Statistical latent-variable models require correlated indicators, but concept definitions require separate, nonredundant attributes.

Evidence

The latent variable approach is virtually, by construction, a unidimensional approach. This is a function of the statistical methodology used. The core requirement of the latent variable approach is that the indicators be of one thing, and hence highly correlated with each other. In contrast, in definitions there is no requirement at all that the defining features be highly correlated. In fact, we have seen that typically they should be separate and nonredundant attributes. — Gary Goertz, Social Science Concepts and Measurement, 349

Categorization

• XREF: Ties directly to the book’s running theme that concept structure (family resemblance, necessary-and-sufficient conditions) is logically distinct from correlational measurement models.
• SEED: Angle: ‘Why your factor analysis is not your definition’ — how statistical tools quietly dictate conceptual choices in the social sciences.