Estimation Strategies for Causal Decomposition Analysis with Allowability Specifications

Causal decomposition analysis (CDA) is an approach for modeling the impact of hypothetical interventions to reduce disparities. It is useful for identifying foci that future interventions, including multilevel and multimodal interventions, could focus on to reduce disparities. Based within the potential outcomes framework, CDA has a causal interpretation when the identifying assumptions are met. CDA also allows an analyst to consider which covariates are allowable (i.e., fair) for defining the disparity in the outcome and in the point of intervention, so that its interpretation is also meaningful. While the incorporation of causal inference and allowability promotes robustness, transparency, and dialogue in disparities research, it can lead to challenges in estimation such as the need to correctly model densities. Also, how CDA differs from commonly used estimators may not be clear, which may limit its uptake. To address these challenges, we provide a tour of estimation strategies for CDA, reviewing existing proposals and introducing novel estimators that overcome key estimation challenges. Among them we introduce what we call “bridging” estimators that avoid directly modeling any density, and weighted sequential regression estimators that are multiply robust. Additionally, we provide diagnostics to assess the quality of the nuisance density models and weighting functions they rely on. We formally establish the estimators’ robustness to model mis-specification, demonstrate their performance through a simulation study based on real data, and apply them to study disparities in hypertension control using electronic health records in a large healthcare system.

💡 Research Summary

This paper addresses the methodological challenges of Causal Decomposition Analysis (CDA), an approach that quantifies how a hypothetical intervention that removes disparity in a “point‑of‑intervention” variable (Z) would change disparity in an outcome (Y). CDA differs from traditional statistical decompositions (Oaxaca‑Blinder, Fairlie) by embedding causal assumptions, explicitly separating covariates that are deemed “allowable” for measuring outcome disparity (Ay) and for defining the intervention (Az), and by using a potential‑outcomes framework to define group‑specific means under both an observational arm and a hypothetical target trial arm.

The authors first review existing CDA estimators and classify them along two dimensions: (1) whether they require direct modeling of conditional densities for Z (or other covariates) and (2) whether they rely on weighting, outcome modeling, or a combination (augmented) approach. They note that many existing estimators are vulnerable because they depend on correctly specifying the density of Z, especially when Z is continuous or multivariate.

To overcome these limitations, the paper introduces two novel families of estimators.

1. Bridging estimators avoid explicit density modeling. They construct synthetic samples that use empirical densities or density‑ratio estimates, thereby “bridge” the observed distribution to the target distribution without fitting a parametric model for p(Z|Az,Ay). This strategy reduces model‑misspecification risk and works for discrete, continuous, or multiple points of intervention.
2. Weighted Sequential Regression (WSR) estimators are multiply robust. They estimate three nuisance components—(a) the conditional density of Z given (Az,Ay), (b) a weighting function ω that re‑weights the sample to mimic the target covariate distribution, and (c) an outcome regression model. The estimator remains consistent if any one of these components is correctly specified, and the authors provide formal proofs of this multiply‑robust property and derive its influence function for variance estimation.

The paper also supplies practical diagnostics. For density models, cross‑validated residual checks and KL‑divergence between estimated and empirical densities are recommended. For weighting functions, the authors propose two balance diagnostics: (i) inspection of mean and variance of the weights within relevant sub‑samples, and (ii) a “target balance” metric that compares weighted covariate distributions to their intended target distributions. These diagnostics enable a design‑based workflow in which analysts can assess and improve estimator quality before reporting results.

Simulation studies, grounded in real electronic health record data on hypertension, compare the new bridging and WSR estimators to existing Augmented Inverse Probability Weighting (AIPW) methods. Results show that the proposed estimators achieve lower mean‑squared error and exhibit far less sensitivity to misspecification of any single nuisance model, confirming their theoretical robustness.

The empirical application examines racial disparities in uncontrolled hypertension within a large health system. The point‑of‑intervention is treatment intensification (starting, increasing dose, or adding a medication). By re‑assigning the treatment intensity for the disadvantaged group to follow the conditional distribution observed in the advantaged group (while respecting allowability specifications for Ay and Az), the analysis estimates that eliminating the treatment‑intensity disparity would reduce the overall hypertension‑control gap by roughly 30 %. The authors also illustrate how different allowability specifications (e.g., including versus excluding socioeconomic covariates in Ay) alter the estimated impact, highlighting the substantive relevance of fairness choices.

In sum, the paper makes three major contributions: (1) a clear conceptual distinction between statistical and causal decomposition, emphasizing the role of allowability; (2) novel estimation strategies—bridging estimators that bypass density modeling and WSR estimators that are multiply robust; and (3) actionable diagnostics for density and weighting quality. These advances provide a robust, transparent, and implementable toolkit for researchers seeking to quantify and ultimately intervene on health and social disparities.