Also, I just took a look your Dropbox README, and part of the issue is
that in your comparison model with just 3 total EVs (grp1 grp2 Behav), the
contrasts [1 0 1] and [0 1 1] are not meaningful contrasts. In such a
model, the meaningful potential contrasts are [1 0 0], [0 1 0], [1 -1 0],
and [0 0 1] (and the negative version of each of those).
cheers,
-MH
--
Michael Harms, Ph.D.
-----------------------------------------------------------
Conte Center for the Neuroscience of Mental Disorders
Washington University School of Medicine
Department of Psychiatry, Box 8134
660 South Euclid Ave. Tel: 314-747-6173
St. Louis, MO 63110 Email: [log in to unmask]
On 7/29/13 9:09 PM, "Harms, Michael" <[log in to unmask]> wrote:
>
>The two group interaction model is equivalent to fitting two lines to the
>data.
>The first two EVs fit the intercept for the two groups. And the betas for
>EVs 3 and 4 fit the slope (for each group) as a function of the behavioral
>variable.
>Contrast [1 -1 0 0] then tests whether the groups differ at the intercept.
>Contrast [0 0 1 -1] tests whether the slopes differ.
>Those two contrasts are testing completely different things, so depending
>on what you are calling "group" vs. "interaction" images, the results you
>report are certainly possible.
>
>With that visual picture in mind, hopefully the model makes sense.
>
>cheers,
>-MH
>
>--
>Michael Harms, Ph.D.
>
>-----------------------------------------------------------
>Conte Center for the Neuroscience of Mental Disorders
>Washington University School of Medicine
>Department of Psychiatry, Box 8134
>660 South Euclid Ave. Tel: 314-747-6173
>St. Louis, MO 63110 Email: [log in to unmask]
>
>
>
>
>On 7/29/13 6:16 PM, "Benjamin Philip" <[log in to unmask]> wrote:
>
>>Let me focus on the empirical reason, because it's much more
>>certain/convincing, even to me:
>>
>>I think it's incorrect because it doesn't hold up to the reality-check of
>>observing the two groups independently. In some areas the two
>>single-group images show drastically different z-statistic values (e.g.
>>at X=61, Y=50, Z=68, Patient = 16.9, Ctl = 7.5), yet the interaction
>>image is not significant (interaction zstat=1.4). Conversely, there are
>>areas where the two single-group images have nearly identical values
>>(e.g. 29/52/68, Patient = 23.6, Ctl = 23.1) yet the interaction image is
>>significant (interaction zstat = 7.7).
>>
>>Whatever interaction effect that model produces, it isn't (P*b) - (C*b).
>>
>>Files are at https://www.dropbox.com/sh/jajnm7gjnmfa615/ASA5he3gcg . In
>>there you can find .png files for a quick-and-dirty show, and nifti files
>>to demonstrate this at the voxel of your choice.
>>
>>Thanks,
>>-Benjamin Philip
>>
>>On Mon, 29 Jul 2013 18:11:28 +0000, Harms, Michael <[log in to unmask]>
>>wrote:
>>
>>>Hi Benjamin,
>>>The model that you need is the interaction one that you noted:
>>>http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Two_Groups_with_continuous_cov
>>>a
>>>ri
>>>ate_interaction
>>>
>>>
>>>Why are you saying that it isn't correct?
>>>
>>>You can't look for an interaction between groups unless your model
>>>includes separate Behavioral EVs for each group.
>>>
>>>cheers,
>>>-MH
>>>
>>>--
>>>Michael Harms, Ph.D.
>>>
>>>-----------------------------------------------------------
>>>Conte Center for the Neuroscience of Mental Disorders
>>>Washington University School of Medicine
>>>Department of Psychiatry, Box 8134
>>>660 South Euclid Ave. Tel: 314-747-6173
>>>St. Louis, MO 63110 Email: [log in to unmask]
>>>
>>>
>>>
>>>
>>>On 7/29/13 1:03 PM, "Benjamin Philip" <[log in to unmask]> wrote:
>>>
>>>>I'm having trouble organizing a FEAT third-level analysis correctly. I
>>>>have 2 groups, and a behavioral measurement for each participant. What
>>>>I
>>>>really want to see is "Behavior-correlated activity in patients" minus
>>>>"Behavior-correlated activity in controls". (P*b) - (C*b).
>>>>
>>>>I tried the interaction model suggested at
>>>>http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Two_Groups_with_continuous_co
>>>>v
>>>>ar
>>>>iate_interaction but it's not addressing the correct question.
>>>>Theoretically I say this because we're looking for a subtraction, not
>>>>an
>>>>interaction. More convincingly/clear-to-express, it's empirically the
>>>>wrong model, because it produces results wildly different from what we
>>>>see comparing (P*b) and (C*b) by eye.
>>>>
>>>>I've also tried a contrast of [1 -1 1] on the EVs "Patient, Control,
>>>>Behavior" - i.e. like this oversimplified version:
>>>>Pat Ctl Bhvr
>>>> 1 0 .4
>>>> 1 0 -.2
>>>> 0 1 -.3
>>>> 0 1 .1
>>>>...If I do that, [1 -1 1] looks pleasingly different from [-1 1 1], but
>>>>something is deeply wrong with this modeling: [1 -1 1] and [1 -1 -1]
>>>>and
>>>>even [1 -1 0] look nearly identical. The same areas can't be both
>>>>behavior-correlated and behavior-anticorrelated, so I'm doing something
>>>>wrong. [1 1 -1] and [1 1 1] also look near-identical, but at least [0
>>>>0
>>>>1] and [0 0 -1] show different areas.
>>>>
>>>>Any suggestions for how to implement (P*b) - (C*b)?
>>>>
>>>>Thanks,
>>>>
>>>>-Benjamin Philip
>>>
>>>
>>>________________________________
>>>The materials in this message are private and may contain Protected
>>>Healthcare Information or other information of a sensitive nature. If
>>>you are not the intended recipient, be advised that any unauthorized
>>>use, disclosure, copying or the taking of any action in reliance on the
>>>contents of this information is strictly prohibited. If you have
>>>received this email in error, please immediately notify the sender via
>>>telephone or return mail.
>
________________________________
The materials in this message are private and may contain Protected Healthcare Information or other information of a sensitive nature. If you are not the intended recipient, be advised that any unauthorized use, disclosure, copying or the taking of any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error, please immediately notify the sender via telephone or return mail.
|