Hi Mona,
GLM as implemented in SPM (and not only) normally makes use of mixed-effect analysis, so very briefly: does fixed effect analysis on the subject level (more liberal) and random effect at the group level (more conservative because it does take into account inter-subject variability to estimate the population parameters rather than only related to a sample).
So what is important is a reasonable sample size to enable population-based inferences. What is your n (as a rule of thumb I would say it should be above 10 and preferably above 15,)?
Anyway still if you had strong effects with FWE at the first level I would imagine them to survive at the second as well (with the same p threshold), but again it may depends on n value...
Secondly I gather you used voxel-based correction? If it so it could kill even large activated clusters because none of constituent voxels survived the correction razor. Try then cluster-based correction (you loosen up the voxel-based thresholds for large activated clusters; briefly look for the smallest size of significant cluster in your stat table and use it as a threshold). But remember it's the two-edge sword, if you expect small clusters they may not survive this type of thresholding.
I hope it helps,
Iwo
>-----Original Message-----
>From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
>On Behalf Of Mona Moisala
>Sent: 14 August 2013 10:59
>To: [log in to unmask]
>Subject: [SPM] 2nd level analysis
>
>This may be a fairly simple issue, but since I'm new to fMRI and SPM, I cannot
>get my head around it.
>
>I simply want to compare experimental condition A to the zero level (which I
>understand is the mean of the implicit rest from all subjects). I have already
>calculated the contrast A against rest in the 1st level analysis, and I have seen
>clear activation clusters that are topographically similar across subjects and are
>significant after FWE-correction and with a p-value threshold of <0.05.
>However, when I enter the con* images (1per subject) into a one-sample t-
>test in 2nd level analysis, using the same p-value threshold does not give me
>significant activation clusters. The activation clusters are present only if I do
>not correct for multiple comparisons. If the data looks similar across subjects,
>wouldn't averaging just increase the effect observed in the data and the
>results should be even more prominent in the 2nd level analysis? Or is
>intersubject variability in absolute BOLD-signal the issue? Should I somehow
>normalize the amplitude of BOLD signal before 2nd level analysis? Or should I
>be looking at individually defined ROIs instead of whole brain analysis at the
>second level?
>
>Thank you so much for your help!
|