Robust inference on indirect causal effects

Isabel Fulcher, Ilya Shpitser, Stella Marealle, Eric J. Tchetgen Tchetgen

Abstract

Standard methods for inference about direct and indirect effects require stringent no unmeasured confounding assumptions which often fail to hold in practice, particularly in observational studies. This paper introduces a new form of indirect effect, the population intervention indirect effect (PIIE), that can be nonparametrically identified in the presence of an unmeasured common cause of exposure and outcome. This new type of indirect effect captures the extent to which the effect of exposure is mediated by an intermediate variable under an intervention which fixes the component of exposure directly influencing the outcome at its observed value. Interestingly, the identification criterion relaxes Judea Pearl's front-door criterion as it does not require no direct effect of exposure not mediated by the intermediate variable. For inference, the authors develop both parametric and semiparametric methods, including a doubly robust semiparametric locally efficient estimator, that perform very well in simulation studies. Finally, the proposed methods are used to measure the effectiveness of monetary saving recommendations among women enrolled in a maternal health program in Tanzania.

Introduction

In causal inference, mediation analysis is the process of decomposing the total effect of an exposure ($A$) on an outcome ($Y$) into pathways that act through an intermediate variable, or mediator ($Z$). Traditional estimands, such as the Natural Indirect Effect (NIE), rely on the "no unmeasured confounding" assumption. This requires researchers to account for every variable that might influence both the exposure and the outcome—a nearly impossible task in observational research where hidden factors like motivation, literacy, or environmental conditions often lurk in the background.

This paper addresses this fundamental challenge by introducing the Population Intervention Indirect Effect (PIIE). The PIIE provides a robust framework for identifying mediated effects even when unmeasured confounding exists between the exposure and the outcome, provided the mediator itself is well-behaved.

The Population Intervention Indirect Effect (PIIE)

The PIIE is the indirect component of the population intervention effect. Conceptually, it compares the observed average outcome to a hypothetical scenario where the mediator ($Z$) is set to the value it would have taken if the subjects had not been exposed, while the exposure’s direct influence on the outcome remains at its naturally observed level.

Generalizing the Front-Door Criterion

A significant theoretical contribution of this work is the generalization of Judea Pearl’s front-door criterion. The original front-door criterion is a powerful tool for causal identification under confounding, but it carries a restrictive requirement: the exposure must have no direct effect on the outcome except through the mediator ($A \to Z \to Y$). The PIIE relaxes this, allowing for a direct path from exposure to outcome ($A \to Y$). This makes the method applicable to a much broader range of scientific inquiries where multiple causal pathways are likely present.

Doubly Robust Estimation

To make the PIIE practically useful, the authors developed a doubly robust (DR) semiparametric locally efficient estimator. In statistical modeling, "double robustness" refers to a safety mechanism: the estimator remains accurate (consistent) if either:

  1. The models for the outcome and exposure are correctly specified, OR
  2. The model for the mediator density is correctly specified.

This property is invaluable in complex social and medical datasets where the "true" underlying model is unknown. By providing two chances to get the mathematics right, the DR estimator offers a more reliable path to valid causal conclusions than traditional Maximum Likelihood Estimation (MLE).

Empirical Validation

Simulation Studies

The authors tested their estimators using extensive synthetic datasets designed to mimic various conditions of model misspecification and unmeasured confounding. The results showed that the DR estimator consistently outperformed simpler methods. Even when the outcome model was intentionally "broken" or misspecified, the DR estimator stayed on target, whereas standard estimators produced biased results.

Case Study: Maternal Health in Tanzania

The methodology was applied to data from the Safer Deliveries Program in Zanzibar, Tanzania, involving 4,102 pregnant women. The study examined whether a mobile health app's "recommended savings amount" (the mediator) influenced "actual savings" (the outcome) for women categorized as "high risk" (the exposure).

The analysis found that the PIIE was not significantly different from zero. This suggests that simply increasing the savings recommendation for high-risk women did not effectively translate into higher actual savings compared to what they would have saved under a standard recommendation. For program designers, this insight is vital: it indicates that the recommendation alone was not the primary driver of behavior, and other interventions may be necessary to support financial preparation for birth.

Limitations and Considerations

While the PIIE is a major step forward, it is not a universal solution for all confounding issues. The method still requires that there be no unmeasured confounding between the exposure and the mediator ($A$ and $Z$) or between the mediator and the outcome ($Z$ and $Y$). Additionally, while the DR estimator can theoretically incorporate machine learning techniques (like Random Forests), these methods must meet specific convergence rates to ensure valid statistical inference, which remains a high bar in practical applications.

Conclusion

The introduction of the Population Intervention Indirect Effect and its associated doubly robust estimator provides researchers with a powerful new tool for mediation analysis. By relaxing the requirements of the front-door criterion and providing a safeguard against model misspecification, this framework allows for more credible causal claims in the presence of the hidden confounding that characterizes real-world observational data.

Summary updated

Abstract

Standard methods for inference about direct and indirect effects require stringent no unmeasured confounding assumptions which often fail to hold in practice, particularly in observational studies. The goal of this paper is to introduce a new form of indirect effect, the population intervention indirect effect (PIIE), that can be nonparametrically identified in the presence of an unmeasured common cause of exposure and outcome. This new type of indirect effect captures the extent to which the effect of exposure is mediated by an intermediate variable under an intervention which fixes the component of exposure directly influencing the outcome at its observed value. The PIIE is in fact the indirect component of the population intervention effect, introduced by Hubbard and Van der Laan (2008). Interestingly, our identification criterion relaxes Judea Pearl's front-door criterion as it does not require no direct effect of exposure not mediated by the intermediate variable. For inference, we develop both parametric and semiparametric methods, including a doubly robust semiparametric locally efficient estimator, that perform very well in simulation studies. Finally, the proposed methods are used to measure the effectiveness of monetary saving recommendations among women enrolled in a maternal health program in Tanzania.
BibTeX
@article{fulcher2019robust,
  author = {Isabel Fulcher and Ilya Shpitser and Stella Marealle and Eric J. Tchetgen Tchetgen},
  title = {{Robust inference on indirect causal effects}},
  journal = {Journal of the Royal Statistical Society Series B (Statistical Methodology)},
  year = {2019},
  doi = {10.1111/rssb.12345},
}
Figure