Hi Anderson,
Thank you very much for your help. I think this is a very elegant solution.
I tried pasting the design into the glm_gui (see attached), at first
omitting subjects with missing data.
I then got the following warning message: "design matrix uses
different groups (for different variances), but these do not contain
"separable" EVs for the different groups". Did I do something wrong?
The design appropriately tests for B > E and C > F.
I know I said that the two baseline scans (A and D) would be expected
to be the same, but on second thought it would be nice to correct for
the baseline condition as well.
I.e. if a difference is seen between B and E and/or C and F, can this
be explained by a difference between A and D?
Therefore what I would really like to do, I guess, is to test for A -
D > B - E and A - D > C - F.
Do you agree and how could that be implemented into the design?
Thank you so much once again and sorry to keep bothering you with this :)
All the best and happy holidays,
Anders
2014-12-22 0:24 GMT+01:00 Anderson M. Winkler <[log in to unmask]>:
> Hi Anders,
>
> Please, see attached a suggested design (you can open .ods files with
> LibreOffice).
>
> There are actually two designs: the first is considering that there were no
> missing data, and the second with the missing data as you indicated. You can
> then compare and see what changes. The rows marked in orange in the second
> design refer to the missing acquisitions, and should not be included in the
> design, that is, these lines don't actually exist.
>
> Some contrasts (lines 94-98) are marked with a "-". These aren't meant to be
> run, and are there just for clarity, otherwise the contrasts that are
> interesting (C1-C6, and specially C5-C6) might seem too mysterious at first
> :-)
>
> Run this in randomise including the option "-e design.grp", supplying then
> the file with the exchangeability blocks (EB) as indicated in the
> spreadsheet.
>
> Hope this helps!
>
> All the best,
>
> Anderson
>
>
> On 20 December 2014 at 18:07, Anders Hougaard <[log in to unmask]> wrote:
>>
>> Hi Anderson,
>>
>> There are 15 subjects in total.
>>
>> The order of the interventions was randomised and the two scan
>> sessions (active/placebo) were carried out on two different days
>> separated by a wash-out period, so we would expect D to be the same as
>> A.
>>
>> All the best,
>> Anders
>>
>> 2014-12-19 11:15 GMT+01:00 Anderson M. Winkler <[log in to unmask]>:
>> > Hi Anders,
>> >
>> > Two questions first:
>> > - How many subjects do you have?
>> > - Can D be considered the same as A? That is, if the sessions happened
>> > in
>> > the order as shown, A, B, C, D, E and finally F, can we consider that,
>> > by
>> > the time D was acquired, any effect from the active treatment had
>> > already
>> > been cleared/washed out?
>> >
>> > Another thing (just a general question, it won't change the analysis at
>> > this
>> > stage): did you randomise order of the interventions (that is, some
>> > subjects
>> > receiving randomly the treatment first, then placebo, whereas other
>> > receiving placebo first, then treatment)?
>> >
>> > Thanks
>> >
>> > All the best,
>> >
>> > Anderson
>> >
>> >
>> > On 18 December 2014 at 19:17, Anders Hougaard <[log in to unmask]>
>> > wrote:
>> >>
>> >> Dear experts,
>> >>
>> >> I have the following 1st level feat analyses from a group of subjects:
>> >>
>> >> A: Baseline - B: Active treatment - C: Post active treatment
>> >>
>> >> D: Baseline - E: Placebo - F: Post placebo
>> >>
>> >> All subjects are scanned in all conditions (*However, see below)
>> >>
>> >> I want to know the following:
>> >> What is the difference between baseline and the active treatment
>> >> condition?
>> >> What is the difference between baseline and the post-treatment
>> >> condition?
>> >> + I want to correct for placebo effects.
>> >>
>> >> What would be the best approach for a higher level analysis?
>> >> Subtracting placebo from active and then doing triple t-test?
>> >>
>> >> *To make things more complicated, some scans are missing:
>> >> Condition B for subject 1, 2, and 3 - and condition C for subject 2, 3,
>> >> and 4.
>> >>
>> >> The easiest would be to exclude all of these 4 subjects entirely, but
>> >> is there some appropriate way of keeping scans from some of these
>> >> subjects and answer the above questions without losing statistical
>> >> power?
>> >>
>> >> All the best and happy holidays,
>> >> Anders
>
>
|