This paper presents a new method for characterizing brain responses in
both PET and fMRI data. The aim is to capture the correlations betwee
n the scans of an experiment and a set of external predictor variables
that are thought to affect the scans, such as type, intensity, or sha
pe of stimulus response. Its main feature is a Canonical Variates Anal
ysis (CVA) of the estimated effects of the predictors from a multivari
ate linear model (MLM). The advantage of this over current methods is
that temporal correlations can be incorporated into the model, making
the MLM method suitable for fMRI as well as PET data. Moreover, tests
for the presence of any correlation, and inference about the number of
canonical variates needed to capture that correlation, can be based o
n standard multivariate statistics, rather than simulations. When appl
ied to an fMRI data set previously analyzed by another CVA method, the
MLM method reveals a pattern of responses that is closer to that dete
cted in an earlier non-CVA analysis. (C) 1997 Academic Press.