Date of Award

2010

Thesis Type

Masters

Document Type

Dissertation

Divisions

Faculty of Science

Department

Institute of Mathematical Sciences

Institution

Universiti Malaya

Abstract

This study proposes outlier and influential observation detection procedures for Cox proportional hazard model. In the estimation process, the parameters for Cox proportional hazard model are estimated using partial likelihood method, while the baseline hazard estimates are obtained using Nelson-Aalen method. The procedure of outlier detection is based on three types of residuals; deviance, log-odd and normal deviate residuals. We study their properties and compare their performance in detecting outliers via simulation. On the other hand, we propose a procedure of identifying influential observation using forward search method. The method has been shown to be effective in detecting influential observations in linear regression and, more importantly, generalized linear models. The later motivates us to extend the method to Cox proportional hazard model due to their close resemblance. We compare the results with that of case-deletion method. As for illustration, the proposed procedure is applied on the prostate cancer data and two cohorts of local breast cancer patients diagnosed at the Breast Cancer Center, University of Malaya Medical Center from year 1993 to year 2002. In both cases, the proposed procedures have successfully detected outliers and influential observations based on the best fitted Cox proportional hazard model of the full data set.

Initial

khm

Additional Information

Dissertation (M.A.) – Faculty of Science, Universiti Malaya, 2010.

Download

Share

COinS