ABSTRACT

This chapter describes the broad types of possible specification errors in structural equation models with unmeasured variables. It shows that the diagnostic tools now available are better suited for discovering certain types of specification error than for discovering other types. The chapter clarifies the nature and interpretation of the diagnostic indices included in the printout for LISREL V. It suggests how best to use the diagnostic print-out of LISREL V in respecifying a model. In doing this, the chapter draws on the results of simulated data that differ in specified ways from the estimating model. Useful as the diagnostic indices are, the effective respecification of a model remains a matter of substantive judgment, not a task that is readily achieved by "mechanically" following the diagnostics. A structural equation model is "specified" by describing the causal dependencies or other sources of covariation between each pair of variables in the model.