A1267
Title: Feature screening for high-dimensional data with measurement errors using adjusted martingale difference correlation
Authors: Wenbo Wu - University of Texas at San Antonio (United States) [presenting]
Baoying Yang - Southwest Jiaotong University (China)
Abstract: High-dimensional data with measurement errors pose significant challenges for statistical modeling and inference due to the "curse of dimensionality" and the unavailability of direct measurements of variables. To reduce the dimensionality of data, feature screening is an effective method for identifying informative variables among a large number of observed features. While feature screening has been extensively studied in the literature, limited research has focused on feature screening with random variables affected by measurement errors. An adjusted martingale difference correlation (AMDC) is proposed to measure conditional mean dependence in the presence of measurement errors. The proposed AMDC is further extended to measure the dependence of the conditional quantile and conditional k-th central moment through specific transformations of the error-prone variables. Invariance relationships between the AMDC and the original martingale difference correlation are established and different feature screening methods are developed for the conditional mean, conditional quantile, and conditional k-th central moment of the response variable under a measurement error model. Sure screening properties are established for all the proposed procedures and extensive simulation studies are conducted to illustrate their effectiveness and advantages. Finally, microRNA expression data is analyzed using the proposed AMDC-based screening methods to demonstrate their practical usefulness.
