Analysis of Life Regression Models for Censored Data using Pseudo Observations

Abstract

Censoring is a common form for missing data in survival analysis. When a data set is censored, there is only partial knowledge of the survival time of some of the study units. To compensate for this, special techniques and adjusted residuals may be used in analysis. An alternative to this, is to obtain new data sets through pseudo observations from jackknife theory. These new data sets can then be treated as uncensored data sets, and ordinary regression methods can be applied. This master's thesis studies methods for obtaining pseudo observations based the Kaplan-Meier estimator and modeling by accelerated failure time models (AFT models). Three methods are presented, one parametric and two non-parametric. How well the three methods preforms under different levels of censoring and true distributions are studied, and some recommendations on when they are appropriate to use are made. Pseudo observations are also studied for Cox-Snell and standardized residuals of AFT models, and also here we arrive at some recommendations regarding their use. Both pseudo observations and pseudo residuals are then used in residual analysis and model checking. Methods are illustrated with simulated and real data sets.