ideas from Donhauser & Baillet

[behinger]

https://www.cell.com/neuron/fulltext/S0896-6273(19)30893-1

• discussed this with Peter at some point: less timelags the further one goes away from the event onset. They use a log-spacing. The principled way to do it would be with a basis function approach (I think).

• Variance Partitioning: They have some kind of way to get rid of the crossvalidated, negative R2_total but better R2_ca variance issue. I have to check it out once more.

• TRF with impulse and step functions, would be interesting to run some simulation of what kind of signal would require step functions? I have to check out what they do with the differing types. Given that they have ridge regression it might not be so important to them (they could see it as a feature expansion without any theoretical motivaiton)