PYXIDA Institutional Repository
and Digital Library
Collections :

Title :Modelling multivariate time series for count data
Creator :Pedeli, Xanthi
Contributor :Karlis, D. (Επιβλέπων καθηγητής)
Athens University of Economics and Business, Department of Statistics (Degree granting institution)
Type :Text
Extent :216p.
Language :en
Abstract :The study of time series models for count data has become a topic of special interest during the last years. However, while research on univariate time series for counts now flourishes, the literature on multivariate time series models for count data is notably more limited. The main reason for this is that the analysis of multivariate counting processes presents many more difficulties. Specifically, the need to account for both serial and cross–correlation complicates model specification, estimation and inference. This thesis deals with the class of INteger–valued AutoRegressive (INAR) processes, a recently popular class of models for time series of counts. The simple, univariate INAR(1) process is initially extended to the 2–dimensional space. In this way, a bivariate (BINAR(1)) process is introduced. Subsequently, the time invariant BINAR(1) model is generalized to a BINAR(1) regression model. Emphasis is given on models with bivariate Poisson and bivariate negative binomial innovations. The properties of the BINAR(1) model are studied in detail and the methods of moments, Yule-Walker and conditional maximum likelihood are proposed for the estimation of its unknown parameters. The small sample properties of the alternative estimators are examined and compared through a simulation experiment. Issues of diagnostics and forecasting are considered and predictions are produced by means of the conditional forecast distribution. Estimation uncertainty is accommodated by taking advantage of the asymptotic normality of maximum likelihood estimators and constructing appropriate confidence intervals for the h–step–ahead conditional probability mass function. A generalized specification of the BINAR(1) process, where cross–correlation between the two series receives contribution from two different sources, is also discussed. In this case, we mainly focus on a specific parametric case that arises under the assumption that the innovations follow jointly a bivariate Poisson distribution. The resulting joint distribution of the bivariate series is identified as an 8–parameters bivariate Hermite. At a second stage, the BINAR(1) process is extended to the multi–dimensional space. Thus, we define a multivariate integer–valued autoregressive process of order 1 (MINAR(1)) and examine its basic statistical properties. Such an extension is not simple and we emphasize on problems that occur, relating to selecting a reasonable innovation distribution as well as on problems related to inference. Apart from the general specification of the MINAR(1) process, we also study two specific parametric cases that arise under the assumptions of a multivariate Poisson and a multivariate negative binomial distribution for the innovations of the process. To overcome the computational difficulties of the maximum likelihood approach we suggest the method of composite likelihood. The performance of the two methods of estimation (i.e. maximum likelihood and composite likelihood) is compared through a small simulation experiment. Extensions to incorporate covariance information are also discussed. The proposed models are illustrated on multivariate count series from the fields of accident analysis, syndromic surveillance and finance.
Subject :Time series
Statistical data
Research methods
Date Issued :06-2011
Licence :

File: Pedeli_2011.pdf

Type: application/pdf