Hi Michal,
sorry, it seems that I have been unclear: the GLM on the subject/
session modes tests if the 'strength' (as shown in the subject/session
mode) of the spatio-temporal component (as shown by the spatial map
and the associated time course) - I was under the impression that this
is what you intend when you say you're looking for the 'between
subject effect'. The GLM analysis tests the single vector of length N
(where the number of subjects/sessions N=37 it seem) against a simple
design - in your case two groups (A vs B). You could test a different
statistic, e.g. the between group effect of the between condition
effect sizes by analysing the other time courses in the t??.txt files
and deriving a single statistics that is then being analyses using
parametirc or non-parametric approaches.
Wrt your analysis it seems that in SPSS you perform some test on a
vector of length 48, the number of time points I presume. This is
conceptually very different and can only be compared to the GLM on the
time courses from melodic, not the GLM on the subject/session modes.
hth
Christian
> Hello,
>
> Thank you very much for answering my questions, they were
> helpful, however I still would have to read technical documents more
> carefully and maybe refresh my matrix algebra :-) I have one follow
> up uncertainity, which I will try to describe below:
>
> Relating to the idea of performing regular GLM on individual
> subject's timecourses you say that: "If you specify a group design,
> melodic already does what you're proposing on the rank-1 time
> courses". I did this analisys in SPSS and the results are comparable
> to the output of melodic but by no means the same. I am including
> the comparison of outcomes from those two calculations below, I
> also included a contrast matrix in the attachment. As can be seen in
> this example the difference apparently significant in melodic
> (COPE(1) is insignificant in GLM (time * gender). Maybe those Copes
> are calculated more like planned contrasts, but still even using "/
> CONTRAST (gender)= SPECIAL (-1, 1)" in SPSS I get different results.
> Additionally I find it hard to grasp why would one contrast be
> significant and it's reciprocal not. After all the the difference
> between goup means should be the same both ways (only the sign would
> change) and the error term is the same for both contrasts (this is
> with assumtion that only timecourses are calculated). Maybe I didn't
> fully understand you and this melodic glm takes also the spatial
> maps into calculations and not only timecourses? In any case I am
> affaraid that I still do not grasp the basis of melodic inference,
> sorry :-(
>
>
> Best regards
> Mike
>
>
> SPSS command (I have only 48 scans - that is timepoints):
>
> GLM
> var001 to var048 BY gender
> /WSFACTOR = time 48 Polynomial
> /CONTRAST (gender)= SPECIAL (1, -1)
> /METHOD = SSTYPE(3)
> /CRITERIA = ALPHA(.05)
> /WSDESIGN = time
> /DESIGN = gender .
>
> SPSS output:
>
>
> Tests of Within-Subjects Effects
>
>
> Measure: MEASURE_1
>
> Source
>
> Type III Sum of Squares
>
> df
>
> Mean Square
>
> F
>
> Sig.
>
> time
>
> Sphericity Assumed
>
> 2337,294
>
> 47
>
> 49,730
>
> 9,053
>
> ,000
>
> Greenhouse-Geisser
>
> 2337,294
>
> 3,926
>
> 595,377
>
> 9,053
>
> ,000
>
> Huynh-Feldt
>
> 2337,294
>
> 4,550
>
> 513,718
>
> 9,053
>
> ,000
>
> Lower-bound
>
> 2337,294
>
> 1,000
>
> 2337,294
>
> 9,053
>
> ,005
>
> time * gender
>
> Sphericity Assumed
>
> 340,970
>
> 47
>
> 7,255
>
> 1,321
>
> ,073
>
> Greenhouse-Geisser
>
> 340,970
>
> 3,926
>
> 86,855
>
> 1,321
>
> ,265
>
> Huynh-Feldt
>
> 340,970
>
> 4,550
>
> 74,942
>
> 1,321
>
> ,261
>
> Lower-bound
>
> 340,970
>
> 1,000
>
> 340,970
>
> 1,321
>
> ,258
>
> Error(time)
>
> Sphericity Assumed
>
> 9810,565
>
> 1786
>
> 5,493
>
> Greenhouse-Geisser
>
> 9810,565
>
> 149,178
>
> 65,764
>
> Huynh-Feldt
>
> 9810,565
>
> 172,891
>
> 56,744
>
> Lower-bound
>
> 9810,565
>
> 38,000
>
> 258,173
>
>
> Contrast Results (K Matrix)
>
>
> Contrast Results (K Matrix)
>
>
> plec Special Contrast
>
>
> Averaged Variable
>
> MEASURE_1
>
> L1
>
> Contrast Estimate
>
> -2,55E-007
>
> Hypothesized Value
>
> 0
>
> Difference (Estimate - Hypothesized)
>
> -2,55E-007
>
> Std. Error
>
> ,000
>
> Sig.
>
> .
>
> 95% Confidence Interval for Difference
>
> Lower Bound
>
> -2,55E-007
>
> Upper Bound
>
> -2,55E-007
>
>
> ICA OUTPUT:
>
> GLM (OLS) on subject/session-mode
> GLM β's
> F-test on
> full model fit
> Contrasts
> PE(1):
> PE(2):
> 1.99150
> 1.17970
> F = 26.98940
> dof1 = 2; dof2 = 37
> p < 0.00000
> (uncorrected for #comp.)
> COPE(1):
> COPE(2):
> COPE(3):
> COPE(4):
> z =
> z =
> z =
> z =
> 1.75 ;
> -1.75 ;
> 5.22 ;
> 3.32 ;
> p < 0.04015
> p < 0.95985
> p < 0.00000
> p < 0.00046
>
>
> This page produced automatically by MELODIC Version 3.05 - a part of
> FSL - FMRIB Software Library.
>
>
>
>
> > -----Original Message-----
> > From: FSL - FMRIB's Software Library
> > [mailto:[log in to unmask]] On Behalf Of Christian F. Beckmann
> > Sent: Monday, July 28, 2008 11:46 PM
> > To: [log in to unmask]
> > Subject: Re: [FSL] Melodic output questions
> >
> >
> >
> > On 25 Jul 2008, at 19:42, Michal Kuniecki wrote:
> >
> > > Hello,
> > >
> > > Customarily I would like to begin with being sorry for asking
> > > question which has high probability of being trivial to
> well ......
> > > Christian Beckmann for example :-) But anyway I haven't
> > found straight
> > > answer for it neither on the list nor in the documentation.
> > >
> > > Basically I would like to understand MELODIC output
> > better. I perform
> > > Tica analysis on my data additionally specifying both
> > box-car function
> > > (I use simple ABAB block design for all subjects) and
> > between subject
> > > contrast. In the output which I get if one clicks at the graph
> > > representing the timecourse of the particular component one gets
> an
> > > array of numbers. As I understand they represent the timecourse
> > > associated with particular component separately for each
> > subject (also
> > > the timecourse and the model fit are given in first two columns).
> >
> > Yes, the final N columns are the time courses for each one of
> > the input data sets, the first column is the rank-1
> > approximation of these N time courses and (if a design was
> > included in the GUI) a further column (2nd) shows the full
> > model fit of the design.mat to the rank-1 approximation.
> >
> > > I do not understand however
> > > what is the unit or metrics behind those numbers (I understand
> that
> > > they are normalized, but normalized what?).
> > >
> >
> > The time courses (1st column in t??.txt) are normalised to
> > unit standard deviation and all the energy is absorbed into
> > subject/session mode vectors. The data is kept on the
> > original scale (typically unit- less gray value intensities
> > that come off your scanner)
> >
> >
> > >
> > > Similarly in "Session/Subjects mode" we get boxplot
> > accompanied by
> > > another plot representing, as I get it goodness of fit between
> > > particular subject and the timecourse of the particular
> > component. But
> > > again, how is this synthetic measure being calculated and
> > what does it
> > > precisely represent?
> >
> > The session/subject mode is calculated from the selection of
> > all time courses via the rank-1 approximation. The boxplot
> > again is on an almost arbitrary scale - it is still useful to
> > judge if a component is non-zero for all the individual
> > subjects or if the component is an 'outlier' component (i.e.
> > has one or only few very strong subjects and the remaining
> > ones close to 0). If no between-subject model has been
> > included then by default melodic will simply test for average
> > group effect (i.e. use a constant 1 regressor to see if on
> > average the group activated). If a group design has been
> > specified, melodic will calculate a between-subject GLM using
> > the specified design, e.g.
> > testing for group differences using an unpaired or paired t-test.
> >
> > >
> > >
> > > Additionally I would like to ask if it is sound and
> > reasonable to
> > > import the data representing timecourses for particular
> > subjects for
> > > particular component (those described in second paragraph)
> > aggregate
> > > them for rest and active condition (that is in case of my
> > simple ABAB
> > > design) and
> > > perform standard GLM analysis on them (in my case I would have
> > > repeated factor of condition 2 levels and between subjects
> > factor of
> > > gender 2 with levels). I'd like to do this GLM in order to get
> some
> > > more insight into the between subject effects other than this
> > > synthetic measure provided in session/subject mode and COPE
> > estimates.
> > >
> >
> > Yes, that's reasonable but might be a lot of work depending
> > on how you do it. If you specify a group design, melodic
> > already does what you're proposing on the rank-1 time
> > courses. If you'd like to test every single (subject
> > specific) time course separately then the easiest thing to do
> > is to run fsl_glm
> >
> > fsl_glm -i t??.txt -d design.mat -c design.con -o output.txt
> >
> > will do what you want for every single column in t??.txt hth
> Christian
> >
> >
> >
> > > Best regards
> > > Mike
> > >
> > > Michal Kuniecki
> > > Jagiellonian University
> > > Institute of Psychology
> > > Dpt. of Psychophysiology
> > > ul. Ingardena 6, room 605
> > > 30-060 Krakow, Poland
> > >
> > >
> > >
> > >
> >
>
> <between_subjects_design_matrix.png>
|