Detection of Correlated Errors in Longitudinal Data

A study of change in ability between two occasions may employ a number of tests believed to measure the ability in question. Either the same battery of tests is used on both occasions, or equivalent forms are used. For a variety of reasons, correlations may exist between certain errors remaining after eliminating variance due to true scores, and hence the classical factor analysis model is not applicable. A procedure for detecting correlations between errors is discussed. A search strategy is proposed since, even if the number of observed variables is small, the number of possible models is very large. A computer program is described, which produces maximum‐likelihood estimates for the parameters in a factor analytic model in which the error variables may be correlated.