 |
 |
Hi Matt, Looks fine. It means that age and sex are absorbing variance that otherwise would appear in the error term, which would reduce the value of the statistic. So, keep age and sex there as nuisance, exactly as you did, as they explain part of the observed variation in the TBSS data. All the best, Anderson Am 11.04.14 20:46, schrieb Walton, Matt: > Hey Anderson, > > I'm trying to run TBSS on the following set up to check if a single > groups continuous language measures correlate with FA, while > correcting for age, sex. > I'm running the following EV's and contrasts (also demeaning data with > -D, and running --T2) > > EV1. Language Scores > EV2.Age > Ev3. Gender. > EV1 EV2 Ev3 > 1 0 0 > -1 0 0 > When I run this set up I get significant results. However when I > remove the language scores and just run Age and Gender, or even just > age, I no longer get significant results. I feel like this is > backwards (shouldn't I be getting more significance when testing with > less corrections?) > > I'm wondering if I have set something up wrong in randomise. I put 0's > in the contrasts for both age and gender (maybe I should put a 1 > somewhere??). Also I noticed the orthogonlizations tab, which I have > never used. NOt sure if that has anything to do with it. I also have > not used any additional voxel dependant EV's. > > Thanks for your help. > > Matt > ------------------------------------------------------------------------ > *From:* FSL - FMRIB's Software Library [[log in to unmask]] on behalf > of Anderson M. Winkler [[log in to unmask]] > *Sent:* March-29-14 1:19 PM > *To:* [log in to unmask] > *Subject:* Re: [FSL] TBSS and Randomise > > Hi Matt, > > I'm afraid that the options in this case are somewhat limited with the > current version of randomise. If there is a language score for each > scan of the same subject, it's possible to use the option -e, defining > one exchangeability block per subject, but there will be an > unavoidable loss of power. > > If there is just one score for all scans of the same subject, one > could consider averaging the scans, but this creates problems with > exchangeability (the averages of various scans will have smaller > variance, rendering them non-exchangeable with the others). > > Another option would be to shuffle within block, then the blocks of > the same size as a whole. This isn't currently available in randomise, > though. > > The only solution that currently remains is to use just 1 scan per > subject I think. > > All the best, > > Anderson > > > Am 29/03/2014 17:52, schrieb Walton, Matt: >> Hey Anderson, >> >> Thanks for all your help with TBSS and randomise. Sorry it took me so >> long to get back, but I had one additional question. Here is a >> summary of what I am doing incase you deleted our previous thread. >> >> I have FA's for 80 scans from 50 participants and am looking to see >> if language scores correlate. I am running as a single group with the >> following EV's. >> EV1: Language score (continuous scores) >> EV2: Age >> EV3: Sex >> >> About half my participants have multiple scans (ranging from 2 scans >> - 6 scans). I was wondering if it would be possible to model these >> repeated measures in randomise, and how I would go about doing that. >> >> Thanks so much. >> >> Matt >> ------------------------------------------------------------------------ >> *From:* FSL - FMRIB's Software Library [[log in to unmask]] on behalf >> of Anderson M. Winkler [[log in to unmask]] >> *Sent:* March-11-14 4:12 PM >> *To:* [log in to unmask] >> *Subject:* Re: [FSL] TBSS and Randomise >> >> Hi Matt, >> >> The intercept is a column full of ones (just the number 1, or in >> fact, any other value, as long as it's constant). If the column with >> the intercept is included, there is no need to demean any of the >> other regressors. If the intercept is not included, then use the >> option -D in randomise, that will demean both the data and all the >> columns of the design. >> >> About coding sex: any value as long as it's the same for all the >> males, and any other value as long as it's the same for all the >> females, but different than the one used for the males. So, it can be >> 0 1, 1 0, +1 -1, 1 2, pi and sqrt(2), any will work... >> >> About group, yep, all subjects into a single group (I mean, the >> leftmost column in the Glm GUI -- that is all 1). >> >> All the best, >> >> Anderson >> >> >> Am 11.03.14 19:15, schrieb Walton, Matt: >>> Hey Anderson, >>> >>> Thanks for the insight, I do have the raw values for the language >>> scores. I have a few more questions in regards to your proposed >>> design matrix >>> >>> EV1: Language score (continuous scores) >>> EV2: Age >>> EV3: Sex >>> EV4: Intercept >>> >>> When I'm entering the language scores and age into the design >>> matrix, do they need to be demeaned (subtract the mean from the >>> actual value)? >>> >>> Is sex entered as either 0 (male) or 1 (female) or vice versa? >>> >>> Could you clarify a bit on what you mean by intercept for EV4? >>> >>> Would this be running under all subjects under a single group? >>> >>> Sorry for all the questions. I very much appreciate your help. >>> >>> Matthew, >>> >>> ------------------------------------------------------------------------ >>> *From:* FSL - FMRIB's Software Library [[log in to unmask]] on >>> behalf of Anderson M. Winkler [[log in to unmask]] >>> *Sent:* March-11-14 6:18 AM >>> *To:* [log in to unmask] >>> *Subject:* Re: [FSL] TBSS and Randomise >>> >>> Hi Matthew, >>> >>> If you have the actual scores not split into 3 discrete groups, but >>> the raw values, in a continuous scale, it would be far better: it's >>> more powerful and the whole investigation can be summarised into >>> just 2 contrasts. >>> >>> The design matrix would be something as: >>> EV1: Language score (continuous scores) >>> EV2: Age >>> EV3: Sex >>> EV4: Intercept >>> >>> And the contrasts would be: >>> 1 0 0 0 >>> -1 0 0 0 >>> >>> However, if that's not an option, and all what you have are the 3 >>> discrete language categories, then the design would be something as: >>> EV1: Low >>> EV2: Mid >>> EV3: High >>> EV4: Age >>> EV5: Sex >>> >>> And the contrasts would be: >>> [1 -1 0 0 0] F1 >>> [1 0 -1 0 0] F1 >>> >>> If F1 is significant, then look at each of the pairwise possible >>> differences (positive and negative) to see which groups are higher >>> or lower than others. >>> >>> There is no guarantee that the first analysis would give the same >>> results as the second, and I would only use the second if the first >>> were not possible for not having the original scores available, or >>> if the hypothesis were to contemplate something as an U-trajectory >>> (e.g., with the Mid group with highest or lowest FA of all three). >>> >>> All the best, >>> >>> Anderson >>> >>> >>> Am 11.03.14 05:40, schrieb Walton, Matt: >>>> Hello FSL team, >>>> >>>> I'm new to FSL and have some questions regarding tbss analysis and >>>> randomize. >>>> >>>> I have approximately 50 different subjects who underwent DTI >>>> scanning once. The subjects age ranges from 1 year to 6 years old, >>>> and are composed of both male and female subjects. I also have >>>> language scores for all these subjects. >>>> >>>> I'm interested in seeing if there is any correlation between >>>> language score and white matter FA. >>>> Could this be run as an ANOVA? 3 groups (high language scores, >>>> middle language scores, and low language scores), with 2 covariates >>>> (age and sex). >>>> >>>> If so, could you give me a little more specifics about setting this >>>> up in the GLM GUI. >>>> >>>> If not, what do you think the best randomise set up to analyze this >>>> data would be. >>>> >>>> Thanks for the Help >>>> Matthew W, B.Sc, >>> >> >