Comparative Evaluation of Statistical Dependence Measures
Eman Abdel Rahman Ibrahim · Journal of the Arkansas Academy of Science · 2020
Measuring and testing dependence between random variables is of great importance in many scientific fields. In the case of linearly correlated variables, Pearson’s correlation coefficient is a commonly used measure of the correlation strength. In the case of nonlinear correlation, several innovative measures have been proposed, such as distance-based correlation, rank-based correlations, and information theory-based correlation. This thesis focuses on the statistical comparison of several important correlations, including Spearman’s correlation, mutual information, maximal information coefficient, biweight midcorrelation, distance correlation, and copula correlation, under various simulation settings such as correlative patterns and the level of random noise. Furthermore, we apply those correlations with the overall best performance to a real genomic data set, to study the co-expression between genes in serous ovarian cancer.