Dear SPM users,
I need some support with a multisite fMRI analysis using SPM. I first estimated the individual beta maps for a given regressor in the first-level analysis. Then, I used these beta maps (or contrasts, in my case, they are the same) in the second-level analysis to obtain the group statistical maps.
In my case, I need to remove the influence of the covariates and, since this is a multisite study, the covariates are the sites. I read some posts here saying that one needs to create binary vectors to represent those site-related covariates. Those would be vectors with 1s representing the beta maps from the site in question and 0s elsewhere. However, the individual beta estimates vary wildly across sites. Please check the attachment, in which I'm boxplotting the individual beta estimates, within a given ROI, across sites (the columns in the plot). In an ROI analysis using R, I was able to reproduce the pattern of the whole-brain results by using the model: lm(contrast ~ 1 + site, data = data).
However, the final group map is primarily based on the dataset that has the highest average/variance, i.e. the sixth column in the figure. If I remove this dataset, then the group map will be primarily based on the next one with the highest average/variance, i.e. the first column in the figure. And so on... It's strange because, even though the data from the sixth column/site contains 8 subjects, it has a much higher influence in the final result compared to the data from the third column/site, which contains 37 subjects
How can I 2nd-level analyze these data using a better way to remove/weight the covariates using SPM?
Z-normalizing the beta maps wouldn't work: we would introduce artificial activations/deactivations. I also thought about using the spmT maps instead of con files, because they are somehow "normalized", but I've never seen anybody doing it.
I appreciate any kind of input. Thank you very much in advance.
Best wishes,
Gustavo
|