Steve,
thanks. Here is the very basic idea (it's an Event related design with
non-overlapping trials):
I have two types of reasoning tasks (Reasoning_A and Reasoning_B), each
with a matched baseline (Baseline_A, Baseline_B). Of course, I then
subtract each reasoning from its matched baseline.
Now, it so happens that Reasoning_A-Baseline_A yields much more
prefrontal activity than Reasoning_B-Baseline_B. Out of previous
literature and my own previous results I'd be inclined to say that this
reflects the differential cognitive mechanisms underlying the two types
of reasoning.
However, it also so happens that Reasoning_A takes statistically longer
than Reasoning_B (and appears more complicated in terms of # of errors
made -- error trials are discarded), thus it could be argued that the
prefrontal activation is due to time, or sheer complexity, not to the
cognitive processes underlying the Reasoning_A task.
Hence I'm thinking of how to parcel out time/complexity from the design.
Should the prefrontal activations remain after such correction I'd be
more persuaded that it has to do with cognitive processes than sheer
time/difficulty.
If I understand correctly there are 2 possibilities that seem applicable
here:
1) use RT as a covariate. This strategy seems intuitive but my EVs are
defined using RT data (i.e. in each trial I model the event time-locked
to the response: each event is modelled as the 4 volumes leading up to
the subject's response, where the RT is collected). Does this prevent me
from using this strategy? I just have the intuition that it will end up
soaking most of the action in both my reasoning tasks. Is this an
incorrect intuition?
2) I understand you suggest, as an alternative, modulating the height of
each event. This also seems a good possibility, but I will need guidance
on some basic points:
How do people typically transform arbitrary scales (e.g. here the
RT, or any other measures of difficulty) into heights?
What is a "reasonable" interval for event heights; so that I can
guide the transformation of RT milliseconds into event height?
(I don't have an obvious way of separating events into "thinking" and
"action" unfortunately, so I fear that strategy may not be available..)
thanks for the patience and the help!
martin
Steve Smith wrote:
> Hi,
>
> On 21 Jan 2008, at 13:12, Martin M Monti wrote:
>
>> Dear FSL masters,
>>
>> I have a design in which 4 tasks are compared 2 x 2 (e.g. TaskA,
>> BaselineA; TaskB, BaselineB). Now, I have a significant difference in
>> RT between TaskA and TaskB (but not between BaselineA and BaselineB),
>> so to prevent the possibility that differential activation between
>> the two tasks is just due to extra difficulty or activation/timing
>> issues, I'd like to use RT as a covariate.
>>
>> I have some questions though, on its usage.
>>
>> - Is the first level analysis the right place for using the
>> covariate? I have RTs for every trial, so I could just insert an
>> extra EV at every 1st level analysis. There is one thing thought that
>> is not-intuitive to me on this point: my EVs of interest are defined
>> using RTs (to select the on-off periods) so I somehow have the
>> feeling that just everything will be caught in the covariate. I
>> suppose this is an incorrect intuition?
>
> There are several ways that such experiments might be modelled. One
> thing is that some people separately model the thinking process
> leading up to the subject action and the post-action period. Then
> indeed some people have one fixed-height EV for the action, and also
> add in an action EV where the height is modulated by the task
> difficulty etc., or in your case possibly the RT. It's hard to know
> what's the best thing here without knowing a little more about the
> experiment, but you can play with these possibilities.
>
>> - Would using the Average RT for each task in the within subject
>> Second-level analysis (I then do a Second-level analysis aggregating
>> all subs with Random Effx) be a more suitable strategy (I'm taking
>> this idea from the example on the feat5 page of the FSL website)?
>
> Not necessarily; as you suggest, this is a cruder model as it does not
> allow for variable within-subject RTs and modelling. It really depends
> on what question you want to answer / model out with your RT data.
>
>> - To Use the covariate, I have to select "Orthogonalize" EVs, correct?
>
> Depends what you want. If you want it just to safely soak up extra
> variance and not affect the mean fitting then yes you would do this.
>
> Cheers.
>
>
>> thanks,
>>
>> all the best
>>
>> martin
>
>
> ---------------------------------------------------------------------------
>
> Stephen M. Smith, Professor of Biomedical Engineering
> Associate Director, Oxford University FMRIB Centre
>
> FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
> +44 (0) 1865 222726 (fax 222717)
> [log in to unmask] http://www.fmrib.ox.ac.uk/~steve
> ---------------------------------------------------------------------------
>
|