Jointly modelling of cluster dependent pro les of fractional and binary variables from a Bayesian point of view
No hay miniatura disponible
Fecha
2020-10-27
Autores
Título de la revista
ISSN de la revista
Título del volumen
Editor
Pontificia Universidad Católica del Perú
DOI
Resumen
En la presente tesis se proponen modelos de clasificación basados en regresiones beta inflacionadas cero-uno con efectos mixtos para modelar perfiles longitudinales de variables fraccionarias mixtas y variables binarias de forma conjunta con formación de clústeres. Las distintas parametrizaciones de los modelos propuestos permiten modelar distintos efectos, como modelar directamente la media marginal a través de covariables e interpretar fácilmente su efecto sobre ella o modelar la media condicional y las probabilidades de inflación de forma separada. Además, se forman clústeres de grupos de individuos con perfiles longitudinales similares a través de una variable latente, asumiendo que las variables respuesta siguen un modelo de mixtura finita. Debido a la complejidad de los modelos, los parámetros se estiman desde un punto de vista bayesiano, a partir de simulaciones MCMC utilizando el software JAGS en R. Se prueban los modelos propuestos sobre diferentes bases de datos simulados
para medir el desempeño de los mismos y se comparan con otros modelos a fin de verificar cual ajusta mejor los perfiles longitudinales de variables fraccionarias mixtas y variables binarias. Por último, se aplican los modelos propuestos a datos reales de un banco peruano, con información del ratio de uso de tarjetas de crédito en el periodo de un año, estado de default del cliente y otras covariables correspondientes al cliente poseedor de la tarjeta, con el objetivo de obtener clústeres de individuos con similar ratio de uso de tarjeta de crédito y relacionarlos con la probabilidad de caer en default que presenta cada grupo.
The following thesis proposes classi cation models that consist of jointly tting longitudinal pro les of mixed fractional and binary variables modelled by zero-one beta in ated mixed regressions with cluster formation. The distinct proposed parametrizations allow di erent effects to be modelled, such as modelling the marginal mean directly through independent variables and easily interpret its e ect on it or modelling the conditional mean and the in- ation probabilities separately. In addition, individuals with similar fractional longitudinal pro les are grouped into a cluster through a latent variable, assuming that the response variables follow a nite mixture model. Due to the complexity of the models, the parameters are estimated from a Bayesian point of view by simulating a MCMC using JAGS software in R. The proposed models are tted in various simulated datasets and are compared against other models to measure performance in tting fractional longitudinal pro les and binary variables. Finally, an application on real data is conducted, consisting on longitudinal information of credit card utilization ratio and default status as dependants variables and covariates corresponding to client information, aiming to obtain clusters of clients with similar behaviour in evolution of credit card utilization and relate them to their probability of default.
The following thesis proposes classi cation models that consist of jointly tting longitudinal pro les of mixed fractional and binary variables modelled by zero-one beta in ated mixed regressions with cluster formation. The distinct proposed parametrizations allow di erent effects to be modelled, such as modelling the marginal mean directly through independent variables and easily interpret its e ect on it or modelling the conditional mean and the in- ation probabilities separately. In addition, individuals with similar fractional longitudinal pro les are grouped into a cluster through a latent variable, assuming that the response variables follow a nite mixture model. Due to the complexity of the models, the parameters are estimated from a Bayesian point of view by simulating a MCMC using JAGS software in R. The proposed models are tted in various simulated datasets and are compared against other models to measure performance in tting fractional longitudinal pro les and binary variables. Finally, an application on real data is conducted, consisting on longitudinal information of credit card utilization ratio and default status as dependants variables and covariates corresponding to client information, aiming to obtain clusters of clients with similar behaviour in evolution of credit card utilization and relate them to their probability of default.
Descripción
Palabras clave
Estadística bayesiana, Análisis de regresión, Estadística--Modelos matemáticos
Citación
Colecciones
item.page.endorsement
item.page.review
item.page.supplemented
item.page.referenced
Licencia Creative Commons
Excepto se indique lo contrario, la licencia de este artículo se describe como info:eu-repo/semantics/openAccess