Dear Shary,
I am sorry for the delay in replying (you might get a quicker response
if you emailed the SPM help-list on these occasions)
>I am comparing two groups one Healthy and the other Patients with AD
(SPECT).
>When I retrieve the results by clicking on the results button (contrast 1 -1
>for decrease in AD), SPM asks for FWE or FDR for multiple comparison. I
>have done the analyses with both and without any of these. However, I
>remember in the older versions of SPM (some years ago starting in 1998)
>there were no such selections, and we had only height threshold (voxel
>threshold) which would give cluster level corrected and uncorrected p-values
>based on RFT. Here are my questions:
>1)- I select none when prompted for FWE | FDR | None and, when prompted
>for height threshold right after clicking on None, I enter 0.001 (default
>value). Will I still get voxel level and cluster level corrected p-values
>for multiple comparisons based on RFT?
Yes. The questions about "FWE | FDR | None" are simply about the display
threshold for the SPM. The adjusted p-values based on maxima are exactly
the same. However, the display threshold (u) defines the size and number
of supra-threshold clusters and therefore the p-values for these will change
with u.
>2)- Does FWE means that probability of false positive regardless of height
>threshold? My understanding is that RFT is used to calculate corrected FWE
>for voxel and cluster level corrected p-values when we select FWE for
>instance at 0.05 (default value or any other value). Am I correct?
Yes, But as noted above the adjusted p-values for clusters and sets of
clusters will depend on the display threshold u used to define a cluster.
>3)- I read that FDR is proportion of false positives. If set to 0.001 and
>we have 100000 voxels then 100 of them are expected to be false positive.
>How this compares to when we bypass FWE and FDR (in question 1) and set a
>height threshold to 0.001 in the next prompt and yet I get corrected and
>uncorrected p-values for both voxel level and cluster level?
The adjusted or corrected and uncorrected p-values for clusters and sets
of clusters all relate to false positive rates. They can be computed
for any display threshold. This display threshold can be specified using
FWE (controlling false positive rates of maxima) of FDR (false discovery
rate of voxels) or none (uncorrected).
FDR treats the SPM as a collection of discrete voxels and only controls
the false discovery rate of voxels (not maxima or clusters). This means
that of 5% of the voxels in the SPM are, on average, false positives.
However,
the number of false clusters is not controlled so you should be careful when
describing the SPM in terms of discrete regional activations.
I hope this helps - Karl
|