PYXIDA Institutional Repository
and Digital Library
Collections :

Title :Bayesian regression models for mortality data
Creator :Alexopoulos, Angelis N.
Αλεξόπουλος, Αγγελής Ν.
Contributor :Dellaportas, Petros (Επιβλέπων καθηγητής)
Athens University of Economics and Business, Department of Statistics (Degree granting institution)
Type :Text
Extent :61p.
Language :en
Abstract :The main theme of this thesis is to provide forecasts for mortality rates by using two different approaches. First, we employ dynamic non-linear logistic models based on Helingman-Pollard formula. Second, we assume that the dynamics of the mortality rates can be modeled through a Gaussian Markov random field. Both methodologies are tested with past data and are used to forecast mortality rates for UK and Wales up to 30 years ahead. In this thesis we make, firstly Bayesian analysis for mortality data. Many researchers in published papers have noticed that estimation of the parameters of the Heligman-Pollard model, Heligman and Pollard (1980), is problematic because of the overparameterization of the model. They also noted that using the weighted least squares approach for the estimation the numerical instabilities can removed only by fixing two parameters to be constant. The above problems can be tackled with Bayesian methods. Dellaportas et al. (2001) adopt a Bayesian inference approach for the Heligman-Pollard model. This approach has the following advantages. First the use of informative priors solves the problem of overparameterization. Secondly the non-normality of the likelihood surface means that the least squares estimates are inadequate. So, as a starting point, we repeated the work of Dellaportas et al. (2001) for recent mortality data. Then, I focused on the important issue of dynamic modeling of mortality data. The approach adopted till now is to first estimate the parameters of mortality rates model for each year (or 5-year) interval, and then modeling the estimated parameters via a time series model. Clearly, the parameter uncertainty and the within model parameter dependence is ignored. In this thesis we worked on this problem with data collected from the Human mortality database ( Our first attempt was to assume a dynamic non-linear generalized model. The dynamic part is added to account for the dynamic projections of demographic characteristics across years. Inference of such a model is very hard, and in particular only ad-hoc solutions for dynamic linear generalized linear models exist. The othermodel-perspective that we attempted is based on non-isotropic Gaussian Processes. Here, our data (mortality rates), were viewed as coming from a huge multivariate normal distribution, in which we estimated the correlation structure and we made forecasts for the future mortality rates, based on this.
Subject :Mortality data
Bayesian regression
Heligman-Pollard model
Gaussian Markov random field
Dynamic modeling
Date :29-06-2012
Licence :

File: Alexopoulos_2012.pdf

Type: application/pdf