 |
 |
Hi Cyril, Yes, I agree completely. Just to clarify for posterity: -If the regressors are normal regressors (i.e. entered in a multiple regression, say a 2nd level VBM analysis), the beta weights can change with different regressors added, because their independence (degree of orthogonality) will change. -If the regressors are parametric modulators (i.e. entered in a 1st level fMRI analysis), SPM will serially orthogonalize each additional parametric modulator. Thus, the regressors entered after PB should not be correlated with it, and the beta for PB should not change (but the results can still change because the residuals will be different with different parametric modulators). John, I wasn't sure which case your data fell into, but hopefully know you know enough to help. Best regards, Jonathan -- Dr. Jonathan Peelle Center for Cognitive Neuroscience and Department of Neurology University of Pennsylvania 3 West Gates 3400 Spruce Street Philadelphia, PA 19104 USA http://jonathanpeelle.net/ On Wed, May 23, 2012 at 7:55 AM, cyril pernet <[log in to unmask]> wrote: > Hi Jonathan > >> Dear John, I have a design matrix with 3 parametric regressors PA, PB and >> PC (in >>> >>> that order) >>> >>> Now, i create a second DM with the following regressors PA,PB,PD,PE. >>> >>> Whereas the results for PA remain the same, the results for PB change. >>> One reason could be that PE and PB are somehow correlated; >>> however, I thought that the correlation would have an effect on PE and >>> not on PB, given that PB appears first in the DM >>> >>> Could somebody explain why results from PB in the first DM are different >>> from results of PB in the second DM? >> >> >> Most likely you are right - the correlation of PB with other regressors >> (PC vs. PD/PE) is not the same, which will give you different results. >> (Also it's important to know what you mean by "results" - the beta >> estimate for PB, or the t statistic from a contrast on PB? The inclusion of >> additional regressors may not only affect the estimate of the PB beta, but >> also the error term, which would give you a different t statistic image.) > > given the orthogonalization, this has to be the same Beta value, but the > results will differ since residuals will be smaller > >> >> You can look at the correlation between columns in your design matrix by >> clicking the "review" button, and then under the "design" menu select >> "design orthogonality". >> Alternatively you can look at the modulators from your design matrix as >> entered, which is stored in SPM.xX.X. Something like this: >> load SPM >> PB = SPM.xX.X(:,2); >> PC = SPM.xX.X(:,3); >> [r, p] = corrcoef(PB, PC); >> plot(PB, PC, 'o'); > > remember that corrcoef is the mean centered correlation whereas in spm that > is just the cos(angle) between regressors (not such a big difference but > worthwhile remembering - cos(angle) = corr only is the vectors are > normalized) > > Cyril > > > -- > The University of Edinburgh is a charitable body, registered in > Scotland, with registration number SC005336. >