Reading through the replies on this, I am getting a better
understanding of FSL and I think this actually goes back to my haziness
about demeaning. Yes, I was trying to model every time point, or at
least every time point for the three types of stimuli I was analyzing.
Thus, I think that this model was, in fact, rank deficient (further
evidenced by lots of very light space off of the diagonal in the
covariance image on the left and a diagonal that progressed toward two
black squares in the covariance image on the right, and a ratio of
min:max eigen values in the SVD of the matrix of 1.415e-3.)
Another problem in my understanding came about because I was forgetting
that FSL models the delay in hemodynamic response (Steve hinted at this
in an earlier email.) With AIR I was modeling the "r" volume because
this is where I expected the peak of my activation of interest to be.
However, I am really wanting to know "what is going on in the subjects'
minds when they press the answer" which is temporally located in the
"a" volume.
I think that the analysis I want to do in FSL is:
Ca La Ma
Ca - (La & Ma) 2 -1 -1
La - (Ca & Ma) -1 2 -1
Ma - (Ca & La) -1 -1 2
because I am wanting to isolate the unique cognitive aspects that are
attributable to the Ca condition and trying to remove the motor and
sensory activations that are common to La and Ma.
Three questions arise that I wish I had more confident answers to:
1. Does that look more sensible within the context of FSL and given
what demeaning does?
2. Does this appear to be a reasonable replication of the AIR analysis
I did previously given the conceptual differences in FSL? If not, what
would be a better design that would approximate what I was doing with
AIR?
3. How would I use F-tests to further help illuminate these contrasts?
How would I set the F-tests up? How would I interpret the resulting
activations in volume created in the F-test?
Again, thanks for helping me slog through this. I am slowly building
an understanding of how the logic of FSL and its use of statistics maps
onto the domain where I was trained (regression analysis of survey
data.)
Darren
On Thursday, May 22, 2003, at 09:31 AM, Stephen Smith wrote:
> Hi Darren, I think there are two separate issues here.
>
> 1. The design is rank deficient I suspect because you are attempting to
> model every timepoint with some condition. Because the data is demeaned
> you are not allowed to do this - unless you have a natural single
> "rest"
> condition then you need to define one of the conditions as rest and
> work
> with respect to this. OR it maybe that your design is ok, just _nearly_
> rank deficient - what deficiency ratio is it giving you?
>
> 2. With your "C vs L&M" - you are asking
> Cr - mean(Cq+Ca+Lr+Lq+La+Mr+Mq+Ma)
> would
> [ Cr - (Cq+Ca)/2 ] - { [ Lr - (Lq+La)/2 ] - [ Mr - (Mq+Ma)/2 ] } / 2
> be more sensible or am I missing the intension of the question you are
> wanting to ask with this question?
>
> Thanks, Steve.
>
>
>
> On Thu, 22 May 2003, Darren Schreiber wrote:
>
>> I am trying to replicate some event-related analysis in FSL that I did
>> with UCLA's AIR tools and I am getting an error that my design is rank
>> deficient.
>>
>> Here is the setup:
>>
>> I have seven types of questions A,C,L,M,R,T,U. The TR is 4 and I
>> collected three volumes for each question (the question, the answer,
>> and the rest: q, a, r).
>>
>> In the AIR tools, I specified three contrasts of C vs L & M, L vs. C &
>> M, and M vs C & L. I included all three volumes (qar) in the contrast
>> design as follows:
>>
>> C vs L&M:
>> q a r
>> C -1 -1 8
>> L -1 -1 -1
>> M -1 -1 -1
>>
>> L vs C&M
>> q a r
>> C -1 -1 -1
>> L -1 -1 8
>> M -1 -1 -1
>>
>> M vs C&L
>> q a r
>> C -1 -1 -1
>> L -1 -1 -1
>> M -1 -1 8
>>
>> The results fit substantive expectations beautifully. But, I want to
>> take advantage of FSL's mixed effects abilities.
>>
>> So, I specified nine EVs corresponding to Cq, Ca, Cr, Lq, La, Lr, Mq,
>> Ma, Mr. And, I tried to set up the contrasts like I had in the AIR
>> tools.
>>
>> C vs LM -1 -1 8 -1 -1 -1 -1 -1
>> -1
>> L vs CM -1 -1 -1 -1 -1 8 -1 -1
>> -1
>> M vs CL -1 -1 -1 -1 -1 -1 -1 -1
>> 8
>>
>> What am I doing wrong? How do I replicate the AIR results in FSL?
>>
>> I have verified that the custom 3 column files specifying the timing
>> of
>> the EVs is correct.
>>
>> Of course, I know that EV1, EV2, and EV3 are not independent, but I am
>> not sure how to replicate the AIR results without setting up the same
>> contrast.
>>
>> By the time I get to the course in Marina del Rey, I will have a very
>> good understanding of FSL, I think. But, since I have to have my data
>> analysis done for a deadline before the class, I appreciate your
>> indulgence with the ceaseless barrage of questions.
>>
>> Darren
>>
>
> Stephen M. Smith MA DPhil CEng MIEE
> Associate Director, FMRIB and Analysis Research Coordinator
>
> Oxford University Centre for Functional MRI of the Brain
> John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK
> +44 (0) 1865 222726 (fax 222717)
>
> [log in to unmask] http://www.fmrib.ox.ac.uk/~steve
>
|