Dear experts,
I am working with a longitudinal set of DWI images which has a
complete within-subjects design.
For each subject, I have 3 conditions (C, V, and M) and 3 scans per
condition: S1 (pre), S2 (post1), S3 (post2).
I have analyzed the scan effect for each condition with a repeated
measures ANOVA (1 factor, 3 levels) to see if there's any difference
between scans within each condition. I am working with 3D data in a
voxel-based analysis style, not a skeleton/TBSS style. I'm analyzing
Mean Diffusivity maps.
When I look at the TFCE results I get some nice clusters for M
condition with p-FWE<0.05. I was expecting something similar for the V
condition but, unfortunately, no significant clusters came up.
However, when I look at the Fstat image I can see several clusters
that look promising and whose F-stat>6. These magnitudes are similar
to those in clusters that are surviving the TFCE p-val correction in
the M condition.
I understand that it is the null distribution of TFCE values the one
that is going to define which clusters are significant. Is it possible
that this distribution could be so different for clusters with similar
sizes, Fstat values and for the same subjects? I don't see any reason
why TFCE-distributions would be so different between conditions (same
subjects, same scanner, images acquired at a similar time of day,
conditions acquired in a randomized and counterbalanced order to avoid
any biases).
Is it just a matter of power or am I missing something? Are the
default values for E and H parameters ok for this analysis (-T
option)?
I have also tried other statistical toolboxes such as PALM and SwE
with similar results. I've discussed this issue with Tom Nichols in
the SwE toolbox forum and he suggested I should check each of the sets
(C, V, M) for outliers/problems.
I have checked the mean and standard deviation of MD maps per
condition (C, M, and V), and also per subject (within condition), to
see if I have any outliers or if conditions differ in some way. Mean
and SD were calculated from voxels within the same mask I'm using for
statistics (a brain mask). I didn't find any differences on the mean
and SD values between conditions that could explain the difference in
distribution and, hence, my results for M and (lack of results) for V
condition.
For example, I am attaching a figure showing the global mean MD
(within the mask) for all conditions for each subject. Regarding this,
Tom Nichols suggested that my global MD values seem to exhibit excess
kurtosis (heavy tails). He told me to try a rank inverse normal
transformation to make the data more normal since this could make the
linear model more sensitive. However, I tried this and the global
means barely changed. I saw no change in statistical results either.
When I do an analysis of the main effect over both conditions (M+V
average) I get some interesting clusters. Looking at the time-course
within those clusters shows that there are significant changes on both
M and V conditions (meaning, the M+V results on the voxel-based
analysis are not being driven only by an effect of scan session on M).
I got similar results when I did a conjunction analysis of V and M
conditions (Union-intersection test using PALM's NPC).
I am a bit puzzled by these results and I am trying to see if I might
be making a mistake on the statistical analysis or might be doing
something that might reduce the power of the statistical tests.
Any comments and/or suggestions will be very much appreciated.
Thanks!
Regards,
Florencia.
--
Ing. Florencia Jacobacci.
PhD student / Becaria doctoral
Instituto de Fisiología y Biofísica (IFIBIO) - Bernardo Houssay
Laboratorio de Fisiología de la Acción, Facultad de Medicina
Universidad de Buenos Aires
Paraguay 2155, C.A.B.A.(C1121ABG), Argentina
(+5411) 5285-3304
http://www.physiologyofactionlab.info/en/about-the-lab/
https://ar.linkedin.com/in/florenciajacobacci/en
########################################################################
To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1
|
