Applying refutation tests to the Lalonde and IHDP datasets

Import the Dependencies

[1]:

import dowhy
from dowhy import CausalModel
import pandas as pd
import numpy as np

# Config dict to set the logging level
import logging.config
DEFAULT_LOGGING = {
    'version': 1,
    'disable_existing_loggers': False,
    'loggers': {
        '': {
            'level': 'WARN',
        },
    }
}

logging.config.dictConfig(DEFAULT_LOGGING)
# Disabling warnings output
import warnings
from sklearn.exceptions import DataConversionWarning
warnings.filterwarnings(action='ignore', category=DataConversionWarning)

Loading the Dataset

Infant Health and Development Program Dataset (IHDP)

The measurements used are on the child—birth weight, head circumference, weeks bornpreterm, birth order, ﬁrst born, neonatal health index (see Scott and Bauer 1989), sex, twinstatus—as well as behaviors engaged in during the pregnancy—smoked cigarettes, drankalcohol, took drugs—and measurements on the mother at the time she gave birth—age,marital status, educational attainment (did not graduate from high school, graduated fromhigh school, attended some college but did not graduate, graduated from college), whethershe worked during pregnancy, whether she received prenatal care—and the site (8 total) inwhich the family resided at the start of the intervention. There are 6 continuous covariatesand 19 binary covariates.

Reference

Hill, J. L. (2011). Bayesian nonparametric modeling for causal inference. Journal of Computational and Graphical Statistics, 20(1), 217-240. https://doi.org/10.1198/jcgs.2010.08162

[2]:

data = pd.read_csv("https://raw.githubusercontent.com/AMLab-Amsterdam/CEVAE/master/datasets/IHDP/csv/ihdp_npci_1.csv", header = None)
col =  ["treatment", "y_factual", "y_cfactual", "mu0", "mu1" ,]
for i in range(1,26):
    col.append("x"+str(i))
data.columns = col
data = data.astype({"treatment":'bool'}, copy=False)
data.head()

[2]:

	treatment	y_factual	y_cfactual	mu0	mu1	x1	x2	x3	x4	x5	...	x16	x17	x18	x19
0	True	5.599916	4.318780	3.268256	6.854457	-0.528603	-0.343455	1.128554	0.161703	-0.316603	...	1	1	1	1
1	False	6.875856	7.856495	6.636059	7.562718	-1.736945	-1.802002	0.383828	2.244320	-0.629189	...	1	1	1	1
2	False	2.996273	6.633952	1.570536	6.121617	-0.807451	-0.202946	-0.360898	-0.879606	0.808706	...	1	0	1	1
3	False	1.366206	5.697239	1.244738	5.889125	0.390083	0.596582	-1.850350	-0.879606	-0.004017	...	1	0	1	1
4	False	1.963538	6.202582	1.685048	6.191994	-1.045229	-0.602710	0.011465	0.161703	0.683672	...	1	1	1	1

5 rows × 30 columns

Lalonde Dataset

A data frame with 445 observations on the following 12 variables.

age: age in years.
educ: years of schooling.
black: indicator variable for blacks.
hisp: indicator variable for Hispanics.
married: indicator variable for martial status.
nodegr: indicator variable for high school diploma.
re74: real earnings in 1974.
re75: real earnings in 1975.
re78: real earnings in 1978.
u74: indicator variable for earnings in 1974 being zero.
u75: indicator variable for earnings in 1975 being zero.
treat: an indicator variable for treatment status.

References

Dehejia, Rajeev and Sadek Wahba. 1999.``Causal Effects in Non-Experimental Studies: Re-Evaluating the Evaluation of Training Programs.’’ Journal of the American Statistical Association 94 (448): 1053-1062.

LaLonde, Robert. 1986. ``Evaluating the Econometric Evaluations of Training Programs.’’ American Economic Review 76:604-620.

[3]:

import dowhy.datasets

lalonde = dowhy.datasets.lalonde_dataset()

lalonde.head()

[3]:

	treat	age	educ	black	nodegr	re78	u74	u75
0	False	23.0	10.0	1.0	1.0	0.00	1.0	1.0
1	False	26.0	12.0	0.0	0.0	12383.68	1.0	1.0
2	False	22.0	9.0	1.0	1.0	0.00	1.0	1.0
3	False	18.0	9.0	1.0	1.0	10740.08	1.0	1.0
4	False	45.0	11.0	1.0	1.0	11796.47	1.0	1.0

Step 1: Building the model

IHDP

[4]:

# Create a causal model from the data and given common causes

common_causes = []

for i in range(1, 26):
    common_causes += ["x"+str(i)]

ihdp_model = CausalModel(
                data=data,
                treatment='treatment',
                outcome='y_factual',
                common_causes=common_causes
            )
ihdp_model.view_model(layout="dot")
from IPython.display import Image, display
display(Image(filename="causal_model.png"))

../_images/example_notebooks_dowhy_refutation_testing_12_0.png

../_images/example_notebooks_dowhy_refutation_testing_12_1.png

Lalonde

[5]:

lalonde_model = CausalModel(
                data=lalonde,
                treatment='treat',
                outcome='re78',
                common_causes='nodegr+black+hisp+age+educ+married'.split('+')
            )
lalonde_model.view_model(layout="dot")
from IPython.display import Image, display
display(Image(filename="causal_model.png"))

../_images/example_notebooks_dowhy_refutation_testing_14_0.png

../_images/example_notebooks_dowhy_refutation_testing_14_1.png

Step 2: Identification

IHDP

[6]:

#Identify the causal effect for the ihdp dataset
ihdp_identified_estimand = ihdp_model.identify_effect(proceed_when_unidentifiable=True)
print(ihdp_identified_estimand)

Estimand type: EstimandType.NONPARAMETRIC_ATE

### Estimand : 1
Estimand name: backdoor
Estimand expression:
     d
────────────(E[y_factual|x16,x5,x24,x4,x2,x8,x18,x21,x7,x13,x11,x25,x14,x1,x15
d[treatment]


,x12,x19,x6,x20,x22,x9,x17,x10,x3,x23])

Estimand assumption 1, Unconfoundedness: If U→{treatment} and U→y_factual then P(y_factual|treatment,x16,x5,x24,x4,x2,x8,x18,x21,x7,x13,x11,x25,x14,x1,x15,x12,x19,x6,x20,x22,x9,x17,x10,x3,x23,U) = P(y_factual|treatment,x16,x5,x24,x4,x2,x8,x18,x21,x7,x13,x11,x25,x14,x1,x15,x12,x19,x6,x20,x22,x9,x17,x10,x3,x23)

### Estimand : 2
Estimand name: iv
No such variable(s) found!

### Estimand : 3
Estimand name: frontdoor
No such variable(s) found!

Lalonde

[7]:

#Identify the causal effect for the lalonde dataset
lalonde_identified_estimand = lalonde_model.identify_effect(proceed_when_unidentifiable=True)
print(lalonde_identified_estimand)

Estimand type: EstimandType.NONPARAMETRIC_ATE

### Estimand : 1
Estimand name: backdoor
Estimand expression:
   d
────────(E[re78|hisp,age,married,educ,nodegr,black])
d[treat]
Estimand assumption 1, Unconfoundedness: If U→{treat} and U→re78 then P(re78|treat,hisp,age,married,educ,nodegr,black,U) = P(re78|treat,hisp,age,married,educ,nodegr,black)

### Estimand : 2
Estimand name: iv
No such variable(s) found!

### Estimand : 3
Estimand name: frontdoor
No such variable(s) found!

Step 3: Estimation (using propensity score weighting)

IHDP

[8]:

ihdp_estimate = ihdp_model.estimate_effect(
                    ihdp_identified_estimand,
                    method_name="backdoor.propensity_score_weighting"
                )

print("The Causal Estimate is " + str(ihdp_estimate.value))

The Causal Estimate is 4.0287482183898895

Lalonde

[9]:

lalonde_estimate = lalonde_model.estimate_effect(
                        lalonde_identified_estimand,
                        method_name="backdoor.propensity_score_weighting"
                    )

print("The Causal Estimate is " + str(lalonde_estimate.value))

The Causal Estimate is 1639.7944838972553

Step 4: Refutation

IHDP

Add Random Common Cause

[10]:

ihdp_refute_random_common_cause = ihdp_model.refute_estimate(
                                        ihdp_identified_estimand,
                                        ihdp_estimate,
                                        method_name="random_common_cause"
                                    )

print(ihdp_refute_random_common_cause)

Refute: Add a random common cause
Estimated effect:4.0287482183898895
New effect:4.0287482183898895
p value:1.0

Replace Treatment with Placebo

[11]:

ihdp_refute_placebo_treatment = ihdp_model.refute_estimate(
                                    ihdp_identified_estimand,
                                    ihdp_estimate,
                                    method_name="placebo_treatment_refuter",
                                    placebo_type="permute"
                                )

print(ihdp_refute_placebo_treatment)

Refute: Use a Placebo Treatment
Estimated effect:4.0287482183898895
New effect:-0.4445285513480844
p value:0.02

Remove Random Subset of Data

[12]:

ihdp_refute_random_subset = ihdp_model.refute_estimate(
                                    ihdp_identified_estimand,
                                    ihdp_estimate,
                                    method_name="data_subset_refuter",
                                    subset_fraction=0.8
                            )
print(ihdp_refute_random_subset)

Refute: Use a subset of data
Estimated effect:4.0287482183898895
New effect:4.026788563574814
p value:0.94

Lalonde

Add Random Common Cause

[13]:

lalonde_refute_random_common_cause = lalonde_model.refute_estimate(
                                            lalonde_identified_estimand,
                                            lalonde_estimate,
                                            method_name="random_common_cause"
                                        )

print(lalonde_refute_random_common_cause)

Refute: Add a random common cause
Estimated effect:1639.7944838972553
New effect:1639.7944838972544
p value:1.0

Replace Treatment with Placebo

[14]:

lalonde_refute_placebo_treatment = lalonde_model.refute_estimate(
                                        lalonde_identified_estimand,
                                        lalonde_estimate,
                                        method_name="placebo_treatment_refuter",
                                        placebo_type="permute"
                                    )

print(lalonde_refute_placebo_treatment)

Refute: Use a Placebo Treatment
Estimated effect:1639.7944838972553
New effect:-151.04563018505425
p value:0.92

Remove Random Subset of Data

[15]:

lalonde_refute_random_subset = lalonde_model.refute_estimate(
                                    lalonde_identified_estimand,
                                    lalonde_estimate,
                                    method_name="data_subset_refuter",
                                    subset_fraction=0.9
                                )

print(lalonde_refute_random_subset)

Refute: Use a subset of data
Estimated effect:1639.7944838972553
New effect:1605.3237581253386
p value:0.96