 |
 |
Dear Anderson, You are right, as always. After reading some interesting papers and some links, I have understood what you were trying to tell me and I have to say sorry for posting such question regarding covariates. Firstly, because I misunderstood the meaning of ANCOVA and the implicit change that has to add a covariate regarding your objective and hypothesis, and Secondlly, because as you say, adding or not adding covariates depends on a) my objective and hypothesis, and b) the literature regarding the area I am working on...what is something from which I have to take my own decisions. Thank you all the team for your patience and your always helpful answers. Rosalia. 2016-01-04 11:14 GMT+01:00 Rosalia Dacosta Aguayo <[log in to unmask]>: > Hi Anderson, > > Thank you a lot for your answer. Yes, too much variables. What I am doing > is test first if there are interactions between my different variables of > interest regarding the group....this just to consider if it is reasonable > to test for independent correlations per group...To see if Age, TIV and > other variables are interfering....I will test interactions too with > independence of the literature about VBM covariates...I will take a > decision depending on the characteristics of my sample...That is, I think, > what you tried to tell me regarding the two papers about age, TIV and > gender (yes or not) adding covariates in a VBM analysis....all depends on > my sample...see what happens. > > Regarding PCA analysis, maybe I did not explained myself well...this > analysis was done with SPSS software, not with neuroimaging...and yes, you > can get independent components...and their respective variance...see the > file attached. > > Thank you :-) > > Rosalia. > > > 2016-01-04 10:12 GMT+01:00 Anderson M. Winkler <[log in to unmask]>: > >> Hi Rosalia, >> >> Yes, there are lots of EVs here, all of them with implicit interactions >> with group, which doesn't seem reasonable. It would be better to be more >> parsimonious, particularly given the small sample size. >> >> I don't know how you found percentages that one set of variables explains >> of another with PCA, but PCA can be useful in that if various of these >> nuisance variables are highly correlated with each other (e.g., metabolic >> syndrome), these can be replaced by the first few PCs only. >> >> This is generic advice. You are the researcher and it's up to you to >> decide how to proceed in your study given the literature in your area. >> >> All the best, >> >> Anderson >> >> >> On 2 January 2016 at 13:07, Rosalia Dacosta Aguayo <[log in to unmask]> >> wrote: >> >>> Dear Anderson, >>> >>> Thank you a lot for your helping. >>> >>> I tried to run PCA with my data using SPSS and I found that in my >>> sample, Hypertension and Alcohol consumption explained 68% of the variance >>> for my whole sample. Those variables are related withr my variable of >>> interest (blood biomarker). This is the only reason why I decided, finally, >>> to introduce those variables as covariates of no intetest with independence >>> regarding their influence in GM. I think that this point of view should be >>> justifiable...but your expertise surely will make me think about >>> this...because I still think I am correct, statistically speaking...but >>> please, correct me if I am wrong. >>> >>> By the other hand, I think that the big problem is in what I want to >>> test and if my design is really testing that question in a suitable manner. >>> The fact is that, reducing the number o covariates after PCA analysis, >>> yield me to some interesting results...but I think my design is not good. >>> >>> I just want to test positive and negative correlations between one blood >>> biomarker and the whole brain of my sample. One of the errors David Smith >>> kindly commented to me after viewing the design is that I was testing only >>> two contrasts for only my group of patients (=only 12 degrees of freedom), >>> with a template made of my 24 subjects, with HTA, Alcohol Consumption, TIV >>> and age as covariates and that I should run the four contrasts, that is to >>> say, use my whole sample (=24) to get 24 degrees of freedom...if I >>> interpreted him well. The other important thing he commented to me was >>> about the design itself: it is really well modeled and it is really asking >>> my question? He adviced me in order to inspect Jeannette Mumford web, but >>> between the designs she proposes, I have not found the correct one for me >>> because I do not want to test mean differences, slope differences or >>> interactions between my two groups, I just want to test correlations >>> between scan from my patients and HC and how this biomarker is anatomically >>> expressed. In fact, I used a global FA measure that was scaled (all values >>> were multiplied by 1000 in order to avoid too small values). >>> >>> In my design: >>> C1 = tests for positive correlations in the group of patients= P (+) and >>> C2 = tests for negative correlations in my group of patients = P (-). >>> I should add, in this design was not added and this is one of the >>> problems. commented above, C3 and C4 to test the same for my hc group. >>> Apart of this, is my design statistically well modelled? and is asking what >>> I want? >>> >>> Notes: HTA = Hyperyension; OH= Alcohol Consumption; C = Healthy >>> Controls...and P= Patients; all entered variables were previously demeaned >>> taking into account the mean of the whole sample for every variable >>> introduced. Covariates, although Jeannete Mumford's web page comments that >>> they are not necessaryly be demeaned...I saw this later and I believe, >>> please correct me if I am wrong, that this would not interfere in the final >>> results. >>> >>> Here I attache you the desig.mat and design.com text files....that are >>> the same for my TBSS data...but in the TBSS data, instead of using TIV I >>> used a global measure of white matter to adjust my design...but I found an >>> enormous cluster impossible to split into small ones...maybe here I did not >>> explain myself well. I used a threshold of 0.95 in my cluster command: >>> cluster -i data -t 0.95...but after inspection of my result that achieved a >>> p < 0.001....I tried to run cluster -i data -t 0.98...etc in order to splt >>> that enormous cluster that is uninformative...again...I think that problem >>> is in the design.....but not sure how to cope with it. >>> >>> Your advice would be greatly appteciate. >>> With my kind regards, >>> >>> Rosalia. >>> >> >> >