Modelización de dise os split-plot y estructuras de covarianza no estacionarias: un estudio de simulación (Modelling split-plot data and nonstationary covariance structures: a simulation study)

A topic that has aroused great interest among researchers who analyse longitudinal data has been the development, by means of simulation studies, of analytic models that incorporate the covariance structures which best fit the data. When analysing covariance structures within the context of longitudinal data one finds that the variances are not always constant. Indeed, the variances commonly increase over time when the correlations between equally spaced observations are not homogeneous. This paper reports a simulation study which analysed two random coefficient models with nonstationary correlations. The first had constant variances (RC), while the second, given its utility in longitudinal contexts, showed variances with a linear structure (RCL). Once the two covariance matrices (RC and RCL) had been generated, eleven covariance structures were fitted by means of PROC MIXED for the Akaike criterion, thus enabling the best fit to be selected. The aim of the study was to determine the fit percentages of the different covariance matrices and the extent to which the one with the best fit corresponds to the population covariance matrix.