Print

Print


Steven Grant wrote:

*	I have now used MedX to make a simple correlation map of 
*	difference images vs.
*	differences in the performance score.  Specifically, I subtracted
the 
*	control
*	image from the task image for each subject.  I then subtracted the 
*	control
*	performance score from the task performance score.   The 
*	difference images and
*	the difference scores were then correlated using the Functional 
*	toolbox module
*	in MedX.

*	The problem is whereas the MedX results produce high correlations 
*	in a priori
*	predicted regions, the 0 0 1 type of Covariate analysis in SPM96 
*	shows nothing.
*	Furthermore, the MedX results are consistent with an independent 
*	ROI analysis of
*	the a priori regions.

*	BUT, if I use a specific condition fit in SPM96  for the Covariates 
*	and specify
*	the contrasts as 0 0 -1 1, then the SPM96 results _exactly_ match 
*	the MedX
*	correlation map !

Could this be because the 0 0 1 contrast treats the repeat measures as 
separate cases, i.e. suffers from the substantially larger (inter-subject) 
error variance, whereas the 0 0 1 1 contrast explicitly throws out the 
inter-subject variance and only tests for within-subject effects? 
SPM99b seems to partition the variance in this fashion, presumably 
SPM96 does, too.  The 0 0 -1 1 contrast (used with your alternating 
sign co-variate, see below) would then test for (positive) correlations 
significantly different from zero between conditions, equivalent to 
your (positive) correlation of differences with differences.


*	What is even stranger is that the covariates in SPM96 were entered 
*	as specified
*	by Andrew Holmes, i.e., the difference scores were mean centered, 
*	divided by 2,
*	then multiplied by -1 for the control condition and +1 for the
active 
*	condition.
*	Therefore, a 0 0 -1 1 contrast would have inverted the signs of the 
*	Neutral
*	covariate to produce the equivalent of a 0 0 1 1 contrast.  
*	Specifically   -1 *
*	(-1*Neutral_Covariate) + 1*(+1*Active_Covariate) = 
*	+1*Neutral_Covariate  +
*	1*Active_Covariate.   So there should not have been any subtraction 
*	at all.

*	I am sorely confused by this outcome. Thanks for your help.


  ------------- End Forwarded Message -------------


If I am right about SPM96, this should explain it.  Many thanks for 
your examples !


Klaus

Professor KP Ebmeier
University of Edinburgh
Royal Edinburgh Hospital
Edinburgh EH10 5HF
Tel./Fax.: *44-131-5376505
E-mail: [log in to unmask]
Website: www.pst.ed.ac.uk