| |
Abstract:
Independent components analysis (ICA) is a method used to
analyze event related potential (ERP) and functional imaging (fMRI)
data. It decomposes overlapping components into linearly
independent spatial modes with their associated time courses by
finding components that are uncorrelated and independent. Principal
components analysis (PCA) is another method used to parse
overlapping components by uncovering latent variables that account
for the patterns of covariance seen in a data set. In the present
study, ICA and PCA were directly compared when applied to various
simulated ERP data sets. Each simulated data set consisted of 65
midline channels with 65 time points (Dien, 1998). Components were
constructed from half sine cycles, a short P2-like component and a
long P3-like component. To compare performance in the two methods,
an ICA infomax algorithm (Bell & Sejnowski,1995) and a PCA
procedure using varimax and promax rotations were used. We
systematically tested the performance of these two methods under
various parameters (e.g. number of observations used for analysis,
amount of noise, spatial and temporal overlap of components). In
addition, ICA and PCA were applied to a real ERP data set acquired
from a spatial cueing experiment. The implications and caveats of
applying ICA and PCA to ERP data will be discussed.
|
