Search

Several statistical techniques for analyzing data matrices use lower rank approximations to these matrices, for which, in general, the appropriate rank must first be estimated depending on the objective of the study. The estimation can be conducted by cross-validation (CV), but most methods are not designed to cope with the presence of outliers, a very common problem in data matrices. The literature suggests one option to circumvent the problem, namely, the elimination of the outliers, but such information removal should only be performed when it is possible to verify that an outlier effectively corresponds to a collection or typing error. This paper proposes a methodology that combines the robust singular value decomposition (rSVD) with a CV scheme, and this allows outliers to be taken into account without eliminating them. For this, three possible rSVD’s are considered and six resistant criteria are proposed for the choice of the rank, based on three classic statistics used in multivariate statistics. To test the performance of the various methods, a simulation study and an analysis of real data are described, using an exclusively numerical evaluation through Procrustes statistics and critical angles between subspaces of principal components. We conclude that, when data matrices are contaminated with outliers, the best estimation of rank is the one that uses a CV scheme over a robust lower rank approximation (RLRA) containing as many components as possible. In our experiments, the best results were obtained when this RLRA was calculated using an rSVD that minimizes the L2 norm.

We describe imputation strategies resistant to outliers, through modifications of the simple imputation method proposed by Krzanowski and assess their performance. The strategies use a robust singular value decomposition, do not depend on distributional or structural assumptions and have no restrictions as to the pattern or missing data mechanisms. They are tested through the simulation of contamination and unbalance, both in artificially generated matrices and in a matrix of real data from an experiment with genotype-by-environment interaction. Their performance is assessed by means of prediction errors, the squared cosine between matrices, and a quality coefficient of fit between imputations and true values. For small matrices, the best results are obtained by applying robust decomposition directly, while for larger matrices the highest quality is obtained by eliminating the singular values of the imputation equation.

The genotype by environment interaction (GEI)) has an influence on the selection and recommendation of cultivars. The aim of this work is to study the effect of GEI and evaluate the adaptability and stability of productivity (kg/ha) of nine maize genotypes using AMMI model (Additive Main effects and Multiplicative Interaction). The AMMI model is one of the most widely used statistical tools in the analysis of multiple-environment trials. It has two purposes, namely understanding complex GEI and increasing accuracy. Nevertheless, the AMMI model is a widely used tool for the analysis of multiple-environment trials, where the data are represented by a two-way table of GEI means. In the complete tables, least squares estimation for the AMMI model is equivalent to fitting an additive two-way ANOVA model for the main effects and applying a singular value decomposition to the interaction residuals. It assumes equal weights for all GEI means implicitly. The experiments were conducted in twenty environments, and the experimental design was a randomized complete block design with four repetitions. The AMMI model identified the best combinations of genotypes and environments with respect to the response variable. This paper concerns a basic and a common application of AMMI: yield-trial analysis without consideration of special structure or additional data for either genotypes or environments.

A common problem in multi-environment trials arises when some genotypeby- environment combinations are missing. In Arciniegas-Alarcón et al. (2010) we outlined a method of data imputation to estimate the missing values, the computational algorithm for which was a mixture of regression and lower-rank approximation of a matrix based on its singular value decomposition (SVD). In the present paper we provide two extensions to this methodology, by including weights chosen by cross-validation and allowing multiple as well as simple imputation. The three methods are assessed and compared in a simulation study, using a complete set of real data in which values are deleted randomly at different rates. The quality of the imputations is evaluated using three measures: the Procrustes statistic, the squared correlation between matrices and the normalised root mean squared error between these estimates and the true observed values. None of the methods makes any distributional or structural assumptions, and all of them can be used for any pattern or mechanism of the missing values.

This paper proposes five new imputation methods for unbalanced experiments with genotype by-environment interaction (G×E). The methods use cross-validation by eigenvector, based on an iterative scheme with the singular value decomposition (SVD) of a matrix. To test the methods, we performed a simulation study using three complete matrices of real data, obtained from G×E interaction trials of peas, cotton, and beans, and introducing lack of balance by randomly deleting in turn 10%, 20%, and 40% of the values in each matrix. The quality of the imputations was evaluated with the additive main effects and multiplicative interaction model (AMMI), using the root mean squared predictive difference (RMSPD) between the genotypes and environmental parameters of the original data set and the set completed by imputation. The proposed methodology does not make any distributional or structural assumptions and does not have any restrictions regarding the pattern or mechanism of missing values.

Os capítulos podem ser lidos de forma independente. O primeiro é introdutório, focalizando a interação genótipo por ambiente e o enfoque estatístico, apresentando o modelo AMMI, o gráfico Biplot e validação cruzada. O capítulo 2 cobre a distribuição empírica dos autovalores da matriz de interação, utilizando reamostragem bootstrap. O capítulo 3 discute a divergência genética utilizando reamostragem bootstrap e análise de agrupamento. O capítulo 4 trás e aplica um método de correção de autovalores viesados e mostra a eficiência da correção. O capítulo 5 apresenta um teste para confirmar a contribuição de genótipos e ambientes para a interação. O capítulo 6 trata da imputação simples e múltipla para observações ausentes na matriz de interação. O capítulo 7 trás os modelos AMMI bivariados utilizando a análise de procrustes. O capítulo 8 cobre o modelo AMMI no estudo da interação entre QTL e ambiente. O capítulo 9 contém a generalização dos modelos AMMI para três fatores com uso dos modelos PARAFAC e TUCKER, gráficos Joint plot e Triplot. Salientamos que os resultados apresentados aqui foram obtidos pelos autores e portanto, a notação algébrica pode não seguir o mesmo padrão em cada capítulo.

O objetivo deste trabalho foi propor um novo algoritmo de imputação múltipla livre de distribuição, por meio de modificações no método de imputação simples recentemente desenvolvido por Yan para contornar o problema de desbalanceamento de experimentos. O método utiliza a decomposição por valores singulares de uma matriz e foi testado por meio de simulações baseadas em duas matrizes de dados reais completos, provenientes de ensaios com eucalipto e cana-de-açúcar, com retiradas aleatórias de valores em diferentes percentagens. A qualidade das imputações foi avaliada por uma medida de acurácia geral que combina a variância entre imputações e o viés quadrático médio delas em relação aos valores retirados. A melhor alternativa para imputação múltipla é um modelo multiplicativo que inclui pesos próximos a 1 para os autovalores calculados com a decomposição. A metodologia proposta não depende de pressuposições distribucionais ou estruturais e não tem restrições quanto ao padrão ou ao mecanismo de ausência dos dados. The objective of this work was to propose a new distribution-free multiple imputation algorithm, through modifications of the simple imputation method recently developed by Yan in order to circumvent the problem of unbalanced experiments. The method uses the singular value decomposition of a matrix and was tested using simulations based on two complete matrices of real data, obtained from eucalyptus and sugarcane trials, with values deleted randomly at different percentages. The quality of the imputations was evaluated by a measure of overall accuracy that combines the variance between imputations and their mean square deviations in relation to the deleted values. The best alternative for multiple imputation is a multiplicative model that includes weights near to 1 for the eigenvalues calculated with the decomposition. The proposed methodology does not depend on distributional or structural assumptions and does not have any restriction regarding the pattern or the mechanism of the missing data.