Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Ratio of Mediator Probability Weighting for Estimating Natural Direct and Indirect Effects

T0 review · 4 major / 4 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 ratio of mediator probabilities under treatment and control recovers natural direct and indirect effects without specifying an outcome model.

desk verdict Sound identification proof under a useful weighting representation, but the stratification estimator in Section 4.3 lacks a consistency proof and can be asymptotically biased. read the letter →

arxiv 2506.03284 v1 pith:KYL3G6AE submitted 2025-06-03 stat.ME

Guanglei Hong This is my paper
classification stat.ME MSC 62D2062G05
keywords causalinferencemediationanalysisnaturaldirecteffectindirectratioofmediatorprobabilityweightingcounterfactualoutcomestreatment-mediatorinteractionpost-treatmentcovariates
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 a nonparametric method, ratio of mediator probability weighting, for decomposing a total causal effect into natural direct and indirect effects. The key move is to reweight experimental units so that their observed mediator distribution approximates the counterfactual mediator distribution under the control condition; each unit's weight is a ratio of conditional mediator probabilities under control versus treatment, multiplied by an inverse-probability-of-treatment weight. No outcome model is specified, so the method works for any outcome distribution and any outcome-mediator relationship, and it does not require assuming away treatment-mediator interaction. The paper proves that the weighted mean of treated outcomes identifies the counterfactual mean $E[Y_{1Z_0}]$, and shows how a single weighted regression with a duplicated treated sample estimates both natural effects simultaneously. A modified version of the weight also adjusts for post-treatment confounders under an additional independence assumption.

What carries the argument

The machinery is the ratio of mediator probability weight, defined in Equation (2) as $$W(1Z_0)=\frac{q_{(0)}(Z_0=z \mid A=0,X)}{q_{(1)}(Z_1=z \mid A=1,X)}\times\frac{p(A=1)}{p(A=1 \mid X)}.$$ It is the ratio of the conditional probability of a mediator value under the control condition to the conditional probability of the same value under the experimental condition, times the inverse-probability-of-treatment weight $p(A=1)/p(A=1 \mid X)$. The numerator forms the counterfactual mediator distribution; the denominator removes the experimental mediator distribution; the treatment weight balances pretreatment covariates. The paper estimates the component probabilities with propensity-score stratification, and the final estimation step is a weighted regression with a duplicated experimental sample, one copy weighted by $W(1Z_0)$ and one by $W(1Z_1)$, so the natural direct and indirect effects appear as two regression coefficients.

What would settle it

Simulate a randomized trial with a known counterfactual mean $\mathbb{E}[Y_{1Z_0}]$, add an unmeasured variable that affects both the mediator and the outcome differently in the treatment and control arms, apply the RMPW weight, and compare the weighted treated mean with the known target; any systematic gap shows the estimand fails when Assumption 7 is violated.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 1: under Assumptions 1, 2, 4, 6, 7, and 8, $\mathbb{E}[W(1Z_0)Y \mid A=1]$ is an observed-data estimand for the counterfactual outcome mean $\mathbb{E}[Y_{1Z_0}]$, where $W(1Z_0) = [q_{(0)}(Z_0=z \mid A=0,X) / q_{(1)}(Z_1=z \mid A=1,X)] \times [p(A=1)/p(A=1 \mid X)]$. The same logic identifies $E[Y_{aZ_{a'}}]$ in general. By weighting the experimental units so their mediator distribution tracks the control counterfactual mediator distribution, the natural direct effect $\mathbb{E}[Y_{1Z_0}-Y_{0Z_0}]$ and the natural indirect effect $\mathbb{E}[Y_{1Z_1}-Y_{1Z_0}]$ are obtained from one weighted regression of $Y$ on treatment and a duplicated-treatment indicator. Theorem 2 extends the weight to settings with post-treatment covariates by conditioning the mediator probabilities on the union of pretreatment covariates $X_+$ and on post-treatment $L_1$, under the additional cross-world Assumption 9.

Load-bearing premise

The method rests on the untestable assumption that, within levels of the observed covariates, no unmeasured variable confounds the mediator-outcome relationship, within or across treatment conditions; the post-treatment version adds a separate cross-world independence assumption between the counterfactual mediator and post-treatment covariates.

Editorial extensions

If this is right

  • Researchers can estimate mediation effects for binary, count, or skewed outcomes without linearity assumptions linking the outcome to the mediator.
  • Treatment-mediator interactions no longer block identification of natural direct and indirect effects, because no average over controlled direct effects is taken.
  • Large sets of pretreatment covariates can be handled through the treatment and mediator propensity scores, without fitting a high-dimensional outcome model.
  • Under the modified assumptions, post-treatment confounders of the mediator-outcome relationship can be adjusted for, a situation the existing regression-based methods cannot accommodate when interaction is present.
  • Natural direct and indirect effects are estimated simultaneously in a single weighted outcome model, so no separate mediator and outcome models need to be combined.

Reading between the lines

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

  • A practical issue the paper does not examine is the behavior of the estimator when the mediator distributions under treatment and control barely overlap; extreme weights would then drive the estimates, and standard-error and weight-trimming choices would need study.
  • The same weighting logic could be embedded in a doubly robust procedure that also fits an outcome model, trading some of the outcome-model freedom for protection against mediator-model misspecification.
  • Assumption 9, unique to the post-treatment extension, is a cross-world independence condition; it will be most plausible in designs with site-level randomization and rich pretreatment measurement, as in the Head Start example, and least plausible in purely observational mediation studies.
  • Iterating the ratio-weighting idea offers a route to multiple mediators and time-varying treatments, directions the author lists as future work.
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 / 4 minor

Summary. The manuscript proposes a nonparametric weighting approach, Ratio of Mediator Probability Weighting (RMPW), for estimating natural direct and indirect effects without specifying an outcome model and without assuming no treatment-mediator interaction. Under assumptions comparable to sequential ignorability, Theorem 1 derives a weight that reweights experimental units so that the weighted mediator distribution matches the counterfactual mediator distribution under the control condition; Theorem 2 extends this to settings with post-treatment covariates under an additional cross-world independence assumption. Section 4.3 proposes estimating the weights by propensity score stratification, and the paper claims that the method is robust and broadly applicable. The identification proofs in Appendices 1 and 2 are algebraically correct under the stated assumptions, but the paper provides no simulation or empirical example and the proposed stratified estimator is not shown to be consistent.

Significance. If the identification result is taken as given, the weighting representation is an elegant and potentially useful alternative to regression-based mediation methods: it avoids outcome model specification, permits treatment-mediator interactions, and (under Assumption 9) nominally accommodates post-treatment confounders. The proof of Theorem 1 is self-contained and the weighted-mean representation is exact under the stated assumptions, which is a clear strength. However, the practical value of the paper depends on the estimator in Section 4.3, whose consistency is not established, and on the plausibility of the additional post-treatment assumption, which is strong and untestable. The lack of any numerical evaluation, explicitly acknowledged in the conclusion, further limits the paper's contribution as a methods paper. The identification result itself is closely related to the mediation formula of Imai et al. (2010) under the same assumptions, so the novelty lies primarily in the weighting formulation and the post-treatment extension rather than in new identification conditions.

major comments (4)
  1. [§4.3] The proposed stratification estimator is not shown to be consistent for the target estimand. In §4.3, the weight for a unit i is defined as (n_{s0,Z0=1}/n_{s0}) / (n_{s1,Z1=1}/n_{s1}), where the numerator and denominator are proportions within strata defined by the two mediator propensity scores θ_Z0(x) and θ_Z1(x). These proportions consistently estimate the stratum-level averages E[q0(Z0=z|A=0,X) | S0 = s0(i)] and E[q1(Z1=z|A=1,X) | S1 = s1(i)], not the unit-level ratio q0(X_i)/q1(X_i). The ratio of two stratum-level averages does not equal the mean of the unit-level ratios, and because the two strata are different subsets of the covariate space, the probability limit of the weighted mean is generally a coarsened version of E[Y1Z0], not the counterfactual mean itself. No consistency theorem is provided for this estimator, and the cited 'Hong (in press)' result concerns marginal mean weighting through stratification for treatment effects, not ratio-of-mediator weights. Since this estimator is the paper's main practical proposal, this gap is load-bearing.
  2. [§5 (Conclusion)] The manuscript contains no simulation study or empirical application. The concluding section explicitly states that 'the performance of the ratio-of-mediator-weighting method relative to other non-parametric and parametric methods is yet to be assessed through simulations,' and no data example is provided. Without either a formal consistency proof for the proposed estimator or numerical evidence on finite-sample bias and coverage, the practical claim that the method 'promises to increase the robustness' of effect estimates is not supported. This is a major omission for a methods paper.
  3. [§4.2] Assumption 9—that Z_a' is independent of L_a given A=a and X(ix)—is a strong cross-world independence assumption that is unique to this method and not implied by the standard sequential ignorability assumptions. It is untestable from observed data. The claim that this assumption is 'weaker than assuming that such post-treatment covariates do not exist' is not formalized; the two conditions are not nested in an obvious way, and the assertion requires justification or a counterexample. Because the post-treatment extension is one of the paper's claimed advantages over existing methods, this assumption needs more critical discussion.
  4. [§4.3] The outcome model in §4.3 is not scale-invariant for non-continuous outcomes. The natural direct and indirect effects are defined as differences in expected potential outcomes, i.e., on the probability (or additive) scale. The paper first presents the weighted linear regression Y = γ0 + A γ1 + A·D γ2 + e, for which the coefficients equal the desired contrasts when the model is saturated. However, it then suggests using a weighted generalized linear model with a logit link for binary outcomes (e.g., kindergarten retention). For a logit link, the coefficients are log odds ratios, not risk differences, and the stated equality between γ1 and E[Y1Z0 − Y0Z0] no longer holds. The manuscript does not state whether the effects are intended on the logit scale or how to transform them, and it provides no derivation for the logit case. This undermines the claim that the method applies 'regardless of the distribution of the outcome.'
minor comments (4)
  1. [References] The reference to 'van der Lann and Petersen (2005, 2008)' appears to be a misspelling of 'van der Laan'; the same applies to the in-text citation and reference list entry.
  2. [Throughout] The name 'Petersen' is spelled inconsistently; for example, 'Petersen et al (2006)' in the text appears as 'Peterson, M. L.' in the reference list. Please standardize.
  3. [§4.3] The phrase 'non-parametric approach' is used loosely: the proposed stratification procedure still requires parametric models (e.g., logistic regression) for the mediator propensity scores and treatment propensity scores, and the number of strata is fixed at five. The text should clarify that the nonparametric claim applies to the outcome model, not to the weight estimation.
  4. [§4.3] The paper states that robust standard errors or bootstrap can be used for inference, but it does not address the fact that the weights are estimated. The uncertainty in the propensity score and stratum proportion estimates is ignored, which may lead to underestimated standard errors. A reference to a method that accounts for estimated weights would be useful.

Circularity Check

0 steps flagged · score 2.0 of 10

No circularity in the identification proof; one minor self-citation supports a robustness claim but does not load-bear the central derivation.

full rationale

Theorem 1 (and its proof in Appendix 1) is self-contained: under Assumptions 1, 2, 4, 6, 7, and 8, the weight W(1Z0)=[q0/q1]*[p(A=1)/p(A=1|X)] is constructed algebraically by Bayes theorem and the stated conditional-independence assumptions, and no term involving the target counterfactual mean E(Y1Z0) is used as a fitted input. The same holds for Theorem 2 under Assumptions 1, 2*, 4*, 6, 7*, 8, and 9. The stratified estimator in Section 4.3 is a plug-in procedure whose weights are estimated from mediator and covariate data, not from outcomes, so the effects are not predictions forced by construction. The only self-citation is the sentence crediting Hong (in press) for robustness of propensity-score-stratification weighting; that result concerns marginal mean weighting through stratification for treatment assignment, and the paper does not prove the transfer to ratio-of-mediator weights. This is a minor self-citation adjacent to a robustness claim, not a circularity in the identification argument, and it does not raise the score beyond 2. The possible asymptotic bias of the five-strata estimator is a consistency question, not a circularity, because the estimator does not encode the target effect into its weights.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central identification result rests entirely on unverifiable no-confounding assumptions and SUTVA. Assumption 9 is introduced specifically for post-treatment covariates and is the most fragile. No new entities are postulated.

free parameters (2)
  • Number of propensity score strata = 5
    Section 4.3 fixes five strata per mediator propensity score with no data-driven justification; consistency of the stratified weight estimator is not proven.
  • Mediator propensity score model choice = logistic regression or stratification
    The estimator requires estimating q0 and q1; the paper relies on a self-cited in-press result for robustness of stratification, so practical bias depends on this modeling choice.
assumptions (6)
  • domain assumption SUTVA (no interference and consistency)
    Invoked in Section 2.1; needed to define potential outcomes and link observed data to counterfactuals.
  • domain assumption Assumption 8: Y_aZ_a, Y_a'Z_a', Y_aZ_a', Y_a'Z_a independent of A given X
    Section 3.2 and 4.1; ensures treatment assignment is unconfounded relative to all relevant potential outcomes.
  • domain assumption Assumptions 4 and 7: no confounding of the mediator-outcome relationship within and across treatment conditions
    Section 4.1; used in the proof of Theorem 1 and untestable in practice.
  • domain assumption Assumption 6: Z_a independent of A given X
    Section 4.1; needed to equate the distribution of Z0 across treatment groups within covariate levels.
  • domain assumption Positivity Assumptions 1, 2, and 2*
    Section 4.1 and 4.2; required for the weights to be well-defined.
  • ad hoc to paper Assumption 9: Z_a' independent of L_a given A=a and X(ix)
    Section 4.2, introduced specifically to handle post-treatment covariates; a strong cross-world independence assumption with no sensitivity analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ratio of Mediator Probability Weighting for Estimating Natural Direct and Indirect Effects." pith.science (2026). https://pith.science/paper/KYL3G6AE

@misc{pith2026250603284,
  author       = {Pith},
  title        = {Pith review of: Ratio of Mediator Probability Weighting for Estimating Natural Direct and Indirect Effects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KYL3G6AE}},
  note         = {Machine review of arXiv:2506.03284}
}
read the original abstract

Decomposing a total causal effect into natural direct and indirect effects is central to revealing causal mechanisms. Conventional methods achieve the decomposition by specifying an outcome model as a linear function of the treatment, the mediator, and the observed covariates under identification assumptions including the assumption of no interaction between treatment and mediator. Recent statistical advances relax this assumption typically within the linear or nonlinear regression framework. I propose a non-parametric approach that also relaxes the assumption of no treatment-mediator interaction while avoiding the problems of outcome model specification that become particularly acute in the presence of a large number of covariates. The key idea is to estimate the marginal mean of each counterfactual outcome by assigning a weight to every experimental unit such that the weighted distribution of the mediator under the experimental condition approximates the counterfactual mediator distribution under the control condition. The weight is a ratio of the conditional probability of a mediator value under the control condition to that of the same mediator value under the experimental condition. A non-parametric approach to estimating the weight on the basis of propensity score stratification promises to increase the robustness of the direct and indirect effect estimates. The outcome is modeled as a function of the direct and indirect effects with minimal model-based assumptions. This method applies regardless of the distribution of the outcome or the functional relationship between the outcome and the mediator, and is suitable for handling a large number of pretreatment covariates. RMPW software packages are available in Stata (https://ideas.repec.org/c/boc/bocode/s458301.html) and R (https://cran.r-project.org/web/packages/rmpw/index.html).

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Coarsening Bias from Variable Discretization in Causal Functionals

    stat.ME 2026-02 conditional novelty 5.0 of 10

    Discretizing a continuous mediator in causal functionals induces first-order approximation bias; a within-bin mean correction reduces it to second order.

Reference graph

Works this paper leans on

5 extracted references · 5 canonical work pages · cited by 1 Pith paper

  1. [1]

    Path analysis: Sociological examples,

    Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology, 51, 1173-1182. Duncan, O. D. (1966). “Path analysis: Sociological examples,” American Journal of Sociology, 72, 1-16. Efron, B. (1988). Boots...

    work page 1986

  2. [23]

    VanderWeele, T. (2009). Marginal structural models for the estimation of direct and indirect effects. Epidemiology, 20, 18-26. VanderWeele, T., & Vansteelandt, S. (2009). Conceptual issues concerning mediation, interventions, and composition. Statistics and its Interface, 2, 457-468. Zanutto, E., Lu, Bo., & Hornik, R. (2005). Using propensity score subcla...

    work page 2009

  3. [251]

    J., & Peterson, M

    van der Lann, M. J., & Peterson, M. L. (2008). Direct effect models. The International Journal of Biostatistics, 4(1), Article

    work page 2008

  4. [2005]

    Biometrics Section – JSM 2010 2414 Pearl, J. (2010). The mediation formula: A guide to the assessment of causal pathways in non-linear models. Los Angeles, CA: University of California, Los Angeles. Technical report R-363, July

    work page 2010

  5. [2010]

    Comment: Which Ifs Have Causal Answers

    Peterson, M. L., Sinisi, S. E., & van der Laan, M. J. (2006). Estimation of direct causal effects. Epidemiology, 17(3), 276-284. Robins, J. M. (1999). Marginal structural models versus structural nested models as tools for causal inference. In M. Elizabeth Halloran and Donald Berry (Eds.), Statistical Models in Epidemiology, the Environment, and Clinical ...

    work page 2006

Pith tools

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