Hi Pavel,
It seems you expect T1 to be a baseline, T2 to have some effect (say,
increase), and T3 to have that same effect still present (so, stable
compared to T2).
Perhaps you can consider a conjunction: test separately T1 vs T2, and T1 vs
T3, then take the least significant of the two results. If that is still
significant, then you have the evidence you need.
All the best,
Anderson
On 16 January 2018 at 18:03, Pavel Hok <[log in to unmask]> wrote:
> Hi Anderson,
>
> I tend to disagree and I would still be thankful for any help with a
> design that would include the third measurement as well. If we drop
> the third measurement, we cannot really be sure that the effect is due
> to the short-acting treatment. For instance, there is no control group
> without the treatment. These patients may also improve without the
> intervention (since they all undergo physiotherapy as well). Anyone
> willing to contribute can consider it as an exercise :-)
>
> Thanks and kind regards
>
> Pavel
>
>
> On 14 January 2018 at 20:57, Anderson M. Winkler <[log in to unmask]>
> wrote:
> > Hi Pavel,
> >
> > Regarding the last point, it seems to me that the best thing to do is to
> > drop the 3rd timepoint if the effect being sought is really at the 2nd
> > timepoint. That's the actual hypothesis test, and that can be tested.
> >
> > All the best,
> >
> > Anderson
> >
> >
> > On 1 January 2018 at 19:06, Pavel Hok <[log in to unmask]> wrote:
> >>
> >> Hi Anderson,
> >>
> >> thanks for your answer and interesting suggestions. Here is my further
> >> clarification:
> >>
> >> On 26 December 2017 at 05:34, Anderson M. Winkler
> >> <[log in to unmask]> wrote:
> >> >
> >> > Hi Pavel,
> >> >
> >> > I think you have here a great case for NPC. I'll consider below that
> >> > you'd run this design using PALM, even though some of these tests
> could in
> >> > fact be done using randomise. Also, below I make reference to this
> >> > spreadsheet:
> >> > https://s3-us-west-2.amazonaws.com/andersonwinkler/
> mailinglist/design_pavel.ods
> >> >
> >> >
> >> > On 18 December 2017 at 18:35, Pavel Hok <[log in to unmask]> wrote:
> >> >>
> >> >> Dear FSL experts,
> >> >>
> >> >> I know that similar questions have been asked before, but I could not
> >> >> find a suitable solution for my problem.
> >> >>
> >> >> I have data from 30 subjects who all underwent 3 fMRI sessions, each
> >> >> with two identical runs with a classical blocked motor task. The
> three
> >> >> sessions were not equally spaced in time, but exactly at weeks 0, 4
> >> >> and 12. The group is also quite heterogeneous in terms of age and
> >> >> there is an important behavioural measure obtained at each timepoint.
> >> >>
> >> >> Between timepoint 1 and 2 there is an intervention that has an
> >> >> immediate impact on timepoint 2, but not on timepoint 3 (this is
> >> >> strong and biologically valid assumption). The intervention also
> >> >> significantly affects the behavioural measure.
> >> >>
> >> >> Assuming I have already averaged the two identical runs in each
> >> >> session in a middle-level analysis (fixed effects in FEAT), I am now
> >> >> interested in a) average activation per session adjusted for age,
> >> >
> >> >
> >> > When you say you are "interested in a) average activation per session
> >> > adjusted for age", do you mean the average activation of the 2 runs
> per
> >> > session (that you've already averaged in the mid-level) or the
> average of
> >> > the three sessions per subject? I believe it's the first case, given
> what
> >> > you write below, and I'll consider that when answering further down.
> >>
> >> Yes, I mean the first case.
> >>
> >> >
> >> >
> >> >
> >> >>
> >> >> b)
> >> >> the effect of my behavioural score adjusted for age, c) intra-subject
> >> >> ("paired") differences related to the intervention adjusted for any
> >> >> linear effect of time.
> >> >>
> >> >> For the purpose I have thought about several designs for FEAT, but I
> >> >> have serious doubts regarding their validity. I would appreciate any
> >> >> correction or comment.
> >> >>
> >> >> 1) Average adjusted for age
> >> >>
> >> >> I assume that the safest way is to run separate group analyses per
> >> >> session with the same age covariate (EV 1 group mean, EV 2 demeaned
> >> >> covariate).
> >> >
> >> >
> >> > Exactly. Same model for each of the 3 sessions. In the link below,
> this
> >> > "design_a1". To correct for all 3, you could use the option -corrmod
> in
> >> > PALM, entering each session as a separate input 4D file. Something
> like
> >> > this:
> >> >
> >> > palm -i sessA_4D.nii -i sessB_4D.nii -i sessC_4D.nii -d design_a1.csv
> -t
> >> > contrasts_a1.csv -ise -n 5000 -logp [...other options...]
> >> >
> >> >>
> >> >> However, is it possible to run one analysis with 3 EVs
> >> >> (one per session) to model the means and a 1 EV for age covariate
> >> >> (similar to a two-group comparison with covariate)? Such design would
> >> >> expect a constant effect of age, which can make sense.
> >> >
> >> >
> >> > Yes, it's possible. This is a less intuitive model. Please see in the
> >> > link below the "design_a2". Each subject is one exchangeability
> block. You'd
> >> > run this model as this:
> >> >
> >> > palm -i allsessions_4D.nii -d design_a2.csv -t contrasts_a2.csv -eb
> >> > EB.csv -ise -whole -n 5000 -logp [...other options...]
> >> >
> >> >
> >> >>
> >> >>
> >> >> 2) Covariate adjusted for age
> >> >>
> >> >> Again, I could analyse each session separately. However, I would be
> >> >> also interested in the average effect of the covariate across all
> >> >> three sessions. First, I thought that I can expand the previous
> design
> >> >> (3 EVs for means, 1 EV for age) by adding an extra EV for the
> demeaned
> >> >> covariate. Unlike age, the covariate differs both between and
> >> >> within-subject, but it definitely shows some consistent changes
> across
> >> >> sessions. Therefore I also considered that it might be worthwhile to
> >> >> combine it with the third model below.
> >> >>
> >> >
> >> > This requires 2 models, one for the within-subject effects, that in
> >> > principle would use within-block (within-subject) permutations, and
> one for
> >> > between-subject effects, that in principle would use whole-block
> >> > (whole-subject) permutations. However, these can be combined and
> shuffled in
> >> > synchrony. The within-block permutations done in between-subjects
> design
> >> > will have no effect. Likewise, the whole-block permutations done in
> >> > within-subject designs will have no effect. By "no effect" I mean
> will not
> >> > perturb the design in a way that is useful for hypothesis testing.
> However,
> >> > unless the sample size were tiny, the number of possible permutations
> is
> >> > enormous, such that it's possible to run within- and whole-
> permutations
> >> > together, and still perturb both designs in ways that are nearly
> always
> >> > useful to construct the distribution of the test statistic and test
> the
> >> > hypothesis. Then these two can be combined with NPC. That is:
> >> >
> >> > palm -i allsessions_4D.nii -d design_b1.csv -d design_b2.csv -t
> >> > contrast_b1.csv -t contrast_b2.csv -eb EB.csv -whole -within -npccon
> -n 5000
> >> > -logp [...other options...]
> >> >
> >> > Note that design_b1 (within-subject) doesn't include age, as that is
> >> > captured already by the subject-specific EVs (EV5 onwards).
> >> >
> >> >
> >> >>
> >> >> 3) Effect of treatment and time
> >> >>
> >> >> This one is tough. I thought the tripled T-test from the FSL website
> >> >>
> >> >> (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Single-Group.
> 2C_Three_Measurements_.28.22Tripled_T-Test.22.29)
> >> >> could be an option, but my measurements are not equally spaced in
> >> >> time, which probably violates the assumptions for such design.
> >> >> Instead, I thought I could model the treatment effect and the time
> >> >> with the following design:
> >> >>
> >> >> Input EV1 EV2 EV3 ... EVN+2
> >> >> 1 -0.33 -5.33 1 0 0 ...
> >> >> 2 -0.33 -5.33 0 1 0 ...
> >> >> ...
> >> >> 30 -0.33 -5.33 0 0 0 ...
> >> >>
> >> >> 31 0.67 -1.33 1 0 0 ...
> >> >> ...
> >> >> 60 0.67 -1.33 0 0 0 ...
> >> >>
> >> >> 61 -0.33 6.67 1 0 0 ...
> >> >> ...
> >> >> 90 -0.33 6.67 0 0 0 ...
> >> >>
> >> >> Contrasts
> >> >>
> >> >> Treatment 1 0 0 ... F-test
> >> >> -Treatment -1 0 0 ...
> >> >> Time 0 1 0 ... F-test
> >> >> -Time 0 -1 0 ...
> >> >>
> >> >> where EV1 models the effect of treatment (assumed in the second
> >> >> timepoint only, modelled as 0 and 1s and demeaned), EV2 models the
> >> >> time linearly (values 0 4 and 12 demeaned). Is this model OK given my
> >> >> hypotheses? Feat does not complain about the matrix being rank
> >> >> deficient. Moreover, can I add the behavioural covariate
> >> >> (orthogonalised or not) as an additional EV? I know that adding the
> >> >> age covariate will make the design rank deficient.
> >> >
> >> >
> >> > Here you'd need to make some decisions about what exactly you'd like
> to
> >> > test. Earlier you said that treatment has immediate effect on the 2nd
> >> > timepoint, but not on the 3rd. Is that the case? If yes, the 3rd
> timepoint
> >> > could be dropped, which would simplify the design. Or would you
> rather model
> >> > the slope within subjects with all 3 timepoints, then do a
> group-level model
> >> > to see if that is slope is significantly different than zero? Also,
> would
> >> > you want an overall F-test to see if there is any difference over time
> >> > across all 3 timepoints, or something more specific such as whether
> the
> >> > average slope (considering the 3 sessions) is different than zero?
> >> >
> >> > The design would be different depending on the answer.
> >>
> >> Well, the third timepoint is not necessary to see the treatment
> >> effect, but we consider it important to separate the immediate
> >> treatment effect from the time effect (which may include, e.g.,
> >> improvement due to spontaneous recovery or due to concomitant
> >> physiotherapy). First, I am interested in the overall F-test to see if
> >> there is any difference at all. If positive, I would like to separate
> >> any linear effect T1->T2->T3 from an effect that is only expected at
> >> T2. Importantly, the timepoints are not equally spaced, so there
> >> should be some weighting. Does that make sense now?
> >>
> >> Thanks in advance for any further suggestions and, by the way, happy New
> >> Year!
> >>
> >> Cheers
> >>
> >> Pavel
> >>
> >> >
> >> > Hope this helps!
> >> >
> >> > All the best,
> >> >
> >> > Anderson
> >> >
> >> >
> >> >>
> >> >>
> >> >> Thanks in advance for any help!
> >> >>
> >> >> Cheers
> >> >>
> >> >> Pavel
> >> >>
> >> >> BTW: Did anyone consider creating a library with "validated" designs
> >> >> for FSL? The GLM website does not seem to cover all possible cases,
> >> >> not even some of those discussed in this forum.
> >> >
> >> >
> >> > Yes... that would be good, but need to find time to put these
> >> > together...
> >> >
> >> >>
> >> >>
> >> >>
> >> >> --
> >> >>
> >> >> ------------------------------------------------------------
> -------------------
> >> >> -- MUDr. Pavel Hok
> >> >>
> >> >> ------------------------------------------------------------
> -------------------
> >> >> -- Laboratoř funkční magnetické rezonance
> >> >> -- Neurologická klinika
> >> >> -- Lékařská fakulta
> >> >> -- Univerzita Palackého v Olomouci
> >> >> -- Fakultní nemocnice Olomouc
> >> >>
> >> >> ------------------------------------------------------------
> -------------------
> >> >> -- Laboratory of functional magnetic resonance imaging
> >> >> -- Department of Neurology
> >> >> -- Faculty of Medicine and Dentistry
> >> >> -- Palacky University Olomouc
> >> >> -- University Hospital Olomouc
> >> >> -- Czech Republic
> >> >>
> >> >> ------------------------------------------------------------
> -------------------
> >> >> -- I.P. Pavlova 6, 775 20 Olomouc
> >> >> -- web: fmri.upol.cz
> >> >> -- tel.: +420 588 443 418
> >> >> -- e-mail: [log in to unmask]
> >> >>
> >> >> ------------------------------------------------------------
> -------------------
> >> >
> >> >
> >>
> >>
> >>
> >> --
> >>
> >> ------------------------------------------------------------
> -------------------
> >> -- MUDr. Pavel Hok
> >>
> >> ------------------------------------------------------------
> -------------------
> >> -- Laboratoř funkční magnetické rezonance
> >> -- Neurologická klinika
> >> -- Lékařská fakulta
> >> -- Univerzita Palackého v Olomouci
> >> -- Fakultní nemocnice Olomouc
> >>
> >> ------------------------------------------------------------
> -------------------
> >> -- Laboratory of functional magnetic resonance imaging
> >> -- Department of Neurology
> >> -- Faculty of Medicine and Dentistry
> >> -- Palacky University Olomouc
> >> -- University Hospital Olomouc
> >> -- Czech Republic
> >>
> >> ------------------------------------------------------------
> -------------------
> >> -- I.P. Pavlova 6, 775 20 Olomouc
> >> -- web: fmri.upol.cz
> >> -- tel.: +420 588 443 418
> >> -- e-mail: [log in to unmask]
> >>
> >> ------------------------------------------------------------
> -------------------
> >
> >
>
>
>
> --
> ------------------------------------------------------------
> -------------------
> -- MUDr. Pavel Hok
> ------------------------------------------------------------
> -------------------
> -- Laboratoř funkční magnetické rezonance
> -- Neurologická klinika
> -- Lékařská fakulta
> -- Univerzita Palackého v Olomouci
> -- Fakultní nemocnice Olomouc
> ------------------------------------------------------------
> -------------------
> -- Laboratory of functional magnetic resonance imaging
> -- Department of Neurology
> -- Faculty of Medicine and Dentistry
> -- Palacky University Olomouc
> -- University Hospital Olomouc
> -- Czech Republic
> ------------------------------------------------------------
> -------------------
> -- I.P. Pavlova 6, 775 20 Olomouc
> -- web: fmri.upol.cz
> -- tel.: +420 588 443 418
> -- e-mail: [log in to unmask]
> ------------------------------------------------------------
> -------------------
>
