A study on GAM-based approaches to EEG/ERP analysis

[carldelfin]

After Googling a bit about GAMs and whether they'd be suitable for EEG/ERP analysis, I stumbled upon this page on the NeuroKit docs. I contacted Dominique, who suggested a study in the form of a tutorial on how to do analyze ERPs using GAMs (with R and Python).

If you're interested in collaborating on this, let us know! For starters, here's a few questions to ponder that may help us get going:

• What (in relatively simple terms) is a GAM (and/or other spline-based methods)?
• What is the added value of a GAM-based approach vis-a-vis current "standard" approaches (i.e., why should people bother)?
• How can the added value of GAMs be evaluated and presented to the reader?
• Does a GAM-based approach bring any additional (statistical, methodological) assumptions that need to be considered?
• (Bonus: Are Bayesian GAMs based on MCMC sampling feasible? Are other Bayesian approaches feasible?)

[DominiqueMakowski]

Some thoughts regarding the "Bayesian perspective" to discuss the tradeoff of Bayesian sampling:

• Computational power issue with traditional MCMC
  • What are the advantages of it? Better inference?
  • Could variational algorithms be a viable alternative?
• The role of priors
  • It appears as a challenge to find informative prior
  • Maybe priors could be used as a penalization parameter (i.e., have more narrow priors to avoid possible overfitting / parameters going crazy because of some extreme observations - e.g., bad channels or artifacts)

[DominiqueMakowski]

What (in relatively simple terms) is a GAM (and/or other spline-based methods)?

Yes this very important as GAMs are still fairly new, especially in psychology / neuroscience.

So far I'd see two ways of starting, either we start by presenting GAMs and then be like "this appears as well-suited for ERPs", or we start by briefly mentioning the current ERPs typical approaches and their limitations, and then how GAMs could be a potential avenue to tackle these limitations

[DominiqueMakowski]

Something like:

1. GAMs and their relevance to ERP/EEG

• What (in relatively simple terms) is a GAM (and/or other spline-based methods)?
• What is the added value of a GAM-based approach vis-a-vis current "standard" approaches (i.e., why should people bother)?

2. Hands-on example of how to use them, interpret and report them

• How can the added value of GAMs be evaluated and presented to the reader?

3. Methodological Caveats

• Does a GAM-based approach bring any additional (statistical, methodological) assumptions that need to be considered?
• What users should be aware of, what are the important things to check / verify / justify (typically with GAM, the amount of wiggleness allowed is critical)

4. Limitations and future developments

• Bayesian framework: is it feasible and why is it desirable?
  • Computational power issue with traditional MCMC (i.e., it's slow)
  • What are the advantages of it? Better inference?
  • Could variational algorithms be a viable alternative?
  • The role of priors
  • It appears as a challenge to find informative prior
  • Maybe priors could be used as a penalization parameter (i.e., have more narrow priors to avoid possible overfitting / parameters going crazy because of some extreme observations - e.g., bad channels or artifacts)

[carldelfin]

Great, yep, that looks like the most promising structure!

I've started to do some literature search to see whether anyone has actually used GAMs for ERP analysis, and found this, with code available here, this, this, and this. The latter study reads a bit like a tutorial but isn't really that "hands-on". It also contains a fair bit of the math underlying GAMs, which may be useful (e.g., we can give a brief overview and refer to that study, for instance, and then focus more on practicalities).

There's also this study about different ways to analyze EEG data, and these lecture slides. All in all, there's a lot of useful information in these studies that should help us get going with the introduction and methodological caveats/limitations. I believe that our study would be a good fit, with a more practical, hands-on approach compared to previous work.

[stale[bot]]

This issue has been automatically marked as inactive because it has not had recent activity. It will eventually be closed if no further activity occurs.

[DominiqueMakowski]

There is a special issue in the sensor journal on "Brain Signals Acquisition and Processing". It could be a really great fit for that, but the deadline is 15 April 2021 so it might be a bit too soon 😁

[carldelfin]

Sorry for the late reply! I agree, that does sound like a great fit, although yes, the deadline might be a bit tight... but we might as well try to get going? :-) What's the best way forward, just create an .Rmd and start writing?

[stale[bot]]

This issue has been automatically marked as inactive because it has not had recent activity. It will eventually be closed if no further activity occurs.

[stale[bot]]

This issue has been inactive for a long time. We're closing it (but feel free to reopen it if need be).