Structural equation modeling and its application to network analysis in functional brain imaging

Abstract

The analysis of brain imaging data has recently focused on the examination of the covariances of activity among neural regions during different behaviors. We present some of the theoretical and technical issues surrounding one of these covariance‐based methods: structural equation modeling. In structural equation modeling, connections between brain areas are based on known neuroanatomy, and the interregional covariances of activity are used to calculate path coefficients representing the magnitude of the influence of each directional path. The logic behind the use of structural equation modeling stems from the suggestion that brain function is the result of changes in the covariances of activity among neural elements. The technical foundations for neural structural equation models are presented, emphasizing the ability to make inferential comparisons to evaluate the experimental changes in path coefficients. Simulated data sets were used to test the effects of omitted regions and omitted connections. The results suggested that structural modeling algorithms can give hints as to possible external influences and missing paths, but that the final decision as to model modifications requires the guidance of the researcher. The utility of anatomically based models to distinguish between the direct effect of one region on another, and indirect effects of darkness or patterned light on the metabolic activity in the rat visual system. The anatomical framework for the structural equation models revealed that the total impact of ascending thalamocortical influences was modified by corticocortical interactions. Extensions of structural equation modeling to human brain imaging experiments are presented. We conclude by suggesting that neural covariances may be a more accurate way to examine the dynamic functional organization of the central nervous system. ©1994 Wiley‐Liss, Inc.