Show all posts by user
Dear AFNI users-
We are very pleased to announce that the new AFNI Message Board framework is up! Please join us at:
https://discuss.afni.nimh.nih.gov
Existing user accounts have been migrated, so returning users can login by requesting a password reset. New users can create accounts, as well, through a standard account creation process. Please note that these setup emails might initially go to spam folders (esp. for NIH users!), so please check those locations in the beginning.
The current Message Board discussion threads have been migrated to the new framework. The current Message Board will remain visible, but read-only, for a little while.
Sincerely,
AFNI HQ
History of AFNI updates
Results 2251 - 2280 of 2632
Emilie,
I can think of two simple approaches to modeling the RT effect:
1) Making a crude assumption of linearity between the BOLD response and RT. This is typically done in amplitude (or parametric) modulation. This method would work across multiple trials (e.g., -stim_times_AM2). Since you're now trying to obtain the response estimate at each trial, the modulation approach seems unfe
by
gang
-
AFNI Message Board
If the F-statistics for the interactions are the only stuff you're looking for, it would be fine. The underpowering issue would occur if you want the t-statistics for post hoc tests.
by
gang
-
AFNI Message Board
Hmm... previously I didn't really believe the length of the factor/level names would be a serious issue! Now I learned something from your experience. Still the number of subjects in your case is so close to a troubling situation: with one less subject you have an indeterministic (a mathematically insolvable) system. Please do let me know how it pans out.
by
gang
-
AFNI Message Board
Could you provide the big picture of the whole analysis? For instance, is it an event-related experiment? What are you planning to do with those trial-by-trial response estimates? What's the purpose of controlling for RT variability? Any follow up on group analysis?
by
gang
-
AFNI Message Board
> Is it possible to combine these two techniques to first control for RT across all trials, and then extract the per-trial estimates?
Emilie, the question is very confusing... When using -stim_times_IM, you basically estimate the response at each trial. Now if you want to further control for RT across trials, do you mean that you want to estimate the response at each trial while pretending
by
gang
-
AFNI Message Board
Isaac,
Since Diff0 does not provide much information about the BOLD response, the first thing you can try is to remove it from the analysis, and see if that fixes the problem. The next thing is to remove the last one Diff8 by the same token.
If the above approach does not work, another method is to either use the contrast between the two levels of Condition or Run (not both) as input, and g
by
gang
-
AFNI Message Board
Hi Isaac,
You really have done a lot in tracking down the problem!
I suspect that the issue is related to the 5th one listed as the possible reasons for model failure. The cbind() used in lm() is meant to run a multivariate regression analysis; that is, as you guessed, cbind(y, z) ~ x solves y ~ x and z ~ x simultaneously in one (multivariate) model.
Before I could offer solutions, cou
by
gang
-
AFNI Message Board
Camille,
Thanks for providing the details of your experiment, which really helps in clarifying the confusions.
> I want to be able to find frequency selective voxels to use in further analyses.
> I thought the best way to do that would be to do an ANOVA to figure out which
> voxels respond selectively to the different frequencies using an F test.
3dANOVA* programs are not the
by
gang
-
AFNI Message Board
> It turns out that the line following -dataTable should be formatted with each of the
> items (i.e., Subj, the factors, and the Inputfiles) separated by spaces and *not* tabs.
That's something new to me too! Thanks for sharing this.
> running the model on the whole data set (rather the masked set) seemed ok;
> I guess the mask should be 1's and 0's, right?
by
gang
-
AFNI Message Board
Could you elaborate a little bit about your experiment? What are those six levels? It looks like you need to run an individual subject analysis with 3dDeconvolve, for example, not a group analysis.
by
gang
-
AFNI Message Board
Sara,
I just realize that the quantitative variable 'ability' does not vary across the conditions. So it does not make sense to run a model like "cond*ability".
Basically it would be much simpler and faster to just used 3dttest++ (or 3dMEMA if you have the associated t-values available).
by
gang
-
AFNI Message Board
Hold on! 3dANOVA* are group analysis programs with input data from multiple subjects. You only have data from one subject? And what kind of data is file mean+orig? Time series?
by
gang
-
AFNI Message Board
There are a few problems here:
1) The following message on your terminal indicates that the script was not formatted properly. For example, those subject labels and factor (task) levels were not identified properly.
***** Summary information of data structure *****
3 subjects : s1/aud/beta_00750s1/aud/beta_0002.nii s1/not/beta_00250s1/not/beta_0001.nii s2/not/beta_003000s2/not/beta_0003.
by
gang
-
AFNI Message Board
> I see Time in your model but not in your data table. Could that be the problem?
That is indeed a problem. The model should be something like
-model "cond*ability" \
However, I'm still not so sure about the exact source for the error message. Hopefully Rick's suggestion of changing the file names from *.img to *.hdr would fix it.
> It took me an embarrassin
by
gang
-
AFNI Message Board
> Would 3dANOVA2 -type 3 still be my best bet?
Yes, 3dANOVA2 -type 3, not 3dANOVA, is the correct approach.
by
gang
-
AFNI Message Board
Camille, what are the factors in your situation? One or multiple groups?
If you have only one within-subject factor, use 3dANOVA2 -type 3...
by
gang
-
AFNI Message Board
One thing I just noticed is that 'subj' should be 'Subj' in the following line in your script:
subj cond ability InputFile \
Maybe I should make some labels case-insensitive in the future?
by
gang
-
AFNI Message Board
> I can ignore the N/Csim error and that minimum cluster size and the NN in the clusterize option is all I need?
Yes, that's right.
by
gang
-
AFNI Message Board
You need some basic understanding about t-test and the reason you want to run such a test.
One possible situation it worked for you before would be something like this:
3dttest++ -setA JBrestBA6corr+orig JMrestBA6corr+orig
But the problem is, how confident are you when you make a general statement with only two subjects?
by
gang
-
AFNI Message Board
3dttest and 3dttest++ run paired, one- or two-sample test at group level. That is, it makes generalization from multiple subjects (e.g., more than 10) with a Gaussian distribution assumption about the data. So it can't used to simply compare two subjects, and that's why you got that error message.
by
gang
-
AFNI Message Board
Oya,
If group analysis is your ultimate goal, you don't need to correct for multiple comparisons at individual level. In other words, 3dClustSim for each subject is unnecessary. Instead you run 3dClustSim with the average smoothness across subjects, and apply the minimum cluster sizes to the group analysis (e.g. 3dMVM) results. Unlikely 3dDeconvolve, 3dMVM does not store the clusterizatio
by
gang
-
AFNI Message Board
The Monte Carlo simulation approach is based on the assumption of the volel-wise noise at the spatially contiguous voxels in the brain. I'm not sure what specific data-driven analysis you're referring to. If it boils down to one statistical map in the end, the simulation method would still seem to be valid.
by
gang
-
AFNI Message Board
> I don't know how to look into the sub-bricks and check the actual r-value distribution
"r-value distribution" across the brain? If so, no AFNI tools are directly available for that purpose, but you can read those voxel-wise R^2 values into another program such as R or Matlab, and then do whatever you want.
by
gang
-
AFNI Message Board
Qing,
With option -rout in 3dDeconvolve, you get at each voxel the overall R^2 as well as partial and glt R^2.
by
gang
-
AFNI Message Board
>groups are 23,53,27 for the anova
The degrees of freedom are calculated based on the pooled variance of all three groups:
23+53+27 minus 3 (three group averages) = 100
> 23,27 for the ttest.
This should render 23+27 minus 2 (two group averages) = 48 degrees of freedom. I'm not so sure why you got 38. Seems strange to me... Could you provide the full 3dttest++ script?
by
gang
-
AFNI Message Board
How many subjects in each of the three groups? And how did you analyze your data with 3dttest++: comparing the three groups in one 3dttest++ script, or comparing one pair of two groups at a time?
by
gang
-
AFNI Message Board
> Is it the most common way to threshold the F test over the same F test?
What do you mean by this? Are you referring to Threshold vs. Overlay? If so, yes you can put the same F-statistic for both.
by
gang
-
AFNI Message Board
Oya,
You could also use 3dMVM for your group analysis. Regarding the clusters associated with FWE correction, you do need to run 3dClustSim to obtain the cluster thresholds. I will think about the possibility of adding the clustering feature for programs such ad 3dMEMA, 3dMVM and 3dLME.
by
gang
-
AFNI Message Board
> is there a way to extract all the R-square values into a certain data format so
> I can do some descriptive stats on the R-square values?
There are two types of r-value 3dDeconvolve provides: the overall R^2 and the partial R^2 associated with each regressor. Which one are you referring to? And what kind of descriptive statistics are you interested?
by
gang
-
AFNI Message Board
> if I want to obtain the correlation coefficient for my seed time course regressor.
> I need to take the square root of the R^2.
In addition to the square root part, you also need to assign the proper sign, which can be obtained from the corresponding beta or t-value for the seed time series. For example:
3dcalc -a subj01+tlrc'' -b subj01+tlrc'' -expr 'ispos
by
gang
-
AFNI Message Board