Mixed General Linear Model Analysis of Quantitative Electroencephalographic (qEEG) Data

Abstract

This paper describes a mixed general linear analysis of the quantitative electroencephalogram (qEEG). The modeling is similar to regression, which builds a regression or ‘best-fit’ model for the data structure but, in addition, provides for correlations between observations. A mixed linear model states that data consists of two parts: fixed effects and random effects. Fixed effects determine the expected values of the observations, while random effects account for the stochastic deviations from these expected values both between and within individuals. Since errors are independent between subjects, the deviations from the expected values may also be modeled using a repeated measures approach. The term ‘repeated measures’ in this model refers to data with multiple observations from one specific source. It is reasonable to assume that these observations from the same source are correlated, even if only slightly, in some measurable way. Consequently, statistical analysis of repeated measures data gives a more accurate prediction capability when the issue of covariation between these measures is addressed. With mixed model methodology now available (e.g., the mixed procedure [Mixed PROC] of the SAS® system), the covariance structure can be incorporated into the statistical model. Disregarding potential random effects not specific to single individuals and absorbing potential within-subject random effects into the covariance matrix allows one to work with a simplified model. The use of a mixed procedure and its method of modeling the data structure appear to provide an accurate and objective method of analysis resulting in quantifiable equations for testing predictions. Essentially, this method allows the physiological pattern of each individual in the study, not related to any other variable, to be represented and accounted for in the model. Several comparative examples will be used to highlight the information that can be hidden in data structures depending on the type of statistical analysis used.