Regression modelling using priors depending on Fisher information covariance kernels (I-priors)

Jamil, Haziq Md. (2018) Regression modelling using priors depending on Fisher information covariance kernels (I-priors). PhD thesis, London School of Economics and Political Science.

Text - Submitted Version Download (6MB)

Abstract

Regression analysis is undoubtedly an important tool to understand the relationship between one or more explanatory and independent variables of interest. In this thesis, we explore a novel methodology for fitting a wide range of parametric and nonparametric regression models, called the I-prior methodology (Bergsma, 2018). We assume that the regression function belongs to a reproducing kernel Hilbert or Kreĭn space of functions, and by doing so, allows us to utilise the convenient topologies of these vector spaces. This is important for the derivation of the Fisher information of the regression function, which might be infinite dimensional. Based on the principle of maximum entropy, an I-prior is an objective Gaussian process prior for the regression function with covariance function proportional to its Fisher information. Our work focusses on the statistical methodology and computational aspects of fitting I-priors models. We examine a likelihood-based approach (direct optimisation and EM algorithm) for fitting I-prior models with normally distributed errors. The culmination of this work is the R package iprior (Jamil, 2017) which has been made publicly available on CRAN. The normal I-prior methodology is subsequently extended to fit categorical response models, achieved by “squashing” the regression functions through a probit sigmoid function. Estimation of I-probit models, as we call it, proves challenging due to the intractable integral involved in computing the likelihood. We overcome this difficulty by way of variational approximations. Finally, we turn to a fully Bayesian approach of variable selection using I-priors for linear models to tackle multicollinearity. We illustrate the use of I-priors in various simulated and real-data examples. Our study advocates the I-prior methodology as being a simple, intuitive, and comparable alternative to similar leading state-of-the-art models.

Item Type:	Thesis (PhD)
Additional Information:	© 2018 Md. Haziq Md. Jamil
Library of Congress subject classification:	H Social Sciences > HA Statistics
Sets:	Departments > Statistics
Supervisor:	Moustaki, Irini and Bergsma, Wicher
URI:	http://etheses.lse.ac.uk/id/eprint/3828

Actions (login required)

Record administration - authorised staff only