Pith. sign in

REVIEW 4 major objections 6 minor 59 references

A Diffusion-Based Method for Learning the Multi-Outcome Distribution of Medical Treatments

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A diffusion-based model called DIME learns the full joint distribution of multiple medical outcomes under a treatment, not just point estimates of treatment effects.

desk verdict Credible extension of DiffPO to joint multi-outcome interventional distributions, but the real-data evaluation only supports pooled marginal distributions, not the per-patient conditional claim. read the letter →

arxiv 2506.01533 v1 pith:SOJLKM5J submitted 2025-06-02 cs.LG cs.AI

Yuchen Ma , Jonas Schweisthal , Hengrui Zhang , Stefan Feuerriegel This is my paper
classification cs.LGcs.AI
keywords diffusionmodelscausalmachinelearningpotentialoutcomesjointinterventionaldistributionmultiplemixed-typetreatmenteffectestimationuncertaintyquantification
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 proposes DIME, a diffusion-based generative method for learning the joint distribution of multiple medical outcomes under a treatment. The target is the joint interventional distribution $p(Y_1(a),\ldots,Y_k(a)\mid X=x)$, not a single expected outcome or one outcome's marginal density. The authors argue that existing treatment-effect methods estimate point values or single-outcome distributions, so they miss the dependence structure among outcomes such as efficacy and safety endpoints. DIME decomposes the joint distribution into conditional distributions, learns each with a conditional score-based diffusion model under causal masking, and generates outcomes autoregressively at inference. If the method works as described, clinicians could ask how likely a whole pattern of outcomes is for a given patient and treatment, with explicit uncertainty quantification.

What carries the argument

The load-bearing mechanism is a conditional score-based diffusion model trained through a chain-rule decomposition of the joint outcome distribution into $p(Y_i\mid Y_{<i},X,A)$. A binary causal mask separates observed factual outcomes from unobserved counterfactual ones so the diffusion loss is applied only where data exist, and a conditional mask selects which earlier outcomes act as conditioning variables for a given ordering. At inference, DIME samples outcomes autoregressively: first $Y_{\sigma(1)}$, then $Y_{\sigma(2)}$ given $Y_{\sigma(1)}$, and so on, averaging over multiple orderings $\sigma$. This combination is what lets a single framework represent the joint interventional distribution for mixed-type outcomes.

What would settle it

Take a benchmark where the true joint interventional distribution is known for each covariate vector $x$. If DIME's samples match the pooled outcome distribution but fail to reproduce the true conditional distribution of one outcome given the other outcomes and $x$ for individual patients, the claim that it learns the joint conditional distribution is refuted.

Watch

Extended reading notes

Core claim

DIME is claimed to be the first neural method tailored to learn the joint interventional distribution of multiple mixed-type medical outcomes from observational data. Given covariates $x$ and treatment $a$, the method targets $p(Y_1(a),\ldots,Y_k(a)\mid X=x)$ rather than conditional average potential outcomes or marginal densities. The causal masking step addresses the fundamental problem of causal inference by allowing the training loss to use only observed factual outcomes, while the chain-rule decomposition into conditional distributions lets a score-based diffusion model capture interdependencies among binary, categorical, and continuous outcomes. During inference, outcomes are sampled autoregressively and predictions are aggregated over several outcome orderings. In experiments on ACIC, IST, and MIMIC-III, DIME reports lower Wasserstein and KL divergences than the adapted baselines, including DiffPO, GANITE, TARNet, CFRNet, and metalearners.

Load-bearing premise

The method's validity depends on the observational data containing all confounders, and on the reported distribution metrics actually testing the per-patient joint conditional distribution rather than only the pooled marginal distribution.

Editorial extensions

If this is right

  • Clinicians could compute probabilities of joint outcome patterns, such as tumor shrinkage together with a serious adverse event, instead of relying only on average treatment effects.
  • Because the model captures how outcomes co-vary, treatment decisions can weigh correlated efficacy and safety endpoints rather than treating them as independent.
  • The method handles mixed-type outcomes in one framework, so endpoints like 90-day mortality and length of hospital stay can be modeled jointly.
  • DIME can also be used for point estimation tasks; the paper reports lower error on the precision of estimating heterogeneous effects than the compared baselines on multiple outcomes.
  • The learned distribution provides explicit uncertainty quantification, enabling patient-specific risk-benefit thresholds for treatment decisions.

Reading between the lines

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

  • The same autoregressive decomposition could be extended to more than two treatment arms or to continuous treatment doses, since the diffusion conditioning structure does not depend on the treatment being binary.
  • A stronger test than pooled Wasserstein or KL divergence would be per-individual conditional calibration: fix $x$ and one observed outcome, then check whether the model's draws for the other outcome match the known conditional ground truth; this would isolate whether the dependence structure is actually learned.
  • With the full joint distribution, one can evaluate arbitrary medical utility functions, such as the probability that treatment benefit outweighs harm, which point estimates cannot support.
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

4 major / 6 minor

Summary. This paper proposes DIME, a score-based diffusion method for estimating the joint interventional distribution p(Y_1(a),...,Y_k(a)|X=x) of multiple medical outcomes from observational data. DIME factorizes the joint distribution via the chain rule into conditional distributions, trains conditional score-based diffusion models with causal and conditional masking to use only factual outcomes, and generates samples autoregressively. The authors evaluate DIME on a synthetic ACIC dataset and on IST and MIMIC-III, comparing Wasserstein distance, KL divergence, and PEHE against several baselines, and report state-of-the-art performance.

Significance. Conditional on the evaluation being corrected, the paper addresses a real gap: existing CATE methods focus on single outcomes or point estimates, and no existing neural method targets the joint multi-outcome interventional distribution. The chain-rule decomposition is exact, the conditional diffusion formulation is principled under unconfoundedness, the method handles mixed-type outcomes, and the code is released. If the conditional claim were verified with per-individual synthetic metrics, the method would be a useful contribution. However, as reported, the empirical evidence does not yet distinguish conditional performance from pooled-marginal performance, so the significance of the contribution is currently conditional on a revised evaluation.

major comments (4)
  1. [6.1.1 / Tables 1-2 / Fig. 3] The reported Wasserstein and KL metrics are computed between sets of n model samples and n data samples (pooled), not per patient. On IST and MIMIC each test patient has exactly one observed outcome vector, so no per-patient ground-truth conditional distribution p(Y_1(a),...,Y_k(a)|X=x) exists; a method that ignores X and matches only the marginal outcome distribution per treatment arm can achieve small pooled distances, especially on heterogeneous real data. The ACIC synthetic DGP is known and could provide per-patient ground-truth conditional samples, but the paper reports no per-individual metric there either. Therefore the real-data results in Tables 1-2 and the aggregated KDE in Fig. 3 do not support the paper's central conditional claim as written.
  2. [Table 3] The PEHE values for IST and MIMIC are reported without any explanation of how ground-truth individual treatment effects are obtained. PEHE requires counterfactual outcomes at the individual level; on observational MIMIC and even RCT IST, these are not observed and are not identifiable without additional assumptions. Unless the authors specify a valid ground-truth construction (e.g., a synthetic DGP, which is not described for these datasets), Table 3 does not support the CATE/PEHE claim.
  3. [5.4 / 6.1.1 / Tables 1-2] The paper claims support for mixed-type outcomes but does not define how the Wasserstein distance in Eq. (11) and the KDE-based KL divergence in Eq. (15) are computed over a product space of continuous and categorical variables. The Euclidean norm in Eq. (11) presumes a vector space, and density estimation for Eq. (15) over mixed support is not specified. Without this, the numerical magnitudes in Tables 1-2 are not interpretable for datasets with binary and continuous outcomes.
  4. [5.2.3] The statement that predictions are 'aggregated across multiple orderings σ by averaging over them' is undefined for autoregressive sampling. Averaging samples generated under different orderings in data space would not produce samples from the target joint distribution and could destroy dependencies. If the intent is to average densities or scores, that should be stated formally, and the effect on the generated joint distribution should be analyzed.
minor comments (6)
  1. [Abstract] The sentence 'We addresses three challenges relevant in medical practice' contains a subject-verb agreement error and should read 'We address three challenges...'.
  2. [5.2.2] The input mask m_i, target mask m_t, and conditional mask m^σ_c are described verbally but never defined formally in terms of how they enter the loss in Eq. (9) or the network architecture; please provide the exact masking equations.
  3. [6.6 / Fig. 3] The text refers to 'the ground truth distribution' for MIMIC, but no true interventional distribution exists for real data; the right panel should be labeled as the empirical observed outcome distribution, or the claim should be removed.
  4. [References] Reference [28] is the underlying birth/infant death data source, not the ACIC 2018 benchmarking study; please cite the actual ACIC 2018 reference or the relevant data-generation paper.
  5. [6.3.2] The IST outcome description 'probability of new strokes due to blocked vessels' leaves unclear whether this outcome is binary, ordinal, or continuous; the type matters for evaluating the mixed-type claims.
  6. [6.1] The paper claims uncertainty quantification but does not report calibration or coverage diagnostics (e.g., prediction intervals or reliability of the joint distribution); such a measure would strengthen the decision-making claims.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the chain-rule decomposition and conditional score-matching objective are self-contained; the self-citations are minor and non-load-bearing.

full rationale

The paper's derivation chain is not circular. The target estimand, p(Y_1(a),...,Y_k(a)|X=x) (Eq. 3), is defined independently, and the training objective (Eqs. 7-9, with Eq. 10 for discrete outcomes) is a standard conditional denoising score-matching loss applied to p(Y_i|Y_v,X,A); Eq. 4 is an exact chain-rule identity and Eq. 6 is its autoregressive inverse, so no fitted quantity is renamed as a prediction. The synthetic ACIC ground truth (Eq. 16) is generated independently of DIME, and the real-data metrics compare model samples to observed outcome sets, so the experiments do not reduce to the model's own definitions. The paper does contain self-citations with overlapping authors: DiffPO [26] is used as the main baseline and adapted for comparison, footnote 2 cites [29] for the non-identifiability of individual treatment effect distributions, and the architecture follows [55], a paper by a co-author. These are implementation or comparison details, not premises that force the central result, and no uniqueness theorem or prior result by the authors is invoked to forbid alternative choices. The main caveat is that the pooled Wasserstein/KL evaluation may not fully certify the conditional estimand, but that is an evaluation-support concern, not circular reasoning.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The method rests on standard causal assumptions, the chain rule, and neural approximation assumptions. No explicit scientific constants are fitted to data, and no new entities are introduced. The main unspecified choices are the diffusion SDE schedule and network hyperparameters, which are tuned by cross-validation and are not part of the causal claim.

free parameters (1)
  • Diffusion SDE coefficients f, g and loss weighting lambda(t) = not specified in the paper
    The paper states only that f and g are designed to converge to a Gaussian and lambda(t) is typically proportional to g(t)^2 (Sec. 5.3); the exact choice affects training but is standard, so it is a hand-chosen modeling choice rather than a fitted constant.
assumptions (4)
  • domain assumption Standard potential outcomes assumptions: consistency, unconfoundedness, overlap
    Invoked in Sec. 4 to identify interventional distributions from observational data.
  • standard math Chain rule of probability factorization
    Used in Eq. (4) to decompose the joint distribution into conditionals.
  • standard math Diffusion score matching regularity conditions, including tractable terminal distribution
    Needed for the reverse SDE in Eq. (8) to generate samples.
  • domain assumption Universal approximation of the conditional score network
    Assumes the neural network s_theta can represent all conditional distributions p(Y_i|Y_v,X,A) simultaneously.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Diffusion-Based Method for Learning the Multi-Outcome Distribution of Medical Treatments." pith.science (2026). https://pith.science/paper/SOJLKM5J

@misc{pith2026250601533,
  author       = {Pith},
  title        = {Pith review of: A Diffusion-Based Method for Learning the Multi-Outcome Distribution of Medical Treatments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SOJLKM5J}},
  note         = {Machine review of arXiv:2506.01533}
}
read the original abstract

In medicine, treatments often influence multiple, interdependent outcomes, such as primary endpoints, complications, adverse events, or other secondary endpoints. Hence, to make optimal treatment decisions, clinicians are interested in learning the distribution of multi-dimensional treatment outcomes. However, the vast majority of machine learning methods for predicting treatment effects focus on single-outcome settings, despite the fact that medical data often include multiple, interdependent outcomes. To address this limitation, we propose a novel diffusion-based method called DIME to learn the joint distribution of multiple outcomes of medical treatments. We addresses three challenges relevant in medical practice: (i)it is tailored to learn the joint interventional distribution of multiple medical outcomes, which enables reliable decision-making with uncertainty quantification rather than relying solely on point estimates; (ii)it explicitly captures the dependence structure between outcomes; (iii)it can handle outcomes of mixed type, including binary, categorical, and continuous variables. In DIME, we take into account the fundamental problem of causal inference through causal masking. For training, our method decomposes the joint distribution into a series of conditional distributions with a customized conditional masking to account for the dependence structure across outcomes. For inference, our method auto-regressively generates predictions. This allows our method to move beyond point estimates of causal quantities and thus learn the joint interventional distribution. To the best of our knowledge, DIME is the first neural method tailored to learn the joint, multi-outcome distribution of medical treatments. Across various experiments, we demonstrate that our method effectively learns the joint distribution and captures shared information among multiple outcomes.

Figures

Figures reproduced from arXiv: 2506.01533 by the authors.

Figure 1
Figure 1. Motivating example showing insights generated [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of our DIME. We decompose the joint interventional distribution into a series of conditional distributions [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Kernel density estimation plot of outcomes on [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 56 canonical work pages

  1. [1]

    Christopher RS Banerji, Tapabrata Chakraborti, Chris Harbron, and Ben D MacArthur. 2023. Clinical AI tools must convey predictive uncertainty for each individual patient. Nature Medicine 29, 12 (2023), 2996–2998

    work page 2023

  2. [2]

    Undral Byambadalai, Tatsushi Oka, and Shota Yasui. 2024. Estimating Distri- butional Treatment Effects in Randomized Experiments: Machine Learning for Variance Reduction. In International Conference on Machine Learning

    work page 2024

  3. [3]

    Chipman, Edward I

    Hugh A. Chipman, Edward I. George, and Robert E. McCulloch. 2010. BART: Bayesian additive regression trees. The Annals of Applied Statistics 4, 1 (March 2010), 266–298

    work page 2010

  4. [4]

    Alicia Curth and Mihaela van der Schaar. 2021. Nonparametric estimation of het- erogeneous treatment effects: From theory to learning algorithms. InInternational Conference on Artificial Intelligence and Statistics

    work page 2021

  5. [5]

    Alicia Curth and Mihaela van der Schaar. 2021. On inductive biases for hetero- geneous treatment effect estimation. Advances in Neural Information Processing Systems (2021)

    work page 2021

  6. [6]

    Matthieu Doutreligne, Tristan Struja, Judith Abecassis, Claire Morgand, Leo An- thony Celi, and Gaël Varoquaux. 2025. Step-by-step causal analysis of EHRs to ground decision-making. PLOS Digital Health 4, 2 (2025), e0000721

    work page 2025

  7. [7]

    J Feliu, V Heredia-Soto, R Gironés, B Jiménez-Munarriz, J Saldaña, C Guillén- Ponce, and MJ Molina-Garrido. 2020. Management of the toxicity of chemother- apy and targeted therapies in elderly cancer patients. Clinical and Translational Oncology 22 (2020), 457–467

    work page 2020

  8. [8]

    Stefan Feuerriegel, Dennis Frauen, Valentyn Melnychuk, Jonas Schweisthal, Kon- stantin Hess, Alicia Curth, Stefan Bauer, Niki Kilbertus, Isaac S Kohane, and Mihaela van der Schaar. 2024. Causal machine learning for predicting treatment outcomes. Nature Medicine 30, 4 (2024), 958–968

    work page 2024

Show all 59 references
  1. [9]

    Emil Frei III and George P Canellos. 1980. Dose: a critical factor in cancer chemotherapy. The American Journal of Medicine 69, 4 (1980), 585–594

    1980

  2. [10]

    Heckman, Jeffrey Smith, and Nancy Clements

    James J. Heckman, Jeffrey Smith, and Nancy Clements. 1997. Making the most out of programme evaluations and social experiments: Accounting for heterogeneity in programme impacts. The Review of Economic Studies 64, 4 (1997), 487–535

    1997

  3. [11]

    Konstantin Hess, Valentyn Melnychuk, Dennis Frauen, and Stefan Feuerriegel

  4. [12]

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

    2011

  5. [13]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems

    2020

  6. [14]

    Paul W Holland. 1986. Statistics and causal inference. Journal of the American Statistical Association 81, 396 (1986), 945–960

    1986

  7. [15]

    Johansson, Nathan Kallus, Uri Shalit, and David Sontag

    Fredrik D. Johansson, Nathan Kallus, Uri Shalit, and David Sontag. 2018. Learning weighted representations for generalization across designs.arXiv preprint (2018)

    2018

  8. [16]

    Johansson, Uri Shalit, and David Sontag

    Fredrik D. Johansson, Uri Shalit, and David Sontag. 2016. Learning representations for counterfactual inference. In International Conference on Machine Learning

    2016

  9. [17]

    Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. 2016. MIMIC-III, a freely accessible critical care database.Scientific Data 3 (2016), 160035

    2016

  10. [18]

    Mohammad Katout, Hong Zhu, Jessica Rutsky, Parthy Shah, Robert D Brook, Jixin Zhong, and Sanjay Rajagopalan. 2014. Effect of GLP-1 mimetics on blood pressure and relationship to weight loss and glycemia lowering: results of a systematic meta-analysis and meta-regression. Ameri...

    2014

  11. [19]

    Edward H. Kennedy. 2023. Towards optimal doubly robust estimation of hetero- geneous causal effects. Electronic Journal of Statistics 17, 2 (2023), 3008–3049

    2023

  12. [20]

    Edward H Kennedy, Shreya Kangovi, and Nandita Mitra. 2019. Estimating scaled treatment effects with multiple outcomes. Statistical Methods in Medical Research 28, 4 (2019), 1094–1104

    2019

  13. [21]

    Christoph Kern, Unai Fischer-Abaigar, Jonas Schweisthal, Dennis Frauen, Rayid Ghani, Stefan Feuerriegel, Mihaela van der Schaar, and Frauke Kreuter. 2025. Algorithms for reliable decision-making need causal reasoning. Nature Compu- tational Science (2025)

    2025

  14. [22]

    Thomas Kneib, Alexander Silbersdorff, and Benjamin Säfken. 2023. Rage against the mean: a review of distributional regression approaches. Econometrics and Statistics 26 (2023), 99–123

    2023

  15. [23]

    Künzel, Jasjeet S

    Sören R. Künzel, Jasjeet S. Sekhon, Peter J. Bickel, and Bin Yu. 2019. Metalearners for estimating heterogeneous treatment effects using machine learning. Proceed- ings of the National Academy of Sciences 116, 10 (2019), 4156–4165

    2019

  16. [24]

    Chaejeong Lee, Jayoung Kim, and Noseong Park. 2023. CoDi: Co-evolving contrastive diffusion models for mixed-type tabular synthesis. In International Conference on Machine Learning

    2023

  17. [25]

    Xihong Lin, Louise Ryan, Mary Sammel, Daowen Zhang, Chantana Padung- tod, and Xiping Xu. 2000. A scaled linear mixed model for multiple outcomes. Biometrics 56, 2 (2000), 593–601

    2000

  18. [26]

    Yuchen Ma, Valentyn Melnychuk, Jonas Schweisthal, and Stefan Feuerriegel

  19. [27]

    Basil Maag, Stefan Feuerriegel, Mathias Kraus, Maytal Saar-Tsechansky, and Thomas Züger. 2021. Modeling longitudinal dynamics of comorbidities. In Con- ference on Health, Inference, and Learning

    2021

  20. [28]

    MacDorman and J.O

    M.F. MacDorman and J.O. Atkinson. 1998. Infant mortality statistics from the linked birth/infant death data set–1995 period data.Monthly Vital Statistics Report 46, 6 Suppl 2 (1998), 1–22

    1998

  21. [29]

    Valentyn Melnychuk, Stefan Feuerriegel, and Mihaela van der Schaar. 2024. Quantifying aleatoric uncertainty of the treatment effect: a novel orthogonal learner. In Advances in Neural Information Processing Systems

    2024

  22. [30]

    Valentyn Melnychuk, Dennis Frauen, and Stefan Feuerriegel. 2023. Normaliz- ing flows for interventional density estimation. In International Conference on Machine Learning

    2023

  23. [31]

    Christof Naumzik, Stefan Feuerriegel, and Anne Molgaard Nielsen. 2023. Data- driven dynamic treatment planning for chronic diseases. European Journal of Operational Research (2023)

    2023

  24. [32]

    Christof Naumzik, Alice Kongsted, Werner Vach, and Stefan Feuerriegel. 2024. Data-driven subgrouping of patient trajectories with chronic diseases: Evidence from low back pain. In Proceedings of the Conference on Health, Inference, and Learning (CHIL)

    2024

  25. [33]

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

    2021

  26. [34]

    Robert L Page, Cindy L O’Bryant, Davy Cheng, Tristan J Dow, Bonnie Ky, C Michael Stein, Anne P Spencer, Robin J Trupp, and JoAnn Lindenfeld. 2016. Drugs that may cause or exacerbate heart failure: a scientific statement from the American Heart Association. Circulation 134, 6 (...

    2016

  27. [35]

    Jonas Peters, Dominik Janzing, and Bernhard Schölkopf. 2017. Elements of Causal Inference: Foundations and Learning Algorithms . MIT Press, Cambridge, MA

    2017

  28. [36]

    Jason Roy, Xihong Lin, and Louise M Ryan. 2003. Scaled marginal models for multiple continuous outcomes. Biostatistics 4, 3 (2003), 371–383

    2003

  29. [37]

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

    2005

  30. [38]

    Peter AG Sandercock, Maciej Niewada, Anna Członkowska, and International Stroke Trial Collaborative Group. 2011. The international stroke trial database. Trials 12, 1 (2011), 101

    2011

  31. [39]

    Johansson, and David Sontag

    Uri Shalit, Fredrik D. Johansson, and David Sontag. 2017. Estimating individ- ual treatment effect: Generalization bounds and algorithms. In International Conference on Machine Learning

    2017

  32. [40]

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli

  33. [41]

    Yang Song and Stefano Ermon. 2019. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems

    2019

  34. [42]

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2020. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations

    2020

  35. [43]

    David Spiegelhalter. 2017. Risk and uncertainty communication. Annual Review of Statistics and Its Application 4, 1 (2017), 31–60

    2017

  36. [44]

    Armando Teixeira-Pinto and Laura Mauri. 2011. Statistical analysis of noncom- mensurate multiple outcomes. Circulation: Cardiovascular Quality and Outcomes 4, 6 (2011), 650–656

    2011

  37. [45]

    Thai, Benjamin J

    Alesha A. Thai, Benjamin J. Solomon, Lecia V. Sequist, Justin F. Gainor, and Rebecca S. Heist. 2021. Lung cancer. The Lancet 398, 10299 (2021), 535–554

    2021

  38. [46]

    Anne Marthe Van Der Bles, Sander Van Der Linden, Alexandra LJ Freeman, James Mitchell, Ana B Galvao, Lisa Zaval, and David J Spiegelhalter. 2019. Communi- cating uncertainty about facts, numbers and science. Royal Society Open Science 6, 5 (2019), 181870

    2019

  39. [47]

    Stefan Wager and Susan Athey. 2018. Estimation and inference of heteroge- neous treatment effects using random forests. Journal of the American Statistical Association 113, 523 (2018), 1228–1242

    2018

  40. [48]

    Shirly Wang, Matthew BA McDermott, Geeticka Chauhan, Marzyeh Ghassemi, Michael C Hughes, and Tristan Naumann. 2020. Mimic-extract: A data extraction, preprocessing, and representation pipeline for MIMIC-III. In ACM Conference on Health, Inference, and Learning

    2020

  41. [49]

    Janick Weberpals, Stefan Feuerriegel, Mihaela van der Schaar, and Kenneth L. Kehl. 2025. Opportunities for causal machine learning in precision oncology. NEJM AI (2025)

    2025

  42. [50]

    Yong Wu, Mingzhou Liu, Jing Yan, Yanwei Fu, Shouyan Wang, Yizhou Wang, and Xinwei Sun. 2023. The blessings of multiple treatments and outcomes in treatment effect estimation. arXiv preprint (2023)

    2023

  43. [51]

    Leon Yao, Caroline Lo, Israel Nir, Sarah Tan, Ariel Evnine, Adam Lerer, and Alex Peysakhovich. 2022. Efficient heterogeneous treatment effect estimation with multiple experiments and multiple outcomes. arXiv preprint arXiv:2206.04907 (2022). A Diffusion-Based Method for Learni...

    2022 arXiv

  44. [52]

    Frank B Yoon, Garrett M Fitzmaurice, Stuart R Lipsitz, Nicholas J Horton, Nan M Laird, and Sharon-Lise T Normand. 2011. Alternative methods for testing treat- ment effects on the basis of multiple outcomes: simulation and case study. Statis- tics in medicine 30, 16 (2011), 1917–1932

    2011

  45. [53]

    Jinsung Yoon, James Jordon, and Mihaela van der Schaar. 2018. GANITE: Esti- mation of individualized treatment effects using generative adversarial nets. In International Conference on Learning Representations

    2018

  46. [54]

    Fernando G Zampieri, Jonathan D Casey, Manu Shankar-Hari, Frank E Harrell Jr, and Michael O Harhay. 2021. Using Bayesian methods to augment the interpre- tation of critical care trials. An overview of theory and example reanalysis of the alveolar recruitment for acute respirat...

    2021

  47. [55]

    Hengrui Zhang, Liancheng Fang, Qitian Wu, and Philip S Yu. 2024. Diffusion- nested auto-regressive synthesis of heterogeneous tabular data. arXiv preprint arXiv:2410.21523 (2024)

    2024 arXiv

  48. [56]

    𝜇1(𝑋,𝐴) 𝜇2(𝑋,𝐴) # , Σ(𝑋,𝐴) =

    Hengrui Zhang, Jiani Zhang, Balasubramaniam Srinivasan, Zhengyuan Shen, Xiao Qin, Christos Faloutsos, Huzefa Rangwala, and George Karypis. 2024. Mixed- type tabular data synthesis with score-based diffusion in latent space. In Interna- tional Conference on Learning Representat...

    2024

  49. [2015]

    In International Conference on Machine Learning

    Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning

  50. [2023]

    In International Conference on Learning Representations

    Bayesian neural controlled differential equations for treatment effect estimation. In International Conference on Learning Representations

  51. [2024]

    In Advances in Neural Information Processing Systems

    DiffPO: A causal diffusion model for learning distributions of potential outcomes. In Advances in Neural Information Processing Systems

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.