Análise de experimentos fatoriais em parcelas subdivididas sem réplicas com observações faltantes

Abstract

Missing data is a topic of great importance in the analysis of experiments, because if the problem is treated improperly it can affect the analysis. In this work, three methods of estimation of missing observations are studied: The methods of Coons method, Rubin and Haseman and Gaylor. A simulation study is performed so that missing observations are estimated by all methods in simulated data for an experiment in split-plot 2 3 × 2 without replication. Nine scenarios are executed for each method, in which different levels of the factors are analyzed: Position of missing observations, plot variance and amount of active effects, with the objective of determining the conditions for positions of the missing observations, variance of the error associated to the plot, amount of active effects associated to the plot and to the subplot and magnitude of the effects. The results of interest are: Mean error of the estimation, standard deviation of the mean error of the estimation, mean of the effects, variance of the effects and performance as the correct identification of active and inactive effects. To analyze the performance are considered the power, the individual error rate (IER) and the experimentwise error rate (EER). Among the methods studied, the method that stands out with better results is the Rubin method, since the Coons method presents a restriction to a missing observation position and the Haseman and Gaylor method has been affected by the three factors considered than in other methods.