FAQ: What to do with NaNs as predictors

[behinger]

I got the question of what to do with NaNs. This is my answer:

As long as the predictor is not "active" in an event type, it shouldn't matter if it is [] or NaN. The only problem is, if you actually do not know a predictor value.

An example: for whatever reason you do not know a saccade-amplitude for one fixation (thus it is NaN), but generally all fixations are modelled with a saccade amplitude effect e.g. fixation: y~1+sac_amplitude. What should the model use for this fixation as the covariate? In that case, you would need imputation of the amplitude effect, for which there are multiple ways to do in uf_imputeMissing.m .

But if you have another event "button", that one does not need to have a saccade-amplitude e.g. button: 1+condition, then you can put whatever you want.