> > Hello ! > > Concerning the use of Ancova or Scaling in normalization for fMRI time-series, I > observed that both of them could lead to some strange effects when analyzing a very > stable time-series where the main effect (as assessed by Eigenimages analysis) is > related to the paradigm. > In this particular case, I was analyzing a simple Control / Activation study (with > C/A/C/A/C/A/C/A/C/A). With scaling, I observed some significant "deactivations" (C-A > contrast), that were not observed with Ancova. When reanalyzing the data without any > normalization, I observed no deactivations, but more significant activations (A-C > contrast) than in the Ancova or the Scaling analysis. > It indeed seems logical that if the time-series is very stable, using Scaling will > produce artifactual deactivations, simply because the mean of each image of the > time-series is significantly affected by the increase in signal brought about by > strong activation in several voxels. Using Ancova actually produced a function of no > interest that was very similar to my paradigm, and I think this is why activations > were weaker in this analysis than in the no-normalization analysis. > So, my question finally is : is it not better to first analyze the data without any > normalization to assess with eigenimages if there is or not in the data a major > global signal change unrelated to the paradigm that should be accounted for by > normalization ? If there is none, analysis without any kind of normalization is > probably the best. > > Thanks in advance, > Elie Lobel > Service Hospitalier Frederic Joliot > Orsay, France > Dear Elie, your question pertains to the special, and rather problematic, case where changes in your global signal is almost entirely caused by changes on local signal (focal activations) and therfore highly corelated to the latter. As you correctly points out entering this signal as a confound (be it through proportional scaling or ANCOVA) may lead to cancellation of some of your local signal and, more seriously, to false "deactivations". The reason for the global normalisation in the first place was the observation that it accounted for unwanted scan-scan variance due to differences in injected activity, pCO2 level, arousal level etc. in the early PET days. The part of the global variance that were actually attributable to local variance was so relatively small that the problem you describe was not initially observed. In the best of worlds, which we are now with the refinement of neuroimaging techniques approaching, all change in global variance should be attributable to change in local activity and there would be no need for global normalisation. There are still sources of variability such as drifts in the MR scanner and changes in pCO2 levels, but if we assume that these occurr at a slower time scale than our experimental paradigm the low frequency components of the fMRI design matrix should take care of those. Hence, one might actually make a case for the abolishment of global normalisation for fMRI, or at least for some care in it's use, especially looking out for any correlations between global values and experimental design. Clearly, as you suggest, a PCA on the unadjusted data might give useful clues as to whether global normalisation is beneficial or not. If one still wants to use global normalisation on a data set such as that described by you a method has been suggested for dealing with your specific problem. Basically it circumvents the problem by assuming that any areas that are seen as activated in a first analysis step are potential sources of bias when estimating global activity, and another analysis is performed where these areas are excluded when calculating global activity. This process may be repeated iteratively until no changes occurr in subsequent calculations of global activity, at which stage the estimates are presumably unbiased. You may find a full description of the method in Andersson JLR, How to estimate global activity independent of changes in local activity, NeuroImage 1997: 4;237-244. Other methods of dealing with this problem has been suggested, and hopefully they will find their way into the final release of SPM98. Good luck Jesper %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%