Hi all,
Here's my $0.02 on the issue of looking at unthresholded maps.
If you were to ask a random statistician if their data analysis
strategies included completely screening-out and ignoring the results
of all non-significant tests, I think you'd find few if any would say
"yes".
I think it's safe to say that no statistician would recommend blinding
yourself to all results except those corresponding to "Reject Ho".
For example, basic statistics courses cover the issue of practical
significance vs. statistical significance, stressing that one must
look at results in real-world units.
In the context of fMRI, here are two extremes that illustrate the
importance of practical vs. statistical significance. If one obtains a
very significant result but it corresponds to a 0.01% BOLD change, it
should be clearly reported as such, so that readers know your
reporting an incredibly subtle effect.
On the other hand, if a 100% BOLD change was found that was
non-significant, we'd like to know what gave rise to such a result;
most likely there is huge variance, but huge variance tells you that
there was virtually no power to detect a result even if Ho was false!.
(And this then ties in to the need for looking at standard deviation
images, if only to identify regions with high risk of Type II errors.)
-Tom
-- Thomas Nichols -------------------- Department of Biostatistics
http://www.sph.umich.edu/~nichols University of Michigan
[log in to unmask] 1420 Washington Heights
-------------------------------------- Ann Arbor, MI 48109-2029
|