Deconfounding Scores and Representation Learning for Causal Effect Estimation with Weak Overlap

Oscar Clivio, Alexander Nicholas D’Amour, Alexander Franks, David Bruns-Smith, Christopher C. Holmes, Avi Feller
Proceedings of The 29th International Conference on Artificial Intelligence and Statistics, PMLR 300:1288-1296, 2026.

Abstract

Overlap, also known as positivity, is a key condition for causal treatment effect estimation. Many popular estimators suffer from high variance and become brittle when features differ strongly across treatment groups. This is especially challenging in high dimensions: the curse of dimensionality can make overlap implausible. To address this, we propose a class of feature representations called deconfounding scores, which preserve both identification and the target of estimation; the classical propensity and prognostic scores are two special cases. We characterize the problem of finding a representation with better overlap as minimizing an overlap divergence under a deconfounding score constraint. We then derive closed-form expressions for a class of deconfounding scores under a broad family of generalized linear models with Gaussian features and show that prognostic scores are overlap-optimal within this class. We conduct extensive experiments to assess this behavior empirically.

Cite this Paper

BibTeX
@InProceedings{pmlr-v300-clivio26a, title = { Deconfounding Scores and Representation Learning for Causal Effect Estimation with Weak Overlap }, author = {Clivio, Oscar and D'Amour, Alexander Nicholas and Franks, Alexander and Bruns-Smith, David and Holmes, Christopher C. and Feller, Avi}, booktitle = {Proceedings of The 29th International Conference on Artificial Intelligence and Statistics}, pages = {1288--1296}, year = {2026}, editor = {Khan, Emtiyaz and Li, Yingzhen and Solin, Arno and Ramdas, Aaditya}, volume = {300}, series = {Proceedings of Machine Learning Research}, month = {02--05 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v300/main/assets/clivio26a/clivio26a.pdf}, url = {https://proceedings.mlr.press/v300/clivio26a.html}, abstract = { Overlap, also known as positivity, is a key condition for causal treatment effect estimation. Many popular estimators suffer from high variance and become brittle when features differ strongly across treatment groups. This is especially challenging in high dimensions: the curse of dimensionality can make overlap implausible. To address this, we propose a class of feature representations called deconfounding scores, which preserve both identification and the target of estimation; the classical propensity and prognostic scores are two special cases. We characterize the problem of finding a representation with better overlap as minimizing an overlap divergence under a deconfounding score constraint. We then derive closed-form expressions for a class of deconfounding scores under a broad family of generalized linear models with Gaussian features and show that prognostic scores are overlap-optimal within this class. We conduct extensive experiments to assess this behavior empirically. } }
Endnote
%0 Conference Paper %T Deconfounding Scores and Representation Learning for Causal Effect Estimation with Weak Overlap %A Oscar Clivio %A Alexander Nicholas D’Amour %A Alexander Franks %A David Bruns-Smith %A Christopher C. Holmes %A Avi Feller %B Proceedings of The 29th International Conference on Artificial Intelligence and Statistics %C Proceedings of Machine Learning Research %D 2026 %E Emtiyaz Khan %E Yingzhen Li %E Arno Solin %E Aaditya Ramdas %F pmlr-v300-clivio26a %I PMLR %P 1288--1296 %U https://proceedings.mlr.press/v300/clivio26a.html %V 300 %X Overlap, also known as positivity, is a key condition for causal treatment effect estimation. Many popular estimators suffer from high variance and become brittle when features differ strongly across treatment groups. This is especially challenging in high dimensions: the curse of dimensionality can make overlap implausible. To address this, we propose a class of feature representations called deconfounding scores, which preserve both identification and the target of estimation; the classical propensity and prognostic scores are two special cases. We characterize the problem of finding a representation with better overlap as minimizing an overlap divergence under a deconfounding score constraint. We then derive closed-form expressions for a class of deconfounding scores under a broad family of generalized linear models with Gaussian features and show that prognostic scores are overlap-optimal within this class. We conduct extensive experiments to assess this behavior empirically.
APA
Clivio, O., D’Amour, A.N., Franks, A., Bruns-Smith, D., Holmes, C.C. & Feller, A.. (2026). Deconfounding Scores and Representation Learning for Causal Effect Estimation with Weak Overlap . Proceedings of The 29th International Conference on Artificial Intelligence and Statistics, in Proceedings of Machine Learning Research 300:1288-1296 Available from https://proceedings.mlr.press/v300/clivio26a.html.

Related Material