 3. There don't seem to be any comparisons of time consumption with baselines. 4. The loss functions are constructed based on the lower or upper bound of the mutual information, but there is no discussion or measurement regarding the tightness of the bound.

Questions

1. In equation 1, $U_{t+j}$ in the numerator has one subscript, while $U_{l,t+j}$ in the denominator has two subscripts. Could you please make them consistent? Is $U_{t+j}$ information for one specific individual or an aggregate for all individuals? Is InfoNCE calculated for each individual or as the expectation over all individuals? Given that in equation 1, the subscript is $\mathcal{B}$, is it the individual index 2. Could you elaborate on how Theorem 5.2 influences the construction of the loss function in Equation 7? Specifically, does Theorem 5.2 impact the formulation or design of the loss function described in Equation 7? The authors discuss what we can infer if equality holds from lines 189 to 192, but will equality actually hold? The description in this paragraph suggests that equality can be achieved, yet Proposition 5.1 and Theorem 5.2 indicate positive gaps between the two mutual information measures without specifying the magnitude of these gaps, and the construction of the contrastive loss does not appear to be influenced by Theorem 5.2, as long as Proposition 5.1 holds. 3. Could you explain what is $\gamma$ in equation 7? Is $\gamma$ a similar definition of horizons? 4. In line 202, is there any proof provided regarding "The theoretical existence of such a function" 5. Since the goal is to estimate the counterfactual responses given in the equation at the bottom of page 3, could you explain why we still need to predict the treatments as the conditional expectation is given $ W_{t+1:t+\tau} $? Is this training needed because we want balanced learning? If so, could you explain why selection bias should be avoided given the goal of learning the expectation conditioning on potential treatments? 6. In Theorem 5.3, could you explain why or when the “if” condition holds? 7. Does the balanced learning only work for $W_{t+1}$? Why is it not necessary to extend it to $W_{t+\tau}$ for $\forall \tau\geq 1$? 8. In section 5.3, is $\theta_{R}=[\theta_1, \theta_2]$?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have discussed the broader societal impacts.

Reviewer PYKS5/10 · confidence 3/52024-07-13

Summary

This paper presents causal CPC, a framework for predicting counterfactual responses under time-varying treatments. The proposed method consists of two components: an encoder that leverages contrastive predictive coding and infomax principle to learn a representation of the history H, and a decoder that leverages a loss function based on adversarial game to predict outcome and treatment with balanced representations to account for selection bias. The concept of utilizing MI-based principles for casual representation learning in the longitudinal setting is interesting, whose efficiency has been demonstrated by the SOTA. However, I believe the paper could benefit from more streamlined writing with a clearer justification of how the use of MI-based principles enables efficient learning of long-term dependency.

Strengths

1. The concept of utilizing MI-based principles for casual representation learning in the longitudinal setting is interesting. 2. The performance of the model has been validated with a number of datasets, including fully-synthetic and semi-synthetic data. 3. Some learning objectives are backed up with theoretical evidence.

Weaknesses

1. Even thought the paper demonstrates the benefits of using CPC and infomax for learning long-term dependency with RNN, it still remains unclear to me, both in intuition and in methodological details, how this benefit was attained. The paper can benefit from some more insights, particularly in section 5.1, that links the proposed information theoretic methods to their actual benefits. The theoretical analysis seems a bit detached from the context and not so helpful in providing insights for the effectiveness of the proposed method. 2. Following on the second point, in the ablation study, it seems that the removal of the InfoNCE loss is non-significant. This seems somewhat curious to me given the amount of emphasis the authors put on the importance of CPC. 3. Some recent work in counterfactual outcome prediction with time-varying treatments should be discussed in the related work: - Berrevoets et al, Disentangled counterfactual recurrent networks for treatment effect inference over time - Chen et al, A Multi-Task Gaussian Process Model for Inferring Time-Varying Treatment Effects in Panel Data - Wu et al, Counterfactual Generative Models for Time-Varying Treatment - Frauen et al, Estimating average causal effects from patient trajectories

Questions

1. How to decide the relative weighting for the encoder and decoder during fine-tuning? 2. Related to Weakness #2, how to better illustrate the role of L_infoNCE given the relatively mild drop in performance after removing it?

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

2

Limitations

Limitations were mentioned in the appendix, but not much in the main text. Adding more discussion of model limitations in the Conclusion section might help.

Reviewer pP4H6/10 · confidence 4/52024-07-14

Summary

This paper proposes a new method for counterfactual outcome prediction over time, with the goal of avoiding highly complicated models and expensive computation while achieving state-of-the-art prediction accuracy for time series with long-range dependencies. To this end, the presented method employs RRNs for long-range forecasting, rather than more complicated network architecture like Transformers, together with contrastive predictive coding (CPC) to capture long-range dependencies in the presence of time varying confounders and information maximization (InfoMax) to help with learning invertible representation for addressing the identification issue. Experiments with synthetic and semi-synthetic data are conducted to demonstrate the efficacy and superiority of the proposed method.

Strengths

1. The aim to develop less complicated neural network models to tackle the challenging problem of counterfactual regression over time. 2. The proposed approach is technically sound and the experiment result looks promising overall. 3. The writing is very effective, and the authors have done a great job on clearly presenting the complicated process/components of their method.

Weaknesses

1. The efficacy of the proposed method for long-range predictions is only demonstrated by experiments to certain extend, and from the results, it is not clear how the method would perform when $\tao$ is greater than 10 steps. In particular, from the middle and right diagrams in Fig. 3, the advantage of the proposed method over the baselines start to drop and some of the baselines start to outperform the proposed method. The paper would be much stronger if some theoretical evidence of the performance of the proposed method could be provided. 2. There are some doubts with the experiment results or performance of the proposed method (please see the Questions section for more details).

Questions

1. Why a mapping of $\textbf{C}_t$, instead of $\textbf{C}_t$ itself is used as input of the decoder part of the proposed architecture? 2. Could you provide any explanation on why the performance advantage of the proposed method starts to drop when $\tao$ becomes bigger (as shown in the middle and right diagrams in Figure 3, and Table 10)? Moreover, why the baselines in general over outperform the proposed method with small horizon predictions? 3. Table 2: the results shown are for the tumor growth data only and $\gamma=1$. How the number of model parameters, training and testing time when different $\gamma$ values are used? 4. Why MSM is not used/reported in all the experiments? Having poor performance may not be a convincing reason for excluding it.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The paper has indicated some limitations of the proposed method, e.g. in Fig. 3 and Section 6.4, but more explicit discussions on the limitations in terms of decreased performance/superiority when the prediction horizon is small or large and when there are unobserved confounders should be given.

Reviewer mdqp6/10 · confidence 2/52024-07-14

Summary

The paper presents a novel approach to counterfactual regression over time, particularly focusing on long-term predictions. It introduces a method that leverages RNNs combined with CPC and InfoMax. This approach aims to capture long-term dependencies in data and improve computational efficiency without relying on complex transformer models.

Strengths

1. Performance: The proposed model achieves superior performance in long-term prediction tasks compared to existing models. 2. Theoretical foundations: The paper provides a thorough theoretical grounding for the proposed method, including a new information-theoretic perspective on representation balancing.

Weaknesses

1. Complexity: While the method avoids transformers, the combination of CPC and InfoMax introduces its own complexity, which may pose challenges for implementation and replication. Also reported in Table 2 and Table 11, where the proposed method requires more training time with fewer parameters compared to the Causal Transformer and G-Net. 2. Scope of evaluation: Although the evaluation is comprehensive, it could be expanded to include more diverse datasets to further validate the robustness and generalizability of the method. For example, 1) what is the total trajectory length in the synthetic experiment? I wonder about the performance and an additional experiment showing the performance under different trajectory lengths would be appreciated. 2) what is the dimensionality of the variables? I wonder about the performance when dealing with the high-dimensional case.

Questions

1. In the problem formulation, the treatment of W is assumed to be discrete, is it possible to extend it to be continuous and would such an extension violate the theoretical result? 2. What is the performance of the approach when dealing with time series under different trajectory lengths and high-dimensional cases? 3. How should the counterfactual treatment be chosen in the experiment? Randomly chosen from the support or do the authors have some rules? 4. Could the authors explain why the proposed method 'excels at large horizon predictions, it does not outperform SOTA models on short-term predictions'?

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors claimed in the Checklist.

Authorsrebuttal2024-08-10

Correction of a typo in rebuttal

In response to Question 1, line 2: "However, the results should be reported as $\Phi(\mathbf{H\_t}) \perp W\_{t+1}$ ...".

Reviewer 6PYm2024-08-12

Thank you for the detailed answers. My concerns have been addressed. I will maintain the score.

Reviewer PYKS2024-08-12

Thanks for the comments. My concerns are addressed and I will raise my score to 5.

Reviewer mdqp2024-08-12

Thank the authors for their clarification and now I have a better understanding of the details. Although I still have some concerns about the higher training time, it is acceptable to me now and I will raise my score. Also it could be better if the authors could polish the figures, as the current font size is somewhat small and difficult to read clearly.

Authorsrebuttal2024-08-12

Thank you very much for your thoughtful feedback and for raising the score. We appreciate your suggestion about the figures and will work specifically on improving them for clarity in the final version.

Reviewer Nj8s2024-08-12

Rebuttal acknowledgement

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

Reviewer DSyv2024-08-12

I acknowledge that I have read the author's rebuttal. I maintain my score of acceptance.

Reviewer pP4H2024-08-13

Thanks for your detailed responses

Thanks the authors for your detailed and helpful responses. I will maintain my positive rating.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC