Pith. sign in

REVIEW 3 major objections 5 minor 39 references

A Bayesian method shows that observational studies can only sharpen a randomized trial's treatment-effect estimates up to a ceiling set by the prior on the observational bias.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 08:41 UTC pith:ALOUF347

load-bearing objection Solid Bayesian borrowing paper: correct information cap, useful ESS formula, but the 'bias-limited' claim is prior-relative and the saturation experiment is oracle-informed. the 3 major comments →

arxiv 2607.04036 v2 pith:ALOUF347 submitted 2026-07-04 stat.ME

B-CALM: Bias-Limited Bayesian Borrowing for RCT-Anchored Treatment Effects under Covariate Mismatch

Amir Asiaee , Samhita Pal This is my paper
classification stat.ME
keywords Bayesian borrowingcausal inferenceconditional average treatment effectexternal controlsobservational datacovariate mismatchbias-limited informationeffective sample size
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Randomized trials give unbiased but often small samples for estimating treatment effects across patient subgroups; observational databases are large but biased, with only partially overlapping covariates. B-CALM treats the observational study as a noisy, biased measurement of the trial's outcome surfaces, and models the gap with two bias functions, one for baseline outcomes and one for treatment contrasts. The paper's central claim is that the observational data can never contribute more information about the trial treatment-effect function than the prior precision of that comparative-bias function, no matter how large the observational study gets. This turns the prior on the bias into an explicit sensitivity knob: tighter bias priors allow more borrowing; diffuse bias priors push inference back toward the trial alone. If right, B-CALM offers a principled way to borrow external data while keeping the trial-defined estimand and quantifying the limits of that borrowing.

Core claim

The central discovery is bias-limited borrowing. In the finite-basis Gaussian version of the model, the observational Fisher information for the trial treatment-effect coefficients is I_o = (Φ_o)^T(σ²I + Φ_oΣ_b(Φ_o)^T)^{-1}Φ_o ⪯ Σ_b^{-1}, meaning the observational study's information about the trial CATE is bounded above by the precision of the prior placed on the comparative-bias function. An equivalent effective-sample-size formula shows the observational contribution saturates at n_o^eff → σ_o²/σ_b² as n_o → ∞. Practically, an observational study of any size acts like at most a fixed number of extra randomized patients, determined by the ratio of observational noise to prior bias variance

What carries the argument

The load-bearing object is the comparative-bias function bΔ(z), defined through the decomposition ηo(a,z)=μ(z)+aτ(z)+b0(z)+a bΔ(z): it is the difference between the observational treatment contrast and the trial contrast after matching on the shared latent state z. Marginalizing out bΔ under a Gaussian prior inflates the observational contrast likelihood's covariance and caps the Fisher information about τ at the prior precision of bΔ. A scalar version collapses to an effective sample size that saturates; the same mechanism yields a PAC-Bayes/IPM risk decomposition separating trial risk, alignment error, and residual calibration.

Load-bearing premise

The safety guarantee rests on the comparative-bias prior being correctly calibrated to the true, unknown difference between the observational and trial contrasts; if the true bias falls far outside the prior's support, borrowing can still be harmful even though the method reports nominal coverage.

What would settle it

Simulate a scenario where the true comparative bias is drawn from a distribution with variance much larger than the prior's, then grow the observational sample size and check whether B-CALM's coverage drops below nominal while its reported effective sample size still saturates; a coverage collapse would show the cap theorem does not protect against mis-specified bias priors.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Adding more observational units beyond the saturation point yields almost no additional precision for the trial CATE; the interval width plateaus at a floor set by σ_o²/σ_b².
  • The comparative-bias prior standard deviation is a transparent sensitivity knob: prespecified values of σΔ give a curve of posterior CATEs, from pooled to RCT-only.
  • Coverage stays near nominal under comparative bias while pooled and forest-based baselines become overconfident; negative transfer relative to RCT-only is small.
  • A fixed global borrowing weight is replaced by local, function-valued borrowing determined pointwise in latent space by the bias posterior.
  • In the one-arm external-control setting, the baseline-bias prior controls the width reduction, and the method reports a bias-susceptibility ratio that flags intervals narrower than the bias they may absorb.

Where Pith is reading between the lines

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

  • If the bound is tight, then even a perfect, infinitely large observational source is informationally equivalent to a modest number of trial participants; this reframes 'data borrowing' as buying a limited number of virtual patients, not asymptotically free precision.
  • The same machinery suggests a testable extension to non-Gaussian or nonlinear outcome models: the cap should reappear as a constraint on the influence function or on a divergence-based information measure, rather than only on Fisher information.
  • Setting the bias prior by empirical Bayes may inherit the adaptivity dip the paper observes at intermediate bias; a decision-theoretic prior that penalizes recommendation-changing errors could mitigate it without sacrificing the cap.
  • In multi-source settings, the cap per source suggests that diversifying multiple observational sources, each with its own bias prior, may be more effective than enlarging a single biased source.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes B-CALM, a Bayesian method for borrowing information from an observational study (OS) to estimate an RCT-defined conditional average treatment effect (CATE) under partially overlapping covariates. The method introduces baseline-bias b0(z) and comparative-bias bΔ(z) functions on a shared latent space, with priors on bΔ acting as a sensitivity knob. The central theoretical result (Theorem 1) shows that, in a finite-basis Gaussian model, the OS Fisher information about the treatment-effect coefficients is bounded above by the prior precision of the comparative-bias function, and the associated effective sample size saturates at σ_o²/σ_b² (Corollary 1). A PAC-Bayes plus IPM-calibration risk decomposition (Theorem 2) is also stated. The paper reports synthetic, semi-synthetic, and pediatric-obesity external-control experiments, where B-CALM maintains near-nominal coverage and low negative transfer while pooled and causal-forest baselines under-cover under comparative bias.

Significance. If the results stand, the paper makes a useful theoretical contribution by giving a closed-form, falsifiable prediction—that OS contrast information about the RCT contrast is capped by the comparative-bias prior precision, with an explicit ESS saturation formula. The derivation is standard Gaussian marginalization, and the paper is transparent about the model conditions. The paper also ships reproducible code, closed-form posterior computations in the Gaussian instance, and honest limitations in the appendix. The key caveat, which the paper itself acknowledges in Appendix A.9, is that the information cap is a property of the assumed bias prior, not a guarantee against prior misspecification; the safety and coverage claims are therefore conditional on the prior being correctly calibrated. This limits the scope of the headline 'bias-limited borrowing' claim.

major comments (3)
  1. [Abstract; §1; §6.1; Appendix A.9] The central safety claim—'bias-limited borrowing'—is stated as if it were a property of the method, but Theorem 1 bounds the OS information only under the assumed prior on bΔ. If the true bias has a nonzero mean, larger scale, or mass outside the prior support, the OS contrast τ+bΔ injects unaccounted bias into the posterior for τ, while the model reports a narrow interval that is well-calibrated only under the prior. The saturation experiment in Fig. 3 and Table 4 sets mΔ=0.2 to match the true bias template (Appendix A.9), so the empirical confirmation of the ESS formula is partly circular. The paper concedes this in A.9, but the abstract and introduction still present the coverage results as unconditional. Please revise the headline claims to state explicitly that protection is conditional on prior calibration, and add a misspecification experiment (e.g., mΔ shifted or σΔ smaller than
  2. [§5, Theorem 2] Theorem 2 as stated in the main text is missing the formal regularity condition needed for the IPM-calibration bound. The statement says 'the IPM regularity condition stated in Appendix A.4 holds' and gives an informal gloss, but the actual assumption used in Lemma 1 is that the debiased OS conditional risk satisfies e^o_f/L ∈ G. Without this condition, the LΔZ(ρ)+εcal(ρ) term in Eq. (13) is not justified. Please state the assumption explicitly in the theorem (or refer to a numbered assumption), so the bound is self-contained.
  3. [§6.3; Appendix B.5] The real-data analysis selects σ0=0.05 by a prior-weighted score that includes interval width and |ATE| (Appendix B.5), which is a data-dependent criterion, not a calibration check. The statement that B-CALM's interval is 'bias-protected' relies on the untestable assumption that the EHR controls are nearly commensurate with the trial (σ0 around 0.05). The paper does report the σ0 sweep and the skeptical-ratio diagnostic, which is good, but the main-text claim that B-CALM 'keeps sensitivity to external-control bias explicit' would be strengthened by presenting the full σ0 path and by stating that the selected value is not validated by ground truth. This is a presentation-scope issue rather than a technical error.
minor comments (5)
  1. [Eq. (9)] The identity in Theorem 1 requires Φo to have full column rank. This is stated, but please also note what the information bound becomes when Φo is rank-deficient (the PSD bound I_o ⪯ Σ_b^{-1} still holds, but the equality with [Σ_b + σ²((Φo)^TΦo)^{-1}]^{-1} does not).
  2. [§6.1, Table 1] The OSCAR/CALM row merges R-OSCAR, MR-OSCAR, and CALM because they 'coincide to within Monte Carlo noise' in the Gaussian harness. This is plausible, but the same merge is used in the real-data Table 7, where it is less obvious; please justify or split the row there.
  3. [Algorithm 2] In step 5, log p_σ = log N{bτ_o − bτ_r; 0, Σ_r + Σ_o + σ²I} treats σ²I as an added variance for the discrepancy. It would help to define σ² here explicitly as a candidate variance for the bias difference, not as the observational noise, to avoid confusion with σ_o².
  4. [§5, Proposition 1] The statement 'If the prior on bΔ converges to a flat Gaussian process prior ... contributes no finite information to τ' is given as a sketch; the formal version is in Proposition 3/Remark 2. Please add a forward reference to the proposition so the reader knows the precise flat-precision condition required.
  5. [Appendix A.9] The sentence 'its coverage values are conservative and should not be read as evidence about the adaptive procedure' is important and should appear in the main text near Fig. 3, not only in the reproducibility appendix.

Circularity Check

1 steps flagged

Bias-limited borrowing cap is a direct algebraic restatement of the comparative-bias prior; experimental 'confirmation' uses the same prior and a truth-matched prior mean.

specific steps
  1. self definitional [Theorem 1 / Corollary 1, Section 5; Appendix A.9]
    "The positive-semidefinite statement means that the OS information is no larger in any finite-dimensional direction than the comparative-bias prior precision. ... The saturation level is determined by the ratio of observational noise to comparative-bias prior variance."

    Theorem 1 is obtained by marginalizing β_b over its prior β_b ~ N(0,Σ_b); the displayed bound I_o = (Σ_b + σ²((Φ_o)^TΦ_o)^{-1})^{-1} ⪯ Σ_b^{-1} is exactly the prior precision, and Corollary 1's limit n_o^eff → σ_o²/σ_b² is the same prior's variance ratio. The 'bias-limited borrowing' prediction is therefore the prior specification restated, not an independent constraint. The saturation experiment fixes σΔ and sets mΔ = 0.2 to match the true bias template (Appendix A.9), so the empirical/theoretical ESS agreement is a self-consistency check of the algebra, not an external confirmation.

full rationale

The paper's Theorem 1 and Corollary 1 are correct, but the central information cap is not an independent discovery: it is the prior precision Σ_b^{-1} of the bias function after Gaussian marginalization. The paper concedes 'The algebra above is standard Gaussian marginalization' (Section 5), and the saturation experiment uses the same σΔ that defines the ceiling while setting the prior mean to the true bias template (Appendix A.9: 'sets the comparative-bias prior mean to mΔ = 0.2, which matches the location of the true bias template; its coverage values are conservative and should not be read as evidence about the adaptive procedure'). Thus the claimed 'bias-limited borrowing' consequence is built into the chosen prior rather than empirically discovered. No load-bearing self-citation is present: references to R-OSCAR, MR-OSCAR and CALM are comparative baselines, not justifications for Theorem 1. The joint-model and calibration-decomposition contributions retain independent content, so the circularity is partial and concentrated in the headline prediction.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

Everything the central claim rests on that the reader didn't pay for upstream: the Gaussian linear model and the comparative-bias prior IS the result; the experiments confirm the model with itself. The bounds are standard algebra conditioned on the modeling assumptions.

free parameters (5)
  • comparative-bias prior scale σΔ = selected per replicate in factorial grid (candidates 0.01–∞); fixed to {0.1,0.2,0.4,∞} in saturation experiment; diffuse
    The main knob controlling the information cap; in the factorial grid it is tuned by empirical-Bayes model averaging with hand-set weights, and in the saturation experiment it is fixed.
  • comparative-bias prior mean mΔ = 0 (factorial, semi-synthetic); 0.2 (saturation experiment)
    In the saturation experiment mΔ=0.2 matches the true bias template; the saturation measurement is therefore conditional on knowing the bias location.
  • baseline-bias prior scale σ0 = 0.6 (synthetic default); selected 0.05 in real-data
    Controls baseline borrowing; in the real-data analysis selected by a prior-weighted score over the candidate set, so the reported width reduction depends on this hand-specified selection.
  • empirical-Bayes prior weights over σΔ and σ0 = (0.01,0.01,0.42,0.21,0.13,0.22) for σΔ; (0.05,0.20,0.30,0.20,0.10,0.05,0.10) for σ0
    Hand-chosen asymmetric weights with a deliberate mode at moderate borrowing; coverage and negative-transfer results depend on these weights.
  • basis functions for outcome surfaces = linear / balanced / mismatch-heavy polynomial-trigonometric sets
    The basis 'overlaps but does not span' the true surfaces, so the outcome model is misspecified by design; performance results depend on these choices.
axioms (5)
  • domain assumption Assumption 1: trial randomization, positivity, consistency
    Fixes the RCT estimand; standard requirement for the trial arm.
  • domain assumption Assumption 2: existence of encoder and functions μ0, τ0 approximating trial outcome surfaces up to ε_rep
    The shared latent representation is assumed sufficient for the RCT outcome surfaces; if it is not, the CATE integrated over the encoder is not the original trial CATE.
  • domain assumption Assumption 3: source encoders exist with IPM ≤ ε_align
    Assumes the OS can be placed on the same latent coordinate system up to bounded error; never empirically tested in the experiments (encoders are identity maps).
  • ad hoc to paper Theorem 2 extra hypothesis: debiased OS risk is proportional to a function in the discriminator class G (e^o_f/L ∈ G), and bounded loss or sub-Gaussian tails
    Needed for the IPM transfer step; not verifiable from data and the calibration residual ε_cal is non-computable.
  • domain assumption Gaussian outcome model with fixed residual variance σ_y=1
    All simulations and the closed-form instance rely on homoskedastic normal likelihood; real outcomes may not be Gaussian at unit variance.
invented entities (3)
  • Shared latent patient state Z no independent evidence
    purpose: Common coordinate system to compare RCT and OS at matched states
    A modeling construct, not a measured variable; no falsifiable handle outside the paper.
  • Baseline-bias function b0(z) no independent evidence
    purpose: Models OS-vs-RCT difference in control outcomes
    A statistical device analogous to historical-bias terms; identified only through prior assumptions.
  • Comparative-bias function bΔ(z) no independent evidence
    purpose: Models OS-vs-RCT difference in treatment contrasts; its prior sets the borrowing cap
    Not identifiable without prior restrictions; the paper itself states this in the Limitations.

pith-pipeline@v1.3.0-alltime-deepseek · 35233 in / 16050 out tokens · 165443 ms · 2026-08-02T08:41:55.506496+00:00 · methodology

0 comments
read the original abstract

Randomized controlled trials (RCTs) identify treatment effects in the randomized trial population but are often too small for reliable heterogeneity estimation; observational studies (OS) are larger but confounded and measured on only partially overlapping covariates. We develop Bayesian Calibrated ALignment under covariate Mismatch (B-CALM), a Bayesian borrowing method for RCT-defined conditional average treatment effect (CATE) estimation. B-CALM maps source-specific covariates into a shared latent state, jointly models trial and observational outcome surfaces, and uses baseline-bias and comparative-bias functions to represent how the OS departs from the trial estimand. The comparative-bias prior becomes an explicit sensitivity knob: we prove a finite-feature bias-limited information bound showing that observational contrast information about the trial treatment-effect function is capped by the prior precision of this bias function, and derive an effective-sample-size formula showing that the RCT-equivalent information contributed by the OS saturates as OS sample size grows. The theory also combines a PAC-Bayes trial-risk bound with an integral-probability-metric (IPM) alignment and calibration decomposition that separates RCT empirical risk, latent alignment, and residual calibration of the debiased OS surface. In synthetic, semi-synthetic, and pediatric-obesity external-control studies, B-CALM maintains near-nominal average coverage and low negative transfer while pooled and causal-forest baselines can become overconfident under comparative bias.

Figures

Figures reproduced from arXiv: 2607.04036 by Amir Asiaee, Samhita Pal.

Figure 1
Figure 1. Figure 1: B-CALM architecture and causal bookkeeping. Source-specific encoders [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparative-bias sweep on the balanced DGP ( [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Interval-width saturation: mean 90% credible-interval width as [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sensitivity analysis on the STAR-like semi-synthetic benchmark (one replicate). Each line is the posterior CATE at [PITH_FULL_IMAGE:figures/full_fig_p028_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Posterior calibration on the balanced DGP under comparative bias: empirical coverage against nominal credible [PITH_FULL_IMAGE:figures/full_fig_p030_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Real-data external-control augmentation diagnostics. With one-arm EHR controls, [PITH_FULL_IMAGE:figures/full_fig_p031_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Baseline-bias recovery and σ0 sensitivity in the pediatric-obesity application. With Ao j ≡ 0 on every EHR row, σ0 is the identified borrowing knob; σ∆ is held diffuse. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_7.png] view at source ↗

Review history (2 revisions) →

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

39 extracted references · 3 linked inside Pith

  1. [1]

    , journal =

    Asiaee, Amir and Di Gravio, Chiara and Beck, Cole and Mei, Yuting and Pal, Samhita and Huling, Jared D. , journal =. Improving Precision of. 2025 , eprint =

    2025

  2. [2]

    and Asiaee, Amir , journal =

    Pal, Samhita and Huling, Jared D. and Asiaee, Amir , journal =. Improving. 2026 , eprint =

    2026

  3. [3]

    Improving

    Asiaee, Amir and Pal, Samhita , journal =. Improving. 2026 , eprint =

    2026

  4. [4]

    arXiv preprint arXiv:2602.09595 , year =

    Sharp Bounds for Treatment Effect Generalization under Outcome Distribution Shift , author =. arXiv preprint arXiv:2602.09595 , year =. 2602.09595 , archivePrefix =

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2603.27788 , year =

    Omitted-Variable Sensitivity Analysis for Generalizing Randomized Trials , author =. arXiv preprint arXiv:2603.27788 , year =. 2603.27788 , archivePrefix =

    arXiv

  6. [6]

    Advances in Neural Information Processing Systems , volume =

    Transfer Learning on Heterogeneous Feature Spaces for Treatment Effects Estimation , author =. Advances in Neural Information Processing Systems , volume =

  7. [7]

    2024 , eprint =

    Dimitriou, Evangelos and Fong, Edwin and Magelund Tarp, Jens and Diaz-Ordaz, Karla and Lehmann, Brieuc , journal =. 2024 , eprint =

    2024

  8. [8]

    Advances in Neural Information Processing Systems , volume =

    Bayesian Inference of Individualized Treatment Effects using Multi-task Gaussian Processes , author =. Advances in Neural Information Processing Systems , volume =

  9. [9]

    Bayesian Analysis , volume =

    Bayesian Regression Tree Models for Causal Inference: Regularization, Confounding, and Heterogeneous Effects (with Discussion) , author =. Bayesian Analysis , volume =. 2020 , doi =

    2020

  10. [10]

    Statistical Science , volume =

    Power Prior Distributions for Regression Models , author =. Statistical Science , volume =. 2000 , doi =

    2000

  11. [11]

    Biometrics , volume =

    Hierarchical Commensurate and Power Prior Models for Adaptive Incorporation of Historical Information in Clinical Trials , author =. Biometrics , volume =. 2011 , doi =

    2011

  12. [12]

    Bayesian Analysis , volume =

    Commensurate Priors for Incorporating Historical Information in Clinical Trials Using General and Generalized Linear Models , author =. Bayesian Analysis , volume =. 2012 , doi =

    2012

  13. [13]

    Biometrics , volume =

    Robust Meta-Analytic-Predictive Priors in Clinical Trials with Historical Control Information , author =. Biometrics , volume =. 2014 , doi =

    2014

  14. [14]

    , year =

    Harrell, Frank E. , year =. Incorporating Historical Control Data Into an

  15. [15]

    and Stuart, Elizabeth A

    Cole, Stephen R. and Stuart, Elizabeth A. , journal =. Generalizing Evidence from Randomized Clinical Trials to Target Populations: The. 2010 , doi =

    2010

  16. [16]

    Journal of the Royal Statistical Society: Series A , volume =

    The Use of Propensity Scores to Assess the Generalizability of Results from Randomized Trials , author =. Journal of the Royal Statistical Society: Series A , volume =. 2011 , doi =

    2011

  17. [17]

    Statistics in Medicine , volume =

    Extending Inferences from a Randomized Trial to a New Target Population , author =. Statistics in Medicine , volume =. 2020 , doi =

    2020

  18. [18]

    Annual Review of Statistics and Its Application , volume =

    A Review of Generalizability and Transportability , author =. Annual Review of Statistics and Its Application , volume =. 2023 , publisher =

    2023

  19. [19]

    Proceedings of the National Academy of Sciences , volume =

    Causal Inference and the Data-Fusion Problem , author =. Proceedings of the National Academy of Sciences , volume =. 2016 , doi =

    2016

  20. [20]

    Which Causal Measure Is Easier to Generalize? , author =

    Risk Ratio, Odds Ratio, Risk Difference... Which Causal Measure Is Easier to Generalize? , author =. 2023 , eprint =

    2023

  21. [21]

    Statistical Science , volume =

    Causal Inference Methods for Combining Randomized Trials and Observational Studies: A Review , author =. Statistical Science , volume =. 2024 , doi =

    2024

  22. [22]

    Journal of Machine Learning Research , volume =

    A Kernel Two-Sample Test , author =. Journal of Machine Learning Research , volume =

  23. [23]

    and George, Edward I

    Chipman, Hugh A. and George, Edward I. and McCulloch, Robert E. , journal =. 2010 , doi =

    2010

  24. [24]

    Journal of Computational and Graphical Statistics , volume =

    Bayesian Nonparametric Modeling for Causal Inference , author =. Journal of Computational and Graphical Statistics , volume =. 2011 , doi =

    2011

  25. [25]

    Journal of the American Statistical Association , volume =

    Estimation and Inference of Heterogeneous Treatment Effects using Random Forests , author =. Journal of the American Statistical Association , volume =. 2018 , doi =

    2018

  26. [26]

    Biometrika , volume =

    Quasi-Oracle Estimation of Heterogeneous Treatment Effects , author =. Biometrika , volume =. 2021 , doi =

    2021

  27. [27]

    Proceedings of the 34th International Conference on Machine Learning , series =

    Estimating Individual Treatment Effect: Generalization Bounds and Algorithms , author =. Proceedings of the 34th International Conference on Machine Learning , series =

  28. [28]

    Proceedings of the 33rd International Conference on Machine Learning , series =

    Learning Representations for Counterfactual Inference , author =. Proceedings of the 33rd International Conference on Machine Learning , series =

  29. [29]

    Journal of Machine Learning Research , volume =

    Generalization Bounds and Representation Learning for Estimation of Potential Outcomes and Causal Effects , author =. Journal of Machine Learning Research , volume =

  30. [30]

    Advances in Neural Information Processing Systems , volume =

    Removing Hidden Confounding by Experimental Grounding , author =. Advances in Neural Information Processing Systems , volume =

  31. [31]

    ICML 2022 Workshop on Spurious Correlations, Invariance and Stability , year =

    Combining Observational and Randomized Data for Estimating Heterogeneous Treatment Effects , author =. ICML 2022 Workshop on Spurious Correlations, Invariance and Stability , year =. 2202.12891 , archivePrefix =

    Pith/arXiv arXiv 2022

  32. [32]

    Transactions on Machine Learning Research , year =

    Combining Interventional and Observational Data Using Causal Reductions , author =. Transactions on Machine Learning Research , year =. 2103.04786 , archivePrefix =

    Pith/arXiv arXiv

  33. [33]

    Biostatistics , volume =

    Bayesian Hierarchical Modeling Based on Multisource Exchangeability , author =. Biostatistics , volume =. 2018 , doi =

    2018

  34. [34]

    Clinical Trials , volume =

    Summarizing Historical Information on Controls in Clinical Trials , author =. Clinical Trials , volume =. 2010 , doi =

    2010

  35. [35]

    Statistical Methods in Medical Research , volume =

    Including Historical Data in the Analysis of Clinical Trials: Is It Worth the Effort? , author =. Statistical Methods in Medical Research , volume =. 2018 , doi =

    2018

  36. [36]

    Biostatistics , volume =

    Dynamic Borrowing in the Presence of Treatment Effect Heterogeneity , author =. Biostatistics , volume =. 2021 , doi =

    2021

  37. [37]

    Statistics in Biopharmaceutical Research , volume =

    Borrowing from Historical Control Data in Cancer Drug Development: A Cautionary Tale and Practical Guidelines , author =. Statistics in Biopharmaceutical Research , volume =. 2019 , doi =

    2019

  38. [38]

    Proceedings of the 24th International Conference on Artificial Intelligence and Statistics , series =

    Nonparametric Estimation of Heterogeneous Treatment Effects: From Theory to Learning Algorithms , author =. Proceedings of the 24th International Conference on Artificial Intelligence and Statistics , series =

  39. [39]

    Proceedings of the 24th International Conference on Artificial Intelligence and Statistics , series =

    Counterfactual Representation Learning with Balancing Weights , author =. Proceedings of the 24th International Conference on Artificial Intelligence and Statistics , series =