Dear Ashley 

> -----Original Message-----
> From: [log in to unmask] [mailto:[log in to unmask]] 
> Sent: Wednesday, August 15, 2007 2:46 PM
> To: [log in to unmask]
> Subject: Questions regarding your post RE:"PPI error" in SPM 
> Archives..
> 
> Dear Dr. Gitelman,
> 
> Greetings!                       I'm currently trying to run a PPI 
> analysis and get away with the multi-session issue. Through the 
> archive-search, i found your detailed    suggestions posted on May 
> 16,2006 (RE: "PPI error") and find the step-by-step 
> instruction extremely helpful! :)

thank you.

> 
> However, there're a couple points that i find myself 
> struggling with and  would appreciate a lot if you can give 
> me some insights.
> 
> 1. On your post, you mentioned ".....You may also want to 
> remove event onsets that occur right near the end of each 
> session, ie. perhaps within 5-10 seconds or so. Do not remove 
> any of the scans...."  => Could you maybe elaborate more on 'why'?

Let's say you have a TR of 2 secs. Also you have a series of trials with the
last 2 trials in a particular condition occurring 12 seconds and 4 seconds
before the end of a run. If you included the event onset for the trial
occurring 12 seconds before the end its HRF will be modeled fairly well, as
there are 6 data points that follow (6 TR's). However, if you include the
onset for the trial 4 seconds before the end it will include only 2 data
points and is less likely to be modeled well. Hence the suggestion that the
trial onsets occurring close to the end of each run might be left off. I
don't feel strongly about this and cannot provide a very specific guideline.

> 
> 2. Is AR(1) necessary if one were to merge all the sessions into one?

yes. these are separate issues.

> and 3. For the merged big session, how would the 
> session-effect (generally controlled for by using grand mean 
> scaling) be controlled?

 The session or block effect, i.e., the fact that some scans are associated
with run 1, others with run 2, etc. allows modeling the intercept for each
session separately as well as modeling the transition between sessions.

> 
> The reason for these questions are as followed..(please see 
> the attached figures) So,i followed your instruction to put 
> my 4 sessions into one, including
> 3 user-defined regressor. The specified design matrix looks 
> great before i push the "estimate" button, however, after the 
> estimation is done, those 3 user-defined regressor looks very 
> funky (ie. not ones and 
> zeros anymore)...   I tried running the estimation with and without 
> AR(1), and there doesn't seem to be any difference.

It looks funky for several reasons. 1) SPM doesn't not treat the confounds
quite correctly. I think the block effects should really be entered into the
block confounds part of the design, which I believe is referred to as iB in
the code. However, when the block confounds are entered as user defined
regressors they are placed in the actual design. Therefore, they are
high-pass filtered and whitened, which would not have been done had they
been modeled as true block effects. I went through the painful exercise once
of putting the block effects into the model properly and it did not make any
substantial difference in the results for my design, but I don't know if
this would be generally true. 

The reason you see the funkiness (which incidentally affects the regular as
well as the block conditions) is that before you estimate the design, spm
shows you the raw design matrix. Once it has been estimated it shows you the
filtered and whitened design matrix.

so the bottom line is you should use AR(1) and it otherwise looks like the
session effects were entered correctly.

good luck
darren


> 
> I'd be very thankful for any advice or directions you point 
> me to! :) Thank you very much, in advance.
> 
> Best,
> Ashley Chen
> 
> 
>