Structural equation models: Perspectives on and challenges to model selection

Structural equation modeling (SEM) is frequently used for the purpose of evaluating multivariate hypotheses about networks of relationships. Because of both the philosophical perspective associated with the practice of SEM and because path models are frequently unsaturated, a distinctive perspective on model selection is associated with this methodology. The “structural” part of SEM refers to the fact that we seek to relate statistical associations to causal processes. This emphasis on causal processes has an influence on the criteria used in model selection. Also of importance is the fact that path models are commonly unsaturated, which permits evaluations of absolute fit (saturated models such as univariate models are nearly always saturated and not subject to evaluations of absolute fit). In SEM an emphasis is placed on evaluation on theory-based a priori models. SEM practice also aspires to be a confirmatory process and places greater weight on the a priori model(s) in the evaluation process than do descriptive/predictive approaches.