Kotti, E;
(2017)
Bayesian variable selection for probit models with an application to clinical diagnosis.
Doctoral thesis , UCL (University College London).
Preview |
Text
PhD_final.pdf Download (5MB) | Preview |
Abstract
My research focuses on the development of a probabilistic model for the classification using spectral measurements of tissue from different stages in the progression of Barrett's oesophagus (BE) and the implementation in this model of variable selection with the aim of improving the classification accuracy. In Chapter 1, a brief introduction to the BE disease, to spectroscopy, and to importance of variable selection is presented. Chapter 2 focuses on (penalized) likelihood methods for variable selection, including also evaluation measures for the performance of prediction models. Chapter 3 introduces Bayesian variable selection (BVS) using a probit model with binary responses. Then, BVS is studied under different prior assumptions for the coefficients and for the indicator vector (indicating if the variable is important). The next chapter contains the results of applying these different assumptions either on real or on simulated binary datasets. The remaining chapters regard the extension of BVS from binary to multi-class responses. Multi-class classification problems have been studied for pure nominal and pure ordinal responses (Chapter 5). However, there are cases with both types of responses, e.g. BE disease. We develop a BVS approach for which the stages of the disease are a mixture of nominal and ordinal responses. To address this problem we build three probit models based on latent variables: (i) a decomposed approach using two indicator vectors, one for nominal and one for ordinal responses (Chapter 6), (ii) BVS approach using a common indicator vector (Chapter 7), and (iii) BVS approach using an indicator matrix, which is a collection of indicator vectors (Chapter 8). Finally, Chapter 9 contains the results of applying the proposed methods to BE for clinical diagnosis and comparing with existing methods. The last chapter contains the conclusions and suggestions for future directions.
Type: | Thesis (Doctoral) |
---|---|
Title: | Bayesian variable selection for probit models with an application to clinical diagnosis |
Event: | UCL (University College London) |
Open access status: | An open access version is available from UCL Discovery |
Language: | English |
UCL classification: | UCL > Provost and Vice Provost Offices UCL > Provost and Vice Provost Offices > UCL BEAMS UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Maths and Physical Sciences UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Maths and Physical Sciences > Dept of Statistical Science |
URI: | https://discovery.ucl.ac.uk/id/eprint/1551658 |
Archive Staff Only
View Item |