An Empirical Comparison of Several Dimensional Reduction Techniques for DNA Microarray Data

• Activity Number: 475
• Type: Contributed
• Date/Time: Thursday, August 7, 2003 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Computing
• Author(s): Zixing Fang*+ and Yunda Huang and Tao Shi and Steve Horvath
• Companies: University of California, Los Angeles and University of California, Los Angeles and University of California, Los Angeles and University of California, Los Angeles
• Keywords: DNA microarray data ; dimension reduction ; random forest ; tumor classification ; gene filtering ; discriminant analysis

Abstract

Gene expression microarray data are increasingly used for classifying tumor samples. Microarrays allow the simultaneous monitoring of thousands of gene expressions per sample. Standard statistical methodologies for prediction do not work well since the number of covariates is usually much larger than the number of observations. Several standard dimensional reduction methods (e.g. partial least squares, principal components) have been successfully used but untransformed covariates are often preferred due to ease of interpretation. Thus it is common practice to reduce the number of genes by thresholding either a measure of variability (e.g., the coefficient of variation) or a test statistic across the outcome classes (e.g., the F-test). Here we conduct an empirical study to compare the aforementioned dimensional reduction techniques to the following two-step approach: first, filter out genes on the basis of variable importance measures from a random forest predictors (L. Breiman 2002). Second, use the resulting covariates in a linear discriminant or classification tree predictor. We propose several ways of arriving at unbiased estimates of the resulting error rates.