Hi Ged.
Thanks for the reply. Essentially the eigenvariate makes sense on the equal case and makes no sense on the unequal case. I can't really understand it myself.
As far as the eigenvariate being already adjusted, I have not found that to be the case. For example, I run a correlation with two covariates, I get significant clusters. However, if I extract the eigenvariate for a cluster and then run the correlation using corrcoef or some other similar function, I will sometimes get a reasonable r and p or I will other times get an r or p that does not follow the SPM results. I would expect the eigenvariate for a cluster to have at least the same significance level as what I threshholded the SPM results at. (e.g. if I threshhold at p=.01, extract the eigenvariate for a cluster at this threshhold, and run a correlation using that eigenvariate and my covariate of interest that I used in the SPM analysis, I should have a resulting p-value of at least .01). This is not always the case, which is why I expect that the eigenvariate is unadjusted for the nuisance covariates.
Thanks for all the help. I'm much closer to an answer than I was before. The equal vs. unequal variance phenomenon is still confusing however.
-John
--- Ged Ridgway wrote:
Hi John,
--- Start of quoted text:
If I compare the two groups and say in the design
specification that the variances are EQUAL I get the approximately
the same results in the graphics window as I do if I declare the
variances UNEQUAL --- end of quoted text ---
Sounds fair enough.
--- Start of quoted text:
(I am assuming that EQAUL means no correction for
non-sphericity and UNEQUAL corrects for non-sphericity).
--- end of quoted text ---
That's my interpretation too.
--- Start of quoted text:
if I extract the 1st eigenvariate for a cluster for both cases I
get completely different results. For EQUAL variance I get values
that corroborate the results I see in the graphics window, however
for UNEQUAL variance I get the opposite. To better explain if I
take the eigenvariates and do a 2 sample TTEST outside of SPM
comparing the two groups: for EQUAL I get group1>group2, but for
UNEQUAL I get group2>group1.
--- end of quoted text ---
So just to be completely clear, the t-test can itself assume equal or unequal variance, e.g. Matlab's ttest2 can be called as:
ttest2(A, B, 0.05, 'both', 'unequal')
but the sign of the result can't possibly change between the two, since the numerator of the t-test is the difference in means in both cases; only the standard error and the degrees of freedom are modified.
So, I assume you're saying that the eigenvariate you extract is totally different whether you indicate equal or unequal variance, right? I'm afraid I don't understand how that would happen... I've had a quick glance at spm_regions.m (the code that extracts the VOI) and I can't really follow it.
Incidentally, spm_regions does make it sound like the eigenvariate should be adjusted for nuisance variables already (to go back to your earlier question) though I can't really be sure.
I'm afraid to say I think you'll need to get a reply from Will Penny or Volkmar Glauche to be sure what is going on here. I don't think I know enough about spm_regions.m to really help.
Best of luck,
Ged.
--- end of quote ---
|