《bti294.dvi-Vol. 21 no. 9 2005, pages 1979ndash;1986》.pdfVIP

Vol. 21 no. 9 2005, pages 1979–1986 BIOINFORMATICS ORIGINAL PAPER doi:10.1093/bioinformatics/bti29 Gene expression Estimating misclassification error with small samples via bootstrap cross-validation Wenjiang J. Fu∗ , Raymond J. Carroll and Suojin Wang Department of Statistics, Texas A University, 447 Blocker Building, 3143 TAMU, College Station, TX 77843, USA Received on October 20, 2004; revised on January 20, 2005; accepted on January 21, 2005 Advance Access publication February 2, 2005 ABSTRACT bootstrap resampling with training set separated from test set and Motivation: Estimation of misclassification error has received increas- yielded slightly biased estimation. Ambroise and McLachlan (2002) ing attention in clinical diagnosis and bioinformatics studies, especially applied 10-fold CV and BT632+ methods to highly fit microarray in small sample studies with microarray data. Current error estim- data, where the number of genes is huge, in the order of thousands, ation methods are not satisfactory because they either have large while the sample size is relatively small, in the order of tens or hun- variability (such as leave-one-out cross-validation) or large bias (such dreds. Their findings with small samples are encouraging. We are as resubstitution and leave-one-out bootstrap). While small sample thus motivated to investigate whether bootstrap resampling improves size remains one of the key features of costly clinical investigations CV with small samples and study the bootstrap cross-validation or of microarray studies that have limited resources in funding, t