Dear Paula,
I would just subtract the theoretical chance level from the observed
accuracy maps and perform the test accordingly against 0 instead of (in
your case) .5.
There is concern about the normality assumptions made when applying
t-tests. If you share this concern, you may want to have a look at this
randomization-based analysis for whole-brain inference using searchlight
maps. http://dx.doi.org/10.1016/j.neuroimage.2012.09.063
Or, as a compromise, you could use a sign flipping test as implemented
in FSL's randomise function or in BROCCOLI's RandomiseGroupLevel. This
test does not assume normality but only a symmetrical null distribution.
(If you have two classes in your classification, then symmetry is a very
reasonable assumption in my opintion.) Of course these analyses need a
little more time to run but if you have a GPU, then BROCCOLI is very
fast. Btw: also in these sign flipping tests, you will have to subtract
.5 from your accuracy maps.
HTH,
Michael
On 03.12.20 11:22, Paula Maldonado wrote:
> Dear SPMers,
> I would like to perform a one-sample t-test analysis in SPM, using as
> input a set of nilearn searchlight classification maps.
>
> How can I set up SPM to estimate a voxelwise one-sample t-test against a
> constant mean value (my chance level) other than zero (e.g., constant =
> 0.5)?
>
>
> Is this possible at all?
>
> Thank you for your help,
>
> Paula
>
|