The blind, forward-search procedure used by Spirtes, Scheines, and Glymour in their simulation study of computer-aided model specification with EQS does not represent a procedure that is recommended for use in practice, nor does it represent a procedure that is actually implemented by researchers in practice. Thus the implication of their results is unclear. Although TETRAD II represents an important development in procedures for searching for structural models that may be consistent with data, its relative performance under varying conditions, especially with much larger models, under other choices of weight and percentage criteria, and under other model structures is not known. The results on three of the nine models used in the simulation cannot be interpreted, because they do not give unique exact representations of the population covariance matrices: Alternative “true†models can be specified, and EQS and LISREL may have found these models. The differing output from the two procedures that were used, a set of models in TETRAD II, and a single model in EQS, does not permit a fair comparison between the methods. Some alternative methods are summarized.
