pith. sign in

arxiv: 2606.21809 · v1 · pith:52VKLOZMnew · submitted 2026-06-20 · 💻 cs.LG

Causal Gaussian Processes for Robust Treatment Effect Evaluation with Unobserved Confounding

Junzhe Zhang , Jingyuan Chen , Elias Bareinboim This is my paper

Pith reviewed 2026-06-26 12:45 UTC · model grok-4.3

classification 💻 cs.LG
keywords causal inferencegaussian processesunobserved confoundingtreatment effect evaluationcontinuous domainsuniversal discretizationinterventional distributionspolicy evaluation
0
0 comments X

The pith

Causal Gaussian processes can approximate observational and interventional distributions of any causal model with unobserved confounding.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper seeks to enable evaluation of causal effects for continuous treatments and outcomes when unobserved confounding is present, relying only on the temporal order between treatment and outcome. It establishes that any causal model can be approximated to arbitrary accuracy by a universal discretization of the exogenous domains into a finite set of latent states. This approximation property supports the construction of a family of Causal Gaussian Process models that capture both observational and interventional behavior. A reader would care because prior robust methods demand strong prior knowledge or restrict to discrete variables, while this approach targets the continuous case common in applications.

Core claim

The authors introduce a universal discretization of the exogenous domains that approximates the observational and interventional distributions of any causal model with arbitrary accuracy using a finite number of latent states. Building on this newfound universal approximation property, they develop a novel family of Causal Gaussian process (CGP) models that effectively approximate the observational and interventional distributions of any causal model with confounded observations.

What carries the argument

universal discretization of the exogenous domains into a finite number of latent states, which enables Causal Gaussian process (CGP) models to approximate any causal model

If this is right

  • Causal effect evaluation is possible over continuous domains from confounded observations.
  • Only basic temporal ordering between treatment and outcome is needed.
  • Any causal model can be approximated with arbitrary accuracy.
  • Robust evaluation works without detailed prior knowledge of the environment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The discretization approach could be combined with other function approximators beyond Gaussian processes for similar causal tasks.
  • Practical tests could compare CGP-based estimates against ground-truth interventions in simulated continuous confounded systems.
  • The method opens a route to handling mixed continuous-discrete variables by extending the same discretization idea.

Load-bearing premise

There exists a universal discretization of the exogenous domains that approximates the observational and interventional distributions of any causal model with arbitrary accuracy using a finite number of latent states.

What would settle it

A concrete causal model with continuous exogenous variables for which no finite discretization of those variables can make the induced observational and interventional distributions arbitrarily close to the true ones.

Figures

Figures reproduced from arXiv: 2606.21809 by Elias Bareinboim, Jingyuan Chen, Junzhe Zhang.

Figure 1
Figure 1. Figure 1: Samples drawn from the observational P(X, Y ) (blue) and interventional Px(Y ) (orange) distributions defined by various reward functions. These functions include: (a) polynomial function; (b) logistic function; (c) phase function; and (d) linear function. The regression function is obtained by applying a Gaussian Process model to the observed data. gives a different evaluation on the treatment effects Ex … view at source ↗
Figure 2
Figure 2. Figure 2: Causal diagrams of (a) a contextual bandit model [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The illustration of the confounding-robust infer [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: A simple function ap￾proximating the reward func￾tion fY (x, u) in the ground-truth causal model M of Example 1. The observed trajectories and their approximations are high￾lighted in blue and orange. Among the above equations, the first condition in fbY (x, u) ensures that the outcome variable Yb(u) in the canonical model Mc effectively approximates the observation Y (u) in the causal model M. It follows … view at source ↗
Figure 5
Figure 5. Figure 5: (a) Stratified observed data based on the assigned functional types; (b, c) posteriors over selected canonical [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Simulations comparing the derived posterior approximations over various reward functions using our proposed [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
read the original abstract

The presence of confounding bias poses a key challenge in policy evaluation, as the target causal effects of actions are not identifiable (i.e., underdetermined) from observational data. On the other hand, existing confounding-robust evaluation strategies require detailed prior knowledge about the environment or apply only to discrete treatments and outcomes. This paper investigates causal effect evaluation over the continuous domain from confounded observations, while requiring only basic temporal ordering between the treatment and the outcome. We introduce a universal discretization of the exogenous domains that approximates the observational and interventional distributions of any causal model with arbitrary accuracy using a finite number of latent states. Building on this newfound universal approximation property, we develop a novel family of Causal Gaussian process (CGP) models that effectively approximate the observational and interventional distributions of any causal model with confounded observations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The manuscript claims to introduce a universal discretization of exogenous domains that approximates the observational and interventional distributions of any causal model to arbitrary accuracy using a finite number of latent states. It then develops a family of Causal Gaussian Process (CGP) models that leverage this property to approximate those distributions from confounded observations, enabling robust causal effect evaluation over continuous treatment and outcome domains while requiring only temporal ordering between treatment and outcome.

Significance. If the universal approximation property holds, the CGP framework would represent a notable contribution to causal inference by providing a nonparametric approach to continuous-domain treatment effect estimation under unobserved confounding, extending beyond existing methods limited to discrete variables or requiring strong prior knowledge. The approach could enable more flexible modeling in policy evaluation settings.

major comments (1)
  1. [Abstract] Abstract: The claim that there exists a 'universal discretization of the exogenous domains' that approximates 'the observational and interventional distributions of any causal model with arbitrary accuracy using a finite number of latent states' is load-bearing for the entire contribution. This property cannot hold for arbitrary causal models without additional uniformity conditions (e.g., compact support of all exogenous variables or uniform Lipschitz bounds on structural equations across the model class) that are not stated; without them the required resolution cannot be bounded independently of the specific model, as the discretization would need to adapt to arbitrarily differing noise supports or function moduli.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful reading and valuable feedback. We address the single major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The claim that there exists a 'universal discretization of the exogenous domains' that approximates 'the observational and interventional distributions of any causal model with arbitrary accuracy using a finite number of latent states' is load-bearing for the entire contribution. This property cannot hold for arbitrary causal models without additional uniformity conditions (e.g., compact support of all exogenous variables or uniform Lipschitz bounds on structural equations across the model class) that are not stated; without them the required resolution cannot be bounded independently of the specific model, as the discretization would need to adapt to arbitrarily differing noise supports or function moduli.

    Authors: We agree that the stated claim requires regularity conditions to hold uniformly. Our discretization construction relies on the exogenous variables having compact support and the structural equations satisfying a uniform Lipschitz bound over the model class; these ensure a finite partition suffices for arbitrary accuracy independent of any particular model. These conditions are standard for obtaining uniform approximation guarantees in nonparametric settings but were not explicitly listed in the abstract. In the revision we will update the abstract, the introduction, and the formal statement of the universal discretization result to make the assumptions explicit, thereby qualifying the claim to apply to causal models satisfying them. The remainder of the CGP development and the empirical results are unaffected, as they operate within this class. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation introduces new discretization property and builds CGP models independently

full rationale

The paper states a universal discretization of exogenous domains as a new property that approximates any causal model's observational and interventional distributions to arbitrary accuracy with finite latent states, then develops CGP models on this property. No equations or text reduce a prediction or result to a fitted input, self-definition, or self-citation chain by construction. The central approximation claim is presented as a foundational result rather than derived from prior fitted parameters or renamed known results within the paper. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the existence of a universal discretization property and introduces CGP as a new modeling family; no free parameters or invented entities with independent evidence are detailed in the abstract.

axioms (1)
  • domain assumption A universal discretization of the exogenous domains approximates the observational and interventional distributions of any causal model with arbitrary accuracy using a finite number of latent states.
    This property is invoked as the foundation for the CGP models in the abstract.
invented entities (1)
  • Causal Gaussian Process (CGP) models no independent evidence
    purpose: Approximate observational and interventional distributions under unobserved confounding.
    New family of models developed in the paper.

pith-pipeline@v0.9.1-grok · 5664 in / 1125 out tokens · 23479 ms · 2026-06-26T12:45:55.085721+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    Pymc: a modern, and comprehensive probabilistic programming framework in python

    Oriol Abril-Pla, Virgile Andreani, Colin Carroll, Larry Dong, Christopher J Fonnesbeck, Maxim Kochurov, Ravin Kumar, Junpeng Lao, Christian C Luhmann, Osvaldo A Martin, et al. Pymc: a modern, and comprehensive probabilistic programming framework in python. PeerJ Computer Science, 9: 0 e1516, 2023

    2023

  2. [2]

    Angrist, G.W

    J.D. Angrist, G.W. Imbens, and D.B. Rubin. Identification of causal effects using instrumental variables (with comments). Journal of the American Statistical Association, 91 0 (434): 0 444--472, 1996

    1996

  3. [3]

    Measure, integration & real analysis

    Sheldon Axler. Measure, integration & real analysis. Springer Nature, 2020

    2020

  4. [4]

    Gaussian process linking functions for mind, brain, and behavior

    Giwon Bahg, Daniel G Evans, Matthew Galdo, and Brandon M Turner. Gaussian process linking functions for mind, brain, and behavior. Proceedings of the National Academy of Sciences, 117 0 (47): 0 29398--29406, 2020

    2020

  5. [5]

    Balke and J

    A. Balke and J. Pearl. Counterfactual probabilities: Computational methods, bounds, and applications. In R. Lopez de Mantaras and D. Poole, editors, Uncertainty in Artificial Intelligence 10, pages 46--54. Morgan Kaufmann, San Mateo, CA, 1994

    1994

  6. [6]

    Balke and J

    A. Balke and J. Pearl. Bounds on treatment effects from studies with imperfect compliance. Journal of the American Statistical Association, 92 0 (439): 0 1172--1176, September 1997

    1997

  7. [7]

    Doubly robust estimation in missing data and causal inference models

    Heejung Bang and James M Robins. Doubly robust estimation in missing data and causal inference models. Biometrics, 61 0 (4): 0 962--973, 2005

    2005

  8. [8]

    E.\ Bareinboim and J. Pearl. Causal inference and the data-fusion problem. Proceedings of the National Academy of Sciences, 113: 0 7345--7352, 2016

    2016

  9. [9]

    On pearl’s hierarchy and the foundations of causal inference

    E Bareinboim, JD Correa, D Ibeling, and T Icard. On pearl’s hierarchy and the foundations of causal inference. ACM Special Volume in Honor of Judea Pearl, 2020. forthcoming. Also, Technical Report R-60, Causal AI Lab, Columbia University, https://causalai.net/r60.pdf https://causalai.net/r60.pdf

    2020

  10. [10]

    Variational inference: A review for statisticians

    David M Blei, Alp Kucukelbir, and Jon D McAuliffe. Variational inference: A review for statisticians. Journal of the American statistical Association, 112 0 (518): 0 859--877, 2017

    2017

  11. [11]

    The international stroke trial (ist): a randomized trial of aspirin, subcutaneous heparin, both, or neither among 19435 patients with acute ischaemic stroke

    A Carolei et al. The international stroke trial (ist): a randomized trial of aspirin, subcutaneous heparin, both, or neither among 19435 patients with acute ischaemic stroke. The Lancet, 349: 0 1569--1581, 1997

    1997

  12. [12]

    Gaussian process regression for materials and molecules

    Volker L Deringer, Albert P Bart \'o k, Noam Bernstein, David M Wilkins, Michele Ceriotti, and G \'a bor Cs \'a nyi. Gaussian process regression for materials and molecules. Chemical Reviews, 121 0 (16): 0 10073--10141, 2021

    2021

  13. [13]

    Doubly robust policy evaluation and learning

    Miroslav Dud \' k, John Langford, and Lihong Li. Doubly robust policy evaluation and learning. In Proceedings of the 28th International Conference on International Conference on Machine Learning, pages 1097--1104. Omnipress, 2011

    2011

  14. [14]

    Frangakis and D.B

    C.E. Frangakis and D.B. Rubin. Principal stratification in causal inference. Biometrics, 1 0 (58): 0 21--29, 2002

    2002

  15. [15]

    Galles and J

    D. Galles and J. Pearl. Axioms of causal relevance. Artificial Intelligence, 97 0 (1-2): 0 9--43, 1997

    1997

  16. [16]

    Stochastic relaxation, gibbs distributions, and the bayesian restoration of images

    Stuart Geman and Donald Geman. Stochastic relaxation, gibbs distributions, and the bayesian restoration of images. IEEE Transactions on pattern analysis and machine intelligence, 0 (6): 0 721--741, 1984

    1984

  17. [17]

    A tutorial on bayesian nonparametric models

    Samuel J Gershman and David M Blei. A tutorial on bayesian nonparametric models. Journal of Mathematical Psychology, 56 0 (1): 0 1--12, 2012

    2012

  18. [18]

    Huang and M

    Y. Huang and M. Valtorta. Pearl's calculus of intervention is complete. In R. Dechter and T.S. Richardson, editors, Proceedings of the Twenty-Second Conference on Uncertainty in Artificial Intelligence, pages 217--224. AUAI Press, Corvallis, OR, 2006

    2006

  19. [19]

    Confounding-robust policy improvement

    Nathan Kallus and Angela Zhou. Confounding-robust policy improvement. In Advances in neural information processing systems, pages 9269--9279, 2018

    2018

  20. [20]

    Auto-Encoding Variational Bayes

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  21. [21]

    Causalgan: Learning causal implicit generative models with adversarial training

    Murat Kocaoglu, Christopher Snyder, Alexandros G Dimakis, and Sriram Vishwanath. Causalgan: Learning causal implicit generative models with adversarial training. In International Conference on Learning Representations, 2018

    2018

  22. [22]

    Automatic differentiation variational inference

    Alp Kucukelbir, Dustin Tran, Rajesh Ranganath, Andrew Gelman, and David M Blei. Automatic differentiation variational inference. Journal of machine learning research, 18 0 (14): 0 1--45, 2017

    2017

  23. [23]

    Toward minimax off-policy value estimation

    Lihong Li, Remi Munos, and Csaba Szepesvari. Toward minimax off-policy value estimation. In Proceedings of the 18th International Conference on Artificial Intelligence and Statistics, May 2015. URL https://www.microsoft.com/en-us/research/publication/toward-minimax-off-policy-value-estimation/

    2015

  24. [24]

    Stein variational gradient descent: A general purpose bayesian inference algorithm

    Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems, 29, 2016

    2016

  25. [25]

    Monotone instrumental variables with an application to the returns to schooling, 1998

    Charles F Manski and John V Pepper. Monotone instrumental variables with an application to the returns to schooling, 1998

    1998

  26. [26]

    Clustering gene expression time series data using an infinite gaussian process mixture model

    Ian C McDowell, Dinesh Manandhar, Christopher M Vockley, Amy K Schmid, Timothy E Reddy, and Barbara E Engelhardt. Clustering gene expression time series data using an infinite gaussian process mixture model. PLoS computational biology, 14 0 (1): 0 e1005896, 2018

    2018

  27. [27]

    Universal kernels

    Charles A Micchelli, Yuesheng Xu, and Haizhang Zhang. Universal kernels. Journal of Machine Learning Research, 7 0 (12), 2006

    2006

  28. [28]

    Safe and efficient off-policy reinforcement learning

    R \'e mi Munos, Tom Stepleton, Anna Harutyunyan, and Marc Bellemare. Safe and efficient off-policy reinforcement learning. In Advances in Neural Information Processing Systems, 2016

    2016

  29. [29]

    Counterfactual identifiability of bijective causal models

    Arash Nasr-Esfahany, Mohammad Alizadeh, and Devavrat Shah. Counterfactual identifiability of bijective causal models. In International Conference on Machine Learning, pages 25733--25754. PMLR, 2023

    2023

  30. [30]

    J. Pearl. Causal diagrams for empirical research. Biometrika, 82 0 (4): 0 669--710, 1995

    1995

  31. [31]

    J. Pearl. Causality: Models, Reasoning, and Inference. Cambridge University Press, NY, 2000. 2nd edition, 2009

    2000

  32. [32]

    Sticking the landing: Simple, lower-variance gradient estimators for variational inference

    Geoffrey Roeder, Yuhuai Wu, and David K Duvenaud. Sticking the landing: Simple, lower-variance gradient estimators for variational inference. Advances in Neural Information Processing Systems, 30, 2017

    2017

  33. [33]

    Rosenbaum and D

    P. Rosenbaum and D. Rubin. The central role of propensity score in observational studies for causal effects. Biometrika, 70: 0 41--55, 1983

    1983

  34. [34]

    Estimating the dimension of a model

    Gideon Schwarz. Estimating the dimension of a model. The annals of statistics, pages 461--464, 1978

    1978

  35. [35]

    A constructive definition of dirichlet priors

    Jayaram Sethuraman. A constructive definition of dirichlet priors. Statistica sinica, pages 639--650, 1994

    1994

  36. [36]

    Gaussian process cosmography

    Arman Shafieloo, Alex G Kim, and Eric V Linder. Gaussian process cosmography. Physical Review D—Particles, Fields, Gravitation, and Cosmology, 85 0 (12): 0 123530, 2012

    2012

  37. [37]

    Shpitser and J

    I. Shpitser and J. Pearl. Identification of joint interventional distributions in recursive semi- M arkovian causal models. In Proceedings of the Twenty-First National Conference on Artificial Intelligence, pages 1219--1226. 2006

    2006

  38. [38]

    Spirtes, C.N

    P. Spirtes, C.N. Glymour, and R. Scheines. Causation, Prediction, and Search. MIT Press, Cambridge, MA, 2nd edition, 2001

    2001

  39. [39]

    Causation, prediction, and search

    Peter Spirtes, Clark N Glymour, and Richard Scheines. Causation, prediction, and search. MIT press, 2000

    2000

  40. [40]

    Real analysis

    Elias M Stein and Rami Shakarchi. Real analysis. Princeton University Press, 2009

    2009

  41. [41]

    Data-efficient off-policy policy evaluation for reinforcement learning

    Philip Thomas and Emma Brunskill. Data-efficient off-policy policy evaluation for reinforcement learning. In International Conference on Machine Learning, pages 2139--2148, 2016

    2016

  42. [42]

    A widely applicable bayesian information criterion

    Sumio Watanabe. A widely applicable bayesian information criterion. The Journal of Machine Learning Research, 14 0 (1): 0 867--897, 2013

    2013

  43. [43]

    Prediction with gaussian processes: From linear regression to linear prediction and beyond

    Christopher KI Williams. Prediction with gaussian processes: From linear regression to linear prediction and beyond. In Learning in graphical models, pages 599--621. Springer, 1998

    1998

  44. [44]

    P.G. Wright. The Tariff on Animal and Vegetable Oils. The MacMillan Company, New York, NY, 1928

    1928

  45. [45]

    The method of path coefficients

    Sewall Wright. The method of path coefficients. The annals of mathematical statistics, 5 0 (3): 0 161--215, 1934

    1934

  46. [46]

    The causal-neural connection: Expressiveness, learnability, and inference

    Kevin Xia, Kai-Zhan Lee, Yoshua Bengio, and Elias Bareinboim. The causal-neural connection: Expressiveness, learnability, and inference. Advances in Neural Information Processing Systems, 34: 0 10823--10836, 2021

    2021

  47. [47]

    Neural causal models for counterfactual identification and estimation

    Kevin Xia, Yushu Pan, and Elias Bareinboim. Neural causal models for counterfactual identification and estimation. In Proceedings of the 11th Eleventh International Conference on Learning Representations, 2022

    2022

  48. [48]

    Bounding causal effects on continuous outcomes

    Junzhe Zhang and Elias Bareinboim. Bounding causal effects on continuous outcomes. In Proceedings of the 35nd AAAI Conference on Artificial Intelligence, 2021

    2021

  49. [49]

    Partial counterfactual identification from observational and experimental data

    Junzhe Zhang, Jin Tian, and Elias Bareinboim. Partial counterfactual identification from observational and experimental data. In International Conference on Machine Learning, pages 26548--26558. PMLR, 2022

    2022