O efeito de reparametrização em testes de sobrevivência acelerados

Abstract

Accelerated life tests are frequently used in industrial experiments to obtain measures on the reliability of products. In these tests, the units are submited at higher levels of stress than usual and the informations obtained are utilized to inference on the reliability of the product, under normal operating conditions. A practical problem, refers to the accuracy of inferences obtained by asymptotic normal approximation of the maximum likelihood estimator, when the sample size is small or moderate. In this dissertation are described the e¤ects of several reparametrizations on the accuracy of the interval estimation of the parameter of interest, here represented by the mean lifetime. The idea is to obtain a parametrization in order to have independent maximum likelihood estimates with good asymptotic normal approximation. A simulation study presents the coverage probability of the con dence intervals when only small or moderate size datasets are available. The e¤ect of the presence of censoring was also investigated in the simulation study. The methodology is illustrated on a real dataset of an accelerated life test at pressurized containers of Kevlan/Epoxy 49. Under the bayesian perspective, Markov Chain Monte Carlo methods are proposed to evaluate the reliability of products. The e¤ect of reparametrization on convergency of the chains is studied.