import inspect
from importlib import import_module
from typing import Any, Callable, List, Optional, Protocol, Union
from warnings import warn
import numpy as np
import pandas as pd
from numpy.distutils.misc_util import is_sequence
from dowhy.causal_estimator import CausalEstimate, CausalEstimator
from dowhy.causal_identifier import IdentifiedEstimand
from dowhy.utils.api import parse_state
class _EconmlEstimator(Protocol):
def fit(self, *args, **kwargs):
...
def effect(self, *args, **kwargs):
...
def effect_interval(self, *args, **kwargs):
...
def effect_inference(self, *args, **kwargs):
...
def shap_values(self, *args, **kwargs):
...
[docs]class Econml(CausalEstimator):
"""Wrapper class for estimators from the EconML library.
Supports additional parameters as listed below. For init and fit
parameters of each estimator, refer to the EconML docs.
"""
def __init__(
self,
identified_estimand: IdentifiedEstimand,
econml_estimator: Union[_EconmlEstimator, str],
test_significance: bool = False,
evaluate_effect_strength: bool = False,
confidence_intervals: bool = False,
num_null_simulations: int = CausalEstimator.DEFAULT_NUMBER_OF_SIMULATIONS_STAT_TEST,
num_simulations: int = CausalEstimator.DEFAULT_NUMBER_OF_SIMULATIONS_CI,
sample_size_fraction: int = CausalEstimator.DEFAULT_SAMPLE_SIZE_FRACTION,
confidence_level: float = CausalEstimator.DEFAULT_CONFIDENCE_LEVEL,
need_conditional_estimates: Union[bool, str] = "auto",
num_quantiles_to_discretize_cont_cols: int = CausalEstimator.NUM_QUANTILES_TO_DISCRETIZE_CONT_COLS,
**kwargs,
):
"""
:param identified_estimand: probability expression
representing the target identified estimand to estimate.
:param econml_methodname: Fully qualified name of econml estimator
class. For example, 'econml.dml.DML'
:param test_significance: Binary flag or a string indicating whether to test significance and by which method. All estimators support test_significance="bootstrap" that estimates a p-value for the obtained estimate using the bootstrap method. Individual estimators can override this to support custom testing methods. The bootstrap method supports an optional parameter, num_null_simulations. If False, no testing is done. If True, significance of the estimate is tested using the custom method if available, otherwise by bootstrap.
:param evaluate_effect_strength: (Experimental) whether to evaluate the strength of effect
:param confidence_intervals: Binary flag or a string indicating whether the confidence intervals should be computed and which method should be used. All methods support estimation of confidence intervals using the bootstrap method by using the parameter confidence_intervals="bootstrap". The bootstrap method takes in two arguments (num_simulations and sample_size_fraction) that can be optionally specified in the params dictionary. Estimators may also override this to implement their own confidence interval method. If this parameter is False, no confidence intervals are computed. If True, confidence intervals are computed by the estimator's specific method if available, otherwise through bootstrap
:param num_null_simulations: The number of simulations for testing the
statistical significance of the estimator
:param num_simulations: The number of simulations for finding the
confidence interval (and/or standard error) for a estimate
:param sample_size_fraction: The size of the sample for the bootstrap
estimator
:param confidence_level: The confidence level of the confidence
interval estimate
:param need_conditional_estimates: Boolean flag indicating whether
conditional estimates should be computed. Defaults to True if
there are effect modifiers in the graph
:param num_quantiles_to_discretize_cont_cols: The number of quantiles
into which a numeric effect modifier is split, to enable
estimation of conditional treatment effect over it.
:param kwargs: (optional) Additional estimator-specific parameters
"""
super().__init__(
identified_estimand=identified_estimand,
test_significance=test_significance,
evaluate_effect_strength=evaluate_effect_strength,
confidence_intervals=confidence_intervals,
num_null_simulations=num_null_simulations,
num_simulations=num_simulations,
sample_size_fraction=sample_size_fraction,
confidence_level=confidence_level,
need_conditional_estimates=need_conditional_estimates,
num_quantiles_to_discretize_cont_cols=num_quantiles_to_discretize_cont_cols,
econml_estimator=econml_estimator,
**kwargs,
)
if isinstance(econml_estimator, str):
warn(
"Using a string to specify the value for econml_estimator is now deprecated, please provide an instance of a econml object",
DeprecationWarning,
stacklevel=2,
)
estimator_class = self._get_econml_class_object(econml_estimator)
self.estimator = estimator_class(**kwargs["init_params"])
else:
self.estimator = econml_estimator
self.logger.info("INFO: Using EconML Estimator")
self.identifier_method = self._target_estimand.identifier_method
[docs] def fit(
self,
data: pd.DataFrame,
treatment_name: str,
outcome_name: str,
effect_modifier_names: Optional[List[str]] = None,
**kwargs,
):
"""
Fits the estimator with data for effect estimation
:param data: data frame containing the data
:param treatment: name of the treatment variable
:param outcome: name of the outcome variable
:param effect_modifiers: Variables on which to compute separate
effects, or return a heterogeneous effect function. Not all
methods support this currently.
"""
self._set_data(data, treatment_name, outcome_name)
self._set_effect_modifiers(effect_modifier_names)
# Save parameters for later refutter fitting
self._econml_fit_params = kwargs
self._observed_common_causes_names = self._target_estimand.get_backdoor_variables().copy()
# Enforcing this ordering is necessary to feed through the propensity values from dataset
self._observed_common_causes_names = [
c for c in self._observed_common_causes_names if "propensity" not in c
] + sorted([c for c in self._observed_common_causes_names if "propensity" in c])
# For metalearners only--issue a warning if w contains variables not in x
if self.estimator.__module__.endswith("metalearners"):
effect_modifier_names = []
if self._effect_modifier_names is not None:
effect_modifier_names = self._effect_modifier_names.copy()
w_diff_x = [w for w in self._observed_common_causes_names if w not in effect_modifier_names]
if len(w_diff_x) > 0:
self.logger.warn(
"Concatenating common_causes and effect_modifiers and providing a single list of variables to metalearner estimator method, "
+ self.estimator.__class__.__name__
+ ". EconML metalearners accept a single X argument."
)
effect_modifier_names.extend(w_diff_x)
# Override the effect_modifiers set in CausalEstimator.__init__()
# Also only update self._effect_modifiers, and create a copy of self._effect_modifier_names
# the latter can be used by other estimator methods later
self._effect_modifiers = self._data[effect_modifier_names]
self._effect_modifiers = pd.get_dummies(self._effect_modifiers, drop_first=True)
self._effect_modifier_names = effect_modifier_names
self.logger.debug("Effect modifiers: " + ",".join(effect_modifier_names))
if self._observed_common_causes_names:
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
self.logger.debug("Back-door variables used:" + ",".join(self._observed_common_causes_names))
# Instrumental variables names, if present
# choosing the instrumental variable to use
if getattr(self, "iv_instrument_name", None) is None:
self.estimating_instrument_names = self._target_estimand.instrumental_variables
else:
self.estimating_instrument_names = parse_state(self.iv_instrument_name)
if self.estimating_instrument_names:
self._estimating_instruments = self._data[self.estimating_instrument_names]
self._estimating_instruments = pd.get_dummies(self._estimating_instruments, drop_first=True)
else:
self._estimating_instruments = None
self.symbolic_estimator = self.construct_symbolic_estimator(self._target_estimand)
self.logger.info(self.symbolic_estimator)
X = None
W = None # common causes/ confounders
Z = None # Instruments
Y = self._outcome
T = self._treatment
if self._effect_modifiers is not None and len(self._effect_modifiers) > 0:
X = self._effect_modifiers
if self._observed_common_causes_names:
W = self._observed_common_causes
if self.estimating_instrument_names:
Z = self._estimating_instruments
named_data_args = {"Y": Y, "T": T, "X": X, "W": W, "Z": Z}
# Calling the econml estimator's fit method
estimator_argspec = inspect.getfullargspec(inspect.unwrap(self.estimator.fit))
# As of v0.9, econml has some kewyord only arguments
estimator_named_args = estimator_argspec.args + estimator_argspec.kwonlyargs
estimator_data_args = {
arg: named_data_args[arg] for arg in named_data_args.keys() if arg in estimator_named_args
}
self.estimator.fit(**estimator_data_args, **kwargs)
return self
def _get_econml_class_object(self, module_method_name, *args, **kwargs):
# from https://www.bnmetrics.com/blog/factory-pattern-in-python3-simple-version
try:
(module_name, _, class_name) = module_method_name.rpartition(".")
estimator_module = import_module(module_name)
estimator_class = getattr(estimator_module, class_name)
except (AttributeError, AssertionError, ImportError):
raise ImportError(
"Error loading {}.{}. Double-check the method name and ensure that all econml dependencies are installed.".format(
module_name, class_name
)
)
return estimator_class
[docs] def estimate_effect(
self, data: pd.DataFrame = None, treatment_value: Any = 1, control_value: Any = 0, target_units=None, **_
):
"""
data: dataframe containing the data on which treatment effect is to be estimated.
treatment_value: value of the treatment variable for which the effect is to be estimated.
control_value: value of the treatment variable that denotes its absence (usually 0)
target_units: The units for which the treatment effect should be estimated.
It can be a DataFrame that contains values of the effect_modifiers and effect will be estimated only for this new data.
It can also be a lambda function that can be used as an index for the data (pandas DataFrame) to select the required rows.
"""
if data is None:
data = self._data
self._target_units = target_units
self._treatment_value = treatment_value
self._control_value = control_value
n_samples = self._treatment.shape[0]
X = None # Effect modifiers
if self._effect_modifiers is not None and len(self._effect_modifiers) > 0:
X = self._effect_modifiers
X_test = X
if X is not None:
if type(target_units) is pd.DataFrame:
X_test = target_units
elif callable(target_units):
filtered_rows = data.where(target_units)
boolean_criterion = np.array(filtered_rows.notnull().iloc[:, 0])
X_test = X[boolean_criterion]
# Changing shape to a list for a singleton value
self._treatment_value = parse_state(self._treatment_value)
est = self.effect(X_test)
ate = np.mean(est, axis=0) # one value per treatment value
if len(ate) == 1:
ate = ate[0]
if self._confidence_intervals:
self.effect_intervals = self.effect_interval(X_test)
else:
self.effect_intervals = None
estimate = CausalEstimate(
estimate=ate,
control_value=control_value,
treatment_value=treatment_value,
target_estimand=self._target_estimand,
realized_estimand_expr=self.symbolic_estimator,
cate_estimates=est,
effect_intervals=self.effect_intervals,
_estimator_object=self.estimator,
)
estimate.add_estimator(self)
return estimate
def _estimate_confidence_intervals(self, confidence_level=None, method=None):
"""Returns None if the confidence interval has not been calculated."""
return self.effect_intervals
def _do(self, x):
raise NotImplementedError
[docs] def construct_symbolic_estimator(self, estimand):
expr = "b: " + ", ".join(estimand.outcome_variable) + "~"
# TODO -- fix: we are actually conditioning on positive treatment (d=1)
if self.estimator.__module__.endswith("metalearners"):
var_list = estimand.treatment_variable + self._effect_modifier_names
expr += "+".join(var_list)
else:
var_list = estimand.treatment_variable + self._observed_common_causes_names
expr += "+".join(var_list)
expr += " | " + ",".join(self._effect_modifier_names)
return expr
[docs] def shap_values(self, df: pd.DataFrame, *args, **kwargs):
return self.estimator.shap_values(df[self._effect_modifier_names].values, *args, **kwargs)
[docs] def apply_multitreatment(self, df: pd.DataFrame, fun: Callable, *args, **kwargs):
ests = []
assert not isinstance(self._treatment_value, str)
assert is_sequence(self._treatment_value)
if df is None:
filtered_df = None
else:
filtered_df = df[self._effect_modifier_names].values
for tv in self._treatment_value:
ests.append(
fun(
filtered_df,
T0=self._control_value,
T1=tv,
*args,
**kwargs,
)
)
est = np.stack(ests, axis=1)
return est
[docs] def effect(self, df: pd.DataFrame, *args, **kwargs) -> np.ndarray:
"""
Pointwise estimated treatment effect,
output shape n_units x n_treatment_values (not counting control)
:param df: Features of the units to evaluate
:param args: passed through to the underlying estimator
:param kwargs: passed through to the underlying estimator
"""
def effect_fun(filtered_df, T0, T1, *args, **kwargs):
return self.estimator.effect(filtered_df, T0=T0, T1=T1, *args, **kwargs)
return self.apply_multitreatment(df, effect_fun, *args, **kwargs)
[docs] def effect_interval(self, df: pd.DataFrame, *args, **kwargs) -> np.ndarray:
"""
Pointwise confidence intervals for the estimated treatment effect
:param df: Features of the units to evaluate
:param args: passed through to the underlying estimator
:param kwargs: passed through to the underlying estimator
"""
def effect_interval_fun(filtered_df, T0, T1, *args, **kwargs):
return self.estimator.effect_interval(
filtered_df, T0=T0, T1=T1, alpha=1 - self.confidence_level, *args, **kwargs
)
return self.apply_multitreatment(df, effect_interval_fun, *args, **kwargs)
[docs] def effect_inference(self, df: pd.DataFrame, *args, **kwargs):
"""
Inference (uncertainty) results produced by the underlying EconML estimator
:param df: Features of the units to evaluate
:param args: passed through to the underlying estimator
:param kwargs: passed through to the underlying estimator
"""
def effect_inference_fun(filtered_df, T0, T1, *args, **kwargs):
return self.estimator.effect_inference(filtered_df, T0=T0, T1=T1, *args, **kwargs)
return self.apply_multitreatment(df, effect_inference_fun, *args, **kwargs)
[docs] def effect_tt(self, df: pd.DataFrame, *args, **kwargs):
"""
Effect of the actual treatment that was applied to each unit
("effect of Treatment on the Treated")
:param df: Features of the units to evaluate
:param args: passed through to estimator.effect()
:param kwargs: passed through to estimator.effect()
"""
eff = self.effect(df, *args, **kwargs).reshape((len(df), len(self._treatment_value)))
out = np.zeros(len(df))
treatment_value = parse_state(self._treatment_value)
treatment_name = parse_state(self._treatment_name)[0]
eff = np.reshape(eff, (len(df), len(treatment_value)))
# For each unit, return the estimated effect of the treatment value
# that was actually applied to the unit
for c, col in enumerate(treatment_value):
out[df[treatment_name] == col] = eff[df[treatment_name] == col, c]
return pd.Series(data=out, index=df.index)