Hi Jeannette,
Thanks a lot for your prompt reply!
The stimuli are equivalent in both conditions, but you're right that the number of trials for AC and BD varies, depending on whether the subject detected them or not. So, depending on the subject's performance, the efficiency of one EV may differ from that of the other (up to 0.5). So that clarify the difference between the two modeling approaches with this respect.
However, and I should have been more explicit about this, sorry, but I expect both conditions to be associated with activity in partly different areas. Indeed, both consist of one tactile and one visual stimulus, the difference between both conditions beong the side on which each stimulation is presented (i.e. tactile left/visual right vs. visual left/tactile right). So, I obviously expect activity in early sensory areas to differ between the two conditions. However, what I'm interested in is what's common to these two conditions, i.e. the fact of perceiving 2 simultaneous visual and tactile stimuli, regardless of their spatial distribution.
So how do the two approaches differ exactly with respect to the location of activations? My understanding is that modeling the 2 conditions with a single EV asks the model to find the voxels that will respond regardless of the stimuli spatial location, which is good for what I want. But how does that differ from a [1 1] contrast in the first model? I should add that the difference in the pattern of activity obtained with the 2 approaches does not concern only early sensory areas, otherwise it would make more sense to me.
Thanks again for your time and patience!
Best,
Stéphane
On Wed, 6 Mar 2013 09:33:29 -0600, Jeanette Mumford <[log in to unmask]> wrote:
>Hi,
>
>Do you have the same number of trials of each and is their efficiency
>around the same?
>
>The first approach gives equal weight to both stimuli, even if one may be
>better (due to more trials or better efficiency). For example, let's say I
>have the numbers (1,2,3) and in another group just 4. Method 1 does the
>following:
>
>(mean([1,2,3])+4)/2=(2+4)/2=3
>
>whereas the second approach is analogous to
>
>mean([1,2,3,4])= 2
>
>The second approach makes more sense if you have unbalanced
>data/efficiencies between the two stimuli. To reiterate, your model 1 is
>the average of the two activations whereas model 2 is the average
>activation over all trials. Slight, but important difference.
>
>Additionally, approach 1 is a bit of over modeling if you have 0 interest
>in studying the two trials individually. You're allowing a super flexible
>model by splitting the regressors into 2 (one for each trial) so your
>residual variance for model 1 will be lower. That may or may not have an
>impact on the group model. Likely not a big one as the between subject
>variance would dominate at that level anyhow.
>
>Model 2 is the more interpretable result.
>
>Cheers,
>Jeanette
>
>On Wed, Mar 6, 2013 at 8:54 AM, Stephane Jacobs
><[log in to unmask]>wrote:
>
>> Greetings everyone,
>>
>> I have what I imagine to be a simple modeling question, but can't seem to
>> find the answer by myself. Apologies if it's very basic indeed!
>> I have 2 stimulation conditions, each composed of 2 stimuli - AC and BD.
>> I'm not interested in analyzing them separately, but rather in the activity
>> associated with the fact of perceiving 2 stimuli (with respect to only one
>> - other conditions I haven't mentionned for the sake of simplicity),
>> whether it be AC or BD. Now I wonder what is the best approach for this:
>> 1. model AC and BD separately with 2 regressors at the first level, and
>> use a [1 1] contrast to look at the average response?
>> 2. model both conditions using one single EV?
>>
>> I've used both approaches, and they give pretty different outcomes: while
>> with the first approach, both regressors are associated with activity
>> within the same regions, which in turn show up for the average response
>> contrast, I only get a small subset of the same areas with the second
>> approach.
>>
>> I'm assuming the question asked by modeling both conditions together
>> differs from that asked by modeling them separately, but I'd like to
>> understand the difference to chose the best approach. Could someone be kind
>> enough to explain to me what happens in the second case?
>>
>> Many thanks in advance!
>>
>> Stephane
>>
>
|