Balancing Weights for Causal Mediation Analysis

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. 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. Furthermore, the proposed method is applied to a real dataset from a study examining the effects of media framing on immigration attitudes.

💡 Research Summary

This paper addresses significant practical limitations in standard methods for causal mediation analysis, which aims to decompose a treatment effect into direct and indirect (mediated) components. The authors focus on estimating the Natural Direct Effect (NDE) and Natural Indirect Effect (NIE). The widely used Efficient Influence Function-based estimator (EIF) and the Inverse Probability Weighting estimator (IPW) rely on inverse estimated propensity scores (probabilities of treatment assignment). This reliance makes them vulnerable to two critical issues in practice: (1) Instability: When estimated propensity scores are close to 0 or 1, their inverses become extremely large, leading to high-variance, unstable estimates with wide confidence intervals. This problem is exacerbated in mediation because the weights often involve a product of inverse probabilities. (2) Finite-Sample Covariate Imbalance: While these weights guarantee balance between treatment groups in large samples, they often fail to achieve perfect balance in the finite samples researchers actually work with. Imbalances in covariates or the mediator that predict the outcome can introduce substantial bias.

To directly tackle both problems simultaneously, the authors propose a novel weighting scheme based on the “Minimal-Dispersion Approximate Balancing Weights” framework. Their method, termed the “Two-Step Minimal Weights” algorithm, formulates weight estimation as an optimization problem. The objective is to minimize a measure of weight dispersion (e.g., sum of squared weights) to promote stability. Crucially, the optimization includes explicit constraints that force the weighted moments of covariates and the mediator in the treatment group to approximately match those in the control group, within a pre-specified tolerance. This directly enforces balance in finite samples. The “two-step” process first calculates weights to balance the marginal distribution of covariates (X), and then uses these to calculate a second set of weights that balance the joint distribution of the mediator (M) and covariates (X). This design is motivated by a formal bias decomposition showing that imbalances in these specific distributions are the sources of finite-sample bias.

The paper makes substantial theoretical contributions. First, the authors establish the consistency of their proposed weights, proving that as the sample size grows and the number of balancing constraints increases appropriately, their estimated weights converge to the true, infeasible optimal weights (i.e., 1/π₀(X) and ξ₀(M,X)/π₀(X)ξ₁(M,X)). Second, they demonstrate that the resulting IPW-type and EIF-type estimators using these weights are asymptotically normally distributed. Third, and importantly, they prove that these estimators achieve the semiparametric efficiency bound. This means that while the new method offers superior finite-sample performance (addressing instability and imbalance), it does not sacrifice large-sample optimality; it is asymptotically equivalent to the standard efficient estimators.

The empirical performance is rigorously evaluated through extensive Monte Carlo simulations. Using two classic data-generating processes from the mediation literature under both correct model specification and misspecification, the proposed estimator is compared against the standard EIF, IPW, and regression imputation estimators. The results show that the proposed method performs as well as or better than all competitors across scenarios. It demonstrates particular strength in challenging settings with model misspecification, often yielding lower bias and mean squared error. The method is also applied to a real-world study on the effect of media framing on immigration attitudes. The application illustrates the practical benefit: using the proposed balancing weights leads to estimates with smaller standard errors, providing clearer and more precise conclusions about the mediation mechanisms.

In summary, this paper provides a robust and practical solution to well-known problems in causal mediation estimation. By integrating weight stabilization and explicit balance constraints into a single optimization framework, it improves reliability in finite samples while preserving desirable asymptotic properties like efficiency and normality. The work successfully bridges the gap between the causal mediation literature and the growing field of balancing weights, offering a valuable tool for applied researchers.