 |
 |
In actual fact, the best thing to do would be to perform a fixed effects 2nd level analysis on each subject's connectivity data to calculate the difference between sessions for each subject. The cope images from such an analysis would represent the change in connectivity from session 1 to session 2, and could then be entered into a randomise group analysis. This would have the advantage of using the varcopes from the 1st level connectivity analysis. -Tom On Sat, May 12, 2012 at 7:16 AM, bettyann <[log in to unmask]> wrote: > Dear Lorena, > > I am hesitant to use randomize in my simple paired t-test design for these reasons: > > (1) The webpage describing 'randomize' specifically addresses repeated measures > http://www.fmrib.ox.ac.uk/fsl/randomise/index.html > Accounting for Repeated Measures > > Permutation tests do not easily accommodate correlated datasets (e.g., temporally smooth timeseries), as null-hypothesis exchangeability is essential. However, the case of "repeated measurements", or more than one measurement per subject in a multisubject analysis, can sometimes be accommodated. > > randomise allows the definition of exchangeability blocks, as specified by the group_labels option. If specfied, the program will only permute observations within block, i.e., only observations with the same group label will be exchanged. See the repeated measures example below for more detail. > > The 'repeated measures example below' is a 1-factor / 4-level ANOVA. I realize my simple paired t-test is the simplest form of an ANOVA but this example spooked me, particularly this statement: > This will ensure that permutations will only occur within subject, respecting the repeated measures structure of the data. > > (I know that statement is supposed to reassure and comfort me, but.) > And then all that stuff about design.grp and computing total permutations. So I just went old school. > > > (2) In my design, I am interested in the change between pre- and post-condition, ie, a repeated design. I can generate this difference by subtracting: ( pre - post ) or something like: > fslmaths pre.feat/zstat1 -sub post.feat/zstat1 diff/zstat1 > > And that's it. Old school. Of course, this subtraction method only works bc my repeated measures is a simple paired t-test. This subtraction method doesn't work for multiple-level ANOVA. > > But this does bring me back to an earlier question / concern I had: > * Am I throwing away useful (necessary?) statistical information by inputing manually-subtracted zstat1 volumes? > > If I use the method you suggest, Lorena, does randomise take advantage of (need?) the cope1 (beta weights) and varcope1 (variance) volumes and does smarter statistical analysis? > > I would appreciate clarification / correction on my thoughts here. > > Maybe if I have time (ie, get SGE parallel stuff ported over to PBS scheduler), I could try both methods and see what happens. That's inviting trouble, isn't it? > > Thanks, > * ba > > > > >> Hello Tom and Bettyann, >> >> Following the topic described below, Why not to feed randomize with the Zstats from the fist level analysis of each subjects for each condition, specifying a paired t-test in the design matrix used in randomize? And e would specify in the design contrast >> the contrast: >> pre> post >> Post > pre, >> pre_positive correlation mean and >> post_positive correlation mean >> >> in order to mask the positive correlation that we would find using the contrasts of pre-condition and post-condition comparison (pre> post and Post > pre) >> >> I would appreciate if you can help me to better understand if what I am describing it could be a reasonable analysis. >> >> Thank you so much >> >> Lorena >> >> >> >>> Subject: Re: Round 2: Change in functional connectivity between pre- and post-conditions using randomise >>> From: Tom Johnstone <[log in to unmask]> >>> Reply-To: >>> >>> FSL - FMRIB's Software Library <[log in to unmask]> >>> Date: Fri, 11 May 2012 16:40:00 +0200 >>> >>> Reply >>> Agree with Dianne - a more informative set of questions would be hard to find! >>> >>> I think the subtracting z-score maps should work fine - as you say, >>> they are a measure of connectivity, so subtraction will yield change >>> in connectivity maps, which can then feed straight into randomise. >>> >>> -Tom >>> Centre for Integrative Neuroscience & Neurodynamics >>> School of Psychology and CLS >>> University of Reading >>> Ph. +44 (0)118 378 7530 >>> [log in to unmask] >>> http://www.personal.reading.ac.uk/~sxs07itj/index.html >>> >> >> >> On Thu, May 10, 2012 at 9:35 PM, Dianne Patterson <[log in to unmask]> wrote: >>> Bettyann, >>> I can't answer your questions, but I love them! >>> Absolutely clearheaded and useful to read. >>> Thankyou for taking the time to write them. >>> >>> -Dianne >>> >>> >>> On Thu, May 10, 2012 at 11:51 AM, bettyann <[log in to unmask]> wrote: >>>> >>>> Aaaand, we're back! Thanks (in advanced) for your thoughts on Round 1: >>>> z-score v beta weight as a measure of functional connectivity >>>> >>>> Here we are with ... >>>> Round 2 for US$500 / 310 GBP / 386 EUR: >>>> >>>> Assessing change in functional connectivity between pre- and >>>> post-conditions using randomise. >>>> >>>> You may recall from Round 1 that I have a better intuitive feeling for >>>> using z-score as a measure of functional connectivity. I don't yet >>>> understand the advantages of using beta weights instead. >>>> >>>> Now I would like to assess the *change* in functional connectivity between >>>> a pre- and post-condition. >>>> >>>> I have set up a paired t-test design where the lower-level FEAT >>>> directories are the results from GLM analysis that produced the functional >>>> connectivity maps to my seed region's time course, two per subject (one from >>>> the pre-condition; the other from the post-condition). >>>> >>>> 'Ah,' you ask, 'but what are these input functional connectivity maps?' I >>>> ask the same thing. Am I correct in thinking that both the cope's (beta >>>> weights) and varcope's (variance) will be combined in some statistically >>>> sound way to give me a measure of change in functional connectivity (since I >>>> am using a repeated measures / paired t-test design where the inputs >>>> themselves are functional connectivity maps). >>>> >>>> The result of this paired t-test produces z-scores, beta weights (copes) >>>> and variances (varcopes). I won't repeat my question from Round 1 here. >>>> No, instead I want to ask about using randomise for inference analysis. >>>> >>>> I am unsure of how best to use randomise in a repeated measures fashion. >>>> I can deal with the repeated measures part by subtracting pre-condition >>>> from post-condition resulting in a difference map, one per subject. >>>> >>>> Given my current understanding that z-scores reflect correlation, I am >>>> leaning toward subtracting z-score (zstat1) volumes to create a >>>> zstat-difference, one per subject. I would then feed these zstat-difference >>>> volumes into randomise. are z-score differences meaningful? I tell myself >>>> the differences are meaningful because these z-scores do reflect >>>> correlation. (But I tell myself a lot of things.) >>>> >>>> Again I am concerned that I'm not comprehending the strength and beauty of >>>> beta weights. Maybe I should be using the difference in beta weights. But >>>> what about noise ... some of these measurements are noisy, which is >>>> uncontrolled (?) in the betas. >>>> >>>> At this point, I am worried that I've become biased about z-scores. And >>>> that I'm missing something important about beta weights. Add into the mix >>>> the idea of 'difference' and 'change in functional connectivity'. >>>> >>>> Thoughts? Comments? >>>> >>>> Thanks for playing, >>>> Thanks for all, >>>> * ba >>>> >>> >>> >>> >>> -- >>> Dianne Patterson, Ph.D. >>> Research Scientist >>> [log in to unmask] >>> University of Arizona >>> Speech and Hearing Science 314 >>> 1131 E 2nd Street, Building #71 >>> (Just East of Harvill) >>> 621-9877 >>> ============== >>> "I used to think that the brain was the most wonderful organ in my body. >>> Then I realized who was telling me this." >>> - Emo Phillips >>> ============== >>> If you write me (expecting an answer) and I don't respond within a >>> day, then the email may have been lost. >>> You can always write me at [log in to unmask] >>> ============== >> >>