Classification in the missing data

Abstract

Missing data is always a problem when it comes to data analysis. This is especially the case in anthropology when sex determination is one of the primary goals for fossil skull data since many measurements were not available. We expect to find a classifier that can handle the large amount of missingness and improve the ability of prediction/classification as well. These are the objectives of this thesis.

Besides of the crude methods (ignore cases with missingness), three possible techniques in handling of missing values are discussed: bootstrap imputation, weighted-averaging classifier and classification trees. All these methods do make use of all the cases in data and can handle any cases with missingness.

The diabetes data and fossil skull data are used to compare the performance of different methods regarding to misclassification error rate. Each method has its own advantages and certain situations under which better performance will be achieved.