Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Deep Disentangled Representation Network for Treatment Effect Estimation

T0 review · 2 major / 5 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 deep network that softly separates instrumental, confounding, and adjustment factors reports lower individual treatment effect error across benchmarks and a production pop-up campaign.

desk verdict Incremental but solid empirical paper on CATE estimation; the disentanglement mechanism is overstated and Section 4.7 doesn't do what it claims. read the letter →

arxiv 2507.06650 v1 pith:SSKYZEPD submitted 2025-07-09 cs.LG cs.AI

Hui Meng , Keping Yang , Xuyu Peng , Bo Zheng This is my paper
classification cs.LGcs.AI
keywords causalinferencecounterfactualregressionindividualtreatmenteffectdisentangledrepresentationlearningupliftmodelingmixtureofexpertsmulti-headattentionorthogonalregularization
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

This paper tries to establish that individual treatment effects can be estimated more accurately when the observed covariates are softly separated in a shared latent space into an instrumental factor $\Gamma$, a confounding factor $\Delta$, and an adjustment factor $\Upsilon$, rather than hard-splitting covariates or running three isolated networks. The proposed Deep Disentangled Representation Network combines a mixture of experts with multi-head attention (MEMA) to learn the three factors from the same representation, a linear orthogonal regularizer (LOR) that pushes the factor masks toward independence, and an importance-sampling re-weighting scheme that uses only $\Delta$ to correct selection bias. On the IHDP and ACIC 2016 benchmarks the method reports lower out-of-sample PEHE and absolute ATE error than the disentangled-representation baselines it compares with, and on a large production dataset it reports higher AUUC, Qini, and online click-through and daily-active-customer gains. If these results hold, the method offers a practical route to better counterfactual predictions for personalized decisions in healthcare, marketing, and public policy.

What carries the argument

The mechanism is a multi-head-attention mixture-of-experts encoder (MEMA) whose task-specific towers each output one latent factor: $\Gamma$ for instrumental information feeding the treatment-assignment head, $\Delta$ for confounding information feeding both the treatment and outcome heads, and $\Upsilon$ for adjustment information feeding only the outcome heads. Two constraints carry the disentanglement claim. A linear orthogonal regularizer forms $\Gamma = W_t \otimes T(X)$, $\Delta = W_c \otimes C(X)$, $\Upsilon = W_a \otimes A(X)$ and minimizes $L_{\mathrm{lor}} = W_t \cdot W_c + W_c \cdot W_a + W_t \cdot W_a$ to drive the three masks toward orthogonality at $\mathcal{O}(H)$ cost. An importance sampling weight $\omega_i = 1 + \Pr(\Delta_i \mid \neg t_i) / \Pr(\Delta_i \mid t_i)$ is computed from $\Delta$ alone, so the factual regression loss is reweighted without contamination from $\Gamma$. A maximum-mean-discrepancy imbalance loss is applied to $\Upsilon$ versus treatment and $\Gamma$ versus outcome to enforce the conditional-independence constraints of the assumed causal graph.

What would settle it

Train the model on a semi-synthetic dataset constructed with known instrumental, confounding, and adjustment variables, then measure the mutual information between each learned factor ($\Gamma$, $\Delta$, $\Upsilon$) and the true factor assignments. If $\Gamma$ carries confounding information or $\Upsilon$ carries treatment information while the reported PEHE still improves, or if removing the orthogonal regularizer leaves PEHE essentially unchanged, that would show the gains are produced by the balancing and importance-weighting losses rather than by successful disentanglement.

Watch

Extended reading notes

Core claim

The paper's central claim is that DDRN, a counterfactual regression network with MEMA and LOR, outperforms existing individual treatment effect estimators on both semi-synthetic benchmarks and real production data. According to the reported results, the out-of-sample IHDP PEHE drops from $0.529 \pm 0.068$ for the DeR-CFR baseline to $0.411 \pm 0.054$, and the absolute ATE error drops to $0.128 \pm 0.031$; similar improvements are reported on ACIC 2016, on Jobs policy risk, and on the Message Pop-up production dataset's AUUC and Qini. The ablation study indicates that every proposed component—the attention-based mixture of experts, the orthogonal regularizer, the imbalance loss, and the $\Delta$-only importance sampling weight—contributes to the improvements. The paper presents this as evidence that soft latent-space decomposition into $\Gamma$, $\Delta$, and $\Upsilon$ is a workable alternative to hard covariate decomposition for counterfactual regression.

Load-bearing premise

The load-bearing premise is that minimizing dot products among three randomly initialized mask vectors and applying balancing losses produces latent factors that genuinely match the instrumental, confounding, and adjustment roles, so that the reported gains come from true disentanglement rather than from the balancing and reweighting terms alone.

Editorial extensions

If this is right

  • On the IHDP and ACIC 2016 benchmarks with known counterfactual outcomes, the reported PEHE and $\epsilon_{\mathrm{ATE}}$ improvements would mean more accurate personalized treatment predictions under the standard unconfoundedness assumptions.
  • On Jobs, the lower policy risk implies that a decision rule based on the estimated individual treatment effect would choose treatment for a better-ordered set of individuals than the compared baselines.
  • On the production pop-up dataset, higher offline AUUC and Qini with positive online A/B gains imply that the model's uplift ranking translates into measurable business lift when deployed.
  • The ablation on ACIC 2016 implies that removing MEMA, LOR, imbalance loss, or importance-sampling re-weighting each degrades estimation, so the gains are not attributable to a single loss term.
  • Because the regularizer costs $\mathcal{O}(H)$ rather than the three-network cost of the deep orthogonal regularizer baseline, the method is positioned as a cheaper route to the same soft separation goal.

Reading between the lines

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

  • The paper does not validate that the learned $\Gamma$, $\Delta$, $\Upsilon$ correspond to true causal roles; a direct probe comparing each learned factor against known generative factors in semi-synthetic data would test whether disentanglement drives the gains or the balancing and reweighting losses do.
  • The same architecture could be adapted to multi-treatment and continuous-treatment settings, which the paper names as future work; MEMA's attention over experts may reduce the need for hard factor assignment in those settings.
  • The online A/B result measures aggregate click-through rate and daily active customers, not individual effect accuracy, so it supports uplift ranking quality but does not by itself confirm the per-person counterfactual estimates.
  • If orthogonality of the random masks is all that separates factors, a variant using learned soft attention routing instead of fixed mask products might match the results with lower sensitivity to initialization; testing such a variant would separate the regularizer's effect from the architecture's effect.
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

2 major / 5 minor

Summary. The paper proposes the Deep Disentangled Representation Network (DDRN) for estimating individual treatment effects from observational data. The architecture combines a mixture of experts with multi-head attention (MEMA) to produce three latent representations—instrumental (Γ), confounding (Δ), and adjustment (Υ)—together with a linear orthogonal regularizer (LOR) applied to mask vectors, an MMD-based imbalance loss, and an importance-sampling reweighting loss. The model is evaluated on the IHDP, ACIC 2016, and Jobs semi-synthetic benchmarks, and on a production pop-up dataset with AUUC/Qini metrics and an online A/B test. The central claim is that DDRN outperforms state-of-the-art disentangled-representation methods for ITE estimation.

Significance. If the empirical results are reproducible and the disentanglement mechanism is genuine, DDRN would be a practically useful contribution: the ACIC 2016 out-of-sample PEHE of 1.168 ± 0.107 is substantially better than the reported DeR-CFR baseline (1.857 ± 0.109), and the production A/B test shows consistent lifts. Strengths of the paper include open-sourced core code, evaluation on benchmarks with counterfactual ground truth, an ablation study, and a real-world deployment test. However, the central claimed mechanism—that the linear orthogonal regularizer yields independent disentangled factors—is not established: the loss is applied to mask vectors, not to the factors themselves, and no identifiability analysis, factor-label validation, or measurement of achieved factor independence is provided. The empirical comparison also relies on baseline numbers taken from a cited paper and lacks significance tests.

major comments (2)
  1. [§4.7, Eq. (14)–(15)] The linear orthogonal regularizer is claimed to ensure precise, independent disentangled factors, but L_lor penalizes pairwise dot products of the randomly initialized mask vectors W_t, W_c, W_a, not the inner products of the latent factors Γ, Δ, Υ. Since Γ = W_t ⊗ T(X), Δ = W_c ⊗ C(X), and Υ = W_a ⊗ A(X), factor orthogonality would involve terms such as Γ·Δ = Σ_i W_t,i W_c,i T_i(X) C_i(X), which also depend on the learned expert outputs; pairwise orthogonal masks are neither necessary nor sufficient for factor independence. In addition, Eq. (15) minimizes a sum of dot products, which drives the values downward rather than toward zero, so it is not a distance penalty as the text claims. The paper provides no identifiability proof, no experiment with ground-truth factor labels, and no measurement of achieved factor independence. The ablation in Table 2 only shows that removing LOR degrades performance; it does not reveal whether the gain comes from true disentanglement, from the extra mask parameters acting as a regularizer, or from interaction with the other loss terms. The central claim therefore needs either a theoretical characterization of what L_lor actually enforces, a validation that Γ, Δ, Υ correspond to their alleged causal roles, or a reformulated claim that does not assert disentanglement.
  2. [§5.3, Table 1] The main empirical claim is that DDRN clearly outperforms state-of-the-art methods, but the baseline numbers are taken directly from [37] rather than re-run under the same protocol, and no significance tests are reported. The comparison set is also narrow: CFR-ISW, DRCFR, and DeR-CFR are all from the same disentangled-representation line, and no representative non-disentangled baselines (e.g., TARNet, Dragonnet) are included. Because Table 1 is the primary evidence for the abstract's 'clearly demonstrate' statement, the authors should either re-run the baselines on the same splits, report paired significance tests with the standard deviations already given, or explicitly restrict the claim to the compared methods.
minor comments (5)
  1. [Section 7] The introduction states that 'Finally, Section 7 is the ethical statement,' but no Section 7 appears in the manuscript; the paper ends after the references.
  2. [§4.5, Eq. (8)] The text says the MMD is used to calculate dissimilarity between the conditional distributions of Y given t=0 versus t=1, but the formula uses Υ(x_i) conditional on treatment arms; this appears to be a typo and should be corrected to describe distributions of the adjustment factor Υ.
  3. [Algorithm 1, line 6] The pseudocode refers to 'Linear Orthography Regularizer' while the rest of the paper uses 'Linear Orthogonal Regularizer'; please make the terminology consistent.
  4. [Table 2] The checkmark layout of the ablation table is difficult to parse; please clarify which components are included in each row, for example by listing the configuration explicitly in each row.
  5. [Table 1] The labels 'Within-sample' and 'Out-of-sample' are used but never defined in the text; please define them, especially because the out-of-sample PEHE for the proposed method is larger than the within-sample PEHE, which readers may find surprising.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: performance claims rest on external counterfactual benchmarks (IHDP, ACIC 2016) and a live online A/B test; the only self-citation ([5]) is related-work context, not load-bearing.

full rationale

The central claim — DDRN outperforms SOTA ITE estimators — is externally validated, not input-equivalent. On IHDP and ACIC 2016, PEHE and epsilon_ATE (Eqs. 16–17) are computed against counterfactual ground truth supplied by the benchmark simulators (Hill 2011; Dorie et al. 2019): the model never observes the benchmark simulation parameters, so the Table 1 numbers are falsifiable external results. Jobs policy risk (Eq. 18) uses factual employment outcomes; the MPD offline metrics (Eqs. 19–21) and the 7-day online A/B test (Table 4) are production measurements. None is a fitted value renamed as a prediction. The Section 4.7 disentanglement chain is also not circular by construction: Gamma, Delta, Upsilon are masked MEMA outputs (Eq. 14), and L_lor (Eq. 15) constrains the masks. Since Gamma·Delta = sum_i (W_ti*W_ci) T_i(x) C_i(x), pairwise mask orthogonality does force pairwise factor orthogonality whenever the masks are nonnegative. What the paper omits is the nonnegativity assumption on the randomly initialized W vectors, and it slides from achieved 'orthogonality' to claimed 'independence' and to instrument/confounder/adjuster identification without any factor-role validation (Section 4.7, Section 5.4 including the 'Look beyond the outcome' passage, and Section 5.5's 'zeta will guarantee... independent latent factors'). These are missing-support and correctness gaps, not circular reductions: the claimed factor property is not asserted to equal the loss by definition, and the headline performance neither reduces to nor depends on that interpretive attribution. Self-citation: [5] (Cheng et al. 2022, whose senior author Bo Zheng matches this paper's Bo Zheng) appears once, in the related-work critique of vCLUB-based mutual-information disentanglement, and supports none of the design choices or reported results — minor and non-load-bearing. Baseline numbers for CFR-ISW, DRCFR, and DeR-CFR are quoted from external publications [11, 12, 37], and the comparison is against published values, not against fits of this model. Section 6's stated limitation (unsuitable for isomorphic multi-interventions) is a scope caveat, not a circular step. Per the rubric, exactly one minor non-load-bearing self-citation with an otherwise self-contained, externally benchmarked derivation yields score 2, not higher.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method relies on standard causal assumptions plus two unproven architectural heuristics: attention over experts yields a shared latent space, and orthogonality of mask vectors yields disentangled factors. Loss weights are tuned on the data. No new physical or conceptual entities are introduced.

free parameters (3)
  • Objective loss weights alpha, beta, zeta, eta = alpha=1.0, eta=0.25, beta=0.5, zeta=0.5
    Tuned by grid search on ACIC 2016 in Section 5.5; final model performance depends on these values and no separate validation protocol is described.
  • Architecture hyperparameters (hidden dimension, number of heads, expert layers) = hidden dimension 200, heads <= 4, expert output in {100,200,300,400}
    Chosen via search in Table 5 and Section 5.5; contributes to the reported empirical results.
  • Mask vectors W_t, W_c, W_a in LOR = learned during training
    Element-wise masks define Gamma, Delta, Upsilon in Eq. 14; their orthogonality is the paper's disentanglement mechanism, yet no independent identifiability or factor labels validate them.
assumptions (4)
  • domain assumption SUTVA, unconfoundedness, and overlap hold (Assumptions 1-3).
    Standard causal identifiability conditions, stated in Section 3; if unconfoundedness fails, the treatment effect estimates are biased.
  • domain assumption Observed covariates can be decomposed into instrumental, confounding, and adjustment factors Gamma, Delta, Upsilon as in the causal graph of [12].
    Section 1 and Section 4 assume this structure without testing it; if the graph is wrong, balancing and reweighting can introduce bias.
  • ad hoc to paper Orthogonality of learned mask vectors yields disentangled latent factors.
    Section 4.7, Eq. 15: no proof that a dot-product penalty on masks produces statistical independence of representations.
  • ad hoc to paper Multi-head self-attention across experts makes factors share a representational space and improves factor identification.
    Section 4.2: architectural heuristic without formal guarantee that attention over expert outputs yields correct causal separation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Disentangled Representation Network for Treatment Effect Estimation." pith.science (2026). https://pith.science/paper/SSKYZEPD

@misc{pith2026250706650,
  author       = {Pith},
  title        = {Pith review of: Deep Disentangled Representation Network for Treatment Effect Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SSKYZEPD}},
  note         = {Machine review of arXiv:2507.06650}
}
read the original abstract

Estimating individual-level treatment effect from observational data is a fundamental problem in causal inference and has attracted increasing attention in the fields of education, healthcare, and public policy.In this work, we concentrate on the study of disentangled representation methods that have shown promising outcomes by decomposing observed covariates into instrumental, confounding, and adjustment factors. However, most of the previous work has primarily revolved around generative models or hard decomposition methods for covariates, which often struggle to guarantee the attainment of precisely disentangled factors. In order to effectively model different causal relationships, we propose a novel treatment effect estimation algorithm that incorporates a mixture of experts with multi-head attention and a linear orthogonal regularizer to softly decompose the pre-treatment variables, and simultaneously eliminates selection bias via importance sampling re-weighting techniques. We conduct extensive experiments on both public semi-synthetic and real-world production datasets. The experimental results clearly demonstrate that our algorithm outperforms the state-of-the-art methods focused on individual treatment effects.

Figures

Figures reproduced from arXiv: 2507.06650 by the authors.

Figure 1
Figure 1. The intuitive illustration of two distinct causal [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall network architecture of our proposed model. This approach takes user profiles, behavior sequences, and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Uplift curves for different models on Message Pop [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Sensitivity analysis of the hyperparameters { [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Uncertainty-Guided LLM Semantic Augmentation for Heterogeneous Treatment Effect Estimation

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Adding LLM-generated, uncertainty-targeted semantic representations — split into assignment and heterogeneity channels and routed asymmetrically — improves finite-sample CATE estimates for most of ten neural host lear...

Reference graph

Works this paper leans on

41 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [37]

    Anpeng Wu, Junkun Yuan, Kun Kuang, Bo Li, Runze Wu, Qiang Zhu, Yueting Zhuang, and Fei Wu. 2022. Learning decomposed representations for treatment effect estimation. IEEE Transactions on Knowledge and Data Engineering 35, 5 (2022), 4989–5001

    work page 2022

  2. [1]

    Smith and Petra E

    Jeffrey A. Smith and Petra E. Todd. 2005. Does matching overcome LaLonde’s critique of nonexperimental estimators? Journal of Econometrics 125, 1 (2005), 305–353. https://doi.org/10.1016/j.jeconom.2004.04.011 Experimental and non- experimental evaluation of economic policy and models

    doi:10.1016/j.jeconom.2004.04.011 2005

  3. [2]

    Martin Arjovsky, Soumith Chintala, and Léon Bottou. 2017. Wasserstein GAN. arXiv:1701.07875 [stat.ML]

    arXiv 2017

  4. [3]

    Peter C Austin. 2011. An introduction to propensity score methods for reducing the effects of confounding in observational studies. Multivariate behavioral research 46, 3 (2011), 399–424

    work page 2011

  5. [4]

    Nitin Bansal, Xiaohan Chen, and Zhangyang Wang. 2018. Can we gain more from orthogonality regularizations in training deep CNNs?. In Proceedings of the 32nd International Conference on Neural Information Processing Systems (Montréal, Canada) (NIPS’18). Curran Associates Inc., Red Hook, NY, USA, 4266–4276

    work page 2018

  6. [5]

    Mingyuan Cheng, Xinru Liao, Quan Liu, Bin Ma, Jian Xu, and Bo Zheng. 2022. Learning disentangled representations for counterfactual regression via mutual information minimization. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval . 1802–1806

    work page 2022

  7. [6]

    Floris Devriendt, Jente Van Belle, Tias Guns, and Wouter Verbeke. 2022. Learning to Rank for Uplift Modeling.IEEE Transactions on Knowledge and Data Engineering 34, 10 (2022), 4888–4904. https://doi.org/10.1109/TKDE.2020.3048510

    work page arXiv 2022

  8. [7]

    Eustache Diemert, Artem Betlei, Christophe Renaudin, and Massih-Reza Amini

    work page

Show all 41 references
  1. [8]

    Vincent Dorie, Jennifer Hill, Uri Shalit, Marc Scott, and Dan Cervone. 2019. Automated versus Do-It-Yourself Methods for Causal Inference: Lessons Learned from a Data Analysis Competition. Statist. Sci. 34, 1 (2019), 43 – 68. https: //doi.org/10.1214/18-STS667

    2019 doi

  2. [9]

    Borgwardt, Malte J

    Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Schölkopf, and Alexander Smola. 2012. A Kernel Two-Sample Test. Journal of Machine Learning Research 13, 25 (2012), 723–773. http://jmlr.org/papers/v13/gretton12a.html

    2012

  3. [10]

    Jens Hainmueller. 2012. Entropy Balancing for Causal Effects: A Multivariate Reweighting Method to Produce Balanced Samples in Observational Studies. Political Analysis 20, 1 (2012), 25–46. https://doi.org/10.1093/pan/mpr025

    2012 doi

  4. [11]

    Negar Hassanpour and Russell Greiner. 2019. CounterFactual Regression with Importance Sampling Weights. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19 . International Joint Confer- ences on Artificial Intelligence Organi...

    2019

  5. [12]

    Negar Hassanpour and Russell Greiner. 2020. Learning Disentangled Represen- tations for CounterFactual Regression. In International Conference on Learning Representations. https://openreview.net/forum?id=HkxBJT4YvB

    2020

  6. [13]

    Hernan and J.M

    M.A. Hernan and J.M. Robins. 2020. Causal Inference: What If . CRC Press. https://books.google.com/books?id=_KnHIAAACAAJ

    2020

  7. [14]

    Jennifer Hill. 2011. Bayesian Nonparametric Modeling for Causal Inference. Journal of Computational and Graphical Statistics 20 (03 2011), 217–240. https: //doi.org/10.1198/jcgs.2010.08162

    2011 arXiv

  8. [15]

    Imbens and Donald B

    Guido W. Imbens and Donald B. Rubin. 2015.Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction . Cambridge University Press

    2015

  9. [16]

    Marshall Joffe, Thomas Have, Harold Feldman, and Stephen Kimmel. 2004. Model Selection, Confounder Control, and Marginal Structural Models: Review and New Applications. The American Statistician 58 (02 2004), 272–279

    2004

  10. [17]

    Johansson, Uri Shalit, and David Sontag

    Fredrik D. Johansson, Uri Shalit, and David Sontag. 2016. Learning Represen- tations for Counterfactual Inference. In Proceedings of the 33rd International Conference on International Conference on Machine Learning - Volume 48 (New York, NY, USA) (ICML’16). JMLR.org, 3020–3029

    2016

  11. [18]

    Ron Kohavi and Roger Longbotham. 2011. Unexpected results in online controlled experiments. SIGKDD Explor. Newsl. 12, 2 (mar 2011), 31–35. https://doi.org/10. 1145/1964897.1964905

    2011

  12. [19]

    Kun Kuang, Peng Cui, Bo Li, Meng Jiang, Yashen Wang, Fei Wu, and Shiqiang Yang. 2019. Treatment effect estimation via differentiated confounder balancing and regression. ACM Transactions on Knowledge Discovery from Data (TKDD) 14, 1 (2019), 1–25

    2019

  13. [20]

    Kun Kuang, Peng Cui, Bo Li, Meng Jiang, Shiqiang Yang, and Fei Wang. 2017. Treatment Effect Estimation with Data-Driven Variable Decomposition. Pro- ceedings of the AAAI Conference on Artificial Intelligence 31, 1 (Feb. 2017). https://doi.org/10.1609/aaai.v31i1.10480

    2017 doi

  14. [21]

    Robert J. LaLonde. 1986. Evaluating the Econometric Evaluations of Training Programs with Experimental Data. American Economic Review 76 (1986), 604–20

    1986

  15. [22]

    Christos Louizos, Uri Shalit, Joris Mooij, David Sontag, Richard Zemel, and Max Welling. 2017. Causal Effect Inference with Deep Latent-Variable Models. arXiv:1705.08821 [stat.ML]

    2017 arXiv

  16. [23]

    Lunceford and Marie Davidian

    Jared K. Lunceford and Marie Davidian. 2004. Stratification and weighting via the propensity score in estimation of causal treatment effects: a comparative study. Statistics in Medicine 23 (2004). https://api.semanticscholar.org/CorpusID: 11912618

    2004

  17. [24]

    Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H. Chi. 2018. Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture- of-Experts. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (London, Un...

    2018

  18. [25]

    Judea Pearl. 2009. Causal inference in statistics: An overview. Statistics Surveys 3, none (2009), 96 – 146. https://doi.org/10.1214/09-SS057

    2009 doi

  19. [26]

    Judea Pearl. 2009. Causality: Models, Reasoning and Inference (2nd ed.). Cambridge University Press

    2009

  20. [27]

    Pearl, M

    J. Pearl, M. Glymour, and N.P. Jewell. 2016.Causal Inference in Statistics: A Primer. Wiley. https://books.google.com/books?id=L3G-CgAAQBAJ

    2016

  21. [28]

    Rosenbaum and Donald B

    Paul R. Rosenbaum and Donald B. Rubin. 1983. The central role of the propensity score in observational studies for causal effects. Biometrika 70, 1 (04 1983), 41–55. https://doi.org/10.1093/biomet/70.1.41

    1983 doi

  22. [29]

    Donald B Rubin. 1974. Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of educational psychology 66, 5 (1974), 688–701

    1974

  23. [30]

    Donald B Rubin. 1997. Estimating Causal Effects from Large Data Sets Using Propensity Scores. Annals of Internal Medicine 127 (1997), 757–763

    1997

  24. [31]

    Joseph Schafer and Joseph Kang. 2009. Average Causal Effects From Nonrandom- ized Studies: A Practical Guide and Simulated Example. Psychological methods 13 (01 2009), 279–313. https://doi.org/10.1037/a0014268

    2009 doi

  25. [32]

    Johansson, and David Sontag

    Uri Shalit, Fredrik D. Johansson, and David Sontag. 2017. Estimating individual treatment effect: generalization bounds and algorithms. In Proceedings of the 34th International Conference on Machine Learning (ICML’17). Sydney, NSW, Australia

    2017

  26. [33]

    Blei, and Victor Veitch

    Claudia Shi, David M. Blei, and Victor Veitch. 2019. Adapting Neural Networks for the Estimation of Treatment Effects. arXiv:1906.02120 [stat.ML]

    2019 arXiv

  27. [34]

    Hidetoshi Shimodaira. 2000. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of Statistical Planning and Inference 90, 2 (2000), 227–244. https://doi.org/10.1016/S0378-3758(00)00115-4

    2000 doi

  28. [35]

    Stéphane Tufféry. 2011. Data Mining and Statistics for Decision Making . Wiley

    2011

  29. [36]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. Attention Is All You Need. arXiv:1706.03762 [cs.CL]

    2023 arXiv

  30. [38]

    Lu Zhang, Yongkai Wu, and Xintao Wu. 2019. Causal Modeling-Based Dis- crimination Discovery and Removal: Criteria, Bounds, and Algorithms. IEEE Transactions on Knowledge and Data Engineering 31, 11 (2019), 2035–2050. https://doi.org/10.1109/TKDE.2018.2872988

    2019

  31. [39]

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

    2021

  32. [40]

    José Zubizarreta. 2015. Stable Weights that Balance Covariates for Estimation With Incomplete Outcome Data. J. Amer. Statist. Assoc. 110 (04 2015), 0–0. https: //doi.org/10.1080/01621459.2015.1023805

    2015

  33. [2018]

    In Proceedings of the AdKDD and TargetAd Workshop, KDD, London,United Kingdom, August, 20, 2018

    A Large Scale Benchmark for Uplift Modeling. In Proceedings of the AdKDD and TargetAd Workshop, KDD, London,United Kingdom, August, 20, 2018 . ACM, London, United Kingdom

    2018

Pith tools

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