Hi Nick,
It is first of all worth restating the standard way to combine info
from the temp derivative, which is to do an f-test across both the
main response and the temp derivative. You can then get at the sign/
directionality of the response (which an f-test is otherwise blind to)
by using contrast masking between the f-test and the sign of t-
contrast for the main response.
If you want to still use the equation: sign(PE1)*sqrt(PE1^2+PE2^2),
then be wary that the resulting statistic is not going to be t-
distibuted under the null hypothesis (consider what happens with
Rayleigh distibutions). You would need to account for this or use
some sort of non-parametric inference.
Cheers, Mark.
----
Dr Mark Woolrich
EPSRC Advanced Research Fellow University Research Lecturer
Oxford University Centre for Functional MRI of the Brain (FMRIB),
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK.
Tel: (+44)1865-222782 Homepage: http://www.fmrib.ox.ac.uk/~woolrich
On 11 Nov 2009, at 15:40, Michael Harms wrote:
> Hi Nick,
> Perhaps I'm missing something, but how do you code the non-linear
> formula in STEP1 in terms of a contrast vector 'c'? That is, I don't
> see how the variance equation in STEP2 can be the variance of the
> variable computed in STEP1...
>
> cheers,
> Mike H.
>
>
> On Tue, 2009-11-10 at 16:17 -0800, Nick Wymbs wrote:
>> After some additional digging, I think I have the solution to my
>> nonlinear question. Please correct me if there's any error!
>> To reiterate, the formula that I used to run a main effects contrast
>> with the main condition PE and it's temporal derivative is the
>> following:
>>
>> STEP1: sign(PE1)*sqrt(PE1^2+PE2^2) .
>>
>> From here, the varcope must be generated in order to get the correct
>> error term in the denominator. I didn't use prewhitening so there
>> formula is quite simple:
>>
>> STEP 2: varcope == c' * inv(X' * X) * c
>>
>> Where c is the contrast vector and X is the full design matrix.
>>
>> When considering the contrast vector, both the main EV and temporal
>> derivative EV receive an equal weight (1) and everything else is 0.
>> Finally, in order to generate t-stats, the result from STEP 1 is
>> divided by the square root from STEP 2 (error).
>>
>>
>> Cheers
>> Nick
>>
>>
>>
>> On Mon, Nov 9, 2009 at 4:28 PM, Nicholas Wymbs <[log in to unmask]>
>> wrote:
>> Hi FSLers
>>
>> I'm trying out a contrast that uses the Calhoun (2004) method
>> of incorporating the temporal derivative using the formula:
>> sign(PE1)*sqrt(PE1^2+PE2^2) where
>>
>> PE1 == standard EV fit
>> PE2 == temporal dv of EV
>>
>> I then divided by the SEM of the residuals (res4d.nii.gz). The
>> values do not look like t-stats, with the range being
>> something like -250 to 250. Now I'm faced with the really
>> basic question: how are the PEs converted to t's and z's? I
>> know the conversion is supposed to be pretty straight-forward
>> and it states on the feat page that 'To convert estimates of
>> parameter estimates (PEs) into statistical maps, it is
>> necessary to divide the actual PE value by the error in the
>> estimate of this PE value', which is what I thought I did when
>> dividing by the SEM of the res4d file. What's goin' on?
>>
>> Thanks
>> Nick
>>
>
|