Comparação entre modelos para predição do nitrogênio mineralizado: uma abordagem bayesiana

Recent studies use the Bayesian Inference in the most several areas. Therefore, intends in this work to develop a Bayesian boarding to predict of the nitrogen mineralized through nonlinear models, that is, to adjust a model of probability for a group of data and to synthesize the result through a di...

Descripción completa

Detalles Bibliográficos
Autor: Pereira, Janser Moura
Tipo de recurso: tesis doctoral
Estado:Versión publicada
Fecha de publicación:2014
País:Brasil
Institución:Universidade Federal de Lavras (UFLA)
Repositorio:Repositório Institucional da UFLA
Idioma:portugués
OAI Identifier:oai:repositorio.ufla.br:1/4302
Acceso en línea:https://repositorio.ufla.br/handle/1/4302
Access Level:acceso abierto
Palabra clave:Estatística
Inferência bayesiana
Modelos não lineares
Amostrador de Gibbs
Metropolis Hastings
Fator de Bayes
Critério de informação bayesiano
Bayesian inference
Descripción
Sumario:Recent studies use the Bayesian Inference in the most several areas. Therefore, intends in this work to develop a Bayesian boarding to predict of the nitrogen mineralized through nonlinear models, that is, to adjust a model of probability for a group of data and to synthesize the result through a distribution of probability for the parameters of models. The nonlinear models considered to evaluate the mineralization of organic nitrogen and to illustrate the bayesian procedure they were: model of Stanford & Smith, model of Marion and model of Cabrera. The comparison of the models was promoted through the Bayes Factor (FB) and Bayes Information Criterion (BIC). In this work we had used Gibbs Sampling and Metropolis Hastings to accomplish inference of the parameters. A Gibbs Sampling algorithm was implemented on R to get the posterior distributions of the models parameters. The convergence of the chains was monitored through graphic analysis, and for the criteria of Geweke and Raftery & Lewis, implemented in the BOA package, executable in the freeware R. The model that provided better adjustment quality to the group of data was the model of Cabrera, being followed by the model of Stanford & Smith and last the one of Marion. Because of the presented results, it can be attested that the Bayesian methodology presented good results in the estimate of the parameters of the models, in other words, the adjustment of models through distributions complete conditional posteriori constitutes a reliable methodology and accuracy.