Pith. sign in

REVIEW 3 major objections 4 minor 33 references

Consistent Labeling Across Group Assignments: Variance Reduction in Conditional Average Treatment Effect Estimation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A wrapper called CLAGA removes assignment-dependent variance from any CATE estimator by relabeling each instance with a cross-fitted out-of-sample prediction, and the paper shows this lowers PEHE without changing the bias of unbiased…

desk verdict Useful empirical wrapper with a real diagnostic; the theoretical overreach and a misstated equation keep it conditional rather than accept. read the letter →

arxiv 2507.04332 v1 pith:5PN6IK4Y submitted 2025-07-06 cs.LG stat.MEstat.ML

Yi-Fu Fu , Keng-Te Liao , Shou-De Lin This is my paper
classification cs.LGstat.MEstat.ML MSC 62D2062G0568T05
keywords conditionalaveragetreatmenteffectCATEestimationvariancereductionheterogeneityupliftmodelingcross-fittingpseudo-outcomediscrepancyratio
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Many CATE estimation algorithms produce different predictions for the same covariate vector depending on whether the instance was assigned to treatment or control during training, even though the true treatment effect $\tau(x)$ is a deterministic function of $x$ alone. The paper argues that this inconsistency is a structural error source, not a tuning artifact: its error decomposition splits PEHE into training-dependent and training-independent terms, and the training-independent terms grow with within-group target variance and with the gap between groupwise target means. CLAGA is a model-agnostic wrapper that eliminates the gap by replacing each instance's raw learning target with an out-of-sample prediction from a K-fold cross-fitted primary CATE estimator, making the new target assignment-invariant. This removes the inconsistency term and shrinks the variance term while leaving the bias component unchanged for unbiased base estimators. Experiments on synthetic benchmarks and real-world uplift datasets show lower PEHE and better AUUC rankings.

What carries the argument

The load-bearing identity is the PEHE decomposition in Equation (3), which separates the squared error into a model-fitting term, a model-target covariance term, and three target-only terms: $(1-\pi)\mathrm{Var}[\tilde\tau(0)]+\pi\mathrm{Var}[\tilde\tau(1)]$ (WVG), $\pi(1-\pi)(\mathbb{E}[\tilde\tau(0)]-\mathbb{E}[\tilde\tau(1)])^2$ (SDMG), and $\mathbb{E}_W[(\tau-\tilde\tau(W))^2]$. CLAGA's mechanism is a K-fold cross-fitting step: train primary CATE estimators on K−1 folds, read each held-out instance's prediction $\hat\tau^-(X_i)$, and train a secondary regression model on the relabeled dataset $\{(X_i,\hat\tau^-(X_i))\}$. Because the new target does not depend on the instance's group assignment, SDMG vanishes and WVG can be further reduced by ensembling the primary estimators, making the inconsistency term removable by construction rather than by tuning.

What would settle it

For a fixed instance with its group assignment held fixed, bootstrap-resample the training set many times and measure the variance of the pseudo-outcome $\tilde\tau(W)$ that the base algorithm assigns; if that variance rivals the WVG and SDMG terms in Equation (3), the claimed PEHE reduction from CLAGA will not materialize. A direct test is to apply CLAGA with $K=2$ to a deliberately weak linear base estimator on the Zenodo data, where the paper predicts little or no gain.

Watch

Extended reading notes

Core claim

The central claim is that assignment-dependent learning is a real, measurable cause of CATE estimation error. The paper introduces the discrepancy ratio—the fraction of instances whose predicted effects differ significantly under the two group assignments—and shows it persists at one million training samples and rises with model complexity. The theoretical core is an exact decomposition of $\mathbb{E}_{M,W}[(\hat\tau-\tau)^2]$ into five terms; three of them depend only on the surrogate learning target $\tilde\tau(W)$, not on model training, and the terms WVG (weighted within-group target variance) and SDMG (squared difference of groupwise target means) capture the cost of inconsistency. CLAGA replaces $\tilde\tau(W)$ with the out-of-sample prediction $\hat\tau^-(X)$ from a primary estimator trained without that instance, so $\mathbb{E}[\tilde\tau'(0)]=\mathbb{E}[\tilde\tau'(1)]$, which sets SDMG to zero and leaves only a reduced variance term. For unbiased primary estimators the bias term is unchanged, so CLAGA is a variance-reduction layer that can be wrapped around any existing CATE algorithm.

Load-bearing premise

The paper's error decomposition treats the surrogate targets $\tilde\tau(0)$ and $\tilde\tau(1)$ as fixed numbers that depend only on group assignment, but in practice these targets are estimated from the same finite sample and carry their own training randomness; if that randomness is large, the WVG and SDMG terms do not fully capture the error that CLAGA removes.

Editorial extensions

If this is right

  • Wrapping single-model, two-model, X-learner, R-learner, and DR-learner with CLAGA reduces average PEHE by 7.8 percent on ACIC-2016 and 46.3 percent on ACIC-2018.
  • The discrepancy ratio gives practitioners a diagnostic: an estimator with a high ratio is leaving assignment-dependent variance on the table even when its pointwise bias is zero.
  • Because WVG and SDMG do not involve the trained model, hyperparameter tuning, ensembling, or model selection alone cannot remove this error component.
  • CLAGA leaves the bias term of unbiased base estimators unchanged, so it can be safely added to existing CATE pipelines without trading bias for variance.
  • On real-world uplift datasets (x5, Lenta, Criteo), CLAGA improves AUUC by 5.6 to 13.2 percent, showing the effect matters for ranking as well as squared error.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the theory implies CLAGA will matter most in exactly the regimes practitioners tune hardest—large models on modest data—because those maximize the discrepancy ratio; it also predicts weak base models will see little or negative gain, which the paper observes on Zenodo.
  • Editorial inference: the same K-fold out-of-sample relabeling idea could be applied to other estimation targets whose pseudo-outcomes depend on a discrete grouping variable, such as quantile treatment effects or survival treatment effects.
  • Editorial inference: choosing K balances primary-estimator bias against the variance of the relabeled target; the paper fixes K by dataset, and a data-driven rule for K is a natural testable extension.
  • Editorial inference: if surrogate-target randomness is substantial, CLAGA's stated decomposition undercounts the error it removes; a bootstrap study of pseudo-outcome variance would delineate where the method's guarantees start to degrade.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper identifies a phenomenon it calls inconsistent learning across group assignments in CATE estimators: for the same covariate vector, an estimator's training target and predictions can depend systematically on whether the unit was assigned to treatment or control. It introduces a discrepancy-ratio metric to quantify this, presents an error decomposition intended to show which PEHE components are tied to this inconsistency, and proposes CLAGA, a model-agnostic K-fold cross-fitting wrapper that relabels training instances with out-of-sample predictions from primary CATE estimators and then fits a secondary regression. Experiments on synthetic benchmarks (ACIC-2016, ACIC-2018, Zenodo) and real-world uplift datasets (x5, Lenta, Criteo) report PEHE reductions and AUUC improvements relative to the base algorithms without CLAGA.

Significance. If the paper's central claims hold, CLAGA is a simple, model-agnostic variance-reduction wrapper that could be applied to any off-the-shelf CATE estimator, and the discrepancy ratio is a useful diagnostic for a real and underappreciated failure mode. The corrected decomposition in Appendix A, Eq. (9), is a valid algebraic identity, and the empirical evaluation is independent of the theoretical mechanism: PEHE and AUUC are measured directly and show consistent improvements in most settings. The paper also candidly lists limitations, including dependence on the base estimator and computational cost. The main weakness is that the theoretical interpretation of the decomposition is over-stated: the claim that the identified error terms are independent of model training and therefore cannot be addressed by conventional machine learning is not established for the surrogate targets used by the modern CATE algorithms studied.

major comments (3)
  1. [Section 5.1, Eq. (3) and Appendix Eq. (9)] The final term in Eq. (3) is printed as E_W[(τ − τ̃(W))^2], but the full derivation in the appendix, Eq. (9), yields (E_W[τ − τ̃(W)])^2. These are not equal; as printed, Eq. (3) is not an identity. This is not merely cosmetic, because the bullet list in Section 5.1 describes the term as "the average squared difference between surrogate label and true effect," which matches the printed form rather than the appendix form, and Remarks 6.1–6.3 reason about this term's behavior under CLAGA. The main text and appendix must be reconciled.
  2. [Section 5.2.1] The classification of WVG(τ̃) and SDMG(τ̃) as "independent of model training" is not justified for the algorithms studied. For the X-learner, R-learner, and DR-learner, the surrogate targets are fitted quantities: imputed counterfactuals, residuals, and pseudo-outcomes that depend on estimated propensity and outcome models. Those nuisance estimates carry their own training randomness and hyperparameter dependence, so WVG and SDMG can be changed by hyperparameter tuning or by changing the nuisance model class. The valid statement is only that, conditional on a fixed surrogate construction, these terms do not depend on the randomness M of the final prediction model's training. The stronger claim in the abstract and contributions, that the inconsistency "cannot be resolved through conventional machine learning techniques," is therefore not established by the argument as written.
  3. [Section 6.3, Remark 6.3] The claim that the bias term E_W[(τ − τ̃(W))^2] "remains unchanged" under CLAGA is ambiguous. If "unbiased primary estimators" means E[τ̂] = τ, then the new target satisfies E[τ̃'] = τ and the corrected bias term (E_W[τ − τ̃'])^2 should vanish, not remain unchanged. If "unbiased" instead means E[τ̂] = E[τ̃], then the statement is essentially a restatement of the definition of the original surrogate and does not support the conclusion as stated. The remark needs to specify which definition is intended and derive the corresponding conclusion.
minor comments (4)
  1. [Section 7.1, Table 1] The claim of "significant performance improvements" is not supported by significance testing or confidence intervals; for example, the X-learner ratio on ACIC-2016 is 1.0034 ± 0.0692, so the average reduction of 7.8% is driven by some algorithms and not others. Paired tests or confidence intervals would strengthen the empirical claim.
  2. [Section 6.3] The statement that the out-of-sample prediction τ̂⁻ "does not depend on the group assignment" should be made precise: it does not depend on the held-out unit's own assignment W_i, but it does depend on the assignments of other units through the fitted primary estimator. Under i.i.d. sampling the conditional expectation over the other units may still be invariant, but the current wording overstates the independence.
  3. [Section 5.1, bullet list] If Eq. (9) is the intended identity, the bullet describing "Bias of Learning Target" should say "the squared difference of the expected surrogate error, (E_W[τ − τ̃(W)])^2," rather than "the average squared difference between surrogate label and true effect," to match the appendix.
  4. [Throughout] There are several typographical issues: "In In Figure 3(b)" in Section 7.1, "A vg." in Tables 1 and 2, and "Goldschmnidt" in the reference list. These should be corrected.

Circularity Check

2 steps flagged · score 4.0 of 10

Definitional circularity in the theoretical framing, but the empirical PEHE/AUUC evaluations are externally benchmarked and independent.

  1. self definitional [Section 5.2 / 5.2.1, 'Insights into the Error Components']
    "Specifically, we classify any component involving τ̂ as training-dependent, and those involving only τ̃ or τ as independent of model training. ... Consequently, their magnitudes cannot be mitigated by better model training or hyperparameter tuning—they must be addressed through improved learning target construction."

    The taxonomy defines 'training-dependent' as 'involving τ̂'; hence WVG and SDMG, which involve only τ̃, are classified as training-independent by definition. For the X-, R-, and DR-learners studied, the surrogate target τ̃ is itself produced by models fitted on the same sample (pseudo-outcomes, imputed values, residuals), so it carries training randomness and hyperparameter dependence. The conclusion that the inconsistency 'cannot be resolved through conventional machine learning techniques' is therefore an artifact of the definitional split, not a derived property. This is a definitional reduction of the paper's central theoretical claim rather than an independent demonstration.

  2. self definitional [Section 6.3, 'Insight' (and Remark 6.1)]
    "Since τ̂− (and thus τ̃′) does not depend on the group assignment, we have E[τ̃′(0)] = E[τ̃′(1)], meaning that SDMG(τ̃′) = 0."

    SDMG is defined in Eq. (3) as π(1−π)(E[τ̃(0)] − E[τ̃(1)])². CLAGA constructs τ̃′ from out-of-sample predictions made without using the instance's own W, so the equality E[τ̃′(0)] = E[τ̃′(1)] holds by construction. Remark 6.1 therefore restates the algorithm's design rather than providing independent evidence that the inconsistency term drives the observed PEHE improvement. The decomposition explains the mechanism but does not independently predict its magnitude.

full rationale

The paper's central empirical results are not circular: PEHE is measured on synthetic datasets with known ground truth, and AUUC is computed on real-world held-out ranking tasks, so the claimed performance gains are externally benchmarked. The theoretical contribution is an algebraic error decomposition; as an identity it does not fit any parameter, although the main-text Eq. (3) misprints the last term as E_W[(τ − τ̃(W))^2] while the appendix Eq. (9) correctly derives (E_W[τ − τ̃(W)])^2. The genuinely self-referential parts are two definitional moves. First, WVG and SDMG are declared 'training-independent' by a classification rule that only looks at whether the final prediction τ̂ appears, even though for X-, R-, and DR-learners the surrogate targets τ̃ are themselves fitted quantities; the strong claim that conventional ML cannot resolve the inconsistency is thus encoded in the taxonomy. Second, CLAGA zeros SDMG by constructing labels that do not use the instance's own treatment assignment, so the elimination of this term is a design guarantee, not a predicted consequence. These definitional steps are load-bearing for the theoretical narrative, but the independent experimental benchmarking prevents the overall claim from reducing entirely to its inputs. No self-citation chain or imported uniqueness theorem appears, so the circularity is partial and confined to the interpretation of the decomposition.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the potential-outcomes identification assumptions, an implicit assumption that learning targets are fixed rather than estimated, and the CLAGA-specific premise that pseudo-labels are stable enough not to inflate model error.

free parameters (3)
  • Number of folds K = 10 (ACIC-2016/2018), 2 (Zenodo, real-world)
    Chosen per dataset, not derived; controls bias-variance trade-off and compute cost. No sensitivity analysis is reported.
  • Discrepancy-ratio t-test threshold = p < 0.05
    Used to define the headline inconsistency metric; arbitrary but standard.
  • LightGBM hyperparameters = n_estimators 100 to 500, num_leaves 32 to 256, subsample 0.5
    Selected per dataset and model complexity; could influence which algorithms benefit from CLAGA.
assumptions (3)
  • domain assumption Potential outcomes framework: consistency, unconfoundedness, positivity
    Section 2 lists these as standard. The paper says its method does not rely on unconfoundedness, but CATE identification in the observational real-world datasets still requires it.
  • ad hoc to paper Learning targets tau_tilde(W) are deterministic functions of W and x
    Section 5.1 derives WVG and SDMG treating tau_tilde as fixed; in practice targets are estimated and random, which is not acknowledged.
  • ad hoc to paper Secondary regression on pseudo-labels does not systematically inflate model error
    CLAGA's guarantee assumes the secondary model can fit noisy out-of-sample predictions; the paper's Figure 3b admits weak primary estimators make CLAGA harmful.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Consistent Labeling Across Group Assignments: Variance Reduction in Conditional Average Treatment Effect Estimation." pith.science (2026). https://pith.science/paper/5PN6IK4Y

@misc{pith2026250704332,
  author       = {Pith},
  title        = {Pith review of: Consistent Labeling Across Group Assignments: Variance Reduction in Conditional Average Treatment Effect Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5PN6IK4Y}},
  note         = {Machine review of arXiv:2507.04332}
}
read the original abstract

Numerous algorithms have been developed for Conditional Average Treatment Effect (CATE) estimation. In this paper, we first highlight a common issue where many algorithms exhibit inconsistent learning behavior for the same instance across different group assignments. We introduce a metric to quantify and visualize this inconsistency. Next, we present a theoretical analysis showing that this inconsistency indeed contributes to higher test errors and cannot be resolved through conventional machine learning techniques. To address this problem, we propose a general method called \textbf{Consistent Labeling Across Group Assignments} (CLAGA), which eliminates the inconsistency and is applicable to any existing CATE estimation algorithm. Experiments on both synthetic and real-world datasets demonstrate significant performance improvements with CLAGA.

Figures

Figures reproduced from arXiv: 2507.04332 by the authors.

Figure 1
Figure 1. (a) The discrepancy ratio decreases as data size increases, suggesting reduced overfitting. (b) The [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. An overview of CLAGA with K=2. 6 CONSISTENT LABELING ACROSS GROUP ASSIGNMENTS 6.1 Motivation In Section 5.2, we identified two types of error com￾ponents. While model training techniques can help reduce the errors related to the training process, our focus here is on improving the inherent errors caused by the algorithm design, as discussed in Section 5.2.1. We propose a general method that can be applied to any CAT… view at source ↗
Figure 3
Figure 3. PEHE results across data sizes and model complexities with and without CLAGA. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 27 canonical work pages

  1. [1]

    Validating causal inference models via influence functions

    Ahmed Alaa and Mihaela Van Der Schaar. Validating causal inference models via influence functions. In International Conference on Machine Learning, pages 191--201. PMLR, 2019

    work page 2019

  2. [2]

    From real-world patient data to individualized treatment effects using machine learning: current and future methods to address underlying challenges

    Ioana Bica, Ahmed M Alaa, Craig Lambert, and Mihaela Van Der Schaar. From real-world patient data to individualized treatment effects using machine learning: current and future methods to address underlying challenges. Clinical Pharmacology & Therapeutics, 109 0 (1): 0 87--100, 2021

    work page 2021

  3. [3]

    Causalml: Python package for causal machine learning, 2020

    Huigang Chen, Totte Harinen, Jeong-Yoon Lee, Mike Yung, and Zhenyu Zhao. Causalml: Python package for causal machine learning, 2020

    2020

  4. [4]

    Nonparametric estimation of heterogeneous treatment effects: From theory to learning algorithms

    Alicia Curth and Mihaela van der Schaar. Nonparametric estimation of heterogeneous treatment effects: From theory to learning algorithms. In International Conference on Artificial Intelligence and Statistics, pages 1810--1818. PMLR, 2021

    work page 2021

  5. [5]

    A large scale benchmark for uplift modeling

    Diemert Eustache, Betlei Artem , Christophe Renaudin, and Amini Massih-Reza. A large scale benchmark for uplift modeling. In Proceedings of the AdKDD and TargetAd Workshop, KDD, London,United Kingdom, August, 20, 2018. ACM, 2018

    work page 2018

  6. [6]

    Automated versus do-it-yourself methods for causal inference: Lessons learned from a data analysis competition

    Vincent Dorie, Jennifer Hill, Uri Shalit, Marc Scott, and Dan Cervone. Automated versus do-it-yourself methods for causal inference: Lessons learned from a data analysis competition. 2019

    work page 2019

  7. [7]

    Subgroup identification from randomized clinical trial data

    Jared C Foster, Jeremy MG Taylor, and Stephen J Ruberg. Subgroup identification from randomized clinical trial data. Statistics in medicine, 30 0 (24): 0 2867--2880, 2011

    work page 2011

  8. [8]

    Conversion uplift in e-commerce: A systematic benchmark of modeling strategies

    Robin Gubela, Artem Bequ \'e , Stefan Lessmann, and Fabian Gebert. Conversion uplift in e-commerce: A systematic benchmark of modeling strategies. International Journal of Information Technology & Decision Making, 18 0 (03): 0 747--791, 2019

    work page 2019

Show all 33 references
  1. [9]

    Causal inference and uplift modelling: A review of the literature

    Pierre Gutierrez and Jean-Yves G \'e rardy. Causal inference and uplift modelling: A review of the literature. In International Conference on Predictive Applications and APIs, pages 1--13. PMLR, 2017

    2017

  2. [10]

    Incremental value modeling

    Behram Hansotia and Brad Rukstales. Incremental value modeling. Journal of Interactive Marketing, 16 0 (3): 0 35, 2002

    2002

  3. [11]

    Bayesian nonparametric modeling for causal inference

    Jennifer L Hill. Bayesian nonparametric modeling for causal inference. Journal of Computational and Graphical Statistics, 20 0 (1): 0 217--240, 2011

    2011

  4. [12]

    Uplift modeling for clinical trial data

    Maciej Jaskowski and Szymon Jaroszewicz. Uplift modeling for clinical trial data. In ICML Workshop on Clinical Data Analysis, volume 46, 2012

    2012

  5. [13]

    Learning representations for counterfactual inference

    Fredrik Johansson, Uri Shalit, and David Sontag. Learning representations for counterfactual inference. In International conference on machine learning, pages 3020--3029. PMLR, 2016

    2016

  6. [14]

    Lightgbm: A highly efficient gradient boosting decision tree

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decision tree. Advances in neural information processing systems, 30: 0 3146--3154, 2017

    2017

  7. [15]

    Optimal doubly robust estimation of heterogeneous causal effects

    Edward H Kennedy. Optimal doubly robust estimation of heterogeneous causal effects. arXiv preprint arXiv:2004.14497, 2020

    2004 arXiv

  8. [16]

    Double machine learning-based programme evaluation under unconfoundedness

    Michael C Knaus. Double machine learning-based programme evaluation under unconfoundedness. The Econometrics Journal, 25 0 (3): 0 602--627, 2022

    2022

  9. [17]

    o ren R K \

    S \"o ren R K \"u nzel, Jasjeet S Sekhon, Peter J Bickel, and Bin Yu. Metalearners for estimating heterogeneous treatment effects using machine learning. Proceedings of the national academy of sciences, 116 0 (10): 0 4156--4165, 2019

    2019

  10. [18]

    Bigtarget hackathon hosted by lenta and microsoft

    Microsoft Lenta. Bigtarget hackathon hosted by lenta and microsoft. https://www.kaggle.com/datasets/mrmorj/bigtarget, 2020

    2020

  11. [19]

    The true lift model: a novel data mining approach to response modeling in database marketing

    Victor SY Lo. The true lift model: a novel data mining approach to response modeling in database marketing. ACM SIGKDD Explorations Newsletter, 4 0 (2): 0 78--86, 2002

    2002

  12. [20]

    Quasi-oracle estimation of heterogeneous treatment effects

    Xinkun Nie and Stefan Wager. Quasi-oracle estimation of heterogeneous treatment effects. Biometrika, 108 0 (2): 0 299--319, 2021

    2021

  13. [21]

    Using control groups to target on predicted lift: Building and assessing uplift model

    Nicholas Radcliffe. Using control groups to target on predicted lift: Building and assessing uplift model. Direct Marketing Analytics Journal, pages 14--21, 2007

    2007

  14. [22]

    Estimating causal effects of treatments in randomized and nonrandomized studies

    Donald B Rubin. Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of educational Psychology, 66 0 (5): 0 688, 1974

    1974

  15. [23]

    Causal inference using potential outcomes: Design, modeling, decisions

    Donald B Rubin. Causal inference using potential outcomes: Design, modeling, decisions. Journal of the American Statistical Association, 100 0 (469): 0 322--331, 2005

    2005

  16. [24]

    Decision trees for uplift modeling

    Piotr Rzepakowski and Szymon Jaroszewicz. Decision trees for uplift modeling. In 2010 IEEE International Conference on Data Mining, pages 441--450. IEEE, 2010

    2010

  17. [25]

    Synth-validation: Selecting the best causal inference method for a given dataset

    Alejandro Schuler, Ken Jung, Robert Tibshirani, Trevor Hastie, and Nigam Shah. Synth-validation: Selecting the best causal inference method for a given dataset. arXiv preprint arXiv:1711.00083, 2017

    2017 arXiv

  18. [26]

    Benchmarking framework for performance-evaluation of causal inference analysis

    Yishai Shimoni, Chen Yanover, Ehud Karavani, and Yaara Goldschmnidt. Benchmarking framework for performance-evaluation of causal inference analysis. arXiv preprint arXiv:1802.05046, 2018

    2018 arXiv

  19. [27]

    Comparing methods for estimation of heterogeneous treatment effects using observational data from health care databases

    Thierry Wendling, Kenneth Jung, Alison Callahan, Alejandro Schuler, Nigam H Shah, and Blanco Gallego. Comparing methods for estimation of heterogeneous treatment effects using observational data from health care databases. Statistics in medicine, 37 0 (23): 0 3309--3324, 2018

    2018

  20. [28]

    Data of x5 retailhero uplift modeling competition

    X5-Retail-Group. Data of x5 retailhero uplift modeling competition. https://ods.ai/competitions/x5-retailhero-uplift-modeling/data, 2019

    2019

  21. [29]

    Estimating heterogeneous treatment effects with observa- tional data

    Yu Xie, Jennie E Brand, and Ben Jann. Estimating heterogeneous treatment effects with observa- tional data. Sociological methodology, 42 0 (1): 0 314--347, 2012

    2012

  22. [30]

    Representation learning for treatment effect estimation from observational data

    Liuyi Yao, Sheng Li, Yaliang Li, Mengdi Huai, Jing Gao, and Aidong Zhang. Representation learning for treatment effect estimation from observational data. Advances in neural information processing systems, 31, 2018

    2018

  23. [31]

    A unified survey of treatment effect heterogeneity modelling and uplift modelling

    Weijia Zhang, Jiuyong Li, and Lin Liu. A unified survey of treatment effect heterogeneity modelling and uplift modelling. ACM Computing Surveys (CSUR), 54 0 (8): 0 1--36, 2021 a

    2021

  24. [32]

    Treatment effect estimation with disentangled latent factors

    Weijia Zhang, Lin Liu, and Jiuyong Li. Treatment effect estimation with disentangled latent factors. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 10923--10930, 2021 b

    2021

  25. [33]

    Synthetic data for uplift modeling and heterogenous treatment effect with known counterfactuals and ite, March 2022

    Zhenyu Zhao. Synthetic data for uplift modeling and heterogenous treatment effect with known counterfactuals and ite, March 2022. URL https://zenodo.org/record/6342552

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.