A simulation study for evaluating the performance of clustering measures in multilevel logistic regression

Area of Science:

• Health Sciences Research
• Biostatistics
• Epidemiology

Background:

• Multilevel logistic regression is crucial for analyzing clustered health data.
• Quantifying between-cluster heterogeneity is essential for accurate effect estimation.
• Existing measures include ICC, MOR, IOR-80, and SOI.

Purpose of the Study:

• To compare the performance of four heterogeneity measures: ICC, MOR, IOR-80, and SOI.
• To assess their accuracy under varying cluster numbers and sizes.
• To evaluate these measures in a real-world childhood anemia study.

Main Methods:

• Simulated two-level logistic regression datasets were used.
• Performance was evaluated based on bias and accuracy with varying cluster numbers and sizes.
• Empirical analysis of childhood anemia data from Malawi was conducted.

Main Results:

• SOI and ICC estimates were unbiased with ≥10 clusters and ≥20 cluster size.
• MOR and IOR-80 estimates were less accurate with ≤50 clusters.
• Performance improved with increased clusters and cluster size for all measures.
• In the anemia study, community heterogeneity (SOI=56.7%) was more significant than rural/urban effects (OR=1.21).

Conclusions:

• At least 300 clusters with ≥50 subjects per cluster are needed for unbiased heterogeneity estimation.
• SOI is recommended when cluster-level covariates are of interest.
• ICC is sufficient for general multilevel logistic regression analyses without covariate focus.