> From [log in to unmask] Sat Jan 8 18:41:43 2000
> Date: Sat, 8 Jan 2000 13:47:48 -0500
> X-MSMail-Priority: Normal
> X-MimeOLE: Produced By Microsoft MimeOLE V5.00.2314.1300
> Subject: missing time points
Dear Walter,
> Before asking my question I would simply like to describe the
> experimental paradigm:
>
> - Each run within the experiment contained 3 randomly presented
> event types or conditions (A, B and C)
> - Each event type or condition was presented every 13 sec
(stimulus
> duration was 2 sec and ISI was 11)
> - TR for the experiment was 3.25 sec
> - 160 volumes were acquired during the run, but because of buffer
> size problem, every 4th volume was discarded after acquisition, leaving
only
> 120 volumes.
>
> Now the original authors (this experiment was performed by another
> lab within our institution, so I do not quite have all the details) would
> like to analyze the data using an event-related approach, but I advised
them
> that this would be problematic because of the missing data points. My
> question is simple, how would one properly handle the data? Were guessing
> setting it up as a box car model and then using random effects analysis
(20
> subjects were scanned). Any comments or suggestions, as always, are
greatly
> appreciated.
One way to proceed would be to (i) create a design matrix in SPM99b as
if the missing scans were present, (ii) delete the appropriate rows
from the design matrix that correspond to the missing scans:
e.g.
>> spm_fMRI_design
>> load SPM_fMRIDesMtx
>> C = xX.X(:,Sess{1}.col);
>> C(1:4:end,:) = [];
and (iii) re-enter these as regressors of interest in the analysis
proper. You could then use either a first- or second-level analysis.
You should still apply high and low pass filtering as normal but
modeling intrinsic autocorrelations may be rendered inappropriate
because of the missing points in the time-series.
I hope this helps - Karl
|