TMLE prediction models with ML about zepid HOT 4 CLOSED

Alright one issue is that sklearn uses the following format; model.fit(X, y) while statsmodels uses model.predict(y, X). Since I want users to be able to use whatever model in TMLE (like homebrew SuPyLearner), this causes an issue.

Some potential solutions:

1. have the user feed in a fitted model, with the prediction function, like fitted.predict(...). This avoids any issues on my end. Essentially all models are run outside of TMLE in this case. Which might be preferable, since it allows users to run diagnostics and independently fit the models. If they are all inside TMLE class, the fit might be ignored (at least more likely)
2. Haven't thought of an alternative yet

from zepid.

Finalized solution; user feeds in a fitted Class with the predict() function. This avoids the above issue and allows a wider variety of models.

Also note, patsy is now used in the background to do some data filtering as necessary. There is a try-except loop that captures whether the fed in model needs an intercept (potentially statsmodels) vs models that don't (sklearn)

This will be released in v0.3.1 (in the next few days as of this comment)

from zepid.

Going back to change this for v0.4.0. I need to use another way to operationalize this. The necessity is because I want LTMLE to match TMLE. For LTMLE and sequential regression, a model has to be fit EACH time point. Therefore, I need to have LTMLE do the fitting each time. The user needs to pass an unfitted model.

Users should pass some class object containing their predictive function of choice. That class should have a function with fit() and predict(). I will need to have a background process that looks to see what variable specifications exist in fit(). For sklearn and supylearner, X and y are used. statsmodels uses endog and exog. Those libraries should be a sufficient start. I can eventually add more possibilities as requested

from zepid.

v0.4.0 #35 supports sklearn and supylearner as input classes in custom_model. Dropped statsmodels support now for TMLE, since they format is a little hard to use within the function. Also, the machine learning is the key reason for this feature. Targeting the ML libraries is my primary goal. I have added an error for users to open an issue if they have some class they would like to see implemented

from zepid.