Dear Yasuhiro,
| I have a question concerning the statistics after the fMRI random
| effect kit. Now I have 16 EPI series from 4 subjects (2 subjects have
| 3 runs and the other 2 have 5 runs: each run contains 2 conditions (rest
| and act).). I have averaged each condition within each EPI run using
| random effect kit, then I 've got 16 mean images each for rest and act
| conditions. Thereafter I moved to PET routine, and performed 2
| different statistical processes.
|
| First, I selected "multisubject, with replication". Then,
| subject 1; 3 act, 3 rest,
| subject 2; 3 act, 3 rest,
| subject 3; 5 act, 5 rest,
| subject 4; 5 act, 5 rest,
| No global normalization,
| adjusted CBF; 100
| contrast; 1 for act, -1 for rest.
|
| In this analysis, the parameters were "2 conditions+ 0 covariates+ 4
| block+ 0 confound, =6 parameters, having 5 degree of freedom, giving 27
| residual df". And the detected foci of activation were minimal.
|
| Second, I selected "multisubject, different conditions", and treated each
| EPI run as different subjects.
| subject 1; 1 act, 1 rest (first run from Sub. 1)
| subject 2; 1 act, 1 rest (second run from Sub. 1)
| ---
| subject 16; 1 act, 1 rest (5th run from Sub. 4)
| No global norm.
| adjusted CBF; 100
| contrast; 1 for act, -1 for rest.
|
| In this case, I got "2 conditions+ 0 covariates+ 16 block+ 0 confound,
| = 18 parameters, having 17 degree of freedom, giving 15 residual df".
| The resulted activation seems to be representative of the activation of
| each subjects.
|
| My questions are;
| 1) Why does this difference occur?
| 2) Which method is appropriate for the statistical analysis after the
| random effect kit?
The analyses are different because the models are different! In
particular, the variances are different, such that the scope of
inference is different.
In the first model you are assessing the average (across subjects)
difference between the average (across sessions (runs) within subject)
rest and activation conditions. Here the residual variance of the model
is the within-subjects within condition between sessions variance.
Thus, this model is not a random effects analysis suitable for
population inference. Inference pertains to the average effect for
these subjects (generalising across sessions).
In the second model you are assessing the mean (across subjects and
sessions) activation from each run. The residual variance in this model
is a mixture of within-subject between-sessions variance and
between-subjects variance, because you've mixed repeated sessions on
the same subject with different subjects. Unless the variance
between-sessions within-subject variance is equal to the betwen-subject
variance, which is highly unlikely, then this model is wrong, since the
residual variance isn't constant across observations.
That the second model gives more than the first model (with more
degrees of freedom) would indicate that there is substantial
within-subject between session variance (i.e. session by condition
interactions).
An appropriate random effects analysis would be to average the mean
rest and activation images from each session within each subject to
produce a single image per condition per subject, and assess these
using the PET Multi-subject: different conditions design, with no
global normalisation and no grand mean scaling. This will leave you
with only three degrees of freedom, which is a bit impractical, but
merely reflects the fact that you can't say much about the population
of all living, dead and to be born "normal controls" from a sample of
just four.
The key concepts for successful random effects analyses using the
random effects kit are:
(1) For inference to the population, it's the sampling of
subjects that matters most, not the number of scans/sessions
per subject.
(2) For valid population inference using the random effects
kit, you cannot have more than one scan per subject per
condition. (Strictly speaking you should only enter two
conditions too.)
Hope this helps,
-andrew
+- Dr Andrew Holmes [log in to unmask]
| -___ __ __ Wellcome Department of Cognitive Neurology - |
| ( _)( )( ) Functional Imaging Laboratory, Stats & |
| ) _) )( )(__ 12 Queen Square, Systems |
| (_) (__)(____) London. WC1N 3BG. England, UK |
+---------------------------------------http://www.fil.ion.ucl.ac.uk/-+
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
