Counterfactual Outcome Estimation in Time Series via Sub-treatment Group Alignment and Random Temporal Masking

Yiling Liu, Juncheng Dong, Chen Fu, Wei Shi, Ziyang Jiang, Zhigang Hua, Bo Long, David Carlson

Keywords: Counterfactual treatment effect estimation, Time series observational data, Confounding in time series, Sub-treatment Group Alignment, Random Temporal Masking

Abstract: Estimating counterfactual outcomes in time series from observational data is important for effective decision-making in many fields, such as determining the optimal timing for a medical intervention. However, this task is challenging, primarily because of the unobservability of counterfactual outcomes and the complexity of confounding in time series. To this end, we introduce a representation learning-based framework for counterfactual estimation in time series with two novel techniques: **Sub-treatment Group Alignment (SGA)** and **Random Temporal Masking (RTM)**. The first technique focuses on reducing confounding at each time point. While the common approach is to align the distributions of different treatment groups in the latent space, our proposed approach, SGA, first identifies *sub-treatment groups* through Gaussian Mixture Models (GMMs) and subsequently aligns the corresponding sub-groups. We demonstrate that, both theoretically and empirically, SGA achieves improved alignment, thus leading to more effective deconfounding. The second technique, RTM, masks covariates at random time steps with Gaussian noises. This approach promotes the time series models to select information not only important for the outcome estimation at current time point but also crucial for the time points in the future where the covariates are masked out, thus preserving the *causal information* and reducing the risk of overfitting to factual outcomes. We observe in experiments on synthetic and semi-synthetic datasets that applying SGA and RTM individually improves counterfactual outcome estimation, and when combined, they achieve state-of-the-art performance.

Primary Area: causal reasoning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10734