Bayesian hierarchical modeling for a non-randomized, longitudinal fall prevention trial with spatially correlated...

Area of Science:

• Public Health
• Gerontology
• Biostatistics

Background:

• Community-based health interventions frequently utilize non-randomized designs due to feasibility constraints.
• Spatial units like zip codes are common in these designs, raising concerns about unmeasured confounding factors.
• Bayesian hierarchical modeling offers a method to analyze spatial data and quantify unmeasured variability.

Purpose of the Study:

• To describe the Bayesian hierarchical analysis of a large-scale, non-randomized, community-wide healthcare intervention trial.
• To emphasize the spatial and longitudinal characteristics of the intervention's evaluation.
• To compare different modeling approaches using simulations and spatial residual analysis.

Main Methods:

• Bayesian hierarchical modeling was employed to analyze the non-randomized intervention data.
• Spatially correlated residual terms were used to quantify unmeasured variability.
• Posterior predictive simulations and maps of spatial residuals were utilized for model comparison.

Main Results:

• The intervention demonstrated an 11% lower rate of fall-related hospital and emergency department utilization among individuals aged 70 and older.
• A 9% reduction in utilization for serious fall-related injuries was observed in the intervention area.
• Graphical analysis of spatial residuals helped assess potential bias from unmeasured covariates.

Conclusions:

• Bayesian hierarchical modeling is a valuable tool for analyzing non-randomized, spatially-based health interventions.
• The Connecticut Collaboration for Fall Prevention trial successfully reduced fall-related healthcare utilization in older adults.
• Spatial residual analysis aids in understanding and mitigating bias in observational health studies.