Tratamento de dados com Outliers: uma aplicação à análise de Regressão

Abstract

In data analysis, it is common to find observations with characteristics different from the others and that do not seem to belong to the variability pattern formed by other observations, these are called Outliers. The presence of outliers distorts estimates such as the mean, standard deviation, etc. Therefore, it is important to detect them and look for ways to correct or minimize the effects caused by their presence in the sample. Thus, in the present work we intend to estimate linear regression models whose sample data have outliers. For this, the TEMPENTREGA database was used and a regression model was estimated using the Ordinary Least Squares (OEM) method, then, to detect the presence of outliers, the studentized residuals and the values ​​of the main diagonal were used. of the projection matrix and to assess whether the outliers found are or not influential in the regression model, the measures of influence of Cook's Distance, the SDFFITS and the SDFBETA were used, with which two observations were detected as influential outliers. To minimize the effect of outliers observations, two methods were used: the Robust Trimmed Least Squares (MQA) regression method and the OLS method without outliers observations. From the study it was concluded that the regression models given by the MQA and OLS methods without the outliers observations were the best for making predictions because they presented better estimates. Keywords: Outliers, Residues, Measures of influence, Regression model. ( Tradução Nossa )