matrix =		Y				x	X								Z	
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
input = 		cases(subXconXchan) 		x 	time-samples(ERP segment) [what is condensed into PCs]		avg-ref uV
VxF loadings =		var=time-samples		x	Factor								factor loadings for plotting PCs in time
CaseXFact scores =	condXchanXsubj cases		x	Factor								factor scores for plotting topography of a PC pseudo-identified with a time based on peak factor-loading; also employed in GLM instead of mean-amplitude for a defined component window
VariabXVar =		time-samples			x	Factor								eigen-values and explained variance (collapsed across Ss x Cond x Chan)

note1: data reduction, re-referencing, baseline correction, and averaging happen before PCA
note2: only rotated factors that account for >= 1% of the variance are typically retained after varimax rotation
note3: and only the subset of those with plausible time-courses and topographies are retained for analysis
note4: GLM stats happen after PCA-Varimax using factor scores (instead of eg mean-amp)
note5: for a massively univariate approach, effect one GLM for each factor-of-interest ("component") x channel with experimental conditions as levels of the IV (NB: modified to include asymmetry score to assess Hemi x Con)


--------------------------
***from dien's chapter in handy's volume*** 

1. necy to multiply the factor loadings by the variable SDs to transform them into uV

2. if a specific time index is chosen, it's possible to reconstruct the scalp topography (in uV) by multiplying the mean scores (mean of FSs across Ss at a particular channel for a particular condition) by the factor loading and the SD for time index of interest.

3. factor scores (FSs) should not be mean corrected

3. it is important to check that the scalp topography of each condition broadly resembles that of the grand mean following PCA/EFA

4. paradoxically, if time course/latency is the analytic/inferential aim, then SPATIAL pca should be used, b/c latency is a free rather than fixed parameter (and topography is fixed)

5. paradoxically, if topographic/spatial/source analysis is the analytic/inferntial aim, then temporal pca should be used, b/c topography is a free parameter (and latency is fixed)

6. in theory, temporal pca is more accurate because volume conduction guarantees high intercorrelations among variables, which tend to reduce the accuracy of PCA

---------------------------
***from dien 1998a****

1. spatial pca is useful for characterizing temporally overlapping or jittered components, whereas temporal pca is useful for characterizing spatially overlapping components

2. HOWEVER, to the extent that ERP components are defiend as reflecting dissociable, unique source generators (dipoles), spatial PCA is more useful

3. Although Dien doesn't say, this: to the extent that a psychological process is instantiated in a sptially distributed network or circuit of partially functionally dissociable territories working in concert, or to the extent that it is easier to infer function from time course than structure, temporal pca is preferred