Thanks Brad, indeed that makes sense, and I particularly like your comment 
about BET....

Sam, it does sound like you have a good grasp of the norms in 
thresholding. You're right that strictly one should scale p-thresholds 
according to numbers of contrasts tested, though you'll find that people 
rarely do that in the literature....

I think you're more likely to get bitten though by using fixed effects - 
do you really have a good excuse not to use mixed effects? Surely the 
normal arguments for ME hold here. In this case you would want a 3-stage 
analysis like in the manual.

Cheers, Steve.



On Fri, 6 Jan 2006, Bradley Buchsbaum wrote:

> 
> Hi Samuel,
> 
> Choosing thresholds for fMRI analyses is something of an art.  More and 
> more often an experiment is conducted for which there is a lot of 
> previous research and for which the "neural correlates" of a given task 
> paradigm has been pretty well mapped out.  And so the more you know 
> about the neurobiology going in, the greater justification for using a 
> "liberal threshold" (for instance, one often sees P < 0.001 -- 
> uncorrected for multiple comparisons -- these days.  The reason for this 
> is because, even though one generally performs statistical comparions 
> for all voxels in the brain, it's generally the case in an advanced 
> research program, that one has an idea of where one is looking ahead of 
> time.  But ESP research is something of a new frontier -- and here you 
> have no a priori hypotheses -- so that it's essential that you use a 
> "conservative" threshold.  How conservative? I would say more 
> conservative than the generally approved "conservative" threshold.  The 
> thing is, before claiming you've found "the neural correlate of ESP" you 
> want to be pretty darn sure it's not a statistical artifact.
> 
> But wait, if there is a neural correlate of ESP can it then be ESP?  A 
> natural explanation for ESP would seem to be incompatible with the whole 
> idea of "extra-sensory perception".  By the same token, a true ESP 
> believer would say that a null result would actually support the 
> existence of ESP!
> Seems to me it's a catch-22.
> 
> So perhaps you should be focusing most intently on the voxels *outside* 
> the brain -- in which case you must be sure not to run BET on the fMRI 
> images.
> 
> good luck,
> 
> Brad Buchsbaum
> 
> 
> 
> 
> 
> 
> 
> 
> 
> 
> Samuel Moulton wrote:
> 
> >hi FSL gurus.
> >
> >first of all, thank you for designing such an impressive piece of software and for maintaining this 
> >very helpful list.
> >
> >i'm the lone FSL user in my lab and this is my first fMRI study. therefore, i'd really appreciate your 
> >help to ensure that i haven't made any dumb mistakes when setting up my analysis. to make 
> >matters worse, part of the experiment happens to be on a particularly controversial topic: ESP. 
> >because of this, i need to be sure that my analysis adheres to the current standards out there, and 
> >that i am conservative (but not overly so). first let me briefly describe my experiment:
> >
> >16 participants completed in a very simple guessing task during 5 functional runs. participants 
> >were sequentially exposed to two photographs and then had to decide which photograph was 
> >randomly selected by the computer as the "target". after they made their choice, they were shown 
> >the target picture a second time as feedback. this trial sequence continued ad nauseam (but with 
> >different pictures for each trial). for half of the trials, the target picture was presented first 
> >(target1--->decoy-->target2), and for the other half the decoy was presented first (decoy---
> >  
> >
> >>target1--->target2). target assignment was also counterbalanced across participants such that 
> >>    
> >>
> >the pictures in the overall target and decoy sets were identical. by looking for differential 
> >activation associated with target vs. decoy exposures, we were testing for ESP. obviously i left off 
> >some details (like, for example, that participants had an identical twin or relative viewing the 
> >targets for each trial in a separate room), but that's the basic design.
> >
> >this link contains my first and second-level design files and matrices, as well as a histogram i used 
> >to settle on a high pass filter cutoff:
> >
> >http://www.courses.fas.harvard.edu/~psy970dn/temp/
> >
> >my goal in this analysis -- given the topic -- was to be completely uncontroversial. i would greatly 
> >appreciate if you could scan the files above to see if anything strikes you as even a little dodgy.
> >
> >i also have a couple questions. first, do you think 25s is too low a cutoff for high-pass filtering? i 
> >initially used 50s, but revised that downwards after reading some of the posts on this list. as far 
> >as i know, the frequencies in my model (see histrogram) are the only things i should use to select 
> >this parameter.
> >
> >secondly: are the standards of fixed-effects analysis different than the standards of mixed-effects 
> >analysis? because i'm not trying to generalize beyond my subject pool, i have a completely fixed-
> >effects analysis. this seems to be a rare situation in psychology. is there any reason to think that i 
> >should adopt more conservative thresholds (i.e. lower alpha-levels) for a fixed-effect analysis than 
> >a mixed-effects one?
> >
> >thirdly: there are two contrasts that would support claims of ESP: decoy > target1 and target 1 > 
> >decoy. does this mean i should halve my alpha values? right now i have a cluster significant in the 
> >decoy > target1 contrast at p = .04 and nothing in the target1 > decoy contrast, so the answer to 
> >this question unfortunately matters. i had no a priori hypothesis about directionality. i've never 
> >seen this type of bonferroni correction done in fMRI studies, but it seems necessary here.
> >
> >finally: this may be an impossible question, but what thresholding parameters are the least 
> >controversial? right now, i'm using cluster stats with a z-threshold of 3.1 and a p-threshold of 
> >0.05 (this is the closest to a standard that i've been able to find). as i mentioned, i do find a blob 
> >(91 voxels) of significant activation. this goes away with voxel-based thresholding (p = 0.05). it 
> >also disappears when my cluster z-threshold goes below 2.7 or above 3.1. i had no a priori 
> >hypothesis about the size or location of potential activation.
> >
> >ok, i think that's enough for now. i hope some of these questions are of general interest to the 
> >list.
> >
> >thanks
> >sam
> >  
> >
> 
> 

-- 
 Stephen M. Smith, Professor of Biomedical Engineering
 Associate Director, Oxford University FMRIB Centre

 FMRIB, John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK
 +44 (0) 1865 222726  (fax 222717)

 [log in to unmask]  http://www.fmrib.ox.ac.uk/~steve