So, this is what I did:
1) First level analysis for each of the 5 runs for each subject with z>2.3 and p<.0125 (.05/4 contrasts)
2) Higher level FE analysis of the 5 runs for each subject.
3) I haven't done this, but plan to: FLAME ME for all subjects.
Up to number 2 above, I am not seeing any activation. My question is, am I being too conservative with my z
and p-value thresholds? Any suggestions regarding these values in this context?
Thanks !!!
On Sun, Nov 22, 2009 at 2:17 PM, Shal Hat
<[log in to unmask]> wrote:
I'm pretty sure I haven't. However, from what I read, I understood that the SVD algorithm used is very conservative.
Also,it is worth noting that the diagonal on the matrix has a few dark squares (0.000).
On Sun, Nov 22, 2009 at 11:12 AM, David V. Smith
<[log in to unmask]> wrote:
I don't think that error has anything to do with your contrasts. Your EVs must be highly correlated. Perhaps you mistakenly entered the same EV twice?
On Nov 20, 2009, at 5:24 PM, Shal Hat wrote:
I am in the process of analyzing each run/subject with 6 EVs. However am getting the common rank deficient error. I am assuming this is
most probably due to the many contrasts. Is there a way around this.
To reiterate my scanning protocol, I have five acquisitions per subject.
In essence, it is one stimuli split across the 5 scanning sessions, and I currently have 6 EV. Actually I have more, but 6 will do for now.
Thanks !
On Thu, Nov 12, 2009 at 6:03 PM, Jesper Andersson
<[log in to unmask]> wrote:
Hi again Dav,
The one thing that still seems to be a bit of an issue is whether I can correct for multiple contrasts, for example, if I have 1 contrast of interest, or 2 or 100 (in theory only!). If I do higher-level contrasts independently for just 1 contrast vs. 100, I am nearly guaranteed to get spurious significance in the latter case. In my case, I have a handful of contrasts (which are actually largely independent - along the lines of modelling 3 two-level factors and the interactions between them). Thus, I am still a little concerned about correcting for these multiple (but at least partially independent) contrasts.
Or is this handled already by those contrasts having been specified simultaneously at the first two levels?
this is something that is, funnily enough, largely ignored in neuroimaging. If you ask a question through some contrast and threshold the resulting SPM at a FWE-rate (i.e. corrected for multiple comparisons among the voxels in that SPM) of 0.05 you basically say that you accept 1 false positive out of every 20 times you test a contrast.
If you use two different contrasts in the same data the false positive rate pretty much doubles, for the experiment as a whole. And so it goes as you keep coming up with more contrasts.
So you are right that in your average neuroimaging paper the false positive rate is typically much higher than 0.05, for the paper/study as a whole.
This is very easy for you to "fix" yourself. Let's say you are doing a study where you want to test four different contrasts. Test them at a 0.05/4 FWE level instead, and you will have a false positive rate of 0.05 for your paper/study.
Chances are you'll report fewer blobs though ;-)
Good luck Jesper