 |
 |
Hi Rodrigo, Please cc the fsl group so that others can chime in and/or respond. Here is the website I was referring to: http://mumford.fmripower.org/mean_centering/ There are two roughly equivalent ways you can analyze your data: 1) Use the design matrices you proposed (w/o an intercept) in conjunction with the -D option in randomise 2) Add a column of 1's to your design matrix, in which case the -D option in randomise is not needed. I say "roughly equivalent" because removing the mean in the input data prior to fitting a design matrix, and modeling an intercept in the design matrix are not strictly identical (the intercept is not exactly the mean in the context of other regressors), but the two approaches should yield similar results in most cases. I'd recommend approach (2) in which you include an intercept in your design matrix, since that is more similar to standard statistical modeling. As for demeaning regressors -- again, that is a separate issue, and whether or not it matters depends on which particular contrasts you are looking at. Jeanette's web page has very clear examples on this. cheers, -MH On Mon, 2011-11-28 at 19:44 -0600, Rodrigo Perea wrote: > When you refer to Jeanetter Mumford..are you talking about this site: > http://mumford.bol.ucla.edu/setting_up_designs_2009.pdf ? > Also, if I understand correctly, none of my design matrix will work > since I have no mean (intercept). Should I add a column of 1s or 0s > into my design matrices? Also, should I demeaned my data (covariates > and regressor) and also add the -D option? > Sorry for many questions but I would like to truthfully understand > this. > > Thanks, > Rodrigo > > > >>> Michael Harms <[log in to unmask]> 11/28/11 2:05 PM >>> > > If you don't model a mean (intercept) in your design matrix, and you > don't include the -D option, then your results are not interpretable > -- > you have nothing in your model to account for a constant term in your > data. > > As for which design matrix you should use, it shouldn't matter in this > particular case because the slope of a regressor (and its > significance) > is not affected by whether or not you demean that regressor. See > Jeanette Mumford's web page on centering for illustrations. > > cheers, > -MH > > On Mon, 2011-11-28 at 13:52 -0600, Rodrigo Perea wrote: > > Michael, > > Thanks for your response....so what you are suggesting > > me to do is just to run the analysis with my first matrix and the -D > option? If so, > > why didnt I get any significant regions and why did I get some when > I ran it without the -D option? > > Thanks in advance, > > Rodrigo > > > > > > > > Hi Rodrigo, > > The -D option in randomise has to do with whether the input data is > to > > be demeaned -- it has nothing to do with whether the regressors in > the > > design matrix are demeaned. > > > > Quoting the web help page for randomise: > > The -D option tells randomise to demean the data before continuing - > > this is necessary if you are not modelling the mean in the design > matrix > > > > In your case, your design matrices are not modeling a mean, so the - > D > > option must be included. > > > > cheers, > > -MH > > > > On Mon, 2011-11-28 at 19:32 +0000, Rodrigo Perea wrote: > > > Could someone explain me this. So I have a column that I want to > use as a regressor in my TBSS analysis. I created a matrix without > demeaning this value and I got a significant value at p<0.05 of what I > expected. Then I demeaned the regressor and also I used the -D option > and after 5000 permutations no significance was found in any > direction. Why? If someone could help guide me in the right direction, > I would greatly appreciate it. > > > > > > > > > Thanks in advance. > > > Rodrigo > > > > > > My non-demeanded design and contrast matrix are (I used this same > matrix when using the -D option): > > > > > > design.mat (sex and age as covariates): > > > /Matrix > > > > > > Sex Age Regressor Variable > > > 1.000000e+00 7.500000e+01 1.634750e+03 > > > 0.000000e+00 7.500000e+01 1.153250e+03 > > > 0.000000e+00 8.300000e+01 3.151000e+03 > > > 0.000000e+00 7.300000e+01 1.651000e+03 > > > 0.000000e+00 8.100000e+01 1.503250e+03 > > > 1.000000e+00 7.100000e+01 1.325500e+03 > > > . > > > . > > > . > > > > > > design.con: > > > /Matrix > > > 0.000000e+00 0.000000e+00 1.000000e+00 > > > 0.000000e+00 0.000000e+00 -1.000000e+00 > > > > > > > > > > > > My demeaned design matrix and contrast are: > > > > > > > > > design.mat (sex and age as covariates): > > > /Matrix > > > > > > Age Sex Regressor > > > 1.000000e+00 7.500000e+01 -6.570900e+02 > > > 0.000000e+00 7.500000e+01 -1.138590e+03 > > > 0.000000e+00 8.300000e+01 8.591600e+02 > > > 0.000000e+00 7.300000e+01 -6.408400e+02 > > > 0.000000e+00 8.100000e+01 -7.885900e+02 > > > 1.000000e+00 7.100000e+01 -9.663400e+02 > > > . > > > . > > > . > > > ' > > > > > > design.con: > > > /Matrix > > > 0.000000e+00 0.000000e+00 1.000000e+00 > > > 0.000000e+00 0.000000e+00 -1.000000e+00 > > > > Rodrigo Dennis Perea > > Graduate Research Assistant > > [log in to unmask] > > Bioengineering Program > > The University of Kansas > > >