Problems in Estimating Measurement Error From Panel Data

In this paper we apply recently developed models for the estimation of reliability and stability coefficients from panel data to a study of scientific productivity. The models, which assume a first-order autoregressive process among true-score variables, yield either reliability and stability estimates which seem implausible when the statistical fits of the models to the data are good, or poor statistical fits when more plausible estimates are produced. We then examine an alternative model encompassing a latent variable which causes true scores and which is itself governed by a first-order autoregressive process. The results for this model are acceptable, and we conclude that the panel models developed for the estimation of measurement reliability are not appropriate representations of the causal processes involved in scientific productivity. We suggest that this type of misspecification will often occur in the application of models which assume independent disturbances for true-score variables.