Modelling multivariate, overdispersed count data with correlated and non-normal heterogeneity effects

Mixed Poisson models are most relevant to the analysis of longitudinal count data in various disciplines. A conventional specification of such models relies on the normality of unobserved heterogeneity effects. In practice, such an assumptionmay be invalid, and non-normal cases are appealing. In thi...

Descripción completa

Detalles Bibliográficos
Autores: Kazemi, Iraj, Hassanzadeh, Fatemeh
Tipo de recurso: artículo
Fecha de publicación:2020
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/362103
Acceso en línea:https://hdl.handle.net/2117/362103
https://dx.doi.org/10.2436/20.8080.02.105
Access Level:acceso abierto
Palabra clave:Bayesian computation
correlated random effects
hierarchical representation
longitudinal data
multivariate skew-normal distribution
over-dispersion
Distribució (Teoria de la probabilitat)
Estadística matemàtica
Anàlisi multivariable
Classificació AMS::62 Statistics::62E Distribution theory
Classificació AMS::62 Statistics::62J Linear inference, regression
Classificació AMS::62 Statistics::62H Multivariate analysis
Àrees temàtiques de la UPC::Matemàtiques i estadística::Estadística matemàtica
Descripción
Sumario:Mixed Poisson models are most relevant to the analysis of longitudinal count data in various disciplines. A conventional specification of such models relies on the normality of unobserved heterogeneity effects. In practice, such an assumptionmay be invalid, and non-normal cases are appealing. In this paper, we propose a modelling strategy by allowing the vector of effects to follow the multivariate skew-normal distribution. It can produce dependence between the correlated longitudinal counts by imposing several structures of mixing priors. In a Bayesian setting, the estimation process proceeds by sampling variants from the posterior distributions. We highlight the usefulness of our approach by conducting a simulation study and analysing two real-life data sets taken from the German Socioeconomic Panel and the US Centers for Disease Control and Prevention. By a comparative study, we indicate that the new approach can produce more reliable results compared to traditional mixed models to fit correlated count data.