On selecting indicators for multivariate measurement and modeling with latent variables: When "good" indicators are bad and "bad" indicators are good

Selecting indicators is as important for the generalizability of research designs as selecting persons or occasions of measurement. Elaborating on the extant knowledge base regarding indicator selection, the authors examine selection influences on the validity and reliability of multivariate representations. A simulation that systematically varied 4 key dimensions of indicator selection was used to investigate their effects on the fidelity of construct representations and the relative ability of exploratory and confirmatory analyses to recover within- and between-construct information. Confirmatory analyses, for example, yielded valid and unbiased estimates of the relations between constructs, even under conditions of very low internal consistency. Design, procedural, and analysis recommendations based on an expanded taxonomy of indicator selection and the simulation results are provided.