 |
 |
Hi Rosalia, Please, see below: On 8 January 2016 at 10:11, Rosalia Dacosta Aguayo <[log in to unmask]> wrote: > Hi Anderson, > > Thank you a lot for answering splitting the answers make me easier to > interpret.... > > 1. I used Bonferroni Correction because I tested for 5 biomarkers...so > "multiple testing"...and the possibility to find a something is > higher...For VBM analysis my results did not achieved Bonferroni > Correction, but for DTI, I tested my 5 biomarkers and I got a peak voxel > surrounded by other voxels forming the cluster. > Ok, so you corrected the cluster p-values to take into account the 5 tests, one for each biomarker. It's fine. This also means that you have more than one biomarker (in earlier emails and also below it appears you'd have 1 biomarker of interest only). Anyway this isn't a major issue. > > 2. I did Factorial Analysis because PCA was not suitable for my sample > (Barlett KMO did not result significant and the value was <0.5). In this > factorial analysis I got two components, I saved as variables, that resumed > the impact of age, diabetes, dislipemia, alcohol consumption and > hypertension)...except for age, the others are common vascular risk factors > and I do not want to see how they are related with my biomarker...in fact > they are highly correlated but I am just interested in the effect of this > biomarker per se in every group separately regarding cognitive scores and > neurological scores...and the results I got were very interesting and > coherent (for example, with neurological scores). > You can ignore Bartlett and KMO and simply use PCA, though it's so similar with FA that I doubt it'll make much difference. So you found 2 components that summarise most of the variance of these 5 variables. Good, so remove these 5 nuisance variables from the model and include the 2 principal components. Don't split them into groups as I think you mentioned already that there were no significant interaction. Include the biomarker in the design, splitting into two groups as you'd like to test the interaction. Then you can form the contrasts to test biomarker within each group, across groups, and the interaction. > > 3. So, I am trying to see how the change of this biomarker in this peak > voxel, affects cognition and neurological status without the influence of > age, and vascular risk factors....for every group separately...Cognition, > Neurological Status and Biomarker in this region of interaction are my main > variables of interest.... > On the peak voxel: if the biomarker is correlated with cognition, neurological status, etc, then analysing this peak voxel is circular, so be careful with this analysis. It would be better if this voxel or region of interest came from an independent study. All the best, Anderson > > Thank you again for your always helpful answers. > > Kind regards, > > Rosalia. > > 2016-01-08 10:20 GMT+01:00 Anderson M. Winkler <[log in to unmask]>: > >> Hi Rosalia, >> >> I'll split in parts, please see below: >> >> >> On 7 January 2016 at 08:42, Rosalia Dacosta Aguayo <[log in to unmask]> >> wrote: >> >>> Hi Anderson, >>> >>> Thank you for answering. I refer that my biomarker is highly correlated >>> with (Hypertension, Alcohol consumption, Diabetes, Dislipemia and Age). >>> >> >> Ok, so a certain biomarker X is highly correlated with these other >> variables. None of these are behavioural except perhaps alcohol >> consumption. Let's call them generically Z. >> >> >>> So, what I have finally done is the following: search for interaction >>> between groups for that biomarker regarding imaging data...and I saw that >>> for one of my biomarkers, there was interaction and that this interaction >>> with TBSS (not VBM...) >>> >> >> Let's call the imaging data generically as Y. So whatever is the >> association of X with Y, it varies according to group. >> >> >>> survived Bonferroni Correction because I used first cluster to identify >>> the region of interaction and then, I used fslmaths -thr 0.99 and I got a >>> mask with significant tracts. >>> >> >> So these are the regions where there is a significant interaction between >> X and group on what concerns the relationship between X and Y. I don't get >> where Bonferroni entered but let's skip that for now. >> >> >>> Then, I selected the value of the peak voxel by using fslmeant ...-c >>> (coordinates for the peak voxel) and this gave me the values for that voxel >>> for every single subject. >>> >> >> Ok, so you have the original values of Y for the voxel with the most >> significant interaction. >> >> >>> Then I entered those values into SPSS and I make two things: 1) >>> bivariate Spearman correlations between those values and my behavioral >>> scores; >>> >> >> What behavioural scores? How much does the biomarker X correlate with >> this behavioural score? If it is as correlated as the Z above (diabetes, >> dyslipidemia, etc), then this analysis is circular and it isn't correct. >> >> >>> 2) I made my factorial analysis and I extracted two components that >>> resumed all the covariates I told you above and I made partial correlations >>> adjusting for this two components...is it right? >>> >> >> You mean factor analysis, a close cousin of PCA. Yes, this is fine. >> >> All the best, >> >> Anderson >> >> >> >>> >>> Thank you a lot and sorry if I did not explain me well. >>> >>> Rosalia. >>> >>> 2016-01-07 9:18 GMT+01:00 Anderson M. Winkler <[log in to unmask]>: >>> >>>> Hi Rosalia, >>>> >>>> For the EVs of interest that are highly correlated with nuisance EVs, >>>> yes, exactly. However, if it's only the nuisance EVs that are all >>>> correlated to each other, then such correlations aren't a problem, unless, >>>> as you have, too many EVs for a small sample. Then maybe PCA could reduce >>>> the number of EVs there. >>>> >>>> All the best, >>>> >>>> Anderson >>>> >>>> >>>> On 6 January 2016 at 10:49, Rosalia Dacosta Aguayo <[log in to unmask] >>>> > wrote: >>>> >>>>> Regarding what you told me a few days ago about my design...using >>>>> covariates that are highly correlated between them and with the variable of >>>>> interest...I can see now the problem...following Mumford Brain Stats...this >>>>> is a problem with colinearity....right? >>>>> >>>>> Rosalia. >>>>> >>>>> 2016-01-06 11:30 GMT+01:00 Rosalia Dacosta Aguayo <[log in to unmask] >>>>> >: >>>>> >>>>>> Hi Anderson, >>>>>> >>>>>> Thank you for your answer. I am looking at: Mumford brain stats in >>>>>> order to learn more about this issue. >>>>>> >>>>>> By the other hand, is it fair to use values that survived TFCE >>>>>> correction and Bonferroni Correction as well (5 tests I used for TBSS) for >>>>>> example, and use them to see how they are related with other scores >>>>>> (neurological and cognitive) ? >>>>>> >>>>>> All the best, >>>>>> >>>>>> Rosalia. >>>>>> >>>>>> 2016-01-06 10:54 GMT+01:00 Anderson M. Winkler < >>>>>> [log in to unmask]>: >>>>>> >>>>>>> Hi Rosalia, >>>>>>> >>>>>>> A significant interaction means that whatever association (even if >>>>>>> non-significant) is seen in one group between the biomarker and the imaging >>>>>>> data, it isn't the same as in the other group. The association can be >>>>>>> stronger, weaker, or in the opposite direction. >>>>>>> >>>>>>> It isn't possible to say that the biomarker would be more expressed >>>>>>> (or things along these lines) in one group than in the other, but that its >>>>>>> correlation with the imaging data is not the same across both groups. >>>>>>> >>>>>>> If the interaction isn't significant you may well keep only the main >>>>>>> EV in the model, i.e., not splitting the biomarker EV into the two groups). >>>>>>> >>>>>>> Interactions aren't something imaging-specific, though, and there >>>>>>> are plenty of online resources (including Wikipedia) that explain what it >>>>>>> is, and how it can be interpreted. >>>>>>> >>>>>>> All the best, >>>>>>> >>>>>>> Anderson >>>>>>> >>>>>>> >>>>>>> >>>>>>> On 5 January 2016 at 10:21, Rosalia Dacosta Aguayo < >>>>>>> [log in to unmask]> wrote: >>>>>>> >>>>>>>> Dear Anderson and FSL team, >>>>>>>> >>>>>>>> There is a thing that it is being difficult for me to understand. >>>>>>>> If I have found interactions for a single biomarker, in one anatomical >>>>>>>> region...how should I interpret this? >>>>>>>> >>>>>>>> C1 ( 0 0 1 -1) where 1 is for patients and -1 is for healthy >>>>>>>> controls...it is reasonsable to say that in that significant region my >>>>>>>> biomarker is expressed more in my patients that in my controls? As I made >>>>>>>> multiple testing (I tested for 5 interactions, 1 for every biomarker) I had >>>>>>>> to do Bonferroni Correction 0.05/5 = 0.01 and my result is of p = 0.013, so >>>>>>>> I have decided to tell this and go on with this biomarker given the small >>>>>>>> size of my sample....and the fact I have no possibilities to enlarge it. >>>>>>>> >>>>>>>> The other thing is that I did not found interactions between >>>>>>>> age*group and between TIV*group...is this normal? Taking into account the >>>>>>>> existing literature, I should covariate for age and TIV...but those two >>>>>>>> variables do not seem to affect my groups.... >>>>>>>> >>>>>>>> Any helping with this would be greatly appreciate. >>>>>>>> >>>>>>>> Yours sincerely, >>>>>>>> >>>>>>>> Rosalia. >>>>>>>> >>>>>>> >>>>>>> >>>>>> >>>>> >>>> >>> >> >