Do you ever wish you could just use python to pull together the files and code for running FSL's randomise? Me too! I made this: github.com/jmumford/ran... It will even replace the numbers in the file outputs with contrast names of your choosing (and replace corrp with 1minusp).

Sure thing! Slides and poster here: www.dropbox.com/scl/fo/re1bc...
Preprint here: www.biorxiv.org/content/10.1...

Interested in practical tips for assessing collinearity and why common fixes might not help? Come to the #OHBM2025 oral session Friday at 11:30AM (P2)! I’ll share simulation results and @demidenko.bsky.social will present real data. Also, check out my poster Thursday (1051)!

chng.it/gmWgmwwz6n
The Open Science Room needs your help!

Following a lot of hard work and dedicated effort from countless parties, our textbook providing a hands-on introduction to analyzing network data is now available for pre-order from Cambridge University Press www.cambridge.org/core/books/h...

Shameless plug for a useful tool tucked away in Appendix B of our preprint. A VIF calculation for *contrasts*, cVIF. A foolproof way to assess how collinearity is impacting what you're actually interested in! Code available on GitHub!
www.biorxiv.org/content/10.1...
github.com/jmumford/vif...

Setting up an fMRI model and getting hit with collinearity errors? Been there! 😢 Try removing regressors? Orthogonal regressors? Other tricks to resolve collinearity errors? Guess what—those fixes might make things worse. We explore why and propose recommendations.
www.biorxiv.org/content/10.1...

The bigger problem is when orthogonalization changes parameter interpretations such that the "adjusting" implied by the regressors isn't actually occurring. "We adjusted our group difference for age" (but age was modeled as an interaction and centered within each group, so no adjusting occurred)

Also, when looking at interactions I've never understood why people look at the main effects if the interaction is significant. If it isn't significant I omit it from the model, but I know some people don't like to do that.

Centering is the only type of orthog I find helpful, but this is often done wrong (e.g., center within groups instead of across all subjects). The big problem with orthog is people often don't realize they're effectively removing the regressors they orthogonalized with respect to from the model.