from sklearn import linear_model
import pandas as pd
from dowhy.causal_estimator import CausalEstimate
from dowhy.causal_estimator import CausalEstimator
[docs]class PropensityScoreStratificationEstimator(CausalEstimator):
""" Estimate effect of treatment by stratifying the data into bins with
identical common causes.
Straightforward application of the back-door criterion.
"""
def __init__(self, *args, num_strata=50, clipping_threshold=10, **kwargs):
super().__init__(*args, **kwargs)
self.logger.debug("Back-door variables used:" +
",".join(self._target_estimand.backdoor_variables))
self._observed_common_causes_names = self._target_estimand.backdoor_variables
if len(self._observed_common_causes_names)>0:
self._observed_common_causes = self._data[self._observed_common_causes_names]
self._observed_common_causes = pd.get_dummies(self._observed_common_causes, drop_first=True)
else:
self._observed_common_causes= None
error_msg ="No common causes/confounders present. Propensity score based methods are not applicable"
self.logger.error(error_msg)
raise Exception(error_msg)
self.logger.info("INFO: Using Propensity Score Stratification Estimator")
self.symbolic_estimator = self.construct_symbolic_estimator(self._target_estimand)
self.logger.info(self.symbolic_estimator)
if not hasattr(self, 'num_strata'):
self.num_strata = num_strata
if not hasattr(self, 'clipping_threshold'):
self.clipping_threshold = clipping_threshold
def _estimate_effect(self):
propensity_score_model = linear_model.LogisticRegression()
propensity_score_model.fit(self._observed_common_causes, self._treatment)
self._data['propensity_score'] = propensity_score_model.predict_proba(self._observed_common_causes)[:,1]
# sort the dataframe by propensity score
# create a column 'strata' for each element that marks what strata it belongs to
num_rows = self._data[self._outcome_name].shape[0]
self._data['strata'] = (
(self._data['propensity_score'].rank(ascending=True) / num_rows) * self.num_strata
).round(0)
# for each strata, count how many treated and control units there are
# throw away strata that have insufficient treatment or control
# print("before clipping, here is the distribution of treatment and control per strata")
#print(self._data.groupby(['strata',self._treatment_name])[self._outcome_name].count())
self._data['dbar'] = 1 - self._data[self._treatment_name]
self._data['d_y'] = self._data[self._treatment_name] * self._data[self._outcome_name]
self._data['dbar_y'] = self._data['dbar'] * self._data[self._outcome_name]
stratified = self._data.groupby('strata')
clipped = stratified.filter(
lambda strata: min(strata.loc[strata[self._treatment_name] == 1].shape[0],
strata.loc[strata[self._treatment_name] == 0].shape[0]) > self.clipping_threshold
)
# print("after clipping at threshold, now we have:" )
#print(clipped.groupby(['strata',self._treatment_name])[self._outcome_name].count())
# sum weighted outcomes over all strata (weight by treated population)
weighted_outcomes = clipped.groupby('strata').agg({
self._treatment_name: ['sum'],
'dbar': ['sum'],
'd_y': ['sum'],
'dbar_y': ['sum']
})
weighted_outcomes.columns = ["_".join(x) for x in weighted_outcomes.columns.ravel()]
treatment_sum_name = self._treatment_name + "_sum"
weighted_outcomes['d_y_mean'] = weighted_outcomes['d_y_sum'] / weighted_outcomes[treatment_sum_name]
weighted_outcomes['dbar_y_mean'] = weighted_outcomes['dbar_y_sum'] / weighted_outcomes['dbar_sum']
weighted_outcomes['effect'] = weighted_outcomes['d_y_mean'] - weighted_outcomes['dbar_y_mean']
total_treatment_population = weighted_outcomes[treatment_sum_name].sum()
ate = (weighted_outcomes['effect'] * weighted_outcomes[treatment_sum_name]).sum() / total_treatment_population
# TODO - how can we add additional information into the returned estimate?
# such as how much clipping was done, or per-strata info for debugging?
estimate = CausalEstimate(estimate=ate,
target_estimand=self._target_estimand,
realized_estimand_expr=self.symbolic_estimator,
propensity_scores = self._data["propensity_score"])
return estimate
[docs] def construct_symbolic_estimator(self, estimand):
expr = "b: " + ",".join(estimand.outcome_variable) + "~"
# TODO -- fix: we are actually conditioning on positive treatment (d=1)
var_list = estimand.treatment_variable + estimand.backdoor_variables
expr += "+".join(var_list)
return expr