Sensitivity analysis for causal mediation: bridge score, sharp sensitivity bounds, and calibration

2026-05-18

AI Summary

This work addresses the untestable sequential ignorability assumption in causal mediation analysis by proposing a sensitivity analysis framework to assess the impact of unmeasured confounding on natural direct and indirect effects. The authors introduce a “bridge score”—a low-dimensional balancing score for the mediator stage—and derive sharp pointwise sensitivity bounds governed by two interpretable latent parameters. They develop two calibration strategies: benchmark calibration (including a rank-invariant variant) and residual budget calibration. Integrating scalar functional dimension reduction with a Bayesian g-computation algorithm, the method effectively quantifies the influence of unobserved confounders while fully propagating uncertainty through posterior sampling, thereby substantially enhancing the rigor and practical utility of sensitivity analysis in causal mediation settings.

Abstract

Causal mediation analysis decomposes the total treatment effect into a portion operating through a hypothesized mediator and a residual direct portion. Identification of natural direct and indirect effects typically rests on the mediator stage of sequential ignorability, which cannot be empirically verified and requires explicit sensitivity analysis. We introduce the \emph{bridge score}, a low-dimensional vector formed from the two treatment-specific mediator densities at a common mediator value, and show that it is a balancing score for the mediator stage of sequential ignorability. Conditional on the bridge score, we then derive a sharp pointwise envelope on the unidentified mediator-outcome confounding function in terms of two interpretable latent confounding parameters. To make the bound operational for sensitivity analysis, we further introduce two calibration approaches. The first is benchmark calibration against an observed covariate, including a rank-based version that is invariant to monotone re-expressions of the benchmark; the second is residual budget calibration based on residual outcome variation. Finally, we show how the pointwise bound can be operationalized for inference through a scalar functional reduction and a Bayesian g-computation algorithm that propagates all sources of uncertainty into posterior draws of the mediation effect estimates.

Problem

• causal mediation
• sensitivity analysis
• sequential ignorability
• unobserved confounding
• mediation effect

Innovation

• bridge score
• sharp sensitivity bounds
• causal mediation analysis
• benchmark calibration
• Bayesian g-computation