Contribuições sobre o envelope simulado na análise de diagnóstico em modelos de regressão

Abstract

The simulated envelope is a diagnostic analysis method used to evaluate the hypothesis about the probability distribution assumed for the response variable in a regression model. In this work, we describe some procedures to obtain the simulated envelope and, later, we propose a method to decide if we should reject a model using the envelope. In order to compare our procedure with other proposals, we performed a Monte Carlo simulation study in two classes of regression models. The results indicate that the proposed method presents good performance, since it provides stable rejection rates of the model under the correct distribution. About other methodologies, besides having a higher computational cost, the rejection rate under the correct model increases as the sample size rises. In addition, we also compare the full normal plot and the half normal plot with envelope using Monte Carlo simulations studies. The results suggest that, in general, the full normal plot performs better, especially with the proposed procedure. Finally, we apply our decison method and the other proposals to real data from the National Health Survey (Brazil) of 2013. To these data, our method suggested a different decision from that one provided by the other procedures.