Pith. sign in

REVIEW 3 major objections 6 minor 62 references

Bayesian implementation of Targeted Maximum Likelihood Estimation for uncertainty quantification in causal effect estimation

T0 review · 3 major / 6 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 This paper claims that BN-TMLE, a Bayesian-network version of Targeted Maximum Likelihood Estimation, gives higher frequentist coverage of the true average treatment effect than classical TMLE at small sample sizes and comparable coverage…

desk verdict A well-meant Bayesian TMLE paper whose headline small-sample coverage claim rests on a misspecified data generator and a dubious duplicated-outcome likelihood. read the letter →

arxiv 2507.15909 v1 pith:73M6IUR6 submitted 2025-07-21 stat.ME stat.ML

Saideep Nannapaneni , Joseph Sakaya , Kyle Caron , Pedro HM Albuquerque , Zaid Tashman This is my paper
classification stat.MEstat.ML MSC 62F1562D20
keywords causalinferencetargetedmaximumlikelihoodestimationBayesiannetworkuncertaintyquantificationaveragetreatmenteffectdoublyrobustsmall-samplecoverageprobabilisticprogramming
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

The paper sets out to give Targeted Maximum Likelihood Estimation, a doubly robust method for estimating causal effects from observational data, a Bayesian treatment so that the average treatment effect (ATE) comes with a full posterior distribution rather than only a point estimate and a confidence interval. It proposes three Bayesian formulations: two that train the TMLE components in sequence using either means or mean-plus-standard-deviation summaries, and one that trains outcome, propensity, and fluctuation models together inside a Bayesian network (BN-TMLE). In simulation studies with a binary treatment and binary or continuous outcomes, the mean ATE estimates are similar across all methods, but BN-TMLE produces narrower credible intervals than the sequential Bayesian variants and its 95% intervals contain the true ATE more often than classical TMLE at small sample sizes, with performance converging as the sample grows. If this holds, researchers with small observational datasets could obtain honest uncertainty statements about causal effects earlier and at lower data-collection cost.

What carries the argument

The load-bearing object is a Bayesian network that unites the three TMLE components, outcome model, propensity model, and fluctuation model, into one joint posterior, with the propensity score and the clever covariate $H(A,X)=I(A=1)/P(A=1|X)-I(A=0)/P(A=0|X)$ as deterministic nodes and two separate outcome nodes for the outcome and fluctuation likelihoods, as in Equations 24 and 25. The paper duplicates the observed outcome node so that the same data train both the outcome model and the fluctuation model inside the graph; posterior samples of the parameters $\theta_Y$, $\theta_A$, and $\epsilon$ then generate interventional predictions under $do(A=1)$ and $do(A=0)$, and the column-wise difference of these predicted outcomes across posterior samples yields a posterior distribution for the ATE.

What would settle it

A direct check is to simulate data in which the outcome model is exactly correct and the true ATE is known, then compare the frequentist coverage of BN-TMLE's nominal 95% credible intervals with and without the duplicated Y node: if the duplication double-counts information, BN-TMLE intervals will under-cover at moderate sample sizes even though they may show favorable coverage at very small n simply because the intervals are wider there.

Watch

Extended reading notes

Core claim

The central claim is that a Bayesian-network version of TMLE can perform the targeting step, the fluctuation model that removes confounding bias, while respecting the probabilistic nature of the initial outcome predictions and clever covariates, and that this joint training pays off where classical TMLE struggles. Concretely, across 100 replications at each of 12 data sizes and across three model-misspecification regimes, the paper reports that the fraction of replications in which the true ATE falls inside the nominal 95% interval is generally higher for BN-TMLE than for classical TMLE at every sample size, with the gap largest and with non-overlapping confidence intervals at the smallest size of 25 data points. At larger sample sizes the mean estimates, interval widths, and coverage converge, and BN-TMLE's ATE distribution has lower variance than the two sequential Bayesian variants because the network propagates the full predictive distributions through the fluctuation model.

Load-bearing premise

The assumption that carries everything is that it is legitimate to treat the same observed outcome as two independent likelihood terms in one Bayesian update, because the network uses one copy of Y for the outcome model and another for the fluctuation model.

Editorial extensions

If this is right

  • BN-TMLE supplies a full posterior distribution for the average treatment effect, so decision makers can use credible intervals and probability statements rather than a single confidence interval.
  • At very small sample sizes, around 25 data points, BN-TMLE's interval captures the true ATE more often than classical TMLE, supporting earlier causal conclusions when data collection is expensive.
  • At larger sample sizes, above about 200 points, coverage and interval widths converge toward classical TMLE, and BN-TMLE shows lower uncertainty in the interval width itself.
  • The two sequential Bayesian variants give similar mean ATEs but wider intervals than BN-TMLE, indicating that joint training through the Bayesian network is the preferred formulation among the three in these experiments.
  • One-parameter and two-parameter fluctuation models produce similar ATE distributions, so the simpler one-parameter fluctuation model is sufficient in the binary-treatment setting studied.

Reading between the lines

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

  • Beyond the paper, the validity of the duplicated outcome node is the point to watch: if two likelihoods over the same observed outcome are not independent, the reported BN-TMLE interval widths mix genuine uncertainty with double-counted information, and the apparent small-sample advantage could be a width effect rather than an efficiency gain.
  • Beyond the paper, the same Bayesian-network construction should transfer to continuous treatments, survival outcomes, or longitudinal TMLE, where the targeting step faces the same point-observation versus distributional-input mismatch; the paper lists these as future work without testing them.
  • Beyond the paper, a direct comparison against a fully Bayesian doubly robust estimator that avoids duplicating the outcome likelihood would separate the contribution of the network structure from the contribution of the priors.
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 / 6 minor

Summary. The paper proposes three Bayesian implementations of Targeted Maximum Likelihood Estimation (TMLE) for causal effect estimation with binary treatment and binary or continuous outcomes: two sequential approaches (B-TMLE-M and B-TMLE-SS) that propagate mean or summary-statistic predictions into the fluctuation model, and a joint Bayesian network approach (BN-TMLE) that trains the outcome, propensity, and fluctuation models simultaneously. The authors demonstrate the approaches on simulated examples, compare point estimates and interval coverage against classical TMLE, and report an extensive small-sample study of data size and model misspecification. The central claim is that BN-TMLE outperforms classical TMLE in small data regimes on frequentist coverage of the true ATE while performing similarly in larger samples, and that all three Bayesian variants provide a full posterior distribution of the ATE.

Significance. If the claims were supported, the paper would offer a practically useful extension of TMLE by providing Bayesian uncertainty quantification and a credible-interval alternative in small-sample settings, with the BN-TMLE variant being a conceptually interesting joint formulation. The paper is transparent about priors, MCMC settings, and simulation design, and it uses a standard probabilistic programming stack (NumPyro/HMC), which aids reproducibility. However, the headline comparative claim rests on a simulation generator that is not a valid Bernoulli data-generating process for the larger effect size and is itself misspecified even in the labeled 'no misspecification' condition, and the BN-TMLE model duplicates the outcome node and thereby multiplies two likelihoods for the same observation without justification. These issues put the main empirical and methodological conclusions on unsupported ground.

major comments (3)
  1. [Section 4.2, Eqs. (29) and (31)] The 'no model misspecification' (NMS) generator is itself misspecified relative to the fitted first-order logistic model, and the generator is not a valid Bernoulli specification for ψ=0.15. Equations (29) and (31) define P(Y1=1)=P(Y0=1)+ψ, i.e., the treatment effect is additive on the probability scale, while the fitted outcome model is logit-linear in A. The implied logit difference is log((P0+ψ)/(1−P0−ψ)) − logit(P0), which is not a constant coefficient on A unless ψ=0; hence even the NMS condition is misspecified. Moreover, for ψ=0.15 and covariate values such as X1=0, X2=2, and X3≳1.6, P0+ψ exceeds 1, so the Bernoulli sampling in Eqs. (29) and (31) is undefined unless probabilities are clamped, and any clamping changes the induced ATE away from ψ. Since coverage in Eq. (33) is computed against ψ rather than the actual expected counterfactual difference under the generator, the reported coverage percentages and the labels NMS/OMS/OPMS do not measure what they claim. This directly undermines the headline small-sample coverage comparison in Figures 8 and 9.
  2. [Section 3.2, Eqs. (24)–(25), Figures 1–3] The BN-TMLE likelihood duplicates the observed outcome node Y, conditioning it on both Bernoulli(Yinit) and Bernoulli(Yf). This multiplies two likelihood factors for the same data point and is not a valid joint model for the observed Y; the two factors are consistent only when Yinit=Yf. The paper states that 'we duplicate the node Y so we can train the outcome and fluctuation model directly,' but it provides no justification in terms of pseudo-likelihood theory, data augmentation, or posterior calibration. As a result, the BN-TMLE posterior and its credible intervals (e.g., [0.020, 0.065] versus [0.013, 0.074] for B-TMLE-M in Table 1) are not interpretable as Bayesian posterior intervals, and the narrower BN-TMLE intervals may reflect double-counted information rather than statistical efficiency. This is load-bearing for the paper's central uncertainty-quantification claim.
  3. [Section 4.2, Discussion bullets] The paper itself reports that at small data sizes the average width of BN-TMLE intervals is higher than that of classical TMLE ('the average width from BN-TMLE ... is higher' and 'wider confidence intervals of ATE'), and that the coverage advantage is 'evidently more in small data sizes.' When point estimates are nearly identical, wider intervals mechanically increase coverage; the paper has therefore not demonstrated that BN-TMLE's higher coverage reflects better calibration rather than simply more conservative intervals. The 'outperformed' conclusion needs to be supported by a width-adjusted comparison, a proper scoring rule, or an explicit discussion of the coverage-width trade-off, rather than by raw coverage alone.
minor comments (6)
  1. [Eq. (8)] The equation 'Y = h(X; θ + ξ' appears to be missing a closing parenthesis; it should presumably read 'Y = h(X; θ) + ξ'.
  2. [Eq. (33)] The word 'Converage' is a typo for 'Coverage'.
  3. [Sections 3.1.3 and 3.3] The abbreviation 'SB-TMLE' is used in Section 3.3 but the methods are elsewhere called B-TMLE-M and B-TMLE-SS; the terminology should be made consistent.
  4. [Section 3.2, text before Eq. (24)] The phrase 'unlike the the first two approaches' duplicates the word 'the'; it should read 'unlike the first two approaches.'
  5. [Section 5 and Section 4.1, bullet 3] There are spelling errors: 'valies' should be 'values' in the concluding paragraph, and 'appraoaches' should be 'approaches' in the third observation of Section 4.1.
  6. [Section 4.3, caption of Figure 10] The caption refers to 'three datasets' from Section 4.1, but Section 4.1 describes only two generated case studies (binary and continuous outcomes); the intended referent should be clarified.

Circularity Check

2 steps flagged · score 6.0 of 10

Coverage superiority is a restatement of interval width, and BN-TMLE's narrow intervals are built into the duplicated-Y design.

  1. renaming known result [Section 4.2, Discussion (Figures 8-9, coverage panels)]
    "At small data sizes, the average width from BN-TMLE (red dots in width variation plots) is higher when compared to classical TMLE (blue dots). ... This indicates that BN-TMLE at small data sizes, on average, produces wider confidence intervals of ATE when compared to classical implementation. ... BN-TMLE outperformed classical TMLE in coverage, evidently more in small data sizes. ... The wider 95% CI that BN-TMLE produces in small data regimes acts as a safeguard against making an inaccurate causal conclusion in small data regimes."

    Coverage in Equation 33 is the fraction of replications in which the true ATE falls inside the predicted 95% CI. For two intervals with similar centers (both methods give mean ATE near 0.043 in the binary case), the wider interval has higher coverage by construction. The paper's own width panels show that BN-TMLE's average CI width is higher at small data sizes, so the headline conclusion that BN-TMLE outperformed classical TMLE in coverage is simply restating that BN-TMLE produces wider intervals. The claimed small-sample advantage is therefore not an independent empirical discovery; it reduces to a definitional property of interval coverage.

  2. self definitional [Section 3.2, Equations 24-25 and text following Equation 25]
    "Note that since the same outcome data ( Y ) is used for training the outcome and fluctuation models, we duplicate the node Y so we can train the outcome and fluctuation model directly."

    Equations 24 and 25 define two likelihood terms for the same observed outcome: Y|Yinit ~ Bernoulli(Yinit) and Y|Yf ~ Bernoulli(Yf). The BN-TMLE posterior multiplies these terms, so the same data point contributes twice. This duplication is baked into the model definition, and it directly produces the narrower credible intervals reported in Table 1 (BN-TMLE [0.02, 0.065] vs B-TMLE-M [0.013, 0.074]). The paper's conclusion that BN-TMLE provides the most accurate results because of its smaller variance is thus a consequence of the duplicated-node construction, not evidence of superior statistical efficiency.

full rationale

The paper does not fit parameters to the reported truth, and there is no load-bearing self-citation chain: the Bayesian TMLE constructions are assembled from standard TMLE and Bayesian ingredients, and the simulations use known ATE values. The reason the score is not 0 is that the central comparative claim has two by-construction components. First, Section 4.2 presents BN-TMLE's higher coverage as outperforming classical TMLE, but the paper itself reports that BN-TMLE's intervals are wider at small n; since coverage counts whether the truth falls in the interval, a wider interval mechanically raises coverage, so the claimed small-sample superiority is a restatement of interval width rather than an independent empirical finding. Second, BN-TMLE's narrower intervals relative to B-TMLE-M and B-TMLE-SS are built into the model: Equations 24 and 25 both place a likelihood on the same observed Y via the duplicated node, so the posterior double-counts the data; the lower variance in Table 1 is a construction artifact. A separate correctness concern, not counted here as circularity, is that the Section 4.2 no-model-misspecification generator adds psi on the probability scale while the fitted models are logistic, and for psi=0.15 can produce probabilities above 1, so the benchmark labels and coverage target are questionable. These issues affect the validity of the empirical claims, but the core algorithm is not a hidden fit to its own output.

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

The central results rest on hand-chosen Bayesian priors, two standard domain conditions for TMLE, an ad hoc duplicated-outcome likelihood, and an unverified MCMC convergence assumption. The prior hyperparameters are chosen by hand rather than fitted to truth, but they shape the location and width of the reported ATE intervals, and no sensitivity analysis is given. The duplicated outcome node is the only newly introduced modeling device and it carries no independent evidential support.

free parameters (4)
  • Priors for outcome and propensity model coefficients (standard Gaussian, given as MVN(0, I)) = N(0, 1) per coefficient, hand-chosen
    Section 4 states standard Gaussian priors for all parameters. The prior scale controls shrinkage in the outcome and propensity posterior predictions, which feed the fluctuation step and the final ATE distribution. No sensitivity analysis is reported.
  • Fluctuation parameter prior (epsilon ~ Normal(0, 1)) = N(0, 1), hand-chosen
    The prior on epsilon governs how strongly the targeting step can correct the initial outcome predictions. A unit variance is chosen without justification on the logit scale, and it directly shapes the ATE posterior in all three proposed approaches.
  • HalfNormal(1) priors on error standard deviations (sigma_xi, sigma_o, sigma_f) = HalfNormal(scale=1), hand-chosen
    Used for continuous-outcome likelihoods in Equations 15, 20, 22, 26 and 27. The scale affects the outcome and fluctuation likelihoods and therefore the width of the reported ATE credible intervals.
  • Latent input priors in B-TMLE-SS (Yinit,i and H_i) = Posterior predictive means and standard deviations from the outcome and propensity models
    Equations 21 and 22 assign Normal priors to the latent true inputs of the fluctuation model using the posterior predictive moments of models already trained on the same data. This two-stage reuse of data is an approximation whose effect on interval calibration is not assessed.
assumptions (4)
  • domain assumption Double robustness conditions: at least one of the outcome or propensity models is correctly specified, and the propensity score is bounded away from 0 and 1
    Standard TMLE theory cited in Section 2.1 requires these for unbiased ATE estimation. In the OPMS case of Section 4.2 both models are misspecified, so neither classical TMLE nor BN-TMLE is guaranteed consistent; the comparison there measures robustness, not unbiasedness.
  • ad hoc to paper The duplicated outcome node Y in BN-TMLE defines a valid joint likelihood
    Section 3.2 (Equations 24-25, Figures 1-3) multiplies two likelihood terms for the same observed outcome without deriving the model from a coherent generative story. This is the central premise of the BN-TMLE posterior and it is asserted, not justified.
  • domain assumption Convergence of NUTS in all analyses with 2 chains, 1000 burn-in draws, and 2000 posterior draws
    Sections 4.1 and 4.2 report these settings but no R-hat, effective sample size, or trace diagnostics, so the validity of every reported posterior and credible interval assumes convergence across the roughly 7,200 fitted analyses.
  • domain assumption Identifiability of the ATE under the simulated DAG (no unmeasured confounding)
    Section 4.1 generates data from the DAG in Figure 4 where X1, X2, X3 are the only parents of A and Y, and the analysis models use these same confounders, so unconfoundedness holds by construction in the simulations.
invented entities (1)
  • Duplicated outcome node Y in BN-TMLE (the two Y nodes in Figures 1-3)
    purpose: Allows the same outcome data to drive both the outcome-model likelihood (Y|Yinit) and the fluctuation-model likelihood (Y|Yf) during joint Bayesian training
    The paper introduces no new physical or probabilistic entity with external evidence; the duplicated node is a modeling device that uses the outcome twice in the likelihood, which can double-count information and inflate posterior precision. There is no falsifiable handle outside the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bayesian implementation of Targeted Maximum Likelihood Estimation for uncertainty quantification in causal effect estimation." pith.science (2026). https://pith.science/paper/73M6IUR6

@misc{pith2026250715909,
  author       = {Pith},
  title        = {Pith review of: Bayesian implementation of Targeted Maximum Likelihood Estimation for uncertainty quantification in causal effect estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/73M6IUR6}},
  note         = {Machine review of arXiv:2507.15909}
}
read the original abstract

Robust decision making involves making decisions in the presence of uncertainty and is often used in critical domains such as healthcare, supply chains, and finance. Causality plays a crucial role in decision-making as it predicts the change in an outcome (usually a key performance indicator) due to a treatment (also called an intervention). To facilitate robust decision making using causality, this paper proposes three Bayesian approaches of the popular Targeted Maximum Likelihood Estimation (TMLE) algorithm, a flexible semi-parametric double robust estimator, for a probabilistic quantification of uncertainty in causal effects with binary treatment, and binary and continuous outcomes. In the first two approaches, the three TMLE models (outcome, treatment, and fluctuation) are trained sequentially. Since Bayesian implementation of treatment and outcome yields probabilistic predictions, the first approach uses mean predictions, while the second approach uses both the mean and standard deviation of predictions for training the fluctuation model (targeting step). The third approach trains all three models simultaneously through a Bayesian network (called BN-TMLE in this paper). The proposed approaches were demonstrated for two examples with binary and continuous outcomes and validated against classical implementations. This paper also investigated the effect of data sizes and model misspecifications on causal effect estimation using the BN-TMLE approach. Results showed that the proposed BN-TMLE outperformed classical implementations in small data regimes and performed similarly in large data regimes.

Figures

Figures reproduced from arXiv: 2507.15909 by the authors.

Figure 1
Figure 1. Bayesian network combining the outcome, treatment, and fluctuation models of [PITH_FULL_IMAGE:figures/full_fig_p023_1.png] view at source ↗
Figure 2
Figure 2. Bayesian network combining the outcome, treatment, and fluctuation models of [PITH_FULL_IMAGE:figures/full_fig_p025_2.png] view at source ↗
Figure 3
Figure 3. Bayesian network TMLE formulations considering two-parameter fluctuation [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: A directed acyclic graphical model (DAG) showing dependence between con [PITH_FULL_IMAGE:figures/full_fig_p031_4.png]
Figure 5
Figure 5. Figure 5: Distributions of outcome data across control and treatment groups for binary [PITH_FULL_IMAGE:figures/full_fig_p033_5.png]
Figure 6
Figure 6. Figure 6: Distributions of propensity scores across control and treatment groups for binary [PITH_FULL_IMAGE:figures/full_fig_p035_6.png]
Figure 7
Figure 7. Figure 7: Comparison of ATE distribution using the three Bayesian TMLE approaches (B [PITH_FULL_IMAGE:figures/full_fig_p035_7.png]
Figure 8
Figure 8. Figure 8: Comparison of variation of mean estimate, variation in width of 95% confidence [PITH_FULL_IMAGE:figures/full_fig_p040_8.png]
Figure 9
Figure 9. Figure 9: Comparison of variation of mean estimate, variation in width of 95% confidence [PITH_FULL_IMAGE:figures/full_fig_p041_9.png]
Figure 10
Figure 10. Figure 10: We should note that both the one-parameter and two-parameter [PITH_FULL_IMAGE:figures/full_fig_p044_10.png]
Figure 10
Figure 10. Figure 10: Comparison of ATE distributions using the BN-TMLE framework with one [PITH_FULL_IMAGE:figures/full_fig_p045_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 53 canonical work pages

  1. [1]

    Acharki, N., Lugo, R., Bertoncello, A., and Garnier, J. (2023). Comparison of meta-learners for estimating multi-valued treatment heterogeneous effects. In International conference on machine learning , pages 91--132. PMLR

    work page 2023

  2. [2]

    Aho, K., Derryberry, D., and Peterson, T. (2014). Model selection for ecologists: the worldviews of aic and bic. Ecology , 95(3):631--636

    work page 2014

  3. [3]

    Athey, S. (2015). Machine learning and causal inference for policy evaluation. In Proceedings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining , pages 5--6

    work page 2015

  4. [4]

    and Robins, J

    Bang, H. and Robins, J. M. (2005). Doubly robust estimation in missing data and causal inference models. Biometrics , 61(4):962--973

    2005

  5. [5]

    Beaumont, M. A. (2010). Approximate bayesian computation in evolution and ecology. Annual review of ecology, evolution, and systematics , 41:379--406

    work page 2010

  6. [6]

    P., Jankowiak, M., Obermeyer, F., Pradhan, N., Karaletsos, T., Singh, R., Szerlip, P., Horsfall, P., and Goodman, N

    Bingham, E., Chen, J. P., Jankowiak, M., Obermeyer, F., Pradhan, N., Karaletsos, T., Singh, R., Szerlip, P., Horsfall, P., and Goodman, N. D. (2019). Pyro: Deep universal probabilistic programming. Journal of machine learning research , 20(28):1--6

    work page 2019

  7. [7]

    D., Lee, D., Goodrich, B., Betancourt, M., Brubaker, M., Guo, J., Li, P., and Riddell, A

    Carpenter, B., Gelman, A., Hoffman, M. D., Lee, D., Goodrich, B., Betancourt, M., Brubaker, M., Guo, J., Li, P., and Riddell, A. (2017). Stan: A probabilistic programming language. Journal of statistical software , 76:1--32

    work page 2017

  8. [8]

    W., Kumara, S., and Yang, H

    Chen, R., Lu, Y., Witherell, P., Simpson, T. W., Kumara, S., and Yang, H. (2021). Ontology-driven learning of bayesian network for causal inference and quality assurance in additive manufacturing. IEEE Robotics and Automation Letters , 6(3):6032--6038

    work page 2021

Show all 62 references
  1. [9]

    Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., and Robins, J. (2018). Double/debiased machine learning for treatment and structural parameters

    2018

  2. [10]

    Chernozhukov, V., Hansen, C., Kallus, N., Spindler, M., and Syrgkanis, V. (2024). Applied causal inference powered by ml and ai. arXiv preprint arXiv:2403.02467

    2024 arXiv

  3. [11]

    Chib, S. (2001). Markov chain monte carlo methods: computation and inference. Handbook of econometrics , 5:3569--3649

    2001

  4. [12]

    M., Crist \'o bal, V., and Sant \' n, D

    Cordero, J. M., Crist \'o bal, V., and Sant \' n, D. (2018). Causal inference on education policies: A survey of empirical studies using pisa, timss and pirls. Journal of Economic Surveys , 32(3):878--915

    2018

  5. [13]

    Dai, H., Pollock, M., and Roberts, G. O. (2023). Bayesian fusion: Scalable unification of distributed statistical analyses. Journal of the Royal Statistical Society Series B: Statistical Methodology , 85(1):84--107

    2023

  6. [14]

    J., Louis, T

    Franco, C., Little, R. J., Louis, T. A., and Slud, E. V. (2019). Comparative study of confidence intervals for proportions in complex sample surveys. Journal of survey statistics and methodology , 7(3):334--364

    2019

  7. [15]

    Frank, H. A. and Karim, M. E. (2024). Implementing tmle in the presence of a continuous outcome. Research Methods in Medicine & Health Sciences , 5(1):8--19

    2024

  8. [16]

    J., Westreich, D., Wiesen, C., St \"u rmer, T., Brookhart, M

    Funk, M. J., Westreich, D., Wiesen, C., St \"u rmer, T., Brookhart, M. A., and Davidian, M. (2011). Doubly robust estimation of causal effects. American journal of epidemiology , 173(7):761--767

    2011

  9. [17]

    and Lopes, H

    Gamerman, D. and Lopes, H. F. (2006). Markov chain Monte Carlo: stochastic simulation for Bayesian inference . Chapman and Hall/CRC

    2006

  10. [18]

    Gangl, M. (2010). Causal inference in sociological research. Annual review of sociology , 36(1):21--47

    2010

  11. [19]

    Geiger, D., Verma, T., and Pearl, J. (1990). Identifying independence in bayesian networks. Networks , 20(5):507--534

    1990

  12. [20]

    B., Stern, H

    Gelman, A., Carlin, J. B., Stern, H. S., and Rubin, D. B. (1995). Bayesian data analysis . Chapman and Hall/CRC

    1995

  13. [21]

    A., Goodman, S

    Glass, T. A., Goodman, S. N., Hern \'a n, M. A., and Samet, J. M. (2013). Causal inference in public health. Annual review of public health , 34(1):61--75

    2013

  14. [22]

    D., Henzinger, T

    Gordon, A. D., Henzinger, T. A., Nori, A. V., and Rajamani, S. K. (2014). Probabilistic programming. In Future of software engineering proceedings , pages 167--181. ACM

    2014

  15. [23]

    Gray, K., Hampton, B., Silveti-Falls, T., McConnell, A., and Bausell, C. (2015). Comparison of bayesian credible intervals to frequentist confidence intervals. Journal of Modern Applied Statistical Methods , 14(1):8

    2015

  16. [24]

    and van der Laan, M

    Gruber, S. and van der Laan, M. (2012). tmle: an r package for targeted maximum likelihood estimation. Journal of Statistical Software , 51:1--35

    2012

  17. [25]

    Hair Jr, J. F. and Sarstedt, M. (2021). Data, measurement, and causal inferences in machine learning: opportunities and challenges for marketing. Journal of Marketing Theory and Practice , 29(1):65--77

    2021

  18. [26]

    S., Costa, L

    Hespanhol, L., Vallio, C. S., Costa, L. M., and Saragiotto, B. T. (2019). Understanding and interpreting confidence and credible intervals around effect estimates. Brazilian journal of physical therapy , 23(4):290--301

    2019

  19. [27]

    Ho, T.-H., Lim, N., Reza, S., and Xia, X. (2017). Om forum—causal inference models in operations management. Manufacturing & Service Operations Management , 19(4):509--525

    2017

  20. [28]

    D., Blei, D

    Hoffman, M. D., Blei, D. M., Wang, C., and Paisley, J. (2013). Stochastic variational inference. the Journal of machine Learning research , 14(1):1303--1347

    2013

  21. [29]

    D., Gelman, A., et al

    Hoffman, M. D., Gelman, A., et al. (2014). The no-u-turn sampler: adaptively setting path lengths in hamiltonian monte carlo. J. Mach. Learn. Res. , 15(1):1593--1623

    2014

  22. [30]

    and Mailhot, A

    Huard, D. and Mailhot, A. (2006). A bayesian perspective on input uncertainty in model calibration: Application to hydrological model “abc”. Water Resources Research , 42(7)

    2006

  23. [31]

    Kaur, P., Polyzou, A., and Karypis, G. (2019). Causal inference in higher education: Building better curriculums. In Proceedings of the Sixth (2019) ACM Conference on Learning@ Scale , pages 1--4

    2019

  24. [32]

    K., Constantinou, A

    Kitson, N. K., Constantinou, A. C., Guo, Z., Liu, Y., and Chobtham, K. (2023). A survey of bayesian network structure learning. Artificial Intelligence Review , 56(8):8721--8814

    2023

  25. [33]

    Kochenderfer, M. J. (2015). Decision making under uncertainty: theory and application . MIT press

    2015

  26. [34]

    and Friedman, N

    Koller, D. and Friedman, N. (2009). Probabilistic graphical models: principles and techniques . MIT press

    2009

  27. [35]

    R., Sekhon, J

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

    2019

  28. [36]

    Kurz, C. F. (2022). Augmented inverse probability weighting and the double robustness property. Medical Decision Making , 42(2):156--167

    2022

  29. [37]

    Li, M. (2013). Using the propensity score method to estimate causal effects: A review and practical guide. Organizational Research Methods , 16(2):188--226

    2013

  30. [38]

    A., Schomaker, M., Rachet, B., and Schnitzer, M

    Luque-Fernandez, M. A., Schomaker, M., Rachet, B., and Schnitzer, M. E. (2018). Targeted maximum likelihood estimation for a binary treatment: A tutorial. Statistics in medicine , 37(16):2530--2546

    2018

  31. [39]

    Meng, Y., Zhang, S., Sun, Y., Zhang, R., Hu, Z., Zhang, Y., Jia, C., Wang, Z., and Pei, D. (2020). Localizing failure root causes in a microservice through causality inference. In 2020 IEEE/ACM 28th International Symposium on Quality of Service (IWQoS) , pages 1--10. IEEE

    2020

  32. [40]

    Mithas, S., Chen, Y., Lin, Y., and De Oliveira Silveira, A. (2022). On the causality and plausibility of treatment effects in operations management research. Production and Operations Management , 31(12):4558--4571

    2022

  33. [41]

    Naser, M. (2022). Causality, causal discovery, and causal inference in structural engineering. arXiv preprint arXiv:2204.01543

    2022 arXiv

  34. [42]

    Patil, A., Huard, D., and Fonnesbeck, C. J. (2010). Pymc: Bayesian stochastic modelling in python. Journal of statistical software , 35:1--81

    2010

  35. [43]

    Pearl, J. (1994). A probabilistic calculus of actions. In Uncertainty in artificial intelligence , pages 454--462. Elsevier

    1994

  36. [44]

    Phan, D., Pradhan, N., and Jankowiak, M. (2019). Composable effects for flexible and accelerated probabilistic programming in numpyro. arXiv preprint arXiv:1912.11554

    2019 arXiv

  37. [45]

    S., Min, J

    Prosperi, M., Guo, Y., Sperrin, M., Koopman, J. S., Min, J. S., He, X., Rich, S., Wang, M., Buchan, I. E., and Bian, J. (2020). Causal inference and counterfactual prediction in machine learning for actionable healthcare. Nature Machine Intelligence , 2(7):369--375

    2020

  38. [46]

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

    2005

  39. [47]

    P., Xia, T., Watson, H

    Sanchez, P., Voisey, J. P., Xia, T., Watson, H. I., O’Neil, A. Q., and Tsaftaris, S. A. (2022). Causal machine learning for healthcare and precision medicine. Royal Society Open Science , 9(8):220638

    2022

  40. [48]

    Scanagatta, M., Salmer \'o n, A., and Stella, F. (2019). A survey on bayesian network structure learning from data. Progress in Artificial Intelligence , 8(4):425--439

    2019

  41. [49]

    Schuler, M. S. and Rose, S. (2017). Targeted maximum likelihood estimation for causal inference in observational studies. American journal of epidemiology , 185(1):65--73

    2017

  42. [50]

    and Stoltz, G

    Sekkat, I. and Stoltz, G. (2021). Quantifying the mini-batching error in bayesian inference for adaptive langevin dynamics. arXiv preprint arXiv:2105.10347

    2021 arXiv

  43. [51]

    and Norgeot, B

    Shi, J. and Norgeot, B. (2022). Learning causal effects from observational data in healthcare: a review and summary. Frontiers in Medicine , 9:864882

    2022

  44. [52]

    and Kawahara, T

    Shiba, K. and Kawahara, T. (2021). Using propensity scores for causal inference: pitfalls and tips. Journal of epidemiology , 31(8):457--463

    2021

  45. [53]

    Siebert, J. (2023). Applications of statistical causal inference in software engineering. Information and Software Technology , 159:107198

    2023

  46. [54]

    J., Phillips, R

    Smith, M. J., Phillips, R. V., Luque-Fernandez, M. A., and Maringe, C. (2023). Application of targeted maximum likelihood estimation in public health and epidemiological studies: a systematic review. Annals of epidemiology

    2023

  47. [55]

    D., Saurous, R

    Tran, D., Hoffman, M. D., Saurous, R. A., Brevdo, E., Murphy, K., and Blei, D. M. (2017). Deep probabilistic programming. arXiv preprint arXiv:1701.03757

    2017 arXiv

  48. [56]

    C., Houston, A

    Trimmer, P. C., Houston, A. I., Marshall, J. A., Mendl, M. T., Paul, E. S., and McNamara, J. M. (2011). Decision-making under uncertainty: biases and bayesians. Animal cognition , 14:465--476

    2011

  49. [57]

    J., Polley, E

    Van der Laan, M. J., Polley, E. C., and Hubbard, A. E. (2007). Super learner. Statistical applications in genetics and molecular biology , 6(1)

    2007

  50. [58]

    J., Rose, S., et al

    Van der Laan, M. J., Rose, S., et al. (2011). Targeted learning: causal inference for observational and experimental data , volume 4. Springer

    2011

  51. [59]

    Van Der Laan, M. J. and Rubin, D. (2006). Targeted maximum likelihood learning. The international journal of biostatistics , 2(1)

    2006

  52. [60]

    Varian, H. R. (2016). Causal inference in economics and marketing. Proceedings of the National Academy of Sciences , 113(27):7310--7315

    2016

  53. [61]

    Yao, L., Chu, Z., Li, S., Li, Y., Gao, J., and Zhang, A. (2021). A survey on causal inference. ACM Transactions on Knowledge Discovery from Data (TKDD) , 15(5):1--46

    2021

  54. [62]

    and Feng, J

    Zou, C. and Feng, J. (2009). Granger causality vs. dynamic bayesian network inference: a comparative study. BMC bioinformatics , 10:1--17

    2009

Pith tools

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