pith. sign in

arxiv: 2604.18430 · v1 · submitted 2026-04-20 · 📊 stat.ME

Shrinkage through multiple identifiability

Carlos Garc\'ia Meixide , David R\'ios Insua This is my paper

Pith reviewed 2026-05-10 03:35 UTC · model grok-4.3

classification 📊 stat.ME
keywords empirical Bayescausal inferenceidentification functionalsshrinkage estimationmultiple identifiabilityasymptotically linear estimatorsaugmenting RCTs with observational data
0
0 comments X

The pith

An empirical Bayes posterior mean pools estimators from multiple identification functionals to recover a causal effect consistently, even when each is biased if the biases average to zero and their number grows with sample size.

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

The paper develops an empirical Bayes method to aggregate several estimators of the same causal parameter, each coming from a different identification functional. Consistency holds in the exact-identifiability case where every functional targets the identical quantity, and also in a second regime where the functionals are individually biased but the biases average to zero across the collection while the number of functionals increases with the sample size. Dependence among the estimators, which all use the same data, is managed by a working-independence assumption that still delivers consistent point estimates. Inference splits on a latent heterogeneity hyperparameter: it vanishes when all functionals share one target and yields frequentist confidence intervals, but is positive when each functional targets a draw from a mixing distribution and then yields Bayesian prediction intervals for a new functional.

Core claim

We establish consistency of an empirical Bayes aggregator for a scalar causal target in the exact identifiability regime where every functional identifies the same effect, and in the multiple identifiability regime where individual functionals are biased but the identification biases are mean-zero across functionals and the number of functionals grows with sample size. The dependence induced by evaluating all estimators on the same sample is handled through a working independence device that preserves consistency of the point estimator. Inference is organized around a latent heterogeneity hyperparameter: when it vanishes the functionals share a common target and we report frequentist or subs

What carries the argument

The empirical Bayes posterior mean that pools asymptotically linear estimators of the causal target, controlled by a latent heterogeneity hyperparameter that determines whether targets coincide or are drawn from a mixing distribution.

If this is right

  • When the heterogeneity hyperparameter vanishes, sandwich-variance or subsampling intervals are valid for the shared causal target.
  • When the hyperparameter is positive, asymptotically valid Bayesian prediction intervals can be formed for the latent target of a new functional.
  • The framework applies directly to combining randomized controlled trial data with observational evidence for the same causal parameter.
  • The point estimator remains consistent under the working-independence device even though the estimators are dependent through shared data.

Where Pith is reading between the lines

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

  • The same pooling idea could be tested in other settings that supply many alternative identification strategies, such as multiple instruments or proxies whose biases vary but average out.
  • Accounting explicitly for the covariance induced by shared samples, rather than relying on working independence, might improve efficiency without losing consistency.
  • A diagnostic for whether the mean-zero bias condition holds could be constructed by examining the spread of the individual estimators before shrinkage.

Load-bearing premise

The working independence device that preserves consistency of the point estimator despite the dependence induced by evaluating all estimators on the same sample.

What would settle it

A Monte Carlo study in which the identification biases across functionals have a nonzero mean while the number of functionals grows with sample size, checking whether the aggregated estimator converges to the true causal effect.

Figures

Figures reproduced from arXiv: 2604.18430 by Carlos Garc\'ia Meixide, David R\'ios Insua.

Figure 1
Figure 1. Figure 1: Single-run output. Left panel: bootstrap distribution of [PITH_FULL_IMAGE:figures/full_fig_p026_1.png] view at source ↗
read the original abstract

We propose an empirical Bayes framework for aggregating estimators obtained from several identification functionals associated to the same causal parameter. The central object is a posterior mean that pools a collection of asymptotically linear estimators of a scalar causal target. We establish consistency in two non-nested regimes: exact identifiability, in which every functional identifies the same causal effect; and a second regime, in which individual functionals are biased but the identification biases are mean-zero across functionals, and the number of functionals grows with sample size. The dependence induced by evaluating all estimators on the same sample is handled through a working independence device that preserves consistency of the point estimator. Inference is organized around a latent heterogeneity hyperparameter: when it vanishes, the functionals share a common target and we report frequentist confidence intervals for that target via a sandwich variance or subsampling; when it is strictly positive, each functional targets a genuine draw from a mixing distribution and we construct asymptotically valid Bayesian prediction intervals for the latent target of a new functional. The two inferential outputs rest on distinct assumption sets and are, therefore, complementary rather than exclusive. We illustrate the framework in the context of augmenting randomized controlled trials with observational evidence.

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 / 1 minor

Summary. The paper proposes an empirical Bayes framework for aggregating estimators obtained from several identification functionals associated to the same causal parameter. The central object is a posterior mean that pools a collection of asymptotically linear estimators of a scalar causal target. It establishes consistency in two non-nested regimes: exact identifiability, in which every functional identifies the same causal effect; and a second regime, in which individual functionals are biased but the identification biases are mean-zero across functionals, and the number of functionals grows with sample size. The dependence induced by evaluating all estimators on the same sample is handled through a working independence device that preserves consistency of the point estimator. Inference is organized around a latent heterogeneity hyperparameter: when it vanishes, the functionals share a common target and frequentist confidence intervals are reported via sandwich variance or subsampling; when it is strictly positive, each functional targets a genuine draw from a mixing distribution and asymptotically valid Bayesian prediction intervals are constructed for the latent target of a new functional. The framework is illustrated in the context of augmenting randomized controlled trials with observational evidence.

Significance. If the consistency results hold under the stated regimes, particularly the mean-zero bias regime with growing functionals, the framework would offer a principled way to pool causal estimators while adapting inference to the presence or absence of heterogeneity via the latent hyperparameter. The complementary frequentist and Bayesian outputs, along with the application to RCT-observational augmentation, could be useful for meta-analytic settings in causal inference where multiple identification strategies are available.

major comments (2)
  1. [Abstract (second regime) and the section developing the working independence device] The consistency claim in the second regime (mean-zero identification biases with m_n functionals, m_n → ∞) rests on the working independence device for the collection of asymptotically linear estimators. Because every estimator is evaluated on the identical sample, the influence functions are jointly dependent and this dependence does not vanish uniformly as m_n grows. The manuscript must show explicitly that the device controls the cumulative effect of the off-diagonal covariances on the shrinkage weights and bias term so that the posterior mean has no asymptotic bias larger than o_p(1); without a detailed argument or proof sketch addressing this, the central claim remains difficult to assess.
  2. [Inference organization around the latent heterogeneity hyperparameter] In the mean-zero bias regime, the interaction between the estimated latent heterogeneity hyperparameter and the consistency of the point estimator is not fully specified. It is unclear whether the hyperparameter estimation preserves the o_p(1) bias property of the posterior mean or affects the validity of the subsequent Bayesian prediction intervals.
minor comments (1)
  1. [Abstract] The abstract introduces the 'mixing distribution over latent targets' without a brief definition or forward reference; adding one sentence would improve accessibility for readers outside empirical Bayes literature.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address the two major comments below and will incorporate revisions to clarify the technical arguments.

read point-by-point responses

  1. Referee: [Abstract (second regime) and the section developing the working independence device] The consistency claim in the second regime (mean-zero identification biases with m_n functionals, m_n → ∞) rests on the working independence device for the collection of asymptotically linear estimators. Because every estimator is evaluated on the identical sample, the influence functions are jointly dependent and this dependence does not vanish uniformly as m_n grows. The manuscript must show explicitly that the device controls the cumulative effect of the off-diagonal covariances on the shrinkage weights and bias term so that the posterior mean has no asymptotic bias larger than o_p(1); without a detailed argument or proof sketch addressing this, the central claim remains difficult to assess.

    Authors: We agree that an explicit argument is required. The working independence device is introduced precisely to handle the joint dependence while preserving the o_p(1) consistency of the posterior mean under the mean-zero bias regime. In the revision we will add a detailed proof sketch in the relevant section (and reference it from the abstract) that bounds the contribution of the off-diagonal covariance terms to both the shrinkage weights and the bias of the aggregated estimator, showing that these terms remain o_p(1) when m_n grows at the stated rate and the mean-zero condition holds. This will make the control of cumulative dependence fully transparent. revision: yes

  2. Referee: [Inference organization around the latent heterogeneity hyperparameter] In the mean-zero bias regime, the interaction between the estimated latent heterogeneity hyperparameter and the consistency of the point estimator is not fully specified. It is unclear whether the hyperparameter estimation preserves the o_p(1) bias property of the posterior mean or affects the validity of the subsequent Bayesian prediction intervals.

    Authors: We appreciate this observation. The hyperparameter estimator is constructed to be consistent under the mean-zero bias regime at a rate that does not disturb the o_p(1) property of the posterior mean. In the revision we will add a short subsection that explicitly derives the joint convergence of the posterior mean and the hyperparameter estimator, confirming that the o_p(1) bias is preserved and that the Bayesian prediction intervals remain asymptotically valid under the same conditions. This will clarify the interaction between the two inferential outputs. revision: yes

Circularity Check

0 steps flagged

No significant circularity; consistency claims rest on stated assumptions rather than self-referential reductions

full rationale

The abstract presents an empirical Bayes framework that aggregates asymptotically linear estimators via a posterior mean, with consistency established in two regimes under explicit assumptions: exact identifiability or mean-zero biases with growing functionals. The dependence from shared samples is addressed by invoking a working independence device as a modeling choice that preserves consistency, without any equation or step reducing the result to a fitted parameter or prior output by construction. No self-citations are invoked for uniqueness theorems, no ansatz is smuggled, and no known result is merely renamed. The latent hyperparameter and complementary inference procedures (frequentist intervals vs. Bayesian prediction intervals) are introduced as distinct assumption sets rather than derived tautologically. Absent any quoted reduction of the form Eq. X = Eq. Y by construction, the derivation chain is self-contained.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 1 invented entities

The proposal rests on standard asymptotic linearity of estimators and introduces a latent heterogeneity hyperparameter plus a working independence assumption; the mean-zero bias condition across functionals is an ad-hoc modeling choice for the second regime.

free parameters (1)
  • latent heterogeneity hyperparameter
    Controls whether functionals share a common target (value zero) or draw from a mixing distribution (positive value); its value determines the inferential output chosen.
axioms (2)
  • domain assumption Estimators are asymptotically linear
    Invoked to establish consistency of the pooled posterior mean under the two regimes.
  • ad hoc to paper Identification biases are mean-zero across functionals as their number grows
    Central modeling assumption for the second non-nested regime.
invented entities (1)
  • mixing distribution over latent targets no independent evidence
    purpose: Models heterogeneity when the hyperparameter is positive, enabling Bayesian prediction intervals for a new functional
    Postulated without independent evidence outside the framework; used to justify the second inferential mode.

pith-pipeline@v0.9.0 · 5503 in / 1507 out tokens · 31534 ms · 2026-05-10T03:35:31.104219+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

45 extracted references · 45 canonical work pages

  1. [1]

    Andrews, D. W. K. (2000). Inconsistency of the bootstrap when a parameter is on the boundary of the parameter space. Econometrica , 68(2):399--405

    work page 2000

  2. [2]

    Angrist, J. D. and Pischke, J.-S. (2009). Mostly Harmless Econometrics: An Empiricist's Companion . Princeton University Press, Princeton, NJ

    work page 2009

  3. [3]

    F., Cand \`e s, E

    Barber, R. F., Cand \`e s, E. J., Ramdas, A., and Tibshirani, R. J. (2021). The limits of distribution-free conditional predictive inference. Information and Inference: A Journal of the IMA , 10(2):455--482

    work page 2021

  4. [4]

    and Reid, N

    Battey, H. and Reid, N. (2024). On the role of parametrization in models with a misspecified nuisance component

    work page 2024

  5. [5]

    J., Klaassen, C

    Bickel, P. J., Klaassen, C. A. J., Ritov, Y., and Wellner, J. A. (1993). Efficient and Adaptive Estimation for Semiparametric Models . Johns Hopkins University Press, Baltimore

    work page 1993

  6. [6]

    Borovskikh, Y. V. (1996). U-Statistics in Banach Spaces . VSP, Utrecht

    work page 1996

  7. [7]

    Boucheron, S., Lugosi, G., and Massart, P. (2013). Concentration Inequalities: A Nonasymptotic Theory of Independence . Oxford University Press, Oxford

    work page 2013

  8. [8]

    U., Gerlinger, C., Harbron, C., Koch, A., Posch, M., Rochon, J., and Schiel, A

    Burger, H. U., Gerlinger, C., Harbron, C., Koch, A., Posch, M., Rochon, J., and Schiel, A. (2021). The use of external controls: To what extent can it currently be recommended? Pharmaceutical Statistics , 20(6):1002--1016

    work page 2021

  9. [9]

    A., Waller, L

    Busgang, S. A., Waller, L. A., Colicino, E., D'Agostino Jr, R., Hertz-Picciotto, I., and Gennings, C. (2022). Selecting external controls for internal cases using stratification score matching methods. International Journal of Environmental Research and Public Health , 19(5):2549

    work page 2022

  10. [10]

    and Sant'Anna, P

    Callaway, B. and Sant'Anna, P. H. C. (2021). Difference-in-differences with multiple time periods. Journal of Econometrics , 225(2):200--230

    work page 2021

  11. [11]

    Chen, S., Zhang, B., and Ye, T. (2021). Minimax rates and adaptivity in combining experimental and observational data

    work page 2021

  12. [12]

    and Kacker, R

    DerSimonian, R. and Kacker, R. (2007). Random-effects model for meta-analysis of clinical trials: an update. Contemporary Clinical Trials , 28(2):105--114

    work page 2007

  13. [13]

    Efron, B. (2014). Two modeling strategies for empirical bayes estimation. Statistical Science , 29(2):285--301

    work page 2014

  14. [14]

    and R \'i os Insua, D

    French, S. and R \'i os Insua, D. (2000). Statistical Decision Theory , volume 9 of Kendall's Library of Statistics . Arnold, London

    work page 2000

  15. [15]

    Galwey, N. W. (2017). Supplementation of a clinical trial by historical control data: Is the prospect of dynamic borrowing an illusion? Statistics in Medicine , 36(6):899--916

    work page 2017

  16. [16]

    and R \'i os Insua, D

    Garc \'i a Meixide, C. and R \'i os Insua, D. (2025). Domain adaptation under hidden confounding. Electronic Journal of Statistics , 19(2):5805--5842

    work page 2025

  17. [17]

    Godambe, V. P. (1960). An optimum property of regular maximum likelihood estimation. The Annals of Mathematical Statistics , 31(4):1208--1211

    work page 1960

  18. [18]

    Gorbach, T., de Luna, X., Karvanen, J., and Waernbaum, I. (2023). Contrasting identifying assumptions of average causal effects: Robustness and semiparametric efficiency. Journal of Machine Learning Research , 24(344):1--67

    work page 2023

  19. [19]

    Hahn, J. (1998). On the role of the propensity score in efficient semiparametric estimation of average treatment effects. Econometrica , 66(2):315--331

    work page 1998

  20. [20]

    R., Carvalho, C

    Hahn, P. R., Carvalho, C. M., Puelz, D., and He, J. (2018). Regularization and confounding in linear regression for treatment effect estimation. Bayesian Analysis , 13(1):163--182

    work page 2018

  21. [21]

    Hedges, L. V. (1983). A random effects model for effect sizes. Psychological bulletin , 93(2):388

    work page 1983

  22. [22]

    Henckel, L., Perkovic, E., and Maathuis, M. H. (2022). Graphical criteria for efficient total effect estimation via adjustment in causal linear models. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , 84(2):579--599

    work page 2022

  23. [23]

    Kojima, M. (2023). Dynamic borrowing method for historical information using a frequentist approach for hybrid control design. arXiv preprint arXiv:2305.14679

    work page arXiv 2023

  24. [24]

    Law, M., Bühlmann, P., and Ritov, Y. (2023). Distributional robustness and transfer learning through empirical bayes

    work page 2023

  25. [25]

    B., Duan, R., and Luedtke, A

    Li, S., Gilbert, P. B., Duan, R., and Luedtke, A. (2025). Data fusion using weakly aligned sources. Journal of the American Statistical Association , pages 1--11

    work page 2025

  26. [26]

    Massart, P. (1990). The tight constant in the Dvoretzky--Kiefer--Wolfowitz inequality. The Annals of Probability , 18(3):1269--1283

    work page 1990

  27. [27]

    Meixide, C. G. and Insua, D. R. (2025). Predictive posteriors under hidden confounding

    work page 2025

  28. [28]

    Meixide, C. G. and van der Laan, M. J. (2025). Causal inference via implied interventions

    work page 2025

  29. [29]

    Mohammad-Taheri, S., Tewari, V., Kapre, R., et al. (2023). Optimal adjustment sets for causal query estimation in partially observed biomolecular networks. Bioinformatics , 39(Supplement 1):i494--i503

    work page 2023

  30. [30]

    Paule, R. C. and Mandel, J. (1982). Consensus values and weighting factors. Journal of Research of the National Bureau of Standards , 87(5):377--385

    work page 1982

  31. [31]

    Peters, J., B \"u hlmann, P., and Meinshausen, N. (2016). Causal inference by using invariant prediction: Identification and confidence intervals. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , 78(5):947--1012

    work page 2016

  32. [32]

    Politis, D. N. and Romano, J. P. (1994). Large sample confidence regions based on subsamples under minimal assumptions. The Annals of Statistics , 22(4):2031--2050

    work page 1994

  33. [33]

    Robins, J. M. and Rotnitzky, A. (1994). Estimation of treatment effects in randomized trials with noncompliance and a dichotomous outcome using structural mean models. Biometrika , 81(4):763--776

    work page 1994

  34. [34]

    Sant'Anna, P. H. C. and Zhao, J. (2020). Doubly robust difference-in-differences estimators. Journal of Econometrics , 219(1):101--122

    work page 2020

  35. [35]

    Schmidli, H., Gsteiger, S., Roychoudhury, S., O'Hagan, A., Spiegelhalter, D., and Neuenschwander, B. (2014). Robust meta-analytic-predictive priors in clinical trials with historical control information. Biometrics , 70(4):1023--1032

    work page 2014

  36. [36]

    A., Thomas, M., Cassidy, A., Weber, S., and Bretz, F

    Schmidli, H., H \"a ring, D. A., Thomas, M., Cassidy, A., Weber, S., and Bretz, F. (2020). Beyond randomized clinical trials: Use of external controls. Clinical Pharmacology & Therapeutics , 107(4):806--816

    work page 2020

  37. [37]

    Tsiatis, A. A. (2006). Semiparametric Theory and Missing Data . Springer Series in Statistics. Springer, 1 edition

    work page 2006

  38. [38]

    van de Geer, S. A. (2000). Empirical Processes in M -Estimation . Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press, Cambridge

    work page 2000

  39. [39]

    van der Vaart, A. W. (1998). Asymptotic Statistics . Cambridge University Press, Cambridge

    work page 1998

  40. [40]

    Viechtbauer, W. (2005). Bias and efficiency of meta-analytic variance estimators in the random-effects model. Journal of Educational and Behavioral Statistics , 30(3):261--293

    work page 2005

  41. [41]

    G., Kinnersley, N., Lindborg, S., et al

    Viele, K., Berry, S., Neuenschwander, B., Amzal, B., Chen, F., Enas, N., Hobbs, B., Ibrahim, J. G., Kinnersley, N., Lindborg, S., et al. (2014). Use of historical control data for assessing treatment effects in clinical trials. Pharmaceutical Statistics , 13(1):41--54

    work page 2014

  42. [42]

    Vovk, V. (2012). Conditional validity of inductive conformal predictors. In Proceedings of the Asian Conference on Machine Learning , volume 25 of Proceedings of Machine Learning Research , pages 475--490. PMLR

    work page 2012

  43. [43]

    Wooldridge, J. M. (2010). Econometric Analysis of Cross Section and Panel Data . MIT Press, Cambridge, MA, 2nd edition

    work page 2010

  44. [44]

    P., and Blei, D

    Wu, B., Salazar, S., Green, D. P., and Blei, D. M. (2026). The illusion of learning from observational data: An empirical bayes perspective

    work page 2026

  45. [45]

    N., and Blei, D

    Wu, B., Weinstein, E. N., and Blei, D. M. (2024). Bayesian empirical B ayes: Simultaneous inference from probabilistic symmetries. arXiv preprint arXiv:2405.09150

    work page arXiv 2024