Calling Models With Causal Indicators 'Measurement Models' Implies More Than They Can Deliver

摘要

Bainter and Bollen have done the psychometric community a great service in elucidating the role of correct model specification in thinking about interpretational confounding. However, despite the appropriate emphasis on correctly specified models, Bainter and Bollen overreach in requiring causal indicators to measure or identify a latent variable. Causal indicators cannot, by themselves, identify latent variables without a set of effect indicators, and given a set of effect indicators they are superfluous in determining the meaning of the latent variable in the model. Causal indicators are not measures of their common effect, and to treat them as such is to court conceptual confusion. Measurement is an important way of operationalizing constructs, but it is not the only way. For example, for many questionnaires it is common to form composite scores according to a fixed weighting scheme. This of course does not allow for interpretational confounding, because the weights are not identified through the fit of a reflective latent variable model. In this sense, we may consider composites to escape some of the problems that plague causal indicators. But of course nobody would claim that indicator variables "measure" the composite they form. And that is exactly as it should be.