How to get fitted learner from Object

[chungdz]

I try to get the fitted learner from DoubleMLPLR object.

I try learner/models attribute of this object. The former one gives me learners that haven't been fitted, the latter gives me None.

Any solutions please?

[SvenKlaassen]

Thanks for your question.
Could you please post an minimal reproducable example, such that I can replicate the error.

[chungdz]

Copy the example code in document with the error I show:

import numpy as np
from doubleml.datasets import fetch_bonus
from sklearn.base import clone
from sklearn.ensemble import RandomForestRegressor
from doubleml import DoubleMLData, DoubleMLPLR

df_bonus = fetch_bonus('DataFrame')
xinputs = ['female', 'black', 'othrace', 'dep1', 'dep2',
        'q2', 'q3', 'q4', 'q5', 'q6', 'agelt35', 'agegt54',
        'durable', 'lusd', 'husd']
dml_data_bonus = DoubleMLData(df_bonus,
                             y_col='inuidur1',
                             d_cols='tg',
                             x_cols=xinputs)

learner = RandomForestRegressor(n_estimators=500, max_features='sqrt', max_depth= 5)
ml_l_bonus = clone(learner)
ml_m_bonus = clone(learner)

obj_dml_plr_bonus = DoubleMLPLR(dml_data_bonus, ml_l_bonus, ml_m_bonus)
obj_dml_plr_bonus.fit()

obj_dml_plr_bonus.learner['ml_l'].predict(df_bonus[xinputs])

obj_dml_plr_bonus.models.predict(df_bonus[xinputs])

Error messages:

[SvenKlaassen]

Thank you.
The learner attribute is just a copy of the unfitted learner (since they have to be used multiple times).
To store the fitted models you have set the store_models argement in the .fit() method. The fitted models are then accessible under the models attribute.

import numpy as np
from doubleml.datasets import fetch_bonus
from sklearn.base import clone
from sklearn.ensemble import RandomForestRegressor
from doubleml import DoubleMLData, DoubleMLPLR

df_bonus = fetch_bonus('DataFrame')
xinputs = ['female', 'black', 'othrace', 'dep1', 'dep2',
        'q2', 'q3', 'q4', 'q5', 'q6', 'agelt35', 'agegt54',
        'durable', 'lusd', 'husd']
dml_data_bonus = DoubleMLData(df_bonus,
                             y_col='inuidur1',
                             d_cols='tg',
                             x_cols=xinputs)

learner = RandomForestRegressor(n_estimators=500, max_features='sqrt', max_depth= 5)
ml_l_bonus = clone(learner)
ml_m_bonus = clone(learner)

obj_dml_plr_bonus = DoubleMLPLR(dml_data_bonus, ml_l_bonus, ml_m_bonus)
obj_dml_plr_bonus.fit(store_models=True)

obj_dml_plr_bonus.models

This should look like this

{'ml_l': {'tg': [[RandomForestRegressor(max_depth=5, max_features='sqrt', n_estimators=500),
    RandomForestRegressor(max_depth=5, max_features='sqrt', n_estimators=500),
    RandomForestRegressor(max_depth=5, max_features='sqrt', n_estimators=500),
    RandomForestRegressor(max_depth=5, max_features='sqrt', n_estimators=500),
    RandomForestRegressor(max_depth=5, max_features='sqrt', n_estimators=500)]]},
 'ml_m': {'tg': [[RandomForestRegressor(max_depth=5, max_features='sqrt', n_estimators=500),
    RandomForestRegressor(max_depth=5, max_features='sqrt', n_estimators=500),
    RandomForestRegressor(max_depth=5, max_features='sqrt', n_estimators=500),
    RandomForestRegressor(max_depth=5, max_features='sqrt', n_estimators=500),
    RandomForestRegressor(max_depth=5, max_features='sqrt', n_estimators=500)]]}}

Remark that this are a couple of models which you can access via a nested dictionary where the first level is the nuisance element and the second element is the treatment name (potentially multiple treatments). This then includes a nested list as the first index refers to the n_rep argument (if the cross-fitting has been repeated multiple times [default is 1]) and lastly to the n_folds argument (as there is one model for each fold [default is 5]).
To access a single model you can use

obj_dml_plr_bonus.models['ml_l']['tg'][0][0].predict(df_bonus[xinputs])

I hope this works for you.

[kartechbabu]

Dear Sven,

Is it possible to get CI or SE for the predictions?

Thanks
Kartik

[SvenKlaassen]

Hello @kartechbabu,

generally it is not possible to get the CI oder SE for these predictions from our package. For the PLR model one could in principle generate outcome predictions via

$$ \hat{y}:=\hat{\theta}D + \hat{g}(X) $$

but this would include the noisy estimate of $\hat{g}(X)$.

[kartechbabu]

Thank you so much, @SvenKlaassen, for your prompt response!