Hi Doug,
I sidestepped the registration issue because I am performing the analysis on 5 ROIs per subject (5x1x1 volumes), which are aligned by design. I'm doing this using the lab's semi-manual NHP processing stream, so usual assumptions you have about which FS-fast
steps have been performed may not hold.
Just to clarify things, I'll try to outline where I stand more completely.
I have a block design study with 3 subjects, and ~30 runs per subject. Each run contains every condition occurs twice in each run. These runs have been collected over 2-3 days for each subject. Within each subject, I aligned all of the different sessions
and combined them into a single metasession. Within these 3 metasessions, I defined ROIs, and created new ROI volumes by averaging the time courses within each ROI. From here I have taken 2 approaches.
Approach A: Combine all of the ROI volumes (~30 runs X 3 subjects), which are pre-aligned with each other, into a single bold folder. Run selxavg3-sess on this data using my contrasts of interest.
Approach B: Keep the ROI volumes from each subject in their own bold folder, and run selxavg3-sess 3 times (once for each subject). Then I used mri_glmfit to attempt fixed effects meta-analysis. Here there's another split.
Approach B.1: I ran the contrasts for each individual subject, then used mri_glmfit --osgm on the ces and cesvar volumes of each contrast to test for significance of the contrast over the group.
This does not work for F-tests.
Approach B.2: I ran an omnibus contrast on each subject separately, and use the ces and cesvar for each condition in each subject. I assembled this for input to mri_glmfit. The fsgd file looks
roughly like:
GroupDescriptorFile 1
Class Subj1
Class Subj2
Class Subj3
Variables
Condition1 Condition2
Condition3 Condition4
Input
Subj1-Cond1 Subj1
1 0
0
0
Input
Subj1-Cond2
Subj1
0
1
0
0
Input
Subj1-Cond3
Subj1
0
0
1
0
Input
Subj1-Cond4
Subj1
0
0
0
1
Input
Subj1-Cond00
Subj1
0
0
0
0
Input
Subj2-Cond1
Subj2 1
0
0
0
Input
Subj2-Cond2
Subj2
0
1
0
0
Input
Subj2-Cond3
Subj2
0
0
1
0
Input
Subj2-Cond4
Subj2
0
0
0
1
Input
Subj2-Cond00
Subj2
0
0
0
0
Input
Subj3-Cond1
Subj3 1
0
0
0
Input
Subj3-Cond2
Subj3
0
1
0
0
Input
Subj3-Cond3
Subj3
0
0
1
0
Input
Subj3-Cond4
Subj3
0
0
0
1
Input
Subj3-Cond00
Subj3
0
0
0
0
Approach B.2 (continued): I then define contrasts comparing conditions (Condition1 vs Condition2, etc.) and run these contrasts using mri_glmfit --C. With this method, I have no problem defining
F contrasts (what they are actually testing is another question).
My actual experiment is a 6 condition, 2x3 factorial design. The 2 contrasts of interest that I need an F test for are one of the main effects (these contrasts are shown without the 3 leading columns of zeros for the different subjects):
1
1 -1
-1 0
0
1
1 0
0 -1
-1
and the interaction effect:
1
-1 -1
1 0
0
1
-1 0
0 -1
1
I switched to Approach B because I want to scale the condition betas per-subject before running the analysis across all subjects (this is to correct for potentially different levels of a contrast agent). I was worried that if I scaled the raw time courses
that are the input to Approach A, that the normalization step built into selxavg3-sess would interact with my manual normalization in ways that would be difficult to figure out. With approach B, I can directly scale the contrast ces and cesvars (B.1) or betas
(B.2) for each subject.
OK, given all of that background information, what is the best way to approach this? Should I:
a) Try to scale the time courses that go into Approach A
b) Work with Approach B.1, and live without knowing the significance of my main/interaction effects (not a nice thought)
c) Work with Approach B.2
d) Take some other tack
If I take Approach B.2, would any violation within-subject independence decrease my power, or artificially inflate my significance? Some loss of power could be acceptable.
Thanks for reading all of the way through this,
Clark
Clark, it sounds like what you did (ie, combine different subjects into one folder) would be tricky to analyze correctly because FSFAST will try to register them to the same subject. I could think of a way to do it, but it would be complicated. What is the
test that you want to do? An F test is usually something like a logical "or" between two conditions. The RM ANOVA won't do this, but a simple ANOVA would. The problem is that the DOF would be off because it would assume that each data point came from a different
subject, and it would not take into account the covariance in the measurements from a single subject. Another approach would be to do separate tests for each condition at the group level, then do a type of conjunction analysis to implement the OR. Let me know
if you want to pursue that approach.
doug
On 11/26/13 4:02 PM, Clark Fisher wrote:
Thanks again,
I have 3 subjects, and want to tests for effects using data collected across all 3 subjects. The multiple sessions for each subject have already been concatenated into a single BOLD folder (3 folders total: 1 for each subject). I have actually already performed
the fixed effects analysis by concatenating all 3 subjects into a single bold folder and re-running the first level analysis. While that works, I now want to try to normalize the data for each subject to minimize the variability introduced by the contrast
agent we use, and that would be easier to do by normalizing the outputs of the first level analysis. If I wanted to normalize per-subject and then combine them into a first level analysis, what would be the best way to do this?
My naive assumption would have been that using glmfit for this should be very similar to the repeated measures ANOVA example that is on the wiki (http://ftp.nmr.mgh.harvard.edu/fswiki/RepeatedMeasuresAnova).
Rather than time points for each subject, I will have the beta from each condition for each subject. In both cases (multiple time points from a single subject and multiple conditions from a single subject/set of runs), the within-subject measurements are
not independent. How is my situation different?
Cheers,
Clark
On Nov 26, 2013, at 1:41 PM, Douglas N Greve <greve@nmr.mgh.harvard.edu> wrote:
Are the different input to glmfit different runs within a session or across sessions? Either way, for within subject analysis, can you put all of the runs into a single "bold" folder? Then it automatically does the FFX analysis across
runs, which can include an F-test. The problem is that glmfit expects all the inputs to be independent, which conditions within a run will not be.
doug
On 11/25/2013 01:35 PM, Clark Fisher wrote:
Thanks Doug,
I had forgotten that my roadblock was actually in isxconcat_sess, not mri_glmfit. This worked out great.
Following this up, how would you recommend calculating the fixed-effects significance of a complex (within subject) contrast with mri_glmfit? For simple contrasts, I have been calculating the contrast at the single subject level, then performing an osgm fixed-effects
analysis on the output ces and cesvar volumes for each subject. However, I have some 2D F-contrasts that don't output ces volumes (naturally).
One possibility would be to concatenate the ces and cesvar from the omnibus contrasts of each subject, set up an appropriate fsgd (as in
http://ftp.nmr.mgh.harvard.edu/fswiki/RepeatedMeasuresAnova) and then define the contrasts at the level of mri_glmfit with .mtx files. Would that be the way to go?
Thanks,
Clark
--
Douglas N. Greve, Ph.D.
MGH-NMR Center
greve@nmr.mgh.harvard.edu
Phone Number: 617-724-2358
Fax: 617-726-7422
Bugs: surfer.nmr.mgh.harvard.edu/fswiki/BugReporting
FileDrop: https://gate.nmr.mgh.harvard.edu/filedrop2
www.nmr.mgh.harvard.edu/facility/filedrop/index.html
Outgoing: ftp://surfer.nmr.mgh.harvard.edu/transfer/outgoing/flat/greve/
The information in this e-mail is intended only for the person to whom it is
addressed. If you believe this e-mail was sent to you in error and the e-mail
contains patient information, please contact the Partners Compliance HelpLine at
http://www.partners.org/complianceline . If the e-mail was sent to you in error
but does not contain patient information, please contact the sender and properly
dispose of the e-mail.