... Model¹

more correctly referred to simply as 'linear model'

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... paradigm²

choosing components based on correlation or shared peak frequency response with an assumed evoked haemodynamic response function often appears to work well for simple block paradigms like the one here; for more complicated paradigms choosing activation maps becomes a much harder challenge

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... level.³

Strictly speaking these will be $T$- distributed, not normal.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... 'activation'⁴

where in this case 'activation' is to be understood as 'cannot be explained as random correlation coefficient to the associated time course'

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... runs⁵

The ICA decomposition begins with a random unmixing matrix and therefore does not necessarily give the same decomposition every time.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... Theory⁶

We here compare PICA results against results obtained from the GLM with GRF-based inference. Cluster-based thresholding appears to be generally accepted as the method of choice in the case of reasonably sized and well-localised activation patterns like the ones used in this example.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... dimensionality⁷

where the ROC analysis is performed on the spatial map with highest temporal correlation between the true and estimated time courses

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.