Dear Michela,
the distinction between corrected and uncorrected p-values in SPM has to do
with correction for multiple comparisons.
Normally when we say something is significant at a certain p-level what we
really mean is "the risk that the effect I observe is just coincidental is
equal to or lower than p". So for example if I were to report something at
the 0.05 level I really say "there is only 5% risk that in fact nothing
happened".
Now, lets say we have a medicine that we think is really good for
something, only we don't really know for what. So, we put a group of people
on the drug and measure a number (say twenty) parameters like
blood-pressure, blood-lipids etc etc before and after treatment. When we
look at the results we find that it had significantly (at the 0.05 level)
reduced blood pressure, so we conclude it was good for lowering blood
pressure.
What did we do wrong (apart from being really silly)?. We did not correct
for multiple comparisons. Assume that the drug had no effect whatsoever on
anything (i.e. the null-hypothesis was true). If we had measured a single
parameter (e.g. blood lipids) there would have been a 5% risk that we
erroneously concluded that the drug was effective. If we had measured two
parameters the risks would have added up (approximately) so that there
would have been an almost 10% risk of erroneously concluding that the drug
had an effect on one of the parameters. With the twenty parameters we
measured there was a more than 60% risk that we would erroneously conclude
that the drug was effective on one or more of the parameters, while in fact
it was completely useless.
So, whenever we measure more than one response variable we need to "correct
for multiple comparisons". When doing fMRI you are effectively measuring
loads of "response variable". If voxels were independent you would do as
many comparisons as there are voxels in the brain.
So the difference between the corrected and the uncorrected p-values are
that the former have not been corrected for the effects described above,
whereas the latter have. As a general rule you shall always use corrected
values, unless you have a-priori (i.e. well before looking at any SPM's)
decided that you will only look in one single voxel given by co-ordinates
[x,y,z]. In that case you can use the uncorrected value.
>
> I run SPM to analyze a fMRI series and I need help to interpret results.
> The volume summary table shows statistical results for cluster-level and
> voxel-level with p-corrected and p-uncorrected values.
> When uncorrected and corrected values are significant for a cluster?
> When uncorrected and corrected values are significant for a voxel?
> For example I have a cluster with the results:
>
> cluster-level
> P-corrected=0.964 ( > 0.05!)
> K=49
> P-uncorrected=0.012
>
> voxel-level
> P-corrected=0.751 ( > 0.05!)
> T=5.49
> Z=4.43
> P-uncorrected=0.000
>
> Results are more meaningful at the voxel-level or cluster-level?
The cluster level results refer to the cluster as a whole, and the
interpretation is that "somewhere in this cluster there is something
activated". It might be tempting to say "surely the highest peak in the
cluster must be it", but strictly speaking one cannot. Hence, a cluster
level significance has a "poorer localising power" in that you don't really
know which are the activated voxels. On the othet hand, cluster level
inference can sometimes be more sensitive (i.e. have higher statistical
power).
>
> My activation is a false positive or a true activation? And why?
> In there any general criteria to be sure about that?
Unfourtunately there is no general criteria to say if an activation is
"true", or a false positive. In your specific case you do not have a "false
positive", mainly because you have no "positive". Your peak is well below
the 0.05 corrected level. On the other hand, there is no way to say if it
is perhaps a false negative.
Good luck Jesper
|