This depends on what contrasts you are trying to use.
If you have a contrast that only contains "missing" EVs then
you can't have this if you drop these EVs. However, I'm
assuming that the contrasts of interest for the group analysis
are not going to include only missing EVs for any subject or
session. Otherwise there is no information that this subject
or session can contribute to the group analysis in this case
and just shouldn't be included for the higher level analysis.
If you do make contrasts to see certain effects in the lower
level analysis that you don't pass up to the group analysis
(e.g. a 0 0 1 0 0 ... 0 type contrast for each EV) then it doesn't
matter what you put in this place for the lower level analysis
as far as the group analysis is concerned.
The best approach for you is probably to restrict yourself to
first level analyses that only contain the contrast of interest
for the group analysis. This should include at least one
non-missing EV for each session/subject (if not, exclude
that session/subject from the group analysis). Once you've
done this the group analysis is straightforward.
If you are interested in looking at other contrasts in the lower
level, you can always run different contrasts again under
post-stats to have a look (which is quick and easy).
All the best,
On Thursday, June 3, 2004, at 10:42 pm, Sam Harris wrote:
> Hi, Mark
> Thanks for your response. It seems to me, however, that I'm not out of
> the woods yet--because once I drop the missing EVs for any given run,
> that run will then not have the same number of contrasts.
> Have I misunderstood you somewhere?
> On Jun 3, 2004, at 11:15 AM, Mark Jenkinson wrote:
>> What you really need to do is create consistent *contrasts* not EVs.
>> It is contrasts that get fed up into the higher level analyses, not
>> If an EV is missing, then do not have it in the model. However, when
>> formulating your contrasts make sure that the numbering of the
>> is the same in each subject/session so that these feed up
>> You'll have to be careful to select the appropriate EVs each time, as
>> their numbering will change whenever some are missing.
>> For example, if you had a contrast like 1 0 -1 0 0 1 0 and in one case
>> EV2 was missing then the design matrix would only have 6 columns,
>> not 7, and the contrast would become 1 -1 0 0 1 0 in this
>> instance (assuming the other EVs were present).
>> Obviously these missing EVs cannot contribute to the contrast, but
>> as long as each contrast contains at least one EV then you are fine.
>> All the best,
>> Sam Harris wrote:
>>> I'm attempting to analyze event-related data acquired on 14 subjects,
>>> 3 runs each, in which
>>> certain EVs, in any given run, were not represented. For example,
>>> subject #6, in run #2, may not
>>> have provided a single example of EV 8 (while all other subjects did,
>>> as did subject #6 himself in
>>> runs 1 and 3). I'm wondering if there is any way of creating a
>>> "place-holder" model (in three-
>>> column format), so that all subjects and all runs can contain the
>>> same number of EVs, thereby
>>> allowing analysis at the group level. Could I, in the above case,
>>> create a spurious model for EV 8
>>> (in subject #6, run #2), with a fictional onset time and duration,
>>> and scale it at (or near) 0? Or
>>> could I create an extra volume, tacked onto to the end of all scans,
>>> that represented the average
>>> value of each functional run, and then reference this timepoint in my
>>> model as a dummy-EV?
>>> Needless to say, I'm looking for a solution that will produce, in the
>>> worst case, a type 2 error.
>>> Thanks for your help.