Dear Darren,
Thank you for helpful response. I tried what you suggested for a one
session for 21 subjects. Interestingly, both performed equally well,
yielding an average correlation among the original and recovered
eigenvariate of about .84. I've attached a histogram of the
correlations for the two models.
Does this seem like a reasonable value? For transparency, I've cut and
pasted the code I used for each subject (the PPI structures were
computed and saved prior):
original = PPI.Y;
xn = PPI.xn;
xY = PPI.xY;
load SPM
% Setup variables
%-----------------------------------------------
RT = SPM.xY.RT;
dt = SPM.xBF.dt;
NT = RT/dt;
fMRI_T0 = SPM.xBF.T0;
N = length(xY(1).u);
k = 1:NT:N*NT;
% Basis function in microtime
%-----------------------------------------------
hrf = spm_hrf(dt);
% Get estimated neural signal and convolve
%-----------------------------------------------
cxn = conv(xn,hrf);
cxn = cxn((k-1) + fMRI_T0);
% Save correlation among orig. and recovered eigenvar
%-----------------------------------------------
correlation(s) = corr(cxn,original);
On Fri, Sep 16, 2011 at 12:24 PM, Darren Gitelman
<[log in to unmask]> wrote:
> Bob:
>
> The Bayesian deconvolution in the PPI script has generally performed quite
> well, although nothing is infallible. I am not that surprised that modeling
> trials as events vs. boxcars would make this difference since you are in a
> sense telling the deconvolution that the underlying neuronal activity is
> different.
>
> One way to decide if the deconvolution is inaccurate would be to take PPI.xn
> (the deconvolved "neural" activity), convolve it with an hrf and interpolate
> it from microtime to TR time. Then you can compare the resulting
> "reconstituted" BOLD data with your original eigenvariate. I'd be interested
> to hear what you find.
>
> Darren
>
>
>
> On Thu, Sep 15, 2011 at 2:32 AM, Bob Spunt <[log in to unmask]> wrote:
>>
>> Dear SPM experts,
>>
>> When conducting PPI with event-related designs, I've found that the
>> PPI regressors are dramatically affected by how trials are modeled,
>> i.e. either with a delta (duration = 0) or boxcar (duration = stimulus
>> offset minus onset) function. For example, the attached plots show PPI
>> regressors from identical seeds in identical designs, where the only
>> parameter varying is the use of a delta vs. boxcar function to model
>> the stimulation (trials varied in length from about 3-5 seconds). The
>> correlation between these regressors is only about .50. This seems
>> surprisingly low.
>>
>> The reason I bring this up because, when one is merely concerned with
>> modeling the psychological response, the difference between modeling
>> trials as delta vs. boxcar functions is often small (although
>> certainly not insignificant; Grinband et al., 2008, NeuroImage).
>> However, this difference seems to be amplified when doing PPI with
>> event-related designs. Is this a reasonable conclusion?
>>
>> Thanks in advance for any comments,
>> Bob
>>
>>
>> -------------------------------------------------------------------------------
>> Bob Spunt
>> Postdoctoral Fellow
>> Social Cognitive Neuroscience Lab - www.scn.ucla.edu
>> Department of Psychology
>> University of California, Los Angeles
>
>
>
> --
> Darren Gitelman, MD
> Northwestern University
> 710 N. Lake Shore Dr.
> Abbott 11th Floor
> Chicago, IL 60611
> Ph: (312) 908-8614
> Fax: (312) 908-5073
>
>
>
|