A Risk-Ratio-Based Marginal Sensitivity Model for Causal Effects in Observational Studies

TL;DR

This paper introduces a new risk-ratio-based sensitivity analysis framework for assessing the impact of unmeasured confounding on causal effect estimates in observational studies, applicable to binary and multivalued treatments.

Contribution

It proposes a modified marginal sensitivity model using risk ratios, extending to multivalued treatments, and provides efficient methods for constructing confidence intervals.

Findings

Framework performs well with adequate treatment overlap

Simulation studies validate the approach's effectiveness

Applied to maternal education and fertility data in Bangladesh

Abstract

In observational studies, the identification of causal estimands depends on the no unmeasured confounding (NUC) assumption. As this assumption is not testable from observed data, sensitivity analysis plays an important role in observational studies to investigate the impact of unmeasured confounding on the causal conclusions. In this paper, we proposed a risk-ratio-based sensitivity analysis framework by introducing a modified marginal sensitivity model for observational studies with binary treatments. We further extended the proposed framework to the multivalued treatment setting.We then showed how the point estimate intervals and the corresponding percentile bootstrap confidence intervals can be constructed efficiently under the proposed framework. Simulation results suggested that the proposed framework of sensitivity analysis performs well in the presence of adequate overlap among the treatment groups. Lastly, we demonstrated our proposed sensitivity analysis framework by estimating the causal effect of maternal education on female fertility in Bangladesh.