Proximal Inference on Population Intervention Indirect Effect

Yang Bai, Yifan Cui, Baoluo SunarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces a robust framework for estimating the Population Intervention Indirect Effect (PIIE)—a causal measure vital for studying harmful exposures like alcohol or pollution—even when hidden "confounders" influence the mediator. By leveraging observed covariates as proxies for unmeasured variables and applying debiased machine learning, the authors provide a way to calculate ethical causal effects that are both statistically efficient and resilient to model errors.

What Problem Does This Paper Address?

In causal mediation analysis, we want to know how much of an effect a treatment ($A$) has on an outcome ($Y$) through an intermediate mediator ($M$). For example: Does alcohol consumption ($A$) increase depression ($Y$) specifically by triggering depersonalization symptoms ($M$)?

Standard methods like the Natural Indirect Effect (NIE) are problematic for two reasons:

  1. Ethics: They often require imagining a world where we force people into harmful exposures, which is ethically impossible to experiment on.
  2. Hidden Confounders: Most models assume we have measured every variable that influences the relationship between the treatment, mediator, and outcome. In reality, "unmeasured confounders" (like genetic predispositions or personality traits) often lurk in the background, biasing the results.

While the Population Intervention Indirect Effect (PIIE) solves the ethical problem by comparing the current population to one where the harmful exposure is simply removed, existing PIIE methods still struggle when hidden variables affect the mediator. This paper fills that gap.

Methods

The authors combine Proximal Causal Inference with Semiparametric Theory to create a "multiply robust" estimation framework.

  • Proxy Variables: Instead of trying to measure the hidden confounder ($U$) directly, they use two sets of "proxies":
  • $Z$ (Exposure-inducing proxies): Variables related to the treatment but not directly to the outcome.
  • $W$ (Outcome-inducing proxies): Variables related to the outcome but not directly to the treatment.
  • Confounding Bridge Functions: They develop mathematical "bridges" ($h$ and $q$ functions) that allow researchers to identify the causal effect using only these proxies and observed data.
  • Three Identification Strategies: They propose three ways to calculate the effect: an outcome-based approach, an exposure-based approach, and a hybrid of the two.
  • Debiased Machine Learning (DML): To handle high-dimensional data, they integrate DML with a cross-fitting procedure. This ensures that even if the machine learning models used to estimate the "nuisance" parts of the data are slightly off or converge slowly, the final causal estimate remains accurate ($\sqrt{n}$-consistent) and follows a normal distribution.

Datasets

The paper validates its approach using two primary sources:

  1. Synthetic Simulations: A data-generating mechanism based on prior literature (Dukes et al., 2023) but modified to include more complex confounding patterns. This allowed the authors to test how well their estimators performed when parts of the model were intentionally "misspecified" (wrong).
  2. Real-World Application: Data from a cross-sectional online survey of 3,856 Mexican health workers conducted during the COVID-19 pandemic. This dataset included measures of alcohol consumption (AUDIT scale), depression (CESD-47), and burnout/depersonalization dimensions.

Key Insights and Ideas

  • The Power of "Multiply Robust" Estimators: The authors developed a P-MR (Proximal Multiply Robust) estimator. Its "aha moment" is its resilience: it remains consistent as long as at least one of the underlying bridge function models is correctly specified. In complex real-world scenarios where we rarely know the "true" model, this redundancy is a lifesaver.
  • Ethical Realism: By focusing on PIIE, the research aligns causal inference with public health reality. We don't need to know what happens if we force people to drink; we need to know what happens if we prevent hazardous drinking in the existing population.
  • Proxies as a Safety Net: The paper demonstrates that you don't need to measure the "unmeasurable" if you have smart proxies. By splitting covariates into $Z$ and $W$, researchers can "cancel out" the bias from hidden variables.

Key Results

  • Simulation Superiority: In scenarios with hidden confounding, traditional "doubly robust" methods failed significantly, with confidence intervals covering the true value only ~30% of the time. In contrast, the authors' P-MR estimator maintained ~95% coverage even when some of the bridge functions were wrong.
  • Alcohol and Depression: In the Mexican health worker study, the P-MR approach revealed a significant indirect effect (estimate: 0.20). This suggests that hazardous alcohol consumption indirectly increases depression risk by exacerbating depersonalization symptoms—a finding that was statistically significant in their model but "missed" or negated by less robust traditional methods.

Limitations and Future Work

  • Proxy Validity: The entire framework relies on the assumption that the chosen proxies ($Z$ and $W$) are valid. If the proxies are "leaky" (e.g., $Z$ directly affects the outcome), the identification fails. The authors suggest future work should focus on sensitivity analyses to test how much a violation of these proxy rules would change the results.
  • High-Dimensional Proxies: While DML helps, extremely high-dimensional settings where the number of proxies exceeds the sample size still pose a challenge for the "completeness" conditions required for the bridge functions.
  • Extension to Intervening Variables: The authors note their method could be extended to study "intervening variables"—a broader class of causal effects where the treatment itself might be ill-defined or impossible to manipulate directly.

Summary updated

Abstract

Population intervention indirect effect (PIIE) is a novel mediation effect representing the indirect component of the population intervention effect. Unlike traditional mediation measures, such as the natural indirect effect, the PIIE holds particular relevance in observational studies involving unethical exposures, when hypothetical interventions that impose harmful exposures are inappropriate. Although prior research has identified PIIE under unmeasured confounders between exposure and outcome, it has not fully addressed the confounding that affects the mediator. This paper proposes a novel PIIE identification framework in settings where unmeasured confounders influence exposure-outcome, exposure-mediator, and mediator-outcome relationships. Specifically, we leverage observed covariates as proxy variables for unmeasured confounders, constructing three proximal identification strategies. Additionally, we characterize the semiparametric efficiency bound and develop multiply robust and locally efficient estimators. To handle high-dimensional nuisance parameters, we propose a debiased machine learning approach that achieves $\sqrt{n}$-consistency and asymptotic normality to estimate the true PIIE values, even when the machine learning estimators for the nuisance functions do not converge at $\sqrt{n}$-rate. In simulations, our estimators demonstrate higher confidence interval coverage rates than conventional methods across various model misspecifications. In a real data application, our approaches reveal an indirect effect of alcohol consumption on depression risk mediated by depersonalization symptoms.
BibTeX
@misc{bai2025proximal,
  author = {Yang Bai and Yifan Cui and Baoluo Sun},
  title = {{Proximal Inference on Population Intervention Indirect Effect}},
  year = {2025},
  eprint = {2504.11848},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.11848},
  howpublished = {arXiv},
}
Figure