Enfoque estatístico na validação de métodos para teste de germinação de sementes florestais

• Ano de defesa: 2012
• Autor(a) principal: Nomelini, Quintiliano Siqueira Schroden
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; BR; Programa de Pós-graduação em Agronomia; Ciências Agrárias; UFU
• Palavras-chave em Português: Repetitividade; Reprodutibilidade; Modelos lineares; Modelos lineares generalizados; Sementes; Germinação; Árvores - Semente; Modelos lineares (Estatística); Estatística agrícola; Repeatability; Reproducibility; Linear models; Generalized linear models; CNPQ::CIENCIAS AGRARIAS::AGRONOMIA
• Link de acesso: https://repositorio.ufu.br/handle/123456789/12056; https://doi.org/10.14393/ufu.te.2012.8

Interlaboratory results obtained by recommended methods for seed germination testing of 25 native forest species were subjected to different statistical tools. The pre-testing and re-testing methods for the preparation of germination were conducted at the Federal University of Uberlândia. In order to validate these methods, seed lots of different physiological qualities and their randomization diagrams of lots and seeds, and the repetitions in completely randomized design (CRD), were sent to laboratories. Among the various statistical methods in the validation process, data outliers were identified by the study of Boxplots and measure DFFITS. After elimination of the outliers, discrepancies in variances were studied by the methods of Cochran and Levene for the mean and median. The consistency and accuracy of results obtained by the laboratories for each lot was done by the study of repeatability and reproducibility (R&R) by the method of analysis of variance, which estimated the variance components of these measures and the their percentage contribution in relation to the total variation. Where it was possible to identify eventual superestimation of these estimates, the same was done using the tools of control charts for mean and range, excluding the results from laboratories identified outside the expected standards. The remaining sets of observations were used to calculate Mandel's h and k measures. The data obtained from the study of R&R were subjected to analysis of variance in CRD evaluating the effect of factors: laboratories, lots and their interaction (laboratory*lots), and their assumptions tested, and another study by the grouped analysis method. In addition to these, a comparison between the classical analysis of variance, using the Normal distribution, with variance analysis by generalized linear models for the Binomial distribution was done. It was observed that the measure DFFITS found a greater number of outlier points in relation to Boxplot. In general, most species had homogeneous variances. The study of R&R by the method of analysis of variance identified, with the aid of control charts, lots tending to overestimate the variance of repeatability and reproducibility, resulting in a set of consistent information. There were no major differences between the choice of experiment analysis by the classic or grouped analysis of variance. However, the same was not observed when using the Binomial distribution to model the original variable number of normal seedlings, in which this distribution was best suited for most species in contrast with the normal distribution. Therefore, the method was validated for 20 species, from a total of 25, and 18 of these using the Binomial distribution. Moreover, the study of R&R by the method of analysis of variance was more interesting than the statistics Mandel's h e k.