Biblioteca R para a análise bayesiana de dados categorizados usando modelos mistos de limiar

Abstract

In this thesis we evaluate algorithms for Bayesian analysis of ordinal categorical data as well as their implementation in the library Bayesthresh for the R statistical programing environment. Bayesthresh package has a flexible structure to insert mixed models and uses Markov Chain Monte Carlo (MCMC) sampling to approximate posterior distributions for model parameters. Simulation study of the efficiency of the algorithms considered processing time, dependency in the MCMC sampling chains and sensitivity to prior specifications for the variance components. Mean Squared Error (MSE) and average Bias in the marginal posterior distributions where also evaluated. An example is discussed from a breeding experiment for resistance to late blight (Phytophthora infestans) in varieties of tomato. Algorithms described by Nandran and Chen (1996) and derived algorithms were the faster and more accurate, although slightly overestimating intraclass correlation ( ) in experiments with higher parametric values ( = 0:8). In the example useful estimates for plant breeding are presented, as Bayes factor test to decide on latent trait having Gaussian distribution. Gaussian distribution was as likely as Student’s t distribution. An additional paper explores the uses of the modeling tool with an example from sensory analysis of dehydrated banana candy recipes with different sugar content, described by Silva (2008).