Abstract:
|
Traditional multivariate tests for means often fail in high dimensions or with functional data due in part to the rank deficiency of the sample covariance matrix. Accommodations to T-squared-style statistics to accommodate this deficiency are often only applicable with prohibitively large sample sizes. An alternative solution is to replace the classical null hypothesis with a neighborhood hypothesis to test for approximate equality. The existing methods for this require the size of the neighborhood to be specified a priori with respect to a distance measure which may be difficult to interpret practically. Additionally, since sample spaces are not compact for many applications, it can be difficult to determine an appropriate upper bound for the radius of the neighborhood. Hence, we propose a novel modified neighborhood testing procedure that defines a neighborhood as a proportion of the total amount of variation present in the population under study. Here, we present a novel test statistic and describe its properties through asymptotic results and simulation studies. Finally, we apply this framework to address a problem arising in the field of precision medicine.
|