Robust versions of classical multivariate techniques based on the Cauchy likelihood

Fayomi, Aisha Fouad (2013) Robust versions of classical multivariate techniques based on the Cauchy likelihood. PhD thesis, University of Nottingham.

[img] PDF - Repository staff only - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Download (954kB)


Classical multivariate analysis techniques such as principal components analysis (PCA), canonical correlation analysis (CCA) and discriminant analysis (DA) can be badly affected when extreme outliers are present. The purpose of this thesis is to present new robust versions of these methods. Our approach is based on the following observation: the classical approaches to PCA, CCA and DA can all be interpreted as operations on a Gaussian likelihood function. Consequently, PCA, CCA and DA can be robustified by replacing the Gaussian likelihood with a Cauchy likelihood. The performance of the Cauchy version of each of these procedures is studied in detail both theoretically, through calculation of the relevant influence function, and numerically, through numerous examples involving real and simulated data. Our results demonstrate that the new procedures have good robustness properties which are certainly far superior to these of the classical versions.

Item Type: Thesis (University of Nottingham only) (PhD)
Supervisors: Wood, A.T.A.
Brignell, C.J.
Subjects: Q Science > QA Mathematics > QA273 Probabilities
Faculties/Schools: UK Campuses > Faculty of Science > School of Mathematical Sciences
Item ID: 13446
Depositing User: EP, Services
Date Deposited: 11 Dec 2013 15:24
Last Modified: 21 Dec 2017 19:04

Actions (Archive Staff Only)

Edit View Edit View