See comment on microtime below.

On Fri, Sep 16, 2011 at 3:05 PM, Darren Gitelman
<[log in to unmask]> wrote:
> A couple additional answers:
>
>
> On Fri, Sep 16, 2011 at 12:08 PM, Friston, Karl <[log in to unmask]> wrote:
>>
>>  Dear Aize,
>>
>>
>>
>> 1. In my understanding, the BOLD signal in a roi is deconvolved (through
>> HRF) into neural signal level (xn), filtered by psy task (psy * xn), then
>> convolved with HRF, where the eigenvariate will be calculated, this is the
>> ppi signal , which will be future fed into GLM to look at the whole brain
>> activation influenced by this ROI. Am I right?
>
>
>>
>> Yes this is right. I should note that the basic PPI analysis does not
>> depend on this deconvolution step. We introduced the deconvolution in the
>> days before DCM (which provides an optimal hemodynamic deconvolution). In
>> many instances, you may get better results if you simply take the raw
>> (mean-corrected) eigenvariate as xn (using the VOI tool in the results
>> interface) and multiply it by the (mean corrected) psy factor; particularly
>> for slow or block designs. This is because the PPI deconvolution is trying
>> to solve a very difficult problem using fairly old methods (Weiner
>> filtering) and can sometimes give unstable estimates.
>
> I agree the concept of PPI is not dependent on deconvolution since it
> relates to interactions in factorial designs, however, the need for
> deconvolution became apparent when one tried to apply PPI to event related
> data. The PPI deconvolution paper (Gitelman et al., Neuroimage, 2003) showed
> that forming the interaction term without deconvolution led to errors
> particularly for event-related data, but only to a minimal extent with block
> design data.  The deconvolution step in the script uses empirical Bayes
> deconvolution and not Weiner filtering (although it is formally identical to
> Weiner filtering if the prior spectral density is assumed to be 1).  In any
> case I would generally suggest using the PPI machinery to do the
> deconvolution.
>
>
>>
>> 2. For my data, we have 98 scans per run. For example length(PPI.ppi) =
>> 98, but I do know why the length(PPI.xn) = 1568. Any idea?
>>
>> If I remember, this is an estimate of the underlying neuronal time-series
>> in micro-time (with TR/16 time bins). Darren may know - he loved this code J
>
> I did love it. :). Yes Karl is correct it reflects the conversion to
> microtime. The default number of bins per TR = 16 in SPM. This value is
> stored in defaults.stats.fmri.fmri_t or in an SPM.mat structure in
> SPM.xBF.dt. So 16 * 98 = 1598. If you want to convert PPI.xn back to TR time
> you would resample by this vector:
> 1:NT:N*NT
>
> where N = number of scans in the session.
> NT = TR/SPM.xBF.t;
>
> so PPI.xnmacro = PPI.xn(1:NT:N*NT);
>
> The resulting vector may miss onsets if you are using an event related
> design.

While you will miss the "onset", you will not miss the HRF associated
with the event.


>
>
> Darren
>>
>>
>>
>> 3. We have four runs for each subject. It is a kids related study, to look
>> at their word comprehension skill. If I overlay the PPI.ppi in left pulvinar
>> for these four runs under word condition, the ppi signal looks different,
>> please see the attached file. This subject was very still and almost no
>> motion inside the scanner, also his performance accuracy is the same across
>> runs. Any suggestion?
>>
>>
>>
>> I would compare these estimates of xn * psy with those obtained by
>> constructing them using an xn that was not deconvolved (as  above). If the
>> raw PPIs look more stable over sessions, you could proceed with these.
>>
>>
>>
>> 4. What is the unit of PPI.ppi ? The original BOLD signal is very high,
>> but after PPI, the signal in PPI is with in 2 or 3
>>
>>
>>
>> The units will depend on the units of the HRF assumed during deconvolution
>> (because they are neuronal activity times psy units; not BOLD units times
>> psy units). I would just call them arbitrary units, because the psy units
>> could be anything.
>>
>>
>>
>> I hope that this helps - Karl
>
>
> --
> Darren Gitelman, MD
> Northwestern University
> 710 N. Lake Shore Dr.
> Abbott 11th Floor
> Chicago, IL 60611
> Ph: (312) 908-8614
> Fax: (312) 908-5073
>
>
>