Teste de Student-Newman-Keuls bootstrap : proposta, avaliação e aplicação e dados de produtividade da graviola

• Ano de defesa: 2015
• Autor(a) principal: Gonçalves, Bruna De Oliveira
• Orientador(a): Ramos, Patrícia De Siqueira
• Banca de defesa: Bittencourt, Flávio, Nogueira, Denismar Alves
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Alfenas
• Programa de Pós-Graduação: Programa de Pós-Graduação em Estatística Aplicada e Biometria
• Departamento: Instituto de Ciências Exatas
• País: Brasil
• Palavras-chave em Português: Comparações múltiplas (Estatística); Monte Carlo, metodo de
• Área do conhecimento CNPq: PROBABILIDADE E ESTATISTICA::PROBABILIDADE E ESTATISTICA APLICADAS
• Link de acesso: https://repositorio.unifal-mg.edu.br/handle/123456789/537

Multiple Comparisons Procedures (MCP) are used to compare treatment means. There are many tests with this purpose and to choose the best one, two features must be analysed: the control of type I error rate (exact, conservative or liberal tests) and the power. Bootstrap resampling methods have been used in some studies to improve the performance of MCP. The Student-Newman-Keuls (SNK) test shows good statistical qualities that can be improved with the use of bootstrap. Therefore, this study aimed to propose a SNK parametric bootstrap version (SNKB) and compare it with the original SNK test. The performance was evaluated by experimentwise error rates and power using a Monte Carlo simulation study considering normal and non-normal situations. We considered N = 1000 simulations of k treatments (k = 5, 10, 20 e 80) with r repetitions (r = 4, 10 and 20). Under null hypothesis, the means were considered all equal, under H1 the means were all different, but the variance was the same and, under partial H0, we considered two groups with different means. Both tests showed type I error rates values close to the nominal level of 0.05 under H0 and normality. Under H0 and non-normality, both tests controlled the experimentwise error rates in most simulated cases for k=5 and k=10, whereas, for k=20 and k=80, the tests were considered liberal in some scenarios. Under H0 partial, the SNKB test was liberal in all simulated cases, while SNK test was generally conservative for δ ≤ 2 and liberal to other δ values. In general, the power of the proposed test surpassed the power of original test under normality and non-normality. Thus, in practice, if the differences between the treatment means are small (δ ≤ 2), the SNK test works better given that it controls the type I error and the power is satisfactory. In the other cases, the SNKB test is recommended, although both are liberal for δ ≥ 4, if we are under partial H0. Furthermore, the tests were applied to a real experiment designed to evaluate the chemical and mechanical controls of pests soursop in order to compare the results of both tests.