On Partial Identification of the Natural Indirect Effect

Area of Science:

• Epidemiology
• Biostatistics
• Public Health

Background:

• Causal mediation analysis is crucial for understanding indirect effects of interventions.
• Standard methods often assume no unobserved confounding and cross-world counterfactual independence.
• Existing bounds for partial identification are limited, particularly for non-binary mediators.

Purpose of the Study:

• To extend existing bounds for causal mediation analysis to polytomous mediators.
• To provide new bounds under relaxed assumptions, specifically addressing cross-world counterfactual independence.
• To assess the mediation of antiretroviral therapy effects on virological failure by adherence in the PEPFAR program.

Main Methods:

• Developed nonparametric bounds for the natural indirect effect with polytomous mediators.
• Extended existing bounding techniques to scenarios with relaxed confounding assumptions.
• Applied the extended bounds to real-world data from a large-scale HIV/AIDS program.

Main Results:

• Successfully extended partial identification bounds to polytomous mediators.
• Demonstrated the utility of these bounds in a public health context (PEPFAR Nigeria).
• Found that causal inference regarding mediation by adherence is sensitive to underlying model assumptions.

Conclusions:

• The extended bounds provide a more flexible framework for causal mediation analysis with polytomous mediators.
• This work offers valuable tools for researchers when strong identifying assumptions cannot be met.
• Findings highlight the importance of considering assumption sensitivity in mediation studies of health interventions.