Estimating the Relative Excess Risk Due to Interaction in Clustered-Data Settings

Area of Science:

• Biostatistics
• Epidemiology
• Public Health

Background:

• Additive interaction is key for evaluating public health interventions on binary outcomes.
• Standard models for binary outcomes often measure interaction multiplicatively, not additively.
• The relative excess risk due to interaction (RERI) assesses additive interaction but rarely accounts for clustered data.

Purpose of the Study:

• To evaluate the RERI metric in the context of clustered data.
• To compare population-averaged and cluster-conditional models for RERI estimation.
• To develop and assess a Bayesian implementation for cluster-conditional models.

Main Methods:

• Simulation studies comparing RERI estimation using population-averaged and cluster-conditional models.
• Evaluation of frequentist implementations of cluster-conditional models with random intercepts.
• Development and application of a Bayesian implementation of log binomial random-intercept models.

Main Results:

• RERI estimation and inference were straightforward using population-averaged models in simulations.
• Frequentist cluster-conditional models often encountered convergence issues or degenerate variance estimates.
• The Bayesian implementation of log binomial random-intercept models proved effective for cluster-conditional RERI estimation.

Conclusions:

• Population-averaged models offer a straightforward approach for RERI in clustered data.
• Frequentist cluster-conditional models present challenges for RERI estimation with clustered data.
• Bayesian log binomial random-intercept models provide a viable alternative for cluster-conditional RERI analysis.