ЭМПИРИЧЕСКОЕ ИССЛЕДОВАНИЕ МОЩНОСТИ СТАТИСТИЧЕСКОГО КРИТЕРИЯ КОЛМОГОРОВА–СМИРНОВА В ЗАДАЧАХ ПРОВЕРКИ ГИПОТЕЗ О ЗАКОНЕ РАСПРЕДЕЛЕНИЯ

Abstract

Statistical tests that use A.N. Kolmogorov one-sample statistics and N.V. Smirnov two-sample statistics have been widely used for solving applied problems for almost a hundred years. Those criteria are present in many textbooks and implemented in computer software for data analysis. The purpose of the work is to empirically estimate the power of the Kolmogorov–Smirnov criterion on a set of test problems to check hypotheses about the distribution law, as well as to investigate the properties of estimates of the power of the criterion for various types of resampling. To obtain various estimates of the power of the Kolmogorov–Smirnov criterion in solving test problems, the classical bootstrap, nonparametric bootstrap, parametric bootstrap, bootstrap with a random term, and resampling without returns were used. Based on the results of multiple solution of test problems, medians of p-values, medians of estimates of the power of the Kolmogorov–Smirnov criterion were calculated, and medians of biases of estimates on test problems and medians of all pairwise differences of various estimates of the power of the criterion were found. The distribution laws for the considered Kolmogorov–Smirnov power estimates, for the biases of the power estimates, and for all pairwise differences of the power estimates were investigated. For test problems where real data were considered, calculating power estimates without resampling is impossible, but unbiased power estimates can be predicted as the half-sum of two other power estimates, one obtained with a nonparametric bootstrap and the other obtained with resampling without returns. Methods for restructuring sample data and software for estimating the power of a statistical test developed within the study are universal and can be used for analyzing the properties of power estimates of other one-sample statistical tests, as well as for developing the classical methodology for checking statistical hypotheses.