NA handling for predict and predict with new data

[leostimpfle]

Is this behaviour of pyfixest.estimation.models.feols_.Feols.predict intentional? It seems slightly odd that we drop NaNs and singletons in the newdata is None case but not in the newdata is not None case.

import numpy as np

import pyfixest as pf

data = pf.get_data(N=1_000, seed=0, model="Feols")
fit = pf.feols("Y ~ X1", data = data)

np.isnan(fit.predict(newdata=None)).any()  # False (NaNs and singletons are dropped)
np.isnan(fit.predict(newdata=data)).any()  # True (NaNs and singletons are kept)

[s3alfisc]

I think this is matching fixest behavior? Basically, the idea is - if new data = None, then we predict based on the estimation sample for which we made sure to drop NaNs. If we pass new data, this is treated as a new data set (for out of sample prediction) and predict reports that it cannot work on NaNs.