Balancing Weights for Causal Mediation Analysis

arXiv:2512.09337 · stat.ME, econ.EM · Submitted 2025-12-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Balancing Weights for Causal Mediation Analysis".

Jane: Here's the gist of the paper. This paper develops methods for estimating natural direct and indirect effects in causal mediation analysis by proposing a novel two-step minimal weights estimator that addresses finite-sample instability and covariate imbalance inherent in standard estimators like the EIF–based estimator and IPW estimator.

Tom: Pruning redundant reasoning tokens The paper introduces estimators based on weights obtained by an algorithm that directly penalizes weight dispersion while enforcing approximate covariate and mediator balance, thereby improving stability and mitigating bias in finite samples.

Jane: The authors establish the convergence rates of the proposed weights and show that the resulting estimators are asymptotically normal and achieve the semiparametric efficiency bound.

Tom: Monte Carlo simulations demonstrate that the proposed estimator outperforms not only the EIF–based estimator and the IPW estimator but also the regression imputation estimator in challenging scenarios with model misspecification.

Paper summary: Tom: We're moving on to the paper "Balancing Weights for Causal Mediation Analysis" to summarize the main ideas we just touched on. The core thesis is that standard methods like EIF based and IPW estimator are vulnerable because they rely too heavily on the inverse of estimated propensity scores <ref:2512.09337#pg1>.

Jane: Which makes them susceptible to two specific problems in finite samples: instability when those scores get close to zero or one, and covariate imbalance across the different data groups <ref:2512.09337#pg3>.

Lu: The authors introduce a novel weighting estimator based on an algorithm that explicitly penalizes weight dispersion while simultaneously trying to enforce approximate balance on both the covariates and the mediator <ref:2512.09337#pg1>.

Meng: So, the central claim is that this new approach improves stability and reduces bias in finite samples by directly addressing these two specific imbalances <ref:2512.09337#pg4>.

Tom: And they back this up with convergence rates, showing the resulting estimators are asymptotically normal and achieving the semiparametric efficiency bound <ref:2512.09337#pg1>.

Jane: That means theoretically, the new estimator is as good as it can possibly be under these conditions <ref:2512.09337#pg1>.

Lu: They also show that through Monte Carlo simulations that this proposed estimator outperforms the EIF based and IPW estimator even when things get really messy with model misspecification <ref:2512.09337#pg1>.

Meng: So, we're essentially saying this new weighting scheme is more robust than the existing ones in challenging situations <ref:2512.09337#pg1>.

Tom: And they used real data from a study examining the effects of media framing on immigration attitudes to show it works in practice <ref:2512.09337#pg1>.

Jane: It’s about getting those natural direct and indirect effects, NDE and NIE, estimated more reliably than before <ref:2512.09337#pg2>.

Lu: This is the main contribution to the field: developing a novel weighting estimator that enhances finite-sample performance <ref:2512.09337#pg4>.

Meng: So, it’s not just about improving existing tools; it's about fundamentally changing how we handle the weighting process itself <ref:2512.09337#pg4>.

Tom: And they did show that by directly targeting sample-level balance and weight stability can yield substantially more reliable estimates of causal mechanisms <ref:2512.09337#pg4>.

Jane: So, it’s about moving away from methods that are just relying on the inverse propensity score as the central piece <ref:2512.09337#pg3>.

Lu: Which is a key insight for anyone looking to improve mediation analysis <ref:2512.09337#pg4>.

Conclusion: Tom: So we’re wrapping up on "Balancing Weights for Causal Mediation Analysis." The authors Kentaro Kawato and his team are tackling the challenge of estimating natural direct and indirect effects in causal mediation analysis in a way that's more robust to finite sample issues <ref:2512.09337#pg1>.

Jane: They proved that by proposing their two-step minimal weights estimator, they can achieve stability and balance the covariates and mediator distributions [ref:two <ref:2512.09337#pg1>

University of Tokyo

stat.ME, econ.EM

Submitted: 2025-12-10

Updated: 2026-11-16

Comments: 89 pages, 2 figures, 5 tables

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 96/100

The gist: This paper develops methods for estimating natural direct and indirect effects in causal mediation analysis by proposing a novel two-step minimal weights estimator that addresses finite-sample…

Key concepts

Natural Direct Effect (NDE)
This measures the effect of a treatment on an outcome, controlling for the mediator. It is calculated as the difference between the expected outcome when treated and not treated, given that both groups have experienced different levels of a specific mediator.
Inverse Probability Weighting (IPW) Estimator
A standard method used to estimate causal effects by weighting observations based on their propensity scores. This estimator is susceptible to instability and imbalance in finite samples, which the new method aims to fix through sequential weight correction.
Two-Step Minimal Weights Estimator
A proposed algorithm that constructs weights by solving an optimization problem designed to sequentially correct two types of imbalances: first balancing marginal covariate distributions, then balancing the joint distribution of covariates and the mediator.

Terminology

Summary

This paper develops methods for estimating natural direct and indirect effects in causal mediation analysis by proposing a novel two-step minimal weights estimator that addresses finite-sample instability and covariate imbalance inherent in standard estimators like the EIF–based estimator and IPW estimator. p. 0

Where the paper says it

This paper develops methods for estimating the natural direct and indirect effects in causal mediation analysis. The efficient influence function–based estimator (EIF– based estimator) and the inverse probability weighting estimator (IPW estimator), which are standard in causal mediation analysis, both rely on the inverse of the estimated propensity scores, and thus they are vulnerable to two key issues: (i) instability and (ii) finite-sample covariate imbalance.

Pruning redundant reasoning tokens

The paper introduces estimators based on weights obtained by an algorithm that directly penalizes weight dispersion while enforcing approximate covariate and mediator balance, thereby improving stability and mitigating bias in finite samples. The authors establish the convergence rates of the proposed weights and show that the resulting estimators are asymptotically normal and achieve the semiparametric efficiency bound. p. 0

Where the paper says it

We propose estimators based on the weights obtained by an algorithm that directly penalizes weight dispersion while enforcing approximate covariate and mediator balance, thereby improving stability and mitigating bias in finite samples. We establish the convergence rates of the proposed weights and show that the resulting estimators are asymptotically normal and achieve the semiparametric efficiency bound.

Monte Carlo simulations demonstrate that the proposed estimator outperforms not only the EIF–based estimator and the IPW estimator but also the regression imputation estimator in challenging scenarios with model misspecification. p. 0
Where the paper says it

We establish the convergence rates of the proposed weights and show that the resulting estimators are asymptotically normal and achieve the semiparametric efficiency bound. Monte Carlo simulations demonstrate that the proposed estimator outperforms not only the EIF–based estimator and the IPW estimator but also the regression imputation estimator in challenging scenarios with model misspecification.

Furthermore, the proposed method is applied to a real dataset from a study examining the effects of media framing on immigration attitudes. p. 0
Where the paper says it

Monte Carlo simulations demonstrate that the proposed estimator outperforms not only the EIF–based estimator and the IPW estimator but also the regression imputation estimator in challenging scenarios with model misspecification. Furthermore, the proposed method is applied to a real dataset from a study examining the effects of media framing on immigration attitudes.

How the training signal is built

The paper defines key estimands such as the natural direct effect (NDE) and natural indirect effect (NIE), which are defined as NDE(d):= E[Y1Md−Y0Md] and NIE(d):= E[YdM1 − YdM0]. The authors focus on two main weighting estimators: the inverse probability weighting (IPW) estimator and its augmented counterpart, the efficient influence function (EIF)–based estimator. p. 1

Where the paper says it

Dashed arrows represent the possible confounding by covariates. We focus on two main weighting estimators for causal mediation analysis: the inverse probability weighting (IPW) estimator and its augmented counterpart, the efficient influence function (EIF)–based estimator.

Both of these rely on weights based on the inverse of the estimated propensity scores and are therefore susceptible to two key issues: instability and finite-sample covariate imbalance. p. 2
Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS 3 but also wb EIF 1 (Xi) = 1/πb0(Xi), which also depends on the inverse of the estimated propensity score. 1 Both the IPW estimator and the EIF–based estimator rely on the weights based on the inverse of the estimated propensity scores and are therefore susceptible to two key issues noted in previous studies: (i) instability and (ii) finite-sample covariate imbalance (Hirshberg and Zubizarreta (2017); Chattopadhyay et al.

Where the gains come from

The proposed two–step minimal weights method is designed to sequentially correct two types of imbalances inherent in causal mediation analysis. Step 1 balances the marginal distribution of covariates, while Step 2 balances the joint distribution of covariates and the mediator. p. 3

Where the paper says it

We propose a “two-step minimal-dispersion approximate balancing weights ”method (hereafter, two-step minimal weights) for causal mediation. The name “two-step” reflects our strategy to first balance the marginal distribution of the covariates, and then the joint distribution of the covariates and the mediator.

The algorithm constructs weights by solving an optimization problem specifically designed to sequentially correct these two imbalances. p. 3
Where the paper says it

4 BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS and Zubizarreta (2019), as this framework encompasses several existing methods as special cases, including entropy balancing (Hainmueller 2012), stable balancing weights (Zubizarreta 2015), and empirical balancing calibration weights (Chan et al. The method’s name describes the two simultaneous objectives of the underlying optimization problem: “minimal dispersion”, i.

This sequential approach is motivated by analysis revealing that imbalances in these specific distributions generate the bias. p. 3
Where the paper says it

The name “two-step” reflects our strategy to first balance the marginal distribution of the covariates, and then the joint distribution of the covariates and the mediator. This sequential approach is motivated by our analysis in Section 3, which reveals that imbalances in these specific distributions generate the bias.

Empirical application

The empirical application uses the media–framing data of Brader et al. (2008) to assess causal pathways through which framing influences information-seeking behavior. p. 0

Where the paper says it

Many traditional studies focus on estimating the average treatment effect (ATE). However, researchers are now increasingly attempting to uncover the underlying causal pathways by decomposing the ATE into its mediated and unmediated components.

The proposed MW estimator achieves a TASMD of virtually zero for all 20 covariates in the empirical application. p. 5
Where the paper says it

Section 6 presents the results of simulation studies. Section 7 then reports an empirical application.

In setting (B), where outcome regression models use the true covariates, our proposed MW method substantially outperforms all competing estimators, being the only method to simultaneously achieve low bias and variance for both NDE(1) and NDE(0). p. 4
Where the paper says it

In addition, we implement other competing weighting schemes, including covariate balancing propensity scores (CBPS)4 and an oracle variant based on the true propensity scores. The proposed method further outperforms these alternatives, confirming its broad advantages.

Conclusion

The paper addresses the challenge of estimating natural direct and indirect effects in the presence of finite–sample instability and covariate imbalance, which are two critical limitations of the EIF–based estimator and IPW estimator. The authors conclude that directly targeting sample-level balance and weight stability can yield substantially more reliable estimates of causal mechanisms. p. 0

Where the paper says it

The efficient influence function–based estimator (EIF– based estimator) and the inverse probability weighting estimator (IPW estimator), which are standard in causal mediation analysis, both rely on the inverse of the estimated propensity scores, and thus they are vulnerable to two key issues: (i) instability and (ii) finite-sample covariate imbalance. We propose estimators based on the weights obtained by an algorithm that directly penalizes weight dispersion while enforcing approximate covariate and mediator balance, thereby improving stability and mitigating bias in finite samples.

References

Ai, C., Linton, O., Motegi, K., Zhang, Z., 2021. A unified framework for efficient estimation of general treatment models. p. 9

Where the paper says it

Although the EIF–based estimator is multiple robust and semiparametrically efficient, its reliance on the inverse of propensity scores leads to the aforementioned issues of instability and covariate imbalance. To formally investigate the latter, we consider a general EIF–type weighting estimator, which replaces the specific EIF weights wb EIF 2 (Mi , Xi) and wb EIF 1 (Xi) with generic weights wb2(Mi , Xi) and wb1(Xi): θbEIF-type 1,0 = 1 n Xn i=1 Diwb2(Mi , Xi) Yi − µb1(Mi , Xi) (5) + 1 n Xn i=1 (1 − Di)wb1(Xi) µb1(Mi , Xi) − ηb1,0(Xi) + 1 n Xn i=1 ηb1,0(Xi).

Quantitative Economics 12, 779–816

Athey, S., Chetty, R., Imbens, G., Kang, H., 2024. Estimating treatment effects using multiple surrogates: The role of the surrogate score and the surrogate index. p. 4

Where the paper says it

We illustrate the practical value of our approach using the media–framing data of Brader et al. To focus squarely on the role of weighting quality, we consider two generic estimators: an IPW–type estimator obtained by plugging arbitrary weights into estimator (1), and an EIF–type estimator defined as its augmented counterpart with outcome regression.

Working Paper, arXiv: 1603. p. 0
Where the paper says it

All errors are my own. 1 arXiv:2512.

09326 [stat. p. 0
Where the paper says it

1 arXiv:2512. 09337v1 [stat.

ME], ArXiV

Ben-Michael, E., Feller, A., Hirshberg, D. p. 2

Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS 3 but also wb EIF 1 (Xi) = 1/πb0(Xi), which also depends on the inverse of the estimated propensity score. 1 Both the IPW estimator and the EIF–based estimator rely on the weights based on the inverse of the estimated propensity scores and are therefore susceptible to two key issues noted in previous studies: (i) instability and (ii) finite-sample covariate imbalance (Hirshberg and Zubizarreta (2017); Chattopadhyay et al.

A., Zubizarreta, J. p. 2
Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS 3 but also wb EIF 1 (Xi) = 1/πb0(Xi), which also depends on the inverse of the estimated propensity score. 1 Both the IPW estimator and the EIF–based estimator rely on the weights based on the inverse of the estimated propensity scores and are therefore susceptible to two key issues noted in previous studies: (i) instability and (ii) finite-sample covariate imbalance (Hirshberg and Zubizarreta (2017); Chattopadhyay et al.

p. 2
Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS 3 but also wb EIF 1 (Xi) = 1/πb0(Xi), which also depends on the inverse of the estimated propensity score. 1 Both the IPW estimator and the EIF–based estimator rely on the weights based on the inverse of the estimated propensity scores and are therefore susceptible to two key issues noted in previous studies: (i) instability and (ii) finite-sample covariate imbalance (Hirshberg and Zubizarreta (2017); Chattopadhyay et al.

p. 2
Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS 3 but also wb EIF 1 (Xi) = 1/πb0(Xi), which also depends on the inverse of the estimated propensity score. 1 Both the IPW estimator and the EIF–based estimator rely on the weights based on the inverse of the estimated propensity scores and are therefore susceptible to two key issues noted in previous studies: (i) instability and (ii) finite-sample covariate imbalance (Hirshberg and Zubizarreta (2017); Chattopadhyay et al.

p. 2
Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS 3 but also wb EIF 1 (Xi) = 1/πb0(Xi), which also depends on the inverse of the estimated propensity score. 1 Both the IPW estimator and the EIF–based estimator rely on the weights based on the inverse of the estimated propensity scores and are therefore susceptible to two key issues noted in previous studies: (i) instability and (ii) finite-sample covariate imbalance (Hirshberg and Zubizarreta (2017); Chattopadhyay et al.

p. 2
Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS 3 but also wb EIF 1 (Xi) = 1/πb0(Xi), which also depends on the inverse of the estimated propensity score. 1 Both the IPW estimator and the EIF–based estimator rely on the weights based on the inverse of the estimated propensity scores and are therefore susceptible to two key issues noted in previous studies: (i) instability and (ii) finite-sample covariate imbalance (Hirshberg and Zubizarreta (2017); Chattopadhyay et al.

R., 2021. The balancing act in causal inference. p. 2
Where the paper says it

2This example is taken from Ben-Michael et al. 3Although it is possible to assess balance after weight estimation and re-estimate if needed, such an ad hoc iterative procedure can invalidate subsequent inference.

Working Paper, arXiv: 2110. p. 0
Where the paper says it

All errors are my own. 1 arXiv:2512.

14831 [stat. p. 0
Where the paper says it

1 arXiv:2512. 09337v1 [stat.

ME], ArXiV

Boyd, S., Vandenberghe, L., 2004. Convex Optimization. p. 3

Where the paper says it

4 BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS and Zubizarreta (2019), as this framework encompasses several existing methods as special cases, including entropy balancing (Hainmueller 2012), stable balancing weights (Zubizarreta 2015), and empirical balancing calibration weights (Chan et al. The method’s name describes the two simultaneous objectives of the underlying optimization problem: “minimal dispersion”, i.

Cambridge University Press, Cambridge

Brader, T., Valentino, N., Suhay, E., 2008. What triggers public opposition to immigration? p. 0

Where the paper says it

Monte Carlo simulations demonstrate that the proposed estimator outperforms not only the EIF–based estimator and the IPW estimator but also the regression imputation estimator in challenging scenarios with model misspecification. Furthermore, the proposed method is applied to a real dataset from a study examining the effects of media framing on immigration attitudes.

anxiety, group cues, and immigration threat. p. 0
Where the paper says it

Monte Carlo simulations demonstrate that the proposed estimator outperforms not only the EIF–based estimator and the IPW estimator but also the regression imputation estimator in challenging scenarios with model misspecification. Furthermore, the proposed method is applied to a real dataset from a study examining the effects of media framing on immigration attitudes.

American Journal of Political Science 52, 959–978

Bruns-Smith, D., Dukes, O., Feller, A., Ogburn, E.L., 2025. Augmented balancing weights as linear regression. p. 4

Where the paper says it

We illustrate the practical value of our approach using the media–framing data of Brader et al. To focus squarely on the role of weighting quality, we consider two generic estimators: an IPW–type estimator obtained by plugging arbitrary weights into estimator (1), and an EIF–type estimator defined as its augmented counterpart with outcome regression.

Journal of the Royal Statistical Society Series B: Statistical Methodology https://doi.org/10.1093/jrsssb/qkaf019

Celli, V., 2022. Causal mediation analysis in economics: Objectives, assumptions, models. p. 3

Where the paper says it

4 BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS and Zubizarreta (2019), as this framework encompasses several existing methods as special cases, including entropy balancing (Hainmueller 2012), stable balancing weights (Zubizarreta 2015), and empirical balancing calibration weights (Chan et al. The method’s name describes the two simultaneous objectives of the underlying optimization problem: “minimal dispersion”, i.

Journal of Economic Surveys 36, 214–234

Chan, K.C.G., Imai, K., Yam, S.C.P., Zhang, Z., 2016a. Efficient nonparametric estimation of causal mediation effects. p. 0

Where the paper says it

This paper develops methods for estimating the natural direct and indirect effects in causal mediation analysis. The efficient influence function–based estimator (EIF– based estimator) and the inverse probability weighting estimator (IPW estimator), which are standard in causal mediation analysis, both rely on the inverse of the estimated propensity scores, and thus they are vulnerable to two key issues: (i) instability and (ii) finite-sample covariate imbalance.

Working Paper, arXiv: 1601. p. 0
Where the paper says it

All errors are my own. 1 arXiv:2512.

03501 [stat. p. 0
Where the paper says it

1 arXiv:2512. 09337v1 [stat.

ME], ArXiV

Chan, K.C.G., Yam, S.C.P., Zhang, Z., 2016b. Globally efficient non-parametric inference of average treatment effects by empirical balancing calibration weighting. p. 3

Where the paper says it

4 BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS and Zubizarreta (2019), as this framework encompasses several existing methods as special cases, including entropy balancing (Hainmueller 2012), stable balancing weights (Zubizarreta 2015), and empirical balancing calibration weights (Chan et al. The method’s name describes the two simultaneous objectives of the underlying optimization problem: “minimal dispersion”, i.

Journal of the Royal Statistical Society: Series B (Statistical Methodology) 78, 673–700

Chattopadhyay, A., Hase, C.H., Zubizarreta, J.R., 2020. Balancing vs modeling approaches to weighting in practice. p. 5

Where the paper says it

2023), its extension to causal mediation analysis remains limited. Although several optimization-based weighting approaches have been proposed recently, our method remains distinct.

Statistics in Medicine 39, 3227–3254

Chen, J., Ritzwoller, D.M., 2023. Semiparametric estimation of long-term treatment effects. p. 0

Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS KENTARO KAWATO The University of Tokyo Abstract. This paper develops methods for estimating the natural direct and indirect effects in causal mediation analysis.

Journal of Econometrics 237, 105545

Chen, X., 2007. Chapter 76: Large sample sieve estimation of semi-nonparametric models, in: Handbook of Econometrics. p. 3

Where the paper says it

Firstly, by considering a nonparametric setting in which the number of balancing constraints increases with the sample size, we show that the optimization-based weights wb MW 1 (Xi) and wb MW 2 (Mi , Xi) converge, respectively, to ξ0(Mi , Xi)/π0(Xi)ξ1(Mi , Xi) and 1/π0(Xi). The proof of this result relies on the dual formulation and is derived under standard M–estimation assumptions.

Elsevier. volume 6 of Handbook of Econometrics, pp. 5549–5632

Cohn, E.R., Ben-Michael, E. p. 2

Where the paper says it

1 Both the IPW estimator and the EIF–based estimator rely on the weights based on the inverse of the estimated propensity scores and are therefore susceptible to two key issues noted in previous studies: (i) instability and (ii) finite-sample covariate imbalance (Hirshberg and Zubizarreta (2017); Chattopadhyay et al. (2020); Ben-Michael et al.

, Feller, A., Zubizarreta, J.R., 2023. Balancing weights for causal inference, in: Handbook of Matching and Weighting Adjustments for Causal Inference. p. 1
Where the paper says it

2 BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS contributes to this field by developing a novel weighting estimator for causal mediation analysis that enhances the finite–sample performance. We introduce notation and define the estimands.

Chapman and Hall/CRC, pp. 293–312

Dikkala, N., Lewis, G., Mackey, L., Syrgkanis, V., 2020. Minimax estimation of conditional moment models, in: Advances in Neural Information Processing Systems, pp. p. 5

Where the paper says it

However, their approach does not explicitly address the issue of (ii) finite-sample covariate imbalance. Liu and Ghassami (2024) introduced a minimax estimation framework following Dikkala et al.

12248–12262

Fan, J., Imai, K., Lee, I., Liu, H., Ning, Y., Yang, X., 2023. Optimal covariate balancing conditions in propensity score estimation. p. 0

Where the paper says it

Furthermore, the proposed method is applied to a real dataset from a study examining the effects of media framing on immigration attitudes. Keywords: causal mediation analysis; natural direct and indirect effects; propensity score instability; balancing weights; finite–sample covariate imbalance.

Journal of Business & Economic Statistics 41, 97–110

Farbmacher, H., Huber, M. p. 0

Where the paper says it

Reflecting this practical importance, this framework has become widespread across various disciplines, with both theoretical and empirical studies actively pursued (Imai et al. (2010a,b, 2011); VanderWeele (2015); Huber (2020); Celli (2022); Nguyen et al.

, Laff´ers, L., Langen, H., Spindler, M., 2022. Causal mediation analysis with double machine learning. p. 0
Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS KENTARO KAWATO The University of Tokyo Abstract. This paper develops methods for estimating the natural direct and indirect effects in causal mediation analysis.

The Econometrics Journal 25, 277–300

Hainmueller, J., 2012. Entropy balancing for causal effects: A multivariate reweighting method to produce balanced samples in observational studies. p. 0

Where the paper says it

Furthermore, the proposed method is applied to a real dataset from a study examining the effects of media framing on immigration attitudes. Keywords: causal mediation analysis; natural direct and indirect effects; propensity score instability; balancing weights; finite–sample covariate imbalance.

Political Analysis 20, 25–46

Hirano, K., Imbens, G.W., Ridder, G., 2003. Efficient estimation of average treatment effects using the estimated propensity score. p. 0

Where the paper says it

This paper develops methods for estimating the natural direct and indirect effects in causal mediation analysis. The efficient influence function–based estimator (EIF– based estimator) and the inverse probability weighting estimator (IPW estimator), which are standard in causal mediation analysis, both rely on the inverse of the estimated propensity scores, and thus they are vulnerable to two key issues: (i) instability and (ii) finite-sample covariate imbalance.

Econometrica 71, 1161–1189

Hirshberg, D.A., Zubizarreta, J.R., 2017. On two approaches to weighting in causal inference. p. 5

Where the paper says it

2023), its extension to causal mediation analysis remains limited. Although several optimization-based weighting approaches have been proposed recently, our method remains distinct.

Epidemiology 28, 812–816

Huang, L., Huang, W. p. 5

Where the paper says it

Although several optimization-based weighting approaches have been proposed recently, our method remains distinct. For instance, Huang et al.

, Linton, O., Zhang, Z., 2024. Nonparametric estimation of mediation effects with a general treatment. p. 3
Where the paper says it

Firstly, by considering a nonparametric setting in which the number of balancing constraints increases with the sample size, we show that the optimization-based weights wb MW 1 (Xi) and wb MW 2 (Mi , Xi) converge, respectively, to ξ0(Mi , Xi)/π0(Xi)ξ1(Mi , Xi) and 1/π0(Xi). The proof of this result relies on the dual formulation and is derived under standard M–estimation assumptions.

Econometric Reviews, 1–0

Huber, M., 2014. Identifying causal mechanisms (primarily) based on inverse probability weighting. p. 0

Where the paper says it

This paper develops methods for estimating the natural direct and indirect effects in causal mediation analysis. The efficient influence function–based estimator (EIF– based estimator) and the inverse probability weighting estimator (IPW estimator), which are standard in causal mediation analysis, both rely on the inverse of the estimated propensity scores, and thus they are vulnerable to two key issues: (i) instability and (ii) finite-sample covariate imbalance.

Journal of Applied Econometrics 29, 920–943

Huber, M., 2020. Mediation analysis, in: Zimmermann, K. p. 0

Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS KENTARO KAWATO The University of Tokyo Abstract. This paper develops methods for estimating the natural direct and indirect effects in causal mediation analysis.

F. (Ed.), Handbook of Labor, Human Resources and Population Economics. Springer International Publishing, pp. 1–38

Imai, K., Keele, L., Tingley, D., 2010a. A general approach to causal mediation analysis. p. 0

Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS KENTARO KAWATO The University of Tokyo Abstract. This paper develops methods for estimating the natural direct and indirect effects in causal mediation analysis.

Psychological Methods 15, 309–334

Imai, K., Keele, L., Yamamoto, T., 2011. Unpacking the black box of causality: Learning about causal mechanisms from experimental and observational studies. p. 0

Where the paper says it

Introduction Causal mediation analysis has become an essential framework for researchers who seek to understand the mechanisms of a causal effect. Many traditional studies focus on estimating the average treatment effect (ATE).

American Political Science Review 105, 765–789

Imai, K., Ratkovic, M. p. 4

Where the paper says it

2016b; Hirshberg and Zubizarreta 2017; Chattopadhyay et al. 4The CBPS method was originally proposed by Imai and Ratkovic (2014).

, 2014. Covariate balancing propensity score. p. 0
Where the paper says it

Furthermore, the proposed method is applied to a real dataset from a study examining the effects of media framing on immigration attitudes. Keywords: causal mediation analysis; natural direct and indirect effects; propensity score instability; balancing weights; finite–sample covariate imbalance.

Journal of the Royal Statistical Society: Series B (Statistical Methodology) 76, 243–263

Imbens, G., Kallus, N., Mao, X., Wang, Y., 2024. Long-term causal inference under persistent confounding via data combination. p. 1

Where the paper says it

The observed data are (Y, M, D, X) ∈ Y × M × {0, 1} × X . We use potential outcomes to formalize causal mechanisms.

Journal of the Royal Statistical Society: Series B (Statistical Methodology).

Liu, C., Ghassami, A. p. 3

Where the paper says it

This sequential approach is motivated by our analysis in Section 3, which reveals that imbalances in these specific distributions generate the bias. While a similar decomposition of the bias source has been recently discussed by Liu and Ghassami (2024), our study leverages this insight to construct the different weighting algorithm.

, 2024. Two-stage nuisance function estimation for causal mediation analysis. p. 0
Where the paper says it

This paper develops methods for estimating the natural direct and indirect effects in causal mediation analysis. The efficient influence function–based estimator (EIF– based estimator) and the inverse probability weighting estimator (IPW estimator), which are standard in causal mediation analysis, both rely on the inverse of the estimated propensity scores, and thus they are vulnerable to two key issues: (i) instability and (ii) finite-sample covariate imbalance.

Working Paper, arXiv: 2404. p. 0
Where the paper says it

All errors are my own. 1 arXiv:2512.

00735 [stat. p. 0
Where the paper says it

1 arXiv:2512. 09337v1 [stat.

ME], ArXiV

Miles, B.C.H., Shpitser, I. p. 1

Where the paper says it

Notably, these propensity score estimators appear in the denominator, which gives rise to the issues of interest. We also consider an augmented version, the EIF–based estimator of Tchetgen Tchetgen and Shpitser (2012), as introduced in (2), which combines weighting with outcome regression (regression imputation).

p. 1
Where the paper says it

Notably, these propensity score estimators appear in the denominator, which gives rise to the issues of interest. We also consider an augmented version, the EIF–based estimator of Tchetgen Tchetgen and Shpitser (2012), as introduced in (2), which combines weighting with outcome regression (regression imputation).

, Kanki, P., Meloni, S., Tchetgen, E. p. 1
Where the paper says it

Notably, these propensity score estimators appear in the denominator, which gives rise to the issues of interest. We also consider an augmented version, the EIF–based estimator of Tchetgen Tchetgen and Shpitser (2012), as introduced in (2), which combines weighting with outcome regression (regression imputation).

J.T., 2020. On semiparametric estimation of a path-specific effect in the presence of mediator-outcome confounding. p. 1
Where the paper says it

The NDE, defined as NDE(d) := E[Y1Md−Y0Md ], measures the causal effect of changing the treatment from 0 to 1 while fixing the mediator to the distribution it would take under D = d. In this way, the NDE captures the part of the treatment effect that does not operate through the mediator and reflects only the direct path from treatment to outcome.

Biometrika 107, 159–172

Newey, W., 1997. Convergence rates and asymptotic normality for series estimators. p. 0

Where the paper says it

We propose estimators based on the weights obtained by an algorithm that directly penalizes weight dispersion while enforcing approximate covariate and mediator balance, thereby improving stability and mitigating bias in finite samples. We establish the convergence rates of the proposed weights and show that the resulting estimators are asymptotically normal and achieve the semiparametric efficiency bound.

Journal of Econometrics 79, 147–168

Nguyen, T.Q., Ogburn, E.L., Schmid, I., Sarker, E.B., Greifer, N., Koning, I.M., Stuart, E.A., 2023. Causal mediation analysis: From simple to more robust strategies for estimation of marginal natural (in)direct effects. p. 0

Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS KENTARO KAWATO The University of Tokyo Abstract. This paper develops methods for estimating the natural direct and indirect effects in causal mediation analysis.

Statistics Surveys 17, 1–41

Nguyen, T.Q., Schmid, I., Ogburn, E.L., Stuart, E.A., 2022. Clarifying causal mediation analysis: Effect identification via three assumptions and five potential outcomes. p. 5

Where the paper says it

Implementation notes and proofs of the main theorems are provided in the Appendix. Causal Mediation Analysis: Setup and Identification This section introduced notations and assumptions for identification.

Journal of Causal Inference 10, 246–279

Nguyen, T.Q., Schmid, I., Stuart, E.A., 2020. Clarifying causal mediation analysis for the applied researcher: Defining effects based on what we want to learn. p. 0

Where the paper says it

Furthermore, the proposed method is applied to a real dataset from a study examining the effects of media framing on immigration attitudes. Keywords: causal mediation analysis; natural direct and indirect effects; propensity score instability; balancing weights; finite–sample covariate imbalance.

Psychological Methods. p. 0
Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS KENTARO KAWATO The University of Tokyo Abstract. This paper develops methods for estimating the natural direct and indirect effects in causal mediation analysis.

Pearl, J., 2001. Direct and indirect effects, in: Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence (UAI), Morgan Kaufmann, San Francisco. p. 0

Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS KENTARO KAWATO The University of Tokyo Abstract. This paper develops methods for estimating the natural direct and indirect effects in causal mediation analysis.

pp. 411–420

Tchetgen Tchetgen, E.J., Shpitser, I., 2012. Semiparametric theory for causal mediation analysis: Efficiency bounds, multiple robustness, and sensitivity analysis. p. 8

Where the paper says it

6All the following analysis applies similarly to θ0,1 := E[Y0M1 ]. 7Another key property of the EIF–based estimator is its multiple robustness.

Annals of Statistics 40, 1816–1845

Tingley, D., Yamamoto, T., Hirose, K., Keele, L., Imai, K., Trinh, M., Wong, W., 2025. mediation: Causal Mediation Analysis. p. 0

Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS KENTARO KAWATO The University of Tokyo Abstract. This paper develops methods for estimating the natural direct and indirect effects in causal mediation analysis.

URL: https://CRAN.R-project.org/package=mediation

Tropp, J.A., 2015. An introduction to matrix concentration inequalities. p. 0

Where the paper says it

Keywords: causal mediation analysis; natural direct and indirect effects; propensity score instability; balancing weights; finite–sample covariate imbalance. Introduction Causal mediation analysis has become an essential framework for researchers who seek to understand the mechanisms of a causal effect.

Foundations and Trends in Machine Learning 8, 1–230

van der Vaart, A.W., Wellner, J.A., 1996. Weak Convergence and Empirical Processes: With Applications to Statistics. p. 0

Where the paper says it

We propose estimators based on the weights obtained by an algorithm that directly penalizes weight dispersion while enforcing approximate covariate and mediator balance, thereby improving stability and mitigating bias in finite samples. We establish the convergence rates of the proposed weights and show that the resulting estimators are asymptotically normal and achieve the semiparametric efficiency bound.

Springer Series in Statistics, Springer, New York

VanderWeele, T.J., 2015. Explanation in Causal Inference: Methods for Mediation and Interaction. p. 0

Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS KENTARO KAWATO The University of Tokyo Abstract. This paper develops methods for estimating the natural direct and indirect effects in causal mediation analysis.

Oxford University Press, New York

VanderWeele, T.J., Vansteelandt, S., Robins, J.M., 2014. Effect decomposition in the presence of an exposure-induced mediator-outcome confounder. p. 7

Where the paper says it

It ensures the absence of confounding between the mediator and the outcome that is induced by the treatment. Such a confounder, often called a treatment-induced confounder (denoted by L in Figure 2), is a variable affected by D that subsequently affects both M and Y .

Epidemiology 25, 300–306

Wang, Y., Zubizarreta, J.R., 2019. Minimal dispersion approximately balancing weights: asymptotic properties and practical considerations. p. 5

Where the paper says it

Section 4 proposes the two–step minimal–weights algorithm. Section 5 establishes the asymptotic properties of the proposed estimators.

Biometrika 107, 93–105

Wong, R.K.W., Chan, K.C.G., 2018. Kernel-based covariate functional balancing for observational studies. p. 2

Where the paper says it

BALANCING WEIGHTS FOR CAUSAL MEDIATION ANALYSIS 3 but also wb EIF 1 (Xi) = 1/πb0(Xi), which also depends on the inverse of the estimated propensity score. 1 Both the IPW estimator and the EIF–based estimator rely on the weights based on the inverse of the estimated propensity scores and are therefore susceptible to two key issues noted in previous studies: (i) instability and (ii) finite-sample covariate imbalance (Hirshberg and Zubizarreta (2017); Chattopadhyay et al.

Biometrika 105, 199–213

Zubizarreta, J.R., 2015. Stable weights that balance covariates for estimation with incomplete outcome data. p. 1

Where the paper says it

Let X , M, and Y denote the supports of the covariates, mediator, and outcome, respectively. The observed data are (Y, M, D, X) ∈ Y × M × {0, 1} × X .

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Develop an AI system capable of automatically constructing two-step minimal weights by solving a constrained optimization problem that sequentially balances covariates (Step 1) and then balances covariates and mediators (Step 2). This allows the system to explicitly target the two sources of bias identified in Section 3, addressing both instability and finite-sample covariate imbalance.

  2. Enable the AI system to tune the tolerance hyperparameters epsilon and delta using a data-driven procedure based on Algorithm 4.2. This procedure selects tolerances that yield the most stable covariate balance across multiple bootstrap resamples, optimizing the trade-off between bias and variance for causal mediation estimators like EIF or IPW.

  3. Implement a mechanism for assessing covariate balance using the Target Absolute Standardized Mean Difference (TASMD) metric for both first-step and second-step weights. The system can then compare the performance of various weighting methods by visualizing these TASMDs, as shown in Figures 13 and 14, to select the optimal weight configuration.

  4. Allow the AI system to perform robust estimation under model misspecification by comparing its proposed two-step minimal weights estimator against baseline methods like EIF and IPW across settings (A) and (B). This comparison will demonstrate that the MW method substantially outperforms all competing estimators when outcome regression models are misspecified.

  5. Implement a feature where the AI system can identify potential covariate imbalance between weighting schemes by calculating TASMDs for specific covariates, such as those in Table 6, providing a diagnostic tool to ensure the weights achieve near-perfect balance (TASMD of virtually zero).

Abstract

Causal mediation analysis decomposes the average treatment effect into components that operate through intermediate variables and a component that does not. This paper studies weighting estimation of generalized mediation functionals, whose contrasts include natural direct and indirect effects, for any fixed finite number of causally ordered mediators. Inverse propensity score weights are vulnerable to (a) weight instability, because they are products of estimated probability ratios, and (b) finite-sample covariate and mediator imbalance, which can bias the estimator. We instead propose sequential balancing weights, which address (a) by minimizing weight dispersion and (b) by imposing exact balance. Stage one balances covariates between a treatment group and the full sample. Each later stage adds the next mediator and balances a treatment group against the preceding weighted group. A finite-sample bias decomposition of the augmented weighting estimator determines these variables and targets. When exact balance cannot be attained in a small sample, the procedure relaxes the balancing conditions by tolerances selected without using the outcome. Under the stated assumptions, the estimated weights converge at every stage to their population inverse-probability counterparts, with errors from preceding stages entering additively. We also establish asymptotic normality and semiparametric efficiency of the augmented weighting estimator, together with conditions for consistent variance estimation. In the main simulation design with two ordered mediators, the proposed weights outperform the compared methods in bias, root mean squared error, and coverage. In a small media-framing experiment with two ordered mediators, the proposed weights reduce the imbalance left by propensity-score weighting.

Sources

Related papers