RSA Analysis on fMRI - use beta or contrast images as input?


I am running representational similarity analyses for 8 categories of inner-speech with fMRI data. I am wondering whether I should be using the beta images or the contrast images I have produced for each category. These contrast images are for example 0 0 1 0, there is no subtraction of a baseline.

Which should be used as input, betas or contrast .niis and also why? (most tutorials seem to use betas).

I plan to use first pymvpa but also may use MRC-CBU matlab toolbox later as this seems more rigorous.



IIUC it is equivalent to use beta versus contrast images in your case. You may want to use statistics images (t versus raw contrast images), as you most likely use multi-run data.

Hi @bthirion,

Thank you very much for your response! This is really helpful to know that it would be equivalent. I guess the only difference. By statistic images do you mean for example after cluster thresholding - setting p>0.001 for example?

As there are multiple sessions per subject, I have been using the normalised data, I am wondering whether this is appropriate, or I should use unnormalised data with only a small smoothing kernel, and do this separately for the two sessions?

No, I simply mean that you divide effect estimates by the standard deviation of the estimate uncertainty. This probably makes the inputs more comparable across samples.
It certainly makes sense to normalize your data across sessions.