| Author |
 Topic  |
|
|
Dichter
BIAC Faculty
190 Posts |
 Posted - Feb 21 2008 : 09:57:45 AM
|
We are trying to run a 3rd-level model with FLAME 1+2 that reads in 58 first-level copes, and has 8 contrasts and 4 ftests. This is taking days to run and seems to invariably crash (without errors) before completing. As a test, the same model with only 8 first-level copes completed without errors but took ~14 hours.
Any tips for helping these larger models finish/converge?
Thanks. |
Gabriel S. Dichter, PhD
UNC Departments of Psychiatry & Psychology
http://www.can.unc.edu/ |
|
|
dvsmith
Advanced Member
USA
218 Posts |
Posted - Feb 21 2008 : 10:11:52 AM
|
Try using just FLAME 1. The results will be very similar to FLAME 1+2, and it will take a lot less time. I've found FLAME 1 only to be less susceptible to crashing without error.
David
|
 |
|
|
dvsmith
Advanced Member
USA
218 Posts |
|
|
petty
BIAC Staff
USA
453 Posts |
Posted - Mar 06 2008 : 11:17:02 AM
|
| what i took away from the FSL class is that FLAME 1 is run regardless ... but then the model goes back and runs FLAME2 on voxels that are near threshold. so with 1+2 your significant region could be slightly larger due to the additional test, but it could look exactly the same. |
|
|
|
Dichter
BIAC Faculty
190 Posts |
Posted - Mar 06 2008 : 11:20:36 AM
|
| Thanks for the update. The key issue is whether FLAME 1 may cause Type 1 or Type 2 errors, relative to FLAME 1+2. Chris's post suggests that FLAME 1 may (but may not) cause Type 2 errors (i.e., identify active voxels as not active). That is less worrisome than identifying not active voxels as active (in most contexts). |
Gabriel S. Dichter, PhD
UNC Departments of Psychiatry & Psychology
http://www.can.unc.edu/ |
|
|
|
morey
BIAC Faculty
USA
25 Posts |
Posted - Mar 06 2008 : 11:29:07 AM
|
| this is correct that Flame 1 will lead to more Type 2 errors. but this does not mean tht Flame 1+2 will lead to more Type 1 errors. I think the Flame 2 part does an extra Monte Carlo simulation to give potentially more significant voxels, i.e. those that were incorrectly identified as inactive from Flame 1 alone. |
Rajendra Morey MD |
|
|
|
dvsmith
Advanced Member
USA
218 Posts |
Posted - Mar 06 2008 : 2:05:39 PM
|
Hi,
Yeah, most of this information is listed on the FSL website and in the FEAT "balloon help". I've compared FLAME1 and FLAME1+2 several times across a few different studies and scanners and there is often little difference; you might get a little more activation with 1+2, but you might not. The additional MCMC stage in FLAME1+2 can take several hours or days, but it has the potential to identify a few voxels that FLAME1 may have "missed" as their website states and as Chris learned at the course. I think your susceptibility to false positives should be about the same for both.
quote:
from http://www.fmrib.ox.ac.uk/fsl/feat5/detail.html#higher
The higher-level estimation method in FEAT (FLAME) uses the above modelling theory and estimates the higher-level parameter estimates and ME variance using sophisticated estimation techniques. First, the higher-level model is fit using a fast approximation to the final estimation ("FLAME stage 1"). Then, all voxels which are close to threshold (according to the selected contrasts and thresholds) are processed with a much more sophisticated estimation process involving implicit estimation of the ME variance, using MH MCMC (Metropolis-Hastings Markov Chain Monte Carlo sampling) to give the distribution for higher-level contrasts of parameter estimates, to which a general t-distribution is then fit. Hypothesis testing can then be carried out on the fitted t-distribution to give inference based upon the best implicit estimation of the ME variance.
My recommendation would be to run FLAME1 only since FLAME1+2 never converged for you (it likely got stuck in the iterative MCMC stage). If you get activation that's below the default cluster threshold in FSL, try using FDR (http://www.fmrib.ox.ac.uk/fsl/randomise/fdr.html) or permutation testing to see if those thresholding methods yield more sensible results.
Cheers,
David
|
|
|
| |
Topic |
|
Large 3rd-level model with multiple F-tests
