Dear Karl,
Thanks very much for the clarification. I'll copy to the list
for others following this thread.
Best,
Andy
At 12:02 PM 12/3/2001 +0000, you wrote:
>Dear Andy,
>
>Oh - I see. Your approach seems absolutely fine. You are simply
>reporting the parameter estimates about which you made an inference (in
>terms of a contrast of these estimates). Just say that they are
>relative to the mean over all conditions and report them directly for
>each condotion.
>
>With very best wishes,
>
>Karl
>
>
>----- Begin Included Message -----
>
>Date: Mon, 26 Nov 2001 12:47:50 -0500
>To: [log in to unmask] (Karl Friston)
>From: "Andrew J. Saykin" <[log in to unmask]>
>Subject: obtaining condition means from beta and contrast image values
>
>Dear Karl,
> I should have provided a bit more background on the rationale for
>trying to extract these means. What we are really trying to do is
>visualize the regional activation results in hypothesized areas for a 4
>level parametric blocked design task. We have a patient and control group
>that we are comparing. I have analyzed the parametric responses as a
>contrast ( -6 0 1 0 2 0 3 0 ) for each individual (4 conditions
>plus temporal derivatives included in model). I then used a 2nd level
>random effects model to compare the above contrast (coding for parametric
>increase) between groups. The motivation for my questions about computing
>means is that the random effects results don't lend themselves to an
>intuitive graph of the parametric activation effects or group
>differences. Since we didn't want to enter 4 conditions per subject in a
>random effects model because it would violate the 1 scan per subject
>restriction, I thought we could get fitted and adjusted means for the 4
>conditions using the betas (1,3,5,7) to examine the parametric
>responses. Is there a different way that you would suggest approaching
>this analysis and visualization?
>
>Many thanks,
>Andy
>
>
>At 04:36 PM 11/26/2001 +0000, you wrote:
> >Dear Andy,
> >
> > > I know there has been previous correspondence on this topic with
> > > regard to calculating percent change and I wanted to check out my
> > > understanding. I am interested in obtaining condition means for fmri
> > > data on a voxel by voxel basis separately for each of a group of
> > > subjects. Each subject was processed separately. We performed a blocked
> > > analysis where there were 4 conditions (and also included the 4 temporal
> > > derivatives to adjust for the possibility of minor timing
> > > errors). Since SPM automatically includes the intercept there are then
> > > 9 columns in the design matrix for extracting contrasts
> > > (cond1, td1, cond2, td2, cond3, td3, cond4, td4, constant).
> > > I think the answers to the following questions are "yes" but would
> > > appreciate confirmation of this or guidance if I am missing something.
> > >
> > >(1) Does the following contrast test for a condition 1 effect versus zero
> > >(null hypothesis, no slope)?
> > >
> > >cond1 td1 cond2 td2 cond3 td3 cond4 td4 constant
> > > 1 0 0 0 0 0 0 0
> 0
> >
> >Yes it does.
> >
> > >(2) Do the obtained contrast image values indicate the condition 1 mean,
> > >but with the global scaling removed since the intercept was included in
> > >the model?
> >
> >The contrast returns the estimated effects of cond1, discounting any
> >effects that can be explained by the other regressors. Because the
> >other regressors include the constant, the ensuing contrast is
> >'activation' from the mean of this voxel's time-series. Global effects
> >are removed before estimation by proportional scaling to 100, in fMRI.
> >This means the [adimensional] units are percent of whole brain mean.
> >
> > >(3) Is the sum of beta1 (ie, cond1) plus beta9 (ie, constant) the actual
> > >mean for condition 1?
> >
> >Not really. The 'mean of condition 1' does not have any meaning.
> >Beta9 is the mean over all conditions. Beta1 + Beta9 would be an
> >estimate of the fitted response during condition 1 but only because the
> >boxcar for cond1 has unit height. The mean over conditions is Beta9.
> >Note that if you are comparing Beta1 from different subjects it is not
> >necessarily a good idea to divide by Beta9 (implicit in % regional
> >activation). This is because you would be confounding any differences
> >in cond1 with differences in Beta9 (which could have artefactual or
> >technical causes).
> >
> >I hope this helps - Karl
>
>----- End Included Message -----
