Bayesian multilevel multivariate logistic regression for superiority decision-making under observable treatment...

Area of Science:

• Multilevel modeling
• Biostatistics
• Health research methodology

Background:

• Medical, social, and behavioral research frequently involves multilevel data with multiple dependent variables.
• Subpopulations often exhibit heterogeneous intervention effects, yet standard analyses ignore this complexity.
• Ignoring data structure leads to inflated Type I errors and masks crucial insights.

Purpose of the Study:

• To introduce a novel Bayesian multilevel multivariate logistic regression model for comprehensive data analysis.
• To accurately account for clustered data structures and ensure reliable posterior inferences.
• To enable information sharing across subpopulations for robust treatment effect estimation.

Main Methods:

• Developed a Bayesian multilevel multivariate logistic regression model.
• Incorporated methods to handle clustered data and estimate multivariate treatment effects.
• Transformed parameters to posterior success probabilities for enhanced interpretability.

Main Results:

• Numerical evaluations confirmed the multilevel model's accurate Type I error rates compared to single-level alternatives.
• The multilevel model demonstrated increased statistical power with a higher number of clusters.
• Re-analysis of stroke trial data showcased improved understanding of treatment effects by incorporating multilevel and multivariate structures.

Conclusions:

• The proposed model accurately predicts treatment effects and aids decision-making in clustered subpopulations.
• It leverages the full study sample size while properly incorporating uncertainty.
• Bayes factors can assist in selecting appropriate models for complex data structures.