Baseline correction with Unfold Models

[behinger]

Jordan asked a cool and important question on baseline corrections. Paraphrasing: "Do we need to do it? If yes, when? Why after deconvolution is the baselinecorrection sometimes not necessary?"

[behinger]

1. Do we need it

The same as in EEG research, if you have systematic differences in baseline because of orthogonal activity, then it is a good idea to bring all curves to the same baseline level. There is a big discussion in the literature going on though, and what I'd recommend in the rERP framework, is to let the data decide itself at each timepoint, whether it wants to be baseline corrected and if yes, how much. There is this nice paper by Phillip Alday (2019, https://onlinelibrary.wiley.com/doi/full/10.1111/psyp.13451) where he suggests doing exactly that. In my experience that approach works very nicely. What you need to do:

1. have epoched data
2. extract the baseline for each trial
3. add it to the evts DataFrame as a column
4. include bsl as a linear effect (potentially with interactions with your conditions) to your model
5. 🎉

Your data is now regression-baselinecorrected.

Why after deconvolution is the baselinecorrection sometimes not necessary?

What Jordan stumbled upon is the effect we sometimes see in deconvolution models, that the baselines are much cleaner after overlap correction, compared to traditional (r)ERP analyses. In some cases, you dont even need to do bsl correction afterwards. The obvious explanation would be, that then the baseline problem was created by overlap in the first place. But o.c. many paradigms where it is not overlap!