Normalization Issues in Latent Variable Modeling

Structural equation models with latent variables have been popular in the social sciences for some time. Recently such models have been used to make comparisons between groups, within dyads, and across time. However, possible problems with establishing meaningful metrics for unobservable variables in such situations are not universally known and acknowledged. It is shown that, when parameter comparisons are a central focus of the research, the method used to determine the metric of the latent variables (known as normalization) can dramatically and arbitrarily affect conclusions reached. The most commonly proposed solution to this problem, placing equality constraints on corresponding indicator coefficients, is discussed and critiqued. We argue that, in many cases, this approach is both statistically sound and substantively meaningful. We also note that there are several potential problems with this technique that should not be overlooked. Solutions and guidelines for dealing with each of these difficulties are presented.