ABSTRACT
The discovery of associations between gene expression and patient outcome relies on statistical models which typically use the expression measurement directly as a predictor. A more accurate approach would assume that the expression measurement is merely an indication of the underlying true expression level. The proposed model recognizes that gene expression has measurement error. Experiments involving many tumors will often perform only one microarray assay on each tumor because of cost limitations. Thus, the distributions of measurement errors are not directly identifiable. This model extends the additive error-in-variable survival model for Affymetrix data of Tadesse et al. (2005) to two color microarrays with correlated multiplicative errors. The error model is included within the framework of a piecewise exponential survival model. Robustness analyses are performed, and the model is applied to a breast cancer study dataset.
