Discriminant analysis is commonly used to classify an observation into one
of two (or more) populations on the basis of correlated measurements. Class
ical discriminant analysis approaches require complete data for all observa
tions. Our extension enables the use of all available longitudinal data, re
gardless of completeness. Traditionally a linear discriminant function assu
mes a common unstructured covariance matrix for both populations, which may
be taken from a multivariate model. Here, we can model the correlated meas
urements and use a structured covariance in the discriminant function. We i
llustrate cases in which the estimated covariance structure is either compo
und symmetric, heterogeneous compound symmetric or heterogeneous autoregres
sive. Thus a structured covariance is incorporated into the discrimination
process in contrast to standard discriminant analysis methodology. Simulati
ons are performed to obtain a true measure of the effect of structure on th
e error rate. In addition, the usual multivariate expected value structure
is altered. The impact on the discrimination process is contrasted when usi
ng the multivariate and random-effects covariance structures and expected v
alues. The random-effects covariance structure leads to an improvement in t
he error rate in small samples. To illustrate the procedure we consider rep
eated measurements data from a clinical trial comparing two active treatmen
ts; the goal is to determine if the treatment could be unblinded based on r
epeated anxiety score measurements. Copyright (C) 1999 John Wiley & Sons, L
td.