Application of the linear mixed and generalized mixed model as alternatives for analysis in experiments with repeated measures

Abstract

Linear mixed and generalized linear mixed models were applied to an experiment with scarified seeds through the process of endozoochory, as tools for processing and analysis with measures repeated over time. Variables analyzed were plant height and stem thickness. Pearson correlation matrix was calculated to determine the existence of association among sampling days. In the analyzed variables, sphericity criterion was used by Bartlett statistic. For both variables, the assumption of normality was verified by Shapiro-Wilk and Kolmogorov-Smirnov tests. In the variable that fulfilled the assumption of normality, a linear mixed model was used. For the variable that was not fulfilled, the generalized linear
mixed model was applied. The Poisson, Gamma, Log Normal, Normal and Binomial variants were analyzed to determine data distribution. Several variance-covariance structures were tested to select the best fit and the information criteria that obtained the smallest values were considered. The use of these statistical models allows to adequately control the probability of occurrence of type I error, since it provides greater flexibility and information when selecting the best fit model, in addition to allowing to
analyze unbalanced data.
Key words: longitudinal data, covariance structures, information criteria