pith. sign in

arxiv: 2605.19006 · v1 · pith:DAQNUL6Vnew · submitted 2026-05-18 · 📊 stat.ME · stat.ML

Causal Inference with Categorical Unobserved Confounder via Mixture Learning

Aytijhya Saha , Stephen Bates , Devavrat Shah This is my paper

Pith reviewed 2026-05-20 07:53 UTC · model grok-4.3

classification 📊 stat.ME stat.ML
keywords causal inferenceunobserved confoundingproximal causal inferencedeconfoundermixture modelstensor decompositioncategorical variablesidentifiability
0
0 comments X

The pith

Causal effects become identifiable from proxies or multiple treatments when the unobserved confounder is categorical.

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

The paper establishes that causal effects can be identified even when an unobserved confounder is present, provided that confounder takes only finitely many values. It unifies the proxy-variable approach and the multiple-treatment approach by treating the observed data as arising from a mixture whose components correspond to the different values of the confounder. Recovering those mixture components through tensor decomposition yields the latent structure and therefore the causal effect. The resulting estimators are consistent and come with non-asymptotic guarantees. A reader cares because the result removes the need to assign asymmetric roles to proxy variables and supplies theoretical support for the deconfounder method under a discrete confounder.

Core claim

Causal effects are identifiable in both the proximal causal inference setting with proxies and the deconfounder setting with multiple treatments when the unobserved confounder is categorical. The confounding structure is recovered by identifying the corresponding mixture distribution over the finite support of the confounder, using tensor decomposition for consistent estimation with non-asymptotic guarantees.

What carries the argument

The mixture distribution induced by the categorical unobserved confounder, recovered via tensor decomposition of the observed joint distributions under conditional independence.

If this is right

  • Consistent recovery of the latent confounding structure from observed data alone.
  • Non-asymptotic error bounds on the estimated causal effects.
  • Usable estimators for both proxy-variable and multiple-treatment designs.
  • Empirical performance that holds with limited sample sizes in simulations and real data.

Where Pith is reading between the lines

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

  • The same mixture-recovery step could be tested on problems where the number of confounder categories is known but larger than in the current experiments.
  • Extensions might explore how misspecification of the number of categories affects the recovered causal effects.
  • The tensor-decomposition route may link to other latent-variable causal models that rely on discreteness for identification.

Load-bearing premise

The unobserved confounder has only finitely many possible values and the observed variables obey the conditional independence relations that make the resulting mixture distribution identifiable from tensor decomposition.

What would settle it

Apply the procedure to data generated from a continuous unobserved confounder and check whether the estimated causal effect converges to the true value or instead exhibits persistent bias as sample size grows.

Figures

Figures reproduced from arXiv: 2605.19006 by Aytijhya Saha, Devavrat Shah, Stephen Bates.

Figure 1
Figure 1. Figure 1: All the proxies (Z1, Z2, Z3) play a symmetric role in our DAG, but not in the existing proximal causal inference framework. U Z1 Z2 Z3 X A Y (a) U Z1 Z2 Z3 A Y (b) U Z1 Z2 Z3 A Y (c) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Variations of the model in Fig. 1b that are accommodated by our framework but are not addressed by existing proximal causal inference methods. 1. Multi-proxy setting. We consider a causal inference framework with multiple proxies where classification of proxy variables is not required. Instead, we need at least three proxy variables that are conditionally independent given the latent confounder. Specifical… view at source ↗
Figure 3
Figure 3. Figure 3: Our multi-treatment DAG Contributions. In summary, our key contributions are: 4 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Results with 100 independent runs. In the vertical axis, we plot [PITH_FULL_IMAGE:figures/full_fig_p026_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results with 100 independent runs. In the vertical axis, we plot [PITH_FULL_IMAGE:figures/full_fig_p027_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Empirical density of different proxies across the recovered latent classes, where the [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗
read the original abstract

Unobserved confounding is a fundamental challenge for estimating causal effects. To address unobserved confounding, recent literature has turned to two different approaches -- proxy variables and the use of multiple treatments. The first approach, commonly referred to as proximal causal inference, requires proxies to be assigned to specific asymmetric roles: treatment-inducing proxies (negative control exposures), variables that act as common causes of the treatment and outcome, and outcome-inducing proxies (negative control outcomes). In practice, however, identifying variables that satisfy these asymmetric roles can be difficult depending on the application domain. The second approach, commonly referred to as the ``Deconfounder," deals with multiple conditionally independent treatments. There has been limited progress towards developing a consistent estimation method for this setting. As the primary contribution of this work, we establish that causal effects are identifiable in both settings when the unobserved confounder is categorical under suitable conditions. Our approach builds on a mixture learning perspective: we show that the underlying confounding structure can be recovered by identifying the corresponding mixture distribution. We propose an estimation procedure based on tensor decomposition, which allows consistent recovery of the latent structure and comes with non-asymptotic guarantees. Simulation studies and real data experiments demonstrate that the proposed method performs well even with limited data.

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

2 major / 2 minor

Summary. The paper claims that causal effects are identifiable when the unobserved confounder is categorical (finite support) in both proximal causal inference (with treatment- and outcome-inducing proxies) and multiple conditionally independent treatments settings. It recovers the confounding structure by identifying the corresponding mixture distribution over observed variables via tensor decomposition, proposes a consistent estimator with non-asymptotic guarantees, and supports the claims with simulation studies and real-data experiments.

Significance. If the identifiability and recovery results hold under the stated conditions, the work offers a useful unification of proxy-variable and deconfounder approaches for categorical confounders, with a practical tensor-decomposition procedure that avoids direct fitting of the target causal quantity. The non-asymptotic guarantees and empirical validation on limited data are explicit strengths that could aid applications where the confounder has small finite support.

major comments (2)
  1. [§3] §3 (Identifiability results): The central claim that the categorical confounder distribution is uniquely recoverable via tensor decomposition of the observed mixture requires Kruskal-rank or full-column-rank conditions on the factor matrices (conditional distributions given each confounder level). The manuscript states conditional independence and finite support but does not explicitly state or verify these rank conditions for the proximal or multiple-treatment settings; without them, multiple decompositions can yield the same marginals, undermining unique identification of the latent structure and thus the causal effect.
  2. [§4] §4 (Estimation and guarantees): Theorem 4.1 (or equivalent non-asymptotic bound) assumes the observed conditional probability tensors admit unique decomposition, yet the minimal support size or number of observed variables needed to satisfy the rank conditions for K>2 confounder categories is not derived or checked; this is load-bearing because the guarantees reduce to standard tensor decomposition results only when those conditions hold.
minor comments (2)
  1. [Abstract and §1] The abstract and introduction refer to 'suitable conditions' without a compact summary table or list of the precise rank and support requirements; adding this would improve readability.
  2. [§2] Notation for the mixture components and tensor unfoldings is introduced gradually; a single notation table or diagram in §2 would clarify the mapping from observed variables to the three-way tensor.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the conditions underlying our identifiability and estimation results. We address each major comment below.

read point-by-point responses

  1. Referee: [§3] §3 (Identifiability results): The central claim that the categorical confounder distribution is uniquely recoverable via tensor decomposition of the observed mixture requires Kruskal-rank or full-column-rank conditions on the factor matrices (conditional distributions given each confounder level). The manuscript states conditional independence and finite support but does not explicitly state or verify these rank conditions for the proximal or multiple-treatment settings; without them, multiple decompositions can yield the same marginals, undermining unique identification of the latent structure and thus the causal effect.

    Authors: We agree that unique recovery via tensor decomposition requires the Kruskal-rank (or full column rank) conditions on the factor matrices A, B, C corresponding to the conditional distributions given each level of the categorical confounder U. These conditions are implicitly part of the 'suitable conditions' referenced in Section 3, but we acknowledge they were not stated explicitly for the proximal and multiple-treatment settings. In the revised manuscript we will add an explicit statement of the required rank conditions (e.g., k_A + k_B + k_C ≥ 2K + 2) together with a short verification showing that conditional independence plus sufficient observed categories implies the necessary rank properties for both settings. revision: yes

  2. Referee: [§4] §4 (Estimation and guarantees): Theorem 4.1 (or equivalent non-asymptotic bound) assumes the observed conditional probability tensors admit unique decomposition, yet the minimal support size or number of observed variables needed to satisfy the rank conditions for K>2 confounder categories is not derived or checked; this is load-bearing because the guarantees reduce to standard tensor decomposition results only when those conditions hold.

    Authors: We concur that the non-asymptotic guarantees in Theorem 4.1 are valid only when the rank conditions for unique decomposition are satisfied, and that the minimal number of observed categories (or variables) required for K > 2 should be made explicit. In the revision we will insert a corollary (or remark) that translates the Kruskal-rank condition into concrete requirements on the support sizes of the observed variables for both the proximal and multiple-treatment cases, thereby clarifying when the tensor-decomposition guarantees apply. revision: yes

Circularity Check

0 steps flagged

No significant circularity; identifiability derived from standard tensor decomposition of mixtures

full rationale

The paper's central claim establishes identifiability of causal effects for a categorical unobserved confounder by recovering the mixture distribution via tensor decomposition of observed conditional probability tensors. This relies on external algebraic results for unique decomposition under rank conditions (Kruskal rank or full column rank on factor matrices), which are independent of the target causal quantity and not obtained by fitting or self-definition within the paper. The estimation procedure is presented as consistent recovery with non-asymptotic guarantees, without reducing any prediction or effect estimate to a fitted parameter by construction. No load-bearing self-citation chains or ansatz smuggling appear in the provided derivation outline; the approach is self-contained against external benchmarks for latent variable identifiability.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The approach rests on the assumption that the confounder is categorical and that standard conditional independence relations hold so that the observed joint distribution factors as a mixture.

free parameters (1)
  • number of categories of the unobserved confounder
    Must be known or selected to define the mixture rank; not stated as estimated from data in the abstract.
axioms (2)
  • domain assumption Unobserved confounder has finite categorical support
    Required for the mixture representation and tensor decomposition to recover the latent structure.
  • domain assumption Conditional independence of treatments/proxies given the confounder
    Standard in both proximal and deconfounder settings; invoked to make the mixture identifiable.

pith-pipeline@v0.9.0 · 5747 in / 1384 out tokens · 39317 ms · 2026-05-20T07:53:07.445850+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

42 extracted references · 42 canonical work pages · 2 internal anchors

  1. [1]

    Biometrika , volume=

    The central role of the propensity score in observational studies for causal effects , author=. Biometrika , volume=. 1983 , publisher=

    work page 1983

  2. [2]

    Conference on learning theory , pages=

    A method of moments for mixture models and hidden Markov models , author=. Conference on learning theory , pages=. 2012 , organization=

    work page 2012

  3. [3]

    Biometrika , volume=

    Identifying causal effects with proxy variables of an unmeasured confounder , author=. Biometrika , volume=. 2018 , publisher=

    work page 2018

  4. [4]

    arXiv preprint arXiv:2009.10982 , year=

    An introduction to proximal causal learning , author=. arXiv preprint arXiv:2009.10982 , year=

    work page arXiv 2009

  5. [5]

    arXiv preprint arXiv:1810.00283 , year=

    Proxy variables and the identification of causal effects , author=. arXiv preprint arXiv:1810.00283 , year=

    work page arXiv

  6. [6]

    Journal of Machine Learning Research , volume=

    Tensor decompositions for learning latent variable models , author=. Journal of Machine Learning Research , volume=

    work page

  7. [7]

    The Annals of Statistics , volume=

    Identifiability of parameters in latent structure models with many observed variables , author=. The Annals of Statistics , volume=. 2009 , publisher=

    work page 2009

  8. [8]

    Advances in Neural Information Processing Systems , volume=

    A spectral algorithm for latent dirichlet allocation , author=. Advances in Neural Information Processing Systems , volume=

    work page

  9. [9]

    Journal of Computer and System Sciences , volume=

    A spectral algorithm for learning hidden Markov models , author=. Journal of Computer and System Sciences , volume=. 2012 , publisher=

    work page 2012

  10. [10]

    Journal of the American Statistical Association , volume=

    The blessings of multiple causes , author=. Journal of the American Statistical Association , volume=. 2019 , publisher=

    work page 2019

  11. [11]

    blessings of multiple causes

    Comment on “blessings of multiple causes” , author=. Journal of the American Statistical Association , volume=. 2019 , publisher=

    work page 2019

  12. [12]

    the blessings of multiple causes

    Counterexamples to" the blessings of multiple causes" by wang and blei , author=. arXiv preprint arXiv:2001.06555 , year=

    work page arXiv 2001

  13. [13]

    Advances in neural information processing systems , volume=

    Causal inference with noisy and missing covariates via matrix factorization , author=. Advances in neural information processing systems , volume=

    work page

  14. [14]

    Advances in Neural Information Processing Systems , volume=

    Causal effect inference with deep latent-variable models , author=. Advances in Neural Information Processing Systems , volume=

    work page

  15. [15]

    Advances in Neural Information Processing Systems , volume=

    Adapting neural networks for the estimation of treatment effects , author=. Advances in Neural Information Processing Systems , volume=

    work page

  16. [16]

    Advances in Neural Information Processing Systems , volume=

    Identifying causal-effect inference failure with uncertainty-aware models , author=. Advances in Neural Information Processing Systems , volume=

    work page

  17. [17]

    Journal of the American statistical association , volume=

    Identification of causal effects using instrumental variables , author=. Journal of the American statistical association , volume=. 1996 , publisher=

    work page 1996

  18. [18]

    Springer Series in Statistics , year=

    Observational studies , author=. Springer Series in Statistics , year=

    work page

  19. [19]

    arXiv preprint arXiv:2205.10777 , year=

    Causal discovery in the presence of unobserved common causes with multi-view data , author=. arXiv preprint arXiv:2205.10777 , year=

    work page arXiv

  20. [20]

    arXiv preprint arXiv:2011.03154 , year=

    Learning latent causal structures with a redundant set of variables , author=. arXiv preprint arXiv:2011.03154 , year=

    work page arXiv 2011

  21. [21]

    Linear algebra and its applications , volume=

    Three-way arrays: rank and uniqueness of trilinear decompositions, with application to arithmetic complexity and statistics , author=. Linear algebra and its applications , volume=. 1977 , publisher=

    work page 1977

  22. [22]

    International conference on machine learning , pages=

    Proximal causal learning with kernels: Two-stage estimation and moment restriction , author=. International conference on machine learning , pages=. 2021 , organization=

    work page 2021

  23. [23]

    International Conference on Machine Learning , pages=

    Nonparametric estimation of multi-view latent variable models , author=. International Conference on Machine Learning , pages=. 2014 , organization=

    work page 2014

  24. [24]

    The Annals of Mathematical Statistics , volume=

    On the identifiability of finite mixtures , author=. The Annals of Mathematical Statistics , volume=. 1968 , publisher=

    work page 1968

  25. [25]

    arXiv preprint arXiv:2006.07691 , year=

    Synthetic interventions , author=. arXiv preprint arXiv:2006.07691 , year=

    work page arXiv 2006

  26. [26]

    Journal of the American Statistical Association , volume=

    Matrix completion methods for causal panel data models , author=. Journal of the American Statistical Association , volume=. 2021 , publisher=

    work page 2021

  27. [27]

    American Economic Review , volume=

    Synthetic difference-in-differences , author=. American Economic Review , volume=. 2021 , publisher=

    work page 2021

  28. [28]

    Journal of Machine Learning Research , volume=

    Robust synthetic control , author=. Journal of Machine Learning Research , volume=

    work page

  29. [29]

    Journal of the American statistical Association , volume=

    Synthetic control methods for comparative case studies: Estimating the effect of California’s tobacco control program , author=. Journal of the American statistical Association , volume=. 2010 , publisher=

    work page 2010

  30. [30]

    The thirty sixth annual conference on learning theory , pages=

    Causal matrix completion , author=. The thirty sixth annual conference on learning theory , pages=. 2023 , organization=

    work page 2023

  31. [31]

    Random design analysis of ridge regression

    An analysis of random design linear regression , author=. arXiv preprint arXiv:1106.2363 , volume=

    work page internal anchor Pith review Pith/arXiv arXiv

  32. [32]

    The Zero Set of a Real Analytic Function

    The zero set of a real analytic function , author=. arXiv preprint arXiv:1512.07276 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [33]

    Introduction to Nonparametric Estimation , pages=

    Nonparametric estimators , author=. Introduction to Nonparametric Estimation , pages=. 2008 , publisher=

    work page 2008

  34. [34]

    Journal of the American Statistical Association , volume=

    On Simpson's paradox and the sure-thing principle , author=. Journal of the American Statistical Association , volume=. 1972 , publisher=

    work page 1972

  35. [35]

    The American Statistician , volume=

    Simpson's paradox in real life , author=. The American Statistician , volume=. 1982 , publisher=

    work page 1982

  36. [36]

    Technical Report , year=

    Simpson's paradox: An anatomy , author=. Technical Report , year=

    work page

  37. [37]

    2009 , edition=

    Causality: Models, Reasoning, and Inference , author=. 2009 , edition=

    work page 2009

  38. [38]

    2018 , publisher=

    High-dimensional probability: An introduction with applications in data science , author=. 2018 , publisher=

    work page 2018

  39. [39]

    2019 , publisher=

    High-dimensional statistics: A non-asymptotic viewpoint , author=. 2019 , publisher=

    work page 2019

  40. [40]

    2010 , publisher=

    Econometric analysis of cross section and panel data , author=. 2010 , publisher=

    work page 2010

  41. [41]

    Review of Economics and Statistics , volume=

    Instrumental-variable estimation of count data models: Applications to models of cigarette smoking behavior , author=. Review of Economics and Statistics , volume=. 1997 , publisher=

    work page 1997

  42. [42]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Multiply robust causal inference with double-negative control adjustment for categorical unmeasured confounding , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2020 , publisher=

    work page 2020