Please use this identifier to cite or link to this item: http://repositorio.ufla.br/jspui/handle/1/13413
Title: Proposição de testes de normalidade multivariada baseados em distâncias robustas
Other Titles: Multivariate normality tests proposition based on robust distances
Authors: Ferreira, Daniel Furtado
Ferreira, Daniel Furtado
Bueno Filho, Júlio Sílvio de Sousa
Pereira, Tiago Martins
Keywords: Normalidade multivariada
Estimadores robustos
Distância de Mahalanobis
Métodos inferenciais
Multivariate normality
Outliers
Mahalanobis distance
Inferential methods
Issue Date: 21-Jul-2017
Publisher: Universidade Federal de Lavras
Citation: PEREIRA, E. A. Proposição de testes de normalidade multivariada baseados em distâncias robustas. 2017. 105 p. Dissertação (Mestrado em Estatística e Experimentação Agropecuária)-Universidade Federal de Lavras, Lavras, 2017.
Abstract: Multivariate normality is one of the most important assumptions for the realization of many inferential methods. The non-verification of this assumption can influence the reliability of the results. There are many tests in the specialized literature to verify normality. In the case of multivariate normality, the tests, in general, are based on correlation coefficients, asymmetry and kurtosis coefficients and distances. Despite the large number of tests, there is no test in the literature that is uniformly more powerful in all evaluated situations. The tests, in general, presents some restrictions, both in relation to size and to dimension. The presence of outliers in the data can result in bad parameter estimation and even distortions in the distribution fitting, making the tests fail. Therefore, the aim of this work is to propose and evaluate four outlier robust tests: multivariate normality test based on Mahalanobis distance with robust measures of the scale and location parameters (TNMD2RKS), multivariate normality test based on robust beta distance (TNMDbRKS), parametric bootstrap multivariate normality test based on robust distances (TNMD2RBoot) and the parametric bootstrap multivariate normality test based on robust beta distances (TNMDbRBoot). For the four tests it was used the robust scale and location estimators calculated via function CovOgk from the R software. Type I error rates and power of the tests were evaluated by comparing then to the parametric bootstrap multivariate normality test based on the correlation between the order statistics and the expected values and the Royston’s Shapiro-Wilk test, via Monte Carlo simulation. The tests TNMD2RKS, TNMD2RBoot and TNMDbRBoot were successfully proposed, obtaining excellent type I error rate control, especially in samples with the presence of outliers, in which the other tests did not perform well. In terms of power, the four tests performed well in large samples, however, they did not outperform the tests used as references.
URI: http://repositorio.ufla.br/jspui/handle/1/13413
Appears in Collections:Estatística e Experimentação Agropecuária - Mestrado (Dissertações)



Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.