Testes escore corrigidos para modelos lineares generalizados no ambiente R

Abstract

Bartlett's corrections are statistical procedures to improve statistics whose distributions are approximated by the chi-square distribution. An application of this methodology is to improve the score test in generalized linear models. The resulting correction formula depends on the construction of several matrices whose elements are expressions which involve rst and second order derivatives of the mean and of the variance function taken both with respect to the model linear predictor. As a result, di culties inherent to the process to obtain those derivatives, or even to modify them when it is necessary to respecify the random or the systematic model component, may be the primary cause that this correction methodology is not yet seen as useful tools in the applications of the score test. This master's thesis proposes a computer program developed in the statistical software R to implement automatically corrected score tests given the t of a generalized linear model. Technical details and instructions to use the program are explored on the basis of the analyses of a series of real data examples found in the literature. Furthermore, the results of two simulation experiments are discussed in order to compare properties of the uncorrected and corrected tests and to show the versatility of the proposed program used as a computing tool in the experiments.