Dear Torsten,
I must admit I am little confused here. What exactly is it you want
to use randomise for?
The point of randomise is to calculate a threshold for you, and to do
so without any distributional assumptions. So, if/when using
randomise you should not need to worry about how to subsequently
threshold your images.
You should also be aware that the threshold that randomise gives you
will _not_ be identical to that which you would get from any table of
t-statistic. Those (table) values all hinge on your data being iid
normal distributed, whereas the threshold that randomise gives you
will not.
Finally, for imaging data the threshold you are really interested is
that which randomise calculates from the max statistic within each
permutation. That will give you a correction for multiple
comparisons, and will be _considerably_ higher than those you quoted
earlier.
Good Luck Jesper
On 27 Apr 2007, at 14:44, Torsten Ruest wrote:
> Dear all,
>
> I probably make things more complicated than they are. I am
> actually just
> interested in the t-values (tstat images only). After having giving
> it a
> thought, I guess it doesn't really matter if 1- or 2-tailed at this
> stage,
> as I am taking the raw t-maps, and threshold later.
>
> So here is what I did:
> I restriced the view to only show me values in the range of 2.23 to
> 6.21 in
> the tstat_1, applied a LUT, and did the same for tstat_2. I then
> added those
> views together, having red hot LUT for control>patient, and cyan
> hot for
> control<patient t-values. 2.23 is the critical t-value for p=0.05
> (2-tailed
> ttest).
> Is this a reasonable way of doing it, or do I miss something here.
> I.e. how
> are the t-maps using design_ttest2 calculated, with n=n1+n2 (so 12
> in my
> case), or n=n (6 in my case)
>
> Thanks in advance for your opinion.
>
> Torsten
>
>
> On Fri, 27 Apr 2007 14:08:29 +0100, Ged Ridgway
> <[log in to unmask]> wrote:
>
>> Hi Torsten,
>>
>> Perhaps Jesper will correct me, but I think randomise does
>> right-tailed tests only. If you use design_ttest2 it produces two
>> contrasts, so that *tstat1* files are for grp1 > grp2 and tstat2 are
>> for grp1 < grp2 (where the latter is actually a right-tailed test of
>> the contrast grp2-grp1).
>>
>> So if you are interested in the two-tailed alternative, then the
>> p-values are half what they should be (i.e. overly significant). I
>> think the two (1-p)-value images that get stored have zeros where the
>> other contrast is significant, so you could get a single two-tailed
>> (1-p)-value image by adding the p-images together, with a little
>> maths, something like (untested):
>> # for (ba)sh, where blah is something like image_max
>> fnm=blah_tstat
>> # for (tc)sh, where blah is something like image_max
>> set fnm=blah_tstat
>>
>> # then:
>> avwmaths ${fnm}1 -mul -1 -add 1 -mul 2 ${fnm}1_p2
>> avwmaths ${fnm}2 -mul -1 -add 1 -mul 2 ${fnm}2_p2
>> avwmaths ${fnm}1_p2 -add ${fnm}2_p2 ${fnm}_p2
>> avwmaths ${fnm}_p2 -mul -1 -add 1 ${fnm}
>> where the final image is (1-p) for two-sided p-values.
>>
>> Hope that helps,
>> Ged.
>>
>> Torsten Ruest wrote:
>>> Dear Jesper,
>>>
>>> thanks again. I've run design_ttest2 design 6 6 to create my
>>> design files.
>>> In the description it reads "for the two-group unpaired t-test
>>> case". So I
>>> never used the glm gui. And these 2 groups are different, I mean
>>> they are
>>> genetically different, therefore I thought this setup should be
>>> right. But I
>>> am not sure if the output would be 2-tailed or not. I've assumed
>>> so, since
>>> the software cannot assume in which direction it would go, and
>>> because of
>>> the output (negative and positive values), I assumed that it's 2-
>>> tailed.
>>> So my only uncertainty is currently if the output of
>>> design_ttest2 design 6
>>> 6 would be 2-tailed or not.
>>>
>>> Thanks again for efforts.
>>>
>>> Cheers,
>>>
>>> Torsten
>>>
|