pith. sign in

arxiv: 2606.21185 · v1 · pith:XH7M42MEnew · submitted 2026-06-19 · 📊 stat.ML · cs.LG· stat.ME

Two Layers of Instability in Causal Estimation

Alexis Bellot This is my paper

Pith reviewed 2026-06-26 12:52 UTC · model grok-4.3

classification 📊 stat.ML cs.LGstat.ME
keywords causal estimationinstabilitydiscontinuitystructural causal modelsinverse propensity weightingregression estimatorsposterior meansdecision theory
0
0 comments X

The pith

Many standard causal point estimators act as discontinuous summaries of multimodal distributions over structural causal models.

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

Even when causal effects are identifiable from observational data, they can still be discontinuous functions of the data distribution. On top of this first layer of instability, the paper identifies a second layer that depends on the choice of estimator. Many common point estimates, including inverse propensity weighted estimators and regression estimators, can be interpreted as selecting a single point from a multimodal distribution over the space of structural causal models consistent with the data. These selections jump discontinuously when the underlying data distribution changes slightly. In contrast, explicit posterior means and medians remain continuous functions of the data distribution.

Core claim

Many standard point estimates can be read as point summaries of multimodal distributions over the space of structural causal models. As such, estimators can jump discontinuously in the data distribution. This defines a taxonomy of estimators that admits a decision-theoretic reading: stability depends on whether the implicit loss function an estimator optimizes is aligned with the causal effect itself. Specifically, inverse propensity weighted estimators and regression estimators are examples of discontinuous summaries, while explicit posterior means and medians are shown to be continuous.

What carries the argument

Multimodal distribution over structural causal models consistent with the observed data, with point summaries that can be discontinuous functions of the data distribution.

If this is right

  • Inverse propensity weighted estimators produce discontinuous causal effect estimates as functions of the data distribution.
  • Regression estimators produce discontinuous causal effect estimates as functions of the data distribution.
  • Explicit posterior means and medians produce continuous causal effect estimates as functions of the data distribution.
  • Estimator stability depends on alignment between the estimator's implicit loss and the target causal effect.

Where Pith is reading between the lines

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

  • Practitioners facing potential multimodality in the space of causal models may prefer posterior means or medians over IPW or regression for smoother behavior under small data perturbations.
  • The same taxonomy could be applied to other point estimators not examined here to classify their continuity properties.
  • Simulation studies could generate sequences of data distributions approaching a discontinuity boundary to measure jump sizes for different estimators.

Load-bearing premise

The space of structural causal models consistent with the observed data distribution induces multimodal distributions over the estimators under consideration.

What would settle it

A pair of arbitrarily close data distributions where an inverse propensity weighted estimator or regression estimator produces markedly different causal effect values while a posterior mean or median does not.

Figures

Figures reproduced from arXiv: 2606.21185 by Alexis Bellot.

Figure 1
Figure 1. Figure 1: Examples of causal diagrams. 2 Why is observational causal inference unstable? We say that an inference problem is unstable if a tiny perturbation in the data, such as that caused by finite-sample noise, can shift the target (i.e., the implied causal effect) from one value to a radically different one. It is well known that certain causal queries are unstable (Robins and Ritov, 1997). A common example is t… view at source ↗
Figure 2
Figure 2. Figure 2: Stability Hierarchy: Corol. 1 and Example 2. The first term on the l.h.s. is the general insta￾bility factor of Thm. 1, the middle term is the instability factor we would obtain by adjusting for the parents of X in a Markov model (i.e., with no unobserved confounding), and the last term is the instability factor we obtain for the corresponding statistical query. Example 2 (Stability Hierarchy) [PITH_FULL_… view at source ↗
Figure 3
Figure 3. Figure 3: Numerical illustration of Examples 3 and 6 with a pair of SCMs with near-equal fit to the data. We follow the setup of Example 3 with A “ 5, k “ 25, n “ 55. Panel (A) shows the model-selective summary (posterior mode) estimate jumping near λ “ 1 2 while the posterior mean varies smoothly in λ. Panel (B) shows a schematic pushforward Ψ#ΠP at the near-tied mixture λ “ 0.49. Panel (C) shows the posterior mean… view at source ↗
read the original abstract

There is a precise sense in which drawing causal inferences from observational data is hard, even when identifiability is assumed. In particular, Robins and Ritov (1997) and Robins et al. (2003) showed that causal effects can be discontinuous as a function of the data distribution: two arbitrarily close data distributions might correspond to different causal effects. This is a fact independent of the choice of estimator; however, not all estimators are equally unstable. Our contribution is to surface a second layer of instability that depends on the choice of estimator. We show that many standard point estimates can be read as point summaries of multimodal distributions over the space of structural causal models. As such, estimators can jump discontinuously in the data distribution. This defines a taxonomy of estimators that admits a decision-theoretic reading: stability depends on whether the implicit loss function an estimator optimizes is aligned with the causal effect itself. Specifically, inverse propensity weighted estimators and regression estimators are examples of discontinuous summaries, while explicit posterior means and medians are shown to be continuous.

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 manuscript claims that causal effect estimation from observational data exhibits two layers of instability. The first layer is intrinsic: causal effects can be discontinuous functions of the data distribution P (citing Robins and Ritov 1997; Robins et al. 2003). The second layer is estimator-dependent: many standard point estimators are point summaries of multimodal distributions over the space of structural causal models (SCMs) consistent with P. This induces discontinuities for estimators such as inverse propensity weighting (IPW) and regression estimators, while explicit posterior means and medians are continuous. The resulting taxonomy admits a decision-theoretic reading in which stability depends on alignment between the estimator's implicit loss and the causal effect itself.

Significance. If the formalization and continuity arguments hold, the work supplies a principled taxonomy of estimators by stability and a decision-theoretic lens on why certain summaries are robust to small perturbations in P. This could inform estimator selection in applied causal inference and clarifies that instability is not solely a property of the identification problem. The explicit contrast between discontinuous summaries (IPW, regression) and continuous ones (posterior mean/median) is a clear contribution.

major comments (2)
  1. [§4] §4, the construction of the multimodal distribution over SCMs: the manuscript must specify the measure on the space of SCMs and how multimodality is induced from the observed P under standard identifiability assumptions; without an explicit construction or example, the claim that IPW corresponds to a discontinuous summary cannot be verified.
  2. [Theorem 5.3] Theorem 5.3 (continuity of posterior median): the argument that the median commutes with limits in the space of distributions over SCMs requires showing that the topology on the SCM space ensures no mode-crossing discontinuities; the current sketch appears to assume this without proof.
minor comments (2)
  1. [Abstract] The abstract states that 'explicit posterior means and medians are shown to be continuous' but does not name the topology or metric on the space of distributions over SCMs; this should be stated in the introduction.
  2. [§2] Notation for the space of SCMs and the induced distribution is introduced informally in §2; a short formal definition would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments. The two major comments identify places where the presentation of the multimodal construction and the continuity argument can be strengthened with additional detail. We address each below.

read point-by-point responses

  1. Referee: [§4] §4, the construction of the multimodal distribution over SCMs: the manuscript must specify the measure on the space of SCMs and how multimodality is induced from the observed P under standard identifiability assumptions; without an explicit construction or example, the claim that IPW corresponds to a discontinuous summary cannot be verified.

    Authors: We agree that an explicit construction would make the argument easier to verify. In the revised manuscript we will add a self-contained example in §4 that defines a finite discrete measure on the space of SCMs (uniform prior over a small collection of linear SCMs that induce the same observational distribution P but different causal effects). This example will show how the posterior is multimodal and why the IPW functional selects a mode that jumps discontinuously while the posterior mean does not. revision: yes

  2. Referee: [Theorem 5.3] Theorem 5.3 (continuity of posterior median): the argument that the median commutes with limits in the space of distributions over SCMs requires showing that the topology on the SCM space ensures no mode-crossing discontinuities; the current sketch appears to assume this without proof.

    Authors: The referee is correct that the current sketch is brief. We will expand the proof of Theorem 5.3 to explicitly verify that, under the weak topology on the space of probability measures over SCMs (induced by the weak topology on the observational distributions), the cumulative distribution function of the causal effect is continuous in the parameter P. This rules out mode-crossing discontinuities for the median under the maintained identifiability assumptions. The revised version will contain the expanded argument. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's first layer of instability is imported from external citations (Robins and Ritov 1997; Robins et al. 2003) and is not derived internally. The second layer is obtained by observing that point estimators are summaries of the induced distribution over SCMs consistent with P, then checking which summaries (means/medians vs. others) commute with limits; this taxonomy follows directly from the algebraic properties of the summary functions and does not reduce to any fitted parameter, self-definition, or self-citation chain. No load-bearing step equates a claimed result to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract introduces no free parameters, axioms, or invented entities; it relies on the prior literature for the first layer of instability and assumes identifiability without adding new postulates.

pith-pipeline@v0.9.1-grok · 5702 in / 1081 out tokens · 24151 ms · 2026-06-26T12:52:55.405564+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

27 extracted references · 5 canonical work pages · 2 internal anchors

  1. [1]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    A General Identification Condition for Causal Effects , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  2. [2]

    2009 , edition=

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

    2009

  3. [3]

    Biometrika , volume=

    Uniform consistency in causal inference , author=. Biometrika , volume=. 2003 , publisher=

    2003

  4. [4]

    Uncertainty in Artificial Intelligence , pages=

    Condition number bounds for causal inference , author=. Uncertainty in Artificial Intelligence , pages=. 2021 , organization=

    2021

  5. [5]

    Proceedings of the Thirty-Second Conference on Uncertainty in Artificial Intelligence , pages=

    Stability of causal inference , author=. Proceedings of the Thirty-Second Conference on Uncertainty in Artificial Intelligence , pages=. 2016 , organization=

    2016

  6. [6]

    arXiv preprint arXiv:1904.02826 , year=

    What can be estimated? Identifiability, estimability, causal inference and ill-posed inverse problems , author=. arXiv preprint arXiv:1904.02826 , year=

    work page arXiv 1904

  7. [7]

    International conference on machine learning , pages=

    Partial counterfactual identification from observational and experimental data , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  8. [8]

    arXiv preprint arXiv:2108.11342 , year=

    Nonparametric identification is not enough, but randomized controlled trials are , author=. arXiv preprint arXiv:2108.11342 , year=

    work page arXiv

  9. [9]

    Statistics in medicine , volume=

    Toward a curse of dimensionality appropriate (CODA) asymptotic theory for semi-parametric models , author=. Statistics in medicine , volume=. 1997 , publisher=

    1997

  10. [10]

    Probabilistic and causal inference: the works of judea pearl , pages=

    On Pearl’s hierarchy and the foundations of causal inference , author=. Probabilistic and causal inference: the works of judea pearl , pages=

  11. [11]

    AAAI , pages=

    Identification of joint interventional distributions in recursive semi-Markovian causal models , author=. AAAI , pages=

  12. [12]

    On the Validity of Covariate Adjustment for Estimating Causal Effects

    On the validity of covariate adjustment for estimating causal effects , author=. arXiv preprint arXiv:1203.3515 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Probabilistic and causal inference: The works of Judea Pearl , pages=

    External validity: From do-calculus to transportability across populations , author=. Probabilistic and causal inference: The works of Judea Pearl , pages=

  14. [14]

    The annals of Statistics , pages=

    Optimal rates of convergence for nonparametric estimators , author=. The annals of Statistics , pages=. 1980 , publisher=

    1980

  15. [15]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Estimating identifiable causal effects through double machine learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  16. [16]

    Electronic Journal of Statistics , volume=

    Towards optimal doubly robust estimation of heterogeneous causal effects , author=. Electronic Journal of Statistics , volume=. 2023 , publisher=

    2023

  17. [17]

    Proceedings of the Tenth Conference on Uncertainty in Artificial Intelligence , pages=

    Counterfactual Probabilities: Computational Methods, Bounds, and Applications , author=. Proceedings of the Tenth Conference on Uncertainty in Artificial Intelligence , pages=

  18. [18]

    The American Economic Review , volume=

    Nonparametric Bounds on Treatment Effects , author=. The American Economic Review , volume=

  19. [19]

    Inverse problems , volume=

    Inverse problems as statistics , author=. Inverse problems , volume=

  20. [20]

    Annals of Mathematics and Artificial Intelligence , volume=

    On the completeness of an identifiability algorithm for semi-Markovian models , author=. Annals of Mathematics and Artificial Intelligence , volume=. 2008 , publisher=

    2008

  21. [21]

    Ill-posed inverse problems in economics , author=. Annu. Rev. Econ. , volume=. 2014 , publisher=

    2014

  22. [22]

    arXiv preprint arXiv:2311.07259 , year=

    Towards bounding causal effects under markov equivalence , author=. arXiv preprint arXiv:2311.07259 , year=

    work page arXiv

  23. [23]

    Advances in Neural Information Processing Systems , volume=

    Partial transportability for domain generalization , author=. Advances in Neural Information Processing Systems , volume=

  24. [24]

    1977 , publisher=

    Solutions of Ill-posed Problems , author=. 1977 , publisher=

    1977

  25. [25]

    Bayesian Analysis , volume=

    Bayesian regression tree models for causal inference: Regularization, confounding, and heterogeneous effects (with discussion) , author=. Bayesian Analysis , volume=. 2020 , publisher=

    2020

  26. [26]

    Journal of Computational and Graphical Statistics , volume=

    Bayesian nonparametric modeling for causal inference , author=. Journal of Computational and Graphical Statistics , volume=. 2011 , publisher=

    2011

  27. [27]

    A Bayesian Approach to Constraint Based Causal Inference

    A Bayesian approach to constraint based causal inference , author=. arXiv preprint arXiv:1210.4866 , year=

    work page internal anchor Pith review Pith/arXiv arXiv