Methodological study uncovers nonlinear local dependence structures in heterogeneous clinical populations, highlighting patient subgroups missed by conventional correlation metrics.
Key Points
To develop a flexible statistical framework, Generalized Local Kendall’s τ, capable of detecting nonlinear and localized dependencies often overlooked by global dependence measures in heterogeneous populations.
Constructed U-statistic-based and copula model-based estimators supporting diverse localized geometric shapes, including rectangular, square, and extreme-value neighborhoods.
Introduced four quantile dependence measures, novel tail dependence coefficients, and two Archimedean copula surface visualization tools.
Evaluated estimation accuracy via numerical simulation benchmarks and demonstrated clinical utility using real-world Parkinson's disease and COVID-19 datasets.
Local Kendall's τ values showed marked divergence from global estimates, demonstrating that standard global metrics fail to capture localized dependence shifts.
Copula-based estimators achieved superior estimation performance over alternatives when underlying models were correctly specified in simulations.
Application to Parkinson's disease and COVID-19 data successfully identified distinct, subgroup-specific dependence structures essential for precision medicine.