Abstract
Although there are many methods available in the literature to compare the covariance structures of two populations, few are suitable for clinical application due to the inability to account for covariate(s) that affect the dependence structure of the variables being investigated. A common method is to adjust the effect of the covariates via a linear model and work with the resulting residuals. However, removing the effects of the covariates could potentially eliminate valuable information from the analysis. We propose a functional nonparametric covariance matrix estimator to account for any given value in the covariate(s), which allows a comparison of the functional covariance structures of the multivariate data. This comparison is facilitated via a test statistic involving the first eigenvalue of the combined form of covariance matrices of the two groups. Three different approaches, namely, the parametric Tracy-Widom, the semi-parametric Forkman's test, and the nonparametric Permutation method, are used to compute the approximate p-values of the test statistic. We have conducted extensive simulation studies to determine the type I error and power of the proposed hypothesis testing methods and developed practical recommendations for implementing this novel approach. Finally, we apply our methods to the Alzheimer's Disease Neuroimaging Initiative (ADNI) study to compare cerebrospinal fluid (CSF) biomarkers between dementia and non-dementia cohorts, which offers a fascinating insight into the differences between covariance structures of biomarkers amyloid β(1 - 42)(Aβ42) , total tau (tau), and phosphorylated tau (ptau) for given values of age, sex, and years of education.