dowhy.causal_refuters package
Submodules
dowhy.causal_refuters.add_unobserved_common_cause module
- class dowhy.causal_refuters.add_unobserved_common_cause.AddUnobservedCommonCause(*args, **kwargs)[source]
Bases:
CausalRefuter
Add an unobserved confounder for refutation.
Supports additional parameters that can be specified in the refute_estimate() method.
‘confounders_effect_on_treatment’: how the simulated confounder affects the value of treatment. This can be linear (for continuous treatment) or binary_flip (for binary treatment)
‘confounders_effect_on_outcome’: how the simulated confounder affects the value of outcome. This can be linear (for continuous outcome) or binary_flip (for binary outcome)
‘effect_strength_on_treatment’: parameter for the strength of the effect of simulated confounder on treatment. For linear effect, it is the regression coeffient. For binary_flip, it is the probability that simulated confounder’s effect flips the value of treatment from 0 to 1 (or vice-versa).
‘effect_strength_on_outcome’: parameter for the strength of the effect of simulated confounder on outcome. For linear effect, it is the regression coeffient. For binary_flip, it is the probability that simulated confounder’s effect flips the value of outcome from 0 to 1 (or vice-versa).
TODO: Needs scaled version of the parameters and an interpretation module (e.g., in comparison to biggest effect of known confounder)
Initialize the parameters required for the refuter
- Parameters
effect_on_t – str : This is used to represent the type of effect on the treatment due to the unobserved confounder.
effect_on_y – str : This is used to represent the type of effect on the outcome due to the unobserved confounder.
kappa_t – float, numpy.ndarray: This refers to the strength of the confounder on treatment. For a linear effect, it behaves like the regression coeffecient. For a binary flip it is the probability with which it can invert the value of the treatment.
kappa_y – floar, numpy.ndarray: This refers to the strength of the confounder on outcome. For a linear effect, it behaves like the regression coefficient. For a binary flip, it is the probability with which it can invert the value of the outcome.
- generate_confounder_from_residuals(c1, c2, d_y, d_t, X)[source]
This function takes the residuals from the treatment and outcome model and their coefficients and simulates the intermediate random variable U by taking the row wise normal distribution corresponding to each residual value and then debiasing the intermediate variable to get the final variable
- Parameters
c1 – coefficient to the residual from the outcome model
:type float :param c2: coefficient to the residual from the treatment model :type float :param d_y: residuals from the outcome model :type list :param d_t: residuals from the treatment model :type list
- Returns final_U
The simulated values of the unobserved confounder based on the data
:type pandas.core.series.Series
- include_confounders_effect(new_data, kappa_t, kappa_y)[source]
This function deals with the change in the value of the data due to the effect of the unobserved confounder. In the case of a binary flip, we flip only if the random number is greater than the threshold set. In the case of a linear effect, we use the variable as the linear regression constant.
- Parameters
new_data – pandas.DataFrame: The data to be changed due to the effects of the unobserved confounder.
kappa_t – numpy.float64: The value of the threshold for binary_flip or the value of the regression coefficient for linear effect.
kappa_y – numpy.float64: The value of the threshold for binary_flip or the value of the regression coefficient for linear effect.
- Returns
pandas.DataFrame: The DataFrame that includes the effects of the unobserved confounder.
- include_simulated_confounder(convergence_threshold=0.1, c_star_max=1000)[source]
- This function simulates an unobserved confounder based on the data using the following steps:
- It calculates the “residuals” from the treatment and outcome model
i.) The outcome model has outcome as the dependent variable and all the observed variables including treatment as independent variables ii.) The treatment model has treatment as the dependent variable and all the observed variables as independent variables.
U is an intermediate random variable drawn from the normal distribution with the weighted average of residuals as mean and a unit variance U ~ N(c1*d_y + c2*d_t, 1) where
The final U, which is the simulated unobserved confounder is obtained by debiasing the intermediate variable U by residualising it with X
Choosing the coefficients c1 and c2: The coefficients are chosen based on these basic assumptions:
There is a hyperbolic relationship satisfying c1*c2 = c_star
c_star is chosen from a range of possible values based on the correlation of the obtained simulated variable with outcome and treatment.
The product of correlations with treatment and outcome should be at a minimum distance to the maximum correlations with treatment and outcome in any of the observed confounders
The ratio of the weights should be such that they maintain the ratio of the maximum possible observed coefficients within some confidence interval
- Parameters
c_star_max – The maximum possible value for the hyperbolic curve on which the coefficients to the residuals lie. It defaults to 1000 in the code if not specified by the user.
:type int :param convergence_threshold: The threshold to check the plateauing of the correlation while selecting a c_star. It defaults to 0.1 in the code if not specified by the user :type float :returns final_U: The simulated values of the unobserved confounder based on the data :type pandas.core.series.Series
- refute_estimate()[source]
This function attempts to add an unobserved common cause to the outcome and the treatment. At present, we have implemented the behavior for one dimensional behaviors for continueous and binary variables. This function can either take single valued inputs or a range of inputs. The function then looks at the data type of the input and then decides on the course of action.
- Returns
CausalRefuter: An object that contains the estimated effect and a new effect and the name of the refutation used.
dowhy.causal_refuters.bootstrap_refuter module
- class dowhy.causal_refuters.bootstrap_refuter.BootstrapRefuter(*args, **kwargs)[source]
Bases:
CausalRefuter
Refute an estimate by running it on a random sample of the data containing measurement error in the confounders. This allows us to find the ability of the estimator to find the effect of the treatment on the outcome.
It supports additional parameters that can be specified in the refute_estimate() method.
- Parameters
num_simulations (int, optional) – The number of simulations to be run,
CausalRefuter.DEFAULT_NUM_SIMULATIONS
by defaultsample_size (int, optional) – The size of each bootstrap sample and is the size of the original data by default
required_variables (int, list, bool, optional) – The list of variables to be used as the input for
y~f(W)
This isTrue
by default, which in turn selects all variables leaving the treatment and the outcome
An integer argument refers to how many variables will be used for estimating the value of the outcome
A list explicitly refers to which variables will be used to estimate the outcome Furthermore, it gives the ability to explictly select or deselect the covariates present in the estimation of the outcome. This is done by either adding or explicitly removing variables from the list as shown below:
Note
We need to pass required_variables =
[W0,W1]
if we wantW0
andW1
.We need to pass required_variables =
[-W0,-W1]
if we want all variables excludingW0
andW1
.
If the value is True, we wish to include all variables to estimate the value of the outcome.
Warning
A
False
value isINVALID
and will result in anerror
.- Parameters
noise (float, optional) – The standard deviation of the noise to be added to the data and is
BootstrapRefuter.DEFAULT_STD_DEV
by defaultprobability_of_change (float, optional) – It specifies the probability with which we change the data for a boolean or categorical variable It is
noise
by default, only if the value ofnoise
is less than 1.random_state (int, RandomState, optional) – The seed value to be added if we wish to repeat the same random behavior. For this purpose, we repeat the same seed in the psuedo-random generator.
- DEFAULT_NUMBER_OF_TRIALS = 1
- DEFAULT_STD_DEV = 0.1
- DEFAULT_SUCCESS_PROBABILITY = 0.5
dowhy.causal_refuters.data_subset_refuter module
- class dowhy.causal_refuters.data_subset_refuter.DataSubsetRefuter(*args, **kwargs)[source]
Bases:
CausalRefuter
Refute an estimate by rerunning it on a random subset of the original data.
Supports additional parameters that can be specified in the refute_estimate() method.
- Parameters
subset_fraction (float, optional) – Fraction of the data to be used for re-estimation, which is
DataSubsetRefuter.DEFAULT_SUBSET_FRACTION
by default.num_simulations (int, optional) – The number of simulations to be run, which is
CausalRefuter.DEFAULT_NUM_SIMULATIONS
by defaultrandom_state (int, RandomState, optional) – The seed value to be added if we wish to repeat the same random behavior. If we with to repeat the same behavior we push the same seed in the psuedo-random generator
- DEFAULT_SUBSET_FRACTION = 0.8
dowhy.causal_refuters.dummy_outcome_refuter module
- class dowhy.causal_refuters.dummy_outcome_refuter.DummyOutcomeRefuter(*args, **kwargs)[source]
Bases:
CausalRefuter
Refute an estimate by replacing the outcome with a simulated variable for which the true causal effect is known.
In the simplest case, the dummy outcome is an independent, randomly generated variable. By definition, the true causal effect should be zero.
More generally, the dummy outcome uses the observed relationship between confounders and outcome (conditional on treatment) to create a more realistic outcome for which the treatment effect is known to be zero. If the goal is to simulate a dummy outcome with a non-zero true causal effect, then we can add an arbitrary function h(t) to the dummy outcome’s generation process and then the causal effect becomes h(t=1)-h(t=0).
Note that this general procedure only works for the backdoor criterion.
1. We find f(W) for a each value of treatment. That is, keeping the treatment constant, we fit a predictor to estimate the effect of confounders W on outcome y. Note that since f(W) simply defines a new DGP for the simulated outcome, it need not be the correct structural equation from W to y. 2. We obtain the value of dummy outcome as:
y_dummy = h(t) + f(W)
To prevent overfitting, we fit f(W) for one value of T and then use it to generate data for other values of t. Future support for identification based on instrumental variable and mediation.
If we originally started out with W / \ t --->y On estimating the following with constant t, y_dummy = f(W) W / \ t --|->y This ensures that we try to capture as much of W--->Y as possible On adding h(t) W / \ t --->y h(t)
Supports additional parameters that can be specified in the refute_estimate() method.
- Parameters
num_simulations (int, optional) – The number of simulations to be run, which defaults to
CausalRefuter.DEFAULT_NUM_SIMULATIONS
transformation_list (list, optional) –
It is a list of actions to be performed to obtain the outcome, which defaults to
DummyOutcomeRefuter.DEFAULT_TRANSFORMATION
. The default transformation is as follows:[("zero",""),("noise", {'std_dev':1} )]
Each of the actions within a transformation is one of the following types:
function argument: function
pd.Dataframe -> np.ndarray
It takes in a function that takes the input data frame as the input and outputs the outcome variable. This allows us to create an output varable that only depends on the covariates and does not depend on the treatment variable.
string argument
Currently it supports some common estimators like
Linear Regression
K Nearest Neighbours
Support Vector Machine
Neural Network
Random Forest
Or functions such as:
Permute This permutes the rows of the outcome, disassociating any effect of the treatment on the outcome.
Noise This adds white noise to the outcome with white noise, reducing any causal relationship with the treatment.
Zero It replaces all the values in the outcome by zero
- Examples:
The
transformation_list
is of the following form:
If the function
pd.Dataframe -> np.ndarray
is already defined.[(func,func_params),('permute',{'permute_fraction':val}),('noise',{'std_dev':val})]
Every function should be able to support a minimum of two arguments
X_train
andoutcome_train
which correspond to the training data and the outcome that we want to predict, along with additional parameters such as the learning rate or the momentum constant can be set with the help offunc_args
.[(neural_network,{'alpha': 0.0001, 'beta': 0.9}),('permute',{'permute_fraction': 0.2}),('noise',{'std_dev': 0.1})]
The neural network is invoked as
neural_network(X_train, outcome_train, **args)
.If a function from the above list is used
[('knn',{'n_neighbors':5}), ('permute', {'permute_fraction': val} ), ('noise', {'std_dev': val} )]
- Parameters
true_causal_effect (function) – A function that is used to get the True Causal Effect for the modelled dummy outcome. It defaults to
DummyOutcomeRefuter.DEFAULT_TRUE_CAUSAL_EFFECT
, which means that there is no relationship between the treatment and outcome in the dummy data.
The equation for the dummy outcome is given by
y_hat = h(t) + f(W)
where
y_hat
is the dummy outcomeh(t)
is the function that gives the true causal effectf(W)
is the best estimate ofy
obtained keepingt
constant. This ensures that the variation in output of functionf(w)
is not caused byt
.
Note
The true causal effect should take an input of the same shape as the treatment and the output should match the shape of the outcome
- Parameters
required_variables (int, list, bool, optional) – The list of variables to be used as the input for
y~f(W)
This isTrue
by default, which in turn selects all variables leaving the treatment and the outcome
An integer argument refers to how many variables will be used for estimating the value of the outcome
A list explicitly refers to which variables will be used to estimate the outcome Furthermore, it gives the ability to explictly select or deselect the covariates present in the estimation of the outcome. This is done by either adding or explicitly removing variables from the list as shown below:
Note
We need to pass required_variables =
[W0,W1]
if we wantW0
andW1
.We need to pass required_variables =
[-W0,-W1]
if we want all variables excludingW0
andW1
.
If the value is True, we wish to include all variables to estimate the value of the outcome.
Warning
A
False
value isINVALID
and will result in anerror
.Note
These inputs are fed to the function for estimating the outcome variable. The same set of required_variables is used for each instance of an internal estimation function.
- Parameters
bucket_size_scale_factor – For continuous data, the scale factor helps us scale the size of the bucket used on the data. The default scale factor is
DummyOutcomeRefuter.DEFAULT_BUCKET_SCALE_FACTOR
.min_data_point_threshold (int, optional) – The minimum number of data points for an estimator to run. This defaults to
DummyOutcomeRefuter.MIN_DATA_POINT_THRESHOLD
. If the number of data points is too few for a certain category, we make use of theDummyOutcomeRefuter.DEFAULT_TRANSFORMATION
for generaring the dummy outcome
- DEFAULT_BUCKET_SCALE_FACTOR = 0.5
- DEFAULT_NEW_DATA_WITH_UNOBSERVED_CONFOUNDING = None
- DEFAULT_STD_DEV = 0.1
- DEFAULT_TEST_FRACTION = [TestFraction(base=0.5, other=0.5)]
- DEFAULT_TRANSFORMATION = [('zero', ''), ('noise', {'std_dev': 1})]
- DEFAULT_TRUE_CAUSAL_EFFECT()
- MIN_DATA_POINT_THRESHOLD = 30
- SUPPORTED_ESTIMATORS = ['linear_regression', 'knn', 'svm', 'random_forest', 'neural_network']
- noise(outcome, std_dev)[source]
Add white noise with mean 0 and standard deviation = std_dev
‘outcome’: np.ndarray
The outcome variable, to which the white noise is added. - ‘std_dev’: float The standard deviation of the white noise.
- permute(outcome, permute_fraction)[source]
If the permute_fraction is 1, we permute all the values in the outcome. Otherwise we make use of the Fisher Yates shuffle. Refer to https://en.wikipedia.org/wiki/Fisher%E2%80%93Yates_shuffle for more details.
‘outcome’: np.ndarray The outcome variable to be permuted. ‘permute_fraction’: float [0, 1] The fraction of rows permuted.
- preprocess_data_by_treatment()[source]
This function groups data based on the data type of the treatment.
Expected variable types supported for the treatment:
bool
pd.categorical
float
int
- Returns
pandas.core.groupby.generic.DataFrameGroupBy
- process_data(X_train, outcome_train, X_validation, outcome_validation, transformation_list)[source]
We process the data by first training the estimators in the transformation_list on
X_train
andoutcome_train
. We then apply the estimators onX_validation
to get the value of the dummy outcome, which we store inoutcome_validation
.- Parameters
X_train (np.ndarray) – The data of the covariates which is used to train an estimator. It corresponds to the data of a single category of the treatment
outcome_train (np.ndarray) – This is used to hold the intermediate values of the outcome variable in the transformation list
For Example:
[ ('permute', {'permute_fraction': val} ), (func,func_params)]
The value obtained from permutation is used as an input for the custom estimator.
- Parameters
X_validation (np.ndarray) – The data of the covariates that is fed to a trained estimator to generate a dummy outcome
outcome_validation (np.ndarray) – This variable stores the dummy_outcome generated by the transformations
transformation_list (np.ndarray) – The list of transformations on the outcome data required to produce a dummy outcome
dowhy.causal_refuters.placebo_treatment_refuter module
- class dowhy.causal_refuters.placebo_treatment_refuter.PlaceboTreatmentRefuter(*args, **kwargs)[source]
Bases:
CausalRefuter
Refute an estimate by replacing treatment with a randomly-generated placebo variable.
Supports additional parameters that can be specified in the refute_estimate() method.
- Parameters
placebo_type (str, optional) – Default is to generate random values for the treatment. If placebo_type is “permute”, then the original treatment values are permuted by row.
num_simulations (int, optional) – The number of simulations to be run, which is
CausalRefuter.DEFAULT_NUM_SIMULATIONS
by defaultrandom_state (int, RandomState, optional) – The seed value to be added if we wish to repeat the same random behavior. If we want to repeat the same behavior we push the same seed in the psuedo-random generator.
- DEFAULT_MEAN_OF_NORMAL = 0
- DEFAULT_NUMBER_OF_TRIALS = 1
- DEFAULT_PROBABILITY_OF_BINOMIAL = 0.5
- DEFAULT_STD_DEV_OF_NORMAL = 0
dowhy.causal_refuters.random_common_cause module
- class dowhy.causal_refuters.random_common_cause.RandomCommonCause(*args, **kwargs)[source]
Bases:
CausalRefuter
Refute an estimate by introducing a randomly generated confounder (that may have been unobserved).