As you can see the *PSY* negative columns do line up well together,
but the positives do not. Samething with the PPI, where you have the
control condition, the curves line up. Now, if we think about the
concept of the the general linear model (Y=BX)...
Trying to fit the all-control vector to the data is much different
than fitting the condition1-control vector. I think this reiterates
the point I'm making in a paper I'm about to submit on PPI. That is
PPI as it is currently implemented fits these vectors above, rather
than fitting a vector for condition1, condition2, .., and control. In
splitting them up, one is then estimating the relationship of each
component, rather than the joint relationship. My hunch is that if you
were to separate them conditions, you would either eliminate the
average effect OR more likely find out which one is driving the
effect. When you only model 2 conditions, there is a chance that you
attribute the variance of the data to the wrong factor or it ends up
in the error term. Also, when you only model 2 conditions, you are
only modelling the activitation effect of those 2 conditions.
Modelling has shown that the individual model fit is improved when you
separate the conditions.
I'm adding some comments to the code for splitting the vectors and
have termed the approach "a generalizable form of PPI (gPPI)" and
hopefully can provide you with the code later today.
Best Regards, Donald McLaren
=================
D.G. McLaren, Ph.D.
Postdoctoral Research Fellow, GRECC, Bedford VA
Research Fellow, Department of Neurology, Massachusetts General
Hospital and Harvard Medical School
Office: (773) 406-2464
=====================
This e-mail contains CONFIDENTIAL INFORMATION which may contain
PROTECTED HEALTHCARE INFORMATION and may also be LEGALLY PRIVILEGED
and which is intended only for the use of the individual or entity
named above. If the reader of the e-mail is not the intended recipient
or the employee or agent responsible for delivering it to the intended
recipient, you are hereby notified that you are in possession of
confidential and privileged information. Any unauthorized use,
disclosure, copying or the taking of any action in reliance on the
contents of this information is strictly prohibited and may be
unlawful. If you have received this e-mail unintentionally, please
immediately notify the sender via telephone at (773) 406-2464 or
email.
On Thu, Feb 24, 2011 at 6:50 PM, J S Lee <[log in to unmask]> wrote:
> Hi Donald,
> Thank you for your reply!
> I am using SPM8.
> I looked at the SPM.xX.X values for each subject. The values in the SPM.xX.X
> *PSY* columns for each of my models line up perfectly across subjects. For
> the different models within a subject, there also seems to be correspondence
> for the PSY negative values (although there are slight baseline shifts
> between different models). The values in the PPI column, however, do not
> correspond so well. I don't think I would have expected the PPI values to
> correspond across subjects, however, because the VOI values differ from
> subject to subject, so the PPI should also differ as it represents an
> interaction? The PPI values for different models in the same subject also do
> not show perfect correspondence (png attached: It shows 2 subjects' SPM.xX.X
> PPI values over the first 80 scans. The All conds - control model and Cond 1
> - control PPI values are plotted for each subject. I didn't include the Cond
> 2 - control, Cond 3 - control, etc. for clarity).
>
> All models are coming from the same VOI, so I think the adjustment has to be
> the same for all models (all extractions were adjusted for an F contrast of
> the effects of interest at the first-level model). Is this what you meant?
> The voxels are also identical (again, coming from the same VOI I think they
> have to be?).
>
>
> Thank you very much for taking the time to consider my question--even
> looking at the PSY and PPI columns has been useful!
>
> Jamie
>
>>
>>Are you using SPM8? The issue of summing was fixed in one of the later
>>releases of SPM5, so if you have an older version, that could explain
>>some of the issue. You could check to make sure that negative aspects
>>of the SPM.xX.X for the PPI term are the same for all subjects. You
>>could plot them.
>>Are you using the same adjustment for all models?
>>Are you using exactly the same voxels for all models?
>
>
>>
>>On Tue, Feb 22, 2011 at 6:21 PM, J S Lee <[log in to unmask]> wrote:
>>> Dear list,
>>>
>>> I conducted a PPI analysis in an experiment with 6 conditions. To
>>> replicate
>>> a previous study's PPI analysis, I was interested in connectivity
>>> differences between 5 of the conditions compared to the control (6th)
>>> condition, so extracted my VOI (using an all effects of interest
>>> contrast),
>>> then created a PPI model with a [1 1 1 1 1 -1] weighting for the
>>> psychological context regressor. I get a reasonable replication of the
>>> same
>>> PPI effects from the previous study, so the results are sensible.
>>>
>>> However, in that previous study, there were not enough trials of each of
>>> the
>>> 5 conditions to realistically analyze them separately, which is why I
>>> collapsed across them. In this study, there are many more trials, so I
>>> was
>>> hoping to look at which of the 5 conditions were driving the original PPI
>>> results. I was given hope when the initial PPI replicated in this new
>>> study.
>>> However, when I create separate PPI models for each condition versus
>>> control
>>> (e.g., context regressors using [1 0 0 0 0 -1] for model 1, [0 1 0 0 0
>>> -1]
>>> for model 2, etc.), NONE of these analyses show the same pattern as the 1
>>> 1
>>> 1 1 1 -1 model does. Mostly there are no significant (or anywhere near
>>> significant) results, and those random speckles that do show up at low
>>> threshold are not in the same places.
>>>
>>> Is it theoretically possible that 5 conditions vs 1 other can produce a
>>> PPI,
>>> but that none of those conditions singly vs the 1 other can do that? Or
>>> must
>>> there be an error? I have checked the microtime onset files to make the
>>> context is specified correctly, and made sure everything matches up in
>>> terms
>>> of specifying the conditions. Everything about the models looks fine to
>>> me.
>>> I know the 5 conditions vs 1 is a bit unbalanced, but it replicates the
>>> previous study (in which the 5 vs 1 were equal in terms of number of
>>> trials), and I understand that when creating the context variable one
>>> does
>>> NOT sum the vector to zero the way one would in defining a contrast for a
>>> regional activation analysis.
>>>
>>> Many thanks in advance for any thoughts,
>>> Jamie Lee
>
> _______________________________________________________________
> Get the Free email that has everyone talking at http://www.mail2world.com
> Unlimited Email Storage – POP3 – Calendar – SMS – Translator – Much More!
|