Nonparametric Bounds in Causal Mediation Analysis in the Presence of Unmeasured Confounding and Imperfect Compliance

2025-09-02

AI Summary

In randomized experiments with unmeasured confounding and imperfect compliance, the natural direct effect (NDE) and average causal mediation effect (ACME) are generally nonidentifiable. This paper proposes a nonparametric sharp bounding approach to partially identify these effects for the complier subpopulation within the instrumental variable framework. Building on Sjölander’s bounding theory and integrating the Balke–Pearl algorithm, monotonicity assumptions, and potential outcomes modeling, we derive tight, theoretically justified bounds for both NDE and ACME. The method relies solely on observed experimental data and imposes no parametric restrictions on unmeasured confounders. Empirical application to a job-search intervention study demonstrates substantially tightened bounds with clear policy interpretability. Our contribution is the first derivation of sharp bounds for NDE and ACME among compliers under minimal assumptions, providing a practical, robust tool for sensitivity analysis and causal policy evaluation in the presence of unobserved confounding.

Technology Category

• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality
• Multiagent Systems: Mechanism Design

Application Category

• Responsible Web: Human-perceived consequences of algorithmic deployment on the web
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs
• Security and Privacy: Large-scale security measurements

Abstract

The average causal mediation effect (ACME) and the natural direct effect (NDE) are two parameters of primary interest in causal mediation analysis. However, the two causal parameters are not identifiable from randomized experimental data in the presence of outcome-mediator confounding and treatment-assignment noncompliance. Sjölander (2009) addressed the partial identification issue and derived nonparametric bounds of the NDE in randomized controlled trials under a set of monotonicity assumptions based on the Balke-Pearl algorithm. These bounds provide partial information on the parameters and can be used for sensitivity analysis. In this paper, we extend Sjölander's bounds on the NDE as well as the ACME to randomized controlled trials in the presence of noncompliance when the treatment assignment serves as an instrumental variable. Nonparametric sharp bounds for the local causal parameters defined on the subpopulation of treatment-assignment compliers are also established. We demonstrate the practical usefulness of the proposed upper and lower bounds through an application to the randomized experimental dataset on Job Search Intervention Study.

Problem

• Extends nonparametric bounds for causal mediation effects
• Addresses unmeasured confounding and imperfect compliance issues
• Develops sharp bounds for local causal parameters

Innovation

• Extends nonparametric bounds to ACME and NDE
• Uses treatment assignment as instrumental variable
• Establishes sharp bounds for local causal parameters