Out-of-Sample Fusion

Abstract

A novel method, called ``out-of-sample fusion", is proposed in this dissertation. This method utilizes artificial samples along with a real data sample of interest to draw statistical inference assuming a semiparametric density ratio model. These artificial samples do not relate directly to the sample of interest, which differentiates the method from the traditional bootstrap approach which is a ``within-sample'' method. Out-of-sample fusion has been elaborated on through the estimation of threshold probabilities and their confidence intervals. A comparison has been made with the Agresti-Coull and the standard Wald methods in terms of confidence interval estimation. The out-of-sample fusion generates sharper and shorter confidence intervals while the nominal coverage is maintained. The out-of-sample method has been applied to cancer and microarray data. An R package has been developed to facilitate the implementation of the out-of-sample fusion method.