Incorporating unobserved heterogeneity and multiple event types in survival models : a Bayesian approach

Preview

PDF WRAP_THESIS_Vallejos_2014.pdf - Submitted Version - Requires a PDF viewer. Download (20MB) | Preview

Request Changes to record.

Abstract

This thesis covers theoretical and practical aspects of Bayesian inference and survival analysis, which is a powerful tool for the analysis of the time until a certain event of interest occurs. This dissertation focuses on non-standard models inspired by features of real datasets that are not accommodated by conventional models.

Materials are divided in two parts. The first and more extended part relates to the development of flexible parametric lifetime distributions motivated by the presence of anomalous observations and other forms of unobserved heterogeneity. Chapter 2 presents the use of mixture families of lifetime distributions for this purpose. This idea can be interpreted as the introduction of an observation-specific random effect on the survival distribution. Two families generated via this mechanism are studied in Chapter 3. Covariates are introduced through an accelerated failure times representation, for which the interpretation of the regression coefficients is invariant to the distribution of the random effect. The Bayesian model is completed using reasonable (improper) priors that require a minimum input from practitioners. Under mild conditions, these priors induce a well-defined posterior distribution. In addition, the mixture structure is exploited in order to propose a novel method for outlier detection where anomalous observations are identified via the posterior distribution of the individual-specific random effects. The analysis is illustrated in Chapter 4 using three real medical applications.

Chapter 5 comprises the second part of this thesis, which is motivated in the context of university outcomes. The aim of the study is to identify determinants of the length of stay at university and its associated academic outcome for undergraduate students of the Pontificia Universidad Católica de Chile. In this setting, survival times are defined as the time until the end of the enrollment period, which can relate to different reasons - graduation or dropout - that are driven by different processes. Hence, a competing risks model is employed for the analysis. Model uncertainty is handled through Bayesian model averaging, which leads to a better predictive performance than choosing a unique model. The output of this analysis does not account for all features of this complex dataset yet it provides a better understanding of the problem and a starting point for future research.

Finally, Chapter 6 summarizes the main findings of this work and suggests future extensions.

Request changes or add full text files to a record

Repository staff actions (login required)

View Item