Data preprocessings for MVPA analysis ?

[JeanLucAnton]

Is it better to have smoothed or unsmoothed data ?

• In the PRoNTO Manual it's written that you should used normalized & smoothed data but in other (eg. RSA manuals) the opposite for smoothing is recommended.
• For Hebart, neither normalize nor smoothing

• Felipe : do everything in terms of univariate analyses except for smoothing. However, when they smoothed they observed either no effect or actually an improvement (signal/noise). 

• Thierry : Smoothing at the same size as the voxel ? 


[JeanLucAnton]

Independance of the preprocessing between learning and test data ?

The data normalisation across the voxels (mean centering to 0 by substraction by the mean, and variance set to 1 by dividing by the variance) has to be done on the learning data. Then this normalisation has to be applied like this to the test data (without recalculating the mean and variance values).

[SylvainTakerkart]

it might be good to gather here all the references available in the literature about these questions (or at least some ;) ), because there are a good amount of papers on the topic!

everybody: do not hesitate to post links to papers (DOI links are the most convenient!)... this discussion thread can serve as a knowledge base that we share!

[eliefabiani]

Hi, I looked at papers doing MVPA-searchlight analysis in the bilingual field. Most of the papers found run the searchlight in normalized and unsmoothed data (Xu 2017 (DOI: 10.1126/sciadv.1603309), Van de Putte 2017 (https://doi.org/10.1016/j.neuroimage.2017.08.082), Lei 2014 (https://doi.org/10.1016/j.bandl.2014.08.009), Brignoni-Perez 2019 (https://doi.org/10.1016/j.bandl.2019.104725), Lu 2021 (https://doi.org/10.1016/j.neuroscience.2020.10.040). Normalized and unsmoothed seems to be the common preprocessing for MVPA searchlight.

[SylvainTakerkart]

...nothing wrong with this!

(except the waste of computation time if "normalized" implies to be at the resolution of the anat; but the waste of computation time is only a crime for planet earth, not for the nature of your scientific result, ahah ;) )

[SylvainTakerkart]

sorry, I thought I was replying in your other thread, Elie!

sooooooooooo, my comments:

• unsmoothed: yes, it makes sense conceptually! (with the underlying hope that there is information in the high resolution spatial information)... and you can smooth later (on the accuracy map) if needed...
• normalized: this is not mandatory, you can also run your searchlight in native space and normalize your accuracy maps also! both are ok! (some people feel strongly on one or the other, but my feeling is that both are unperfect yet acceptable)

;)

[CarolineLndl]

I feel that it is better to work on the individual space (specificaly in functionnal space). Indeed, we regress our models with regressor estimated in functionnal space (motion, CompCor...) In theory it should not change much but it seems to me cleaner that everything is estimated on the same untransformed image. In addition, some toolboxes tend to resample the images to the reference image (anat or template) which doesn't bring anything to the real resolution of the func data but risks to increase the image size and the computation time drastically! Last but not least, if you decide at the end of your analysis to apply the Dartel transfo or something like that, it is still possible and you will not need to rerun all the first level analyses! I'm rather in favor of applying the coregistration on the final stat' images :)

(btw I'm completely agree that we should run MVPA as well as denoising on unsmoothed data)