Demo for the DoWhy causal API

We show a simple example of adding a causal extension to any dataframe.

import dowhy
dowhy.enable_notebook_rendering()

import dowhy.datasets
import dowhy.api
from dowhy.graph import build_graph_from_str

import numpy as np
import pandas as pd

from statsmodels.api import OLS

data = dowhy.datasets.linear_dataset(beta=5,
        num_common_causes=1,
        num_instruments = 0,
        num_samples=1000,
        treatment_is_binary=True)
df = data['df']
df['y'] = df['y'] + np.random.normal(size=len(df)) # Adding noise to data. Without noise, the variance in Y|X, Z is zero, and mcmc fails.
nx_graph = build_graph_from_str(data["dot_graph"])

treatment= data["treatment_name"][0]
outcome = data["outcome_name"][0]
common_cause = data["common_causes_names"][0]
df

	W0	v0	y
0	1.382710	True	8.060313
1	1.472841	True	8.334425
2	0.306800	True	7.820742
3	-0.430429	False	-0.455280
4	0.618401	True	6.239341
...	...	...	...
995	1.002819	True	9.206626
996	0.213227	True	6.627189
997	3.054601	True	10.629182
998	0.230412	True	6.199779
999	1.725729	True	8.568986

1000 rows × 3 columns

# data['df'] is just a regular pandas.DataFrame
df.causal.do(x=treatment,
             variable_types={treatment: 'b', outcome: 'c', common_cause: 'c'},
             outcome=outcome,
             common_causes=[common_cause],
            ).groupby(treatment).mean().plot(y=outcome, kind='bar')

<Axes: xlabel='v0'>

df.causal.do(x={treatment: 1},
              variable_types={treatment:'b', outcome: 'c', common_cause: 'c'},
              outcome=outcome,
              method='weighting',
              common_causes=[common_cause]
              ).groupby(treatment).mean().plot(y=outcome, kind='bar')

<Axes: xlabel='v0'>

cdf_1 = df.causal.do(x={treatment: 1},
              variable_types={treatment: 'b', outcome: 'c', common_cause: 'c'},
              outcome=outcome,
              graph=nx_graph
              )

cdf_0 = df.causal.do(x={treatment: 0},
              variable_types={treatment: 'b', outcome: 'c', common_cause: 'c'},
              outcome=outcome,
              graph=nx_graph
              )

cdf_0

	W0	v0	y	propensity_score	weight
0	1.857935	False	1.035500	0.044035	22.709352
1	1.441613	False	2.773858	0.080289	12.455070
2	1.161925	False	1.867316	0.118267	8.455410
3	2.289897	False	2.249772	0.023179	43.143088
4	-0.360114	False	-0.769929	0.581352	1.720127
...	...	...	...	...	...
995	0.859367	False	0.020092	0.175907	5.684812
996	0.394338	False	-0.685477	0.303598	3.293825
997	1.611883	False	1.782004	0.062979	15.878280
998	0.136252	False	-2.412841	0.393196	2.543263
999	2.017139	False	2.659729	0.034817	28.721824

1000 rows × 5 columns

cdf_1

	W0	v0	y	propensity_score	weight
0	0.413405	True	5.595959	0.702556	1.423374
1	1.392446	True	7.668865	0.913957	1.094144
2	0.726007	True	6.576239	0.792414	1.261967
3	2.509753	True	8.063277	0.983352	1.016930
4	-0.564910	True	4.925555	0.344610	2.901828
...	...	...	...	...	...
995	0.428467	True	5.023315	0.707367	1.413694
996	2.598522	True	8.081928	0.985443	1.014772
997	1.304708	True	6.729470	0.902753	1.107722
998	2.152605	True	7.396697	0.971536	1.029298
999	1.264880	True	6.376408	0.897250	1.114516

1000 rows × 5 columns

Comparing the estimate to Linear Regression

First, estimating the effect using the causal data frame, and the 95% confidence interval.

(cdf_1['y'] - cdf_0['y']).mean()

$\displaystyle 5.43679388164681$

1.96*(cdf_1['y'] - cdf_0['y']).std() / np.sqrt(len(df))

$\displaystyle 0.144027847284715$

Comparing to the estimate from OLS.

model = OLS(np.asarray(df[outcome]), np.asarray(df[[common_cause, treatment]], dtype=np.float64))
result = model.fit()
result.summary()

OLS Regression Results

Dep. Variable:	y	R-squared (uncentered):	0.974
Model:	OLS	Adj. R-squared (uncentered):	0.974
Method:	Least Squares	F-statistic:	1.864e+04
Date:	Sun, 25 Oct 2026	Prob (F-statistic):	0.00
Time:	21:11:19	Log-Likelihood:	-1394.5
No. Observations:	1000	AIC:	2793.
Df Residuals:	998	BIC:	2803.
Df Model:	2
Covariance Type:	nonrobust

	coef	std err	t	P>|t|	[0.025	0.975]
x1	1.4748	0.035	42.029	0.000	1.406	1.544
x2	5.0245	0.054	93.093	0.000	4.919	5.130

Omnibus:	3.385	Durbin-Watson:	2.003
Prob(Omnibus):	0.184	Jarque-Bera (JB):	3.256
Skew:	0.134	Prob(JB):	0.196
Kurtosis:	3.078	Cond. No.	2.97

Notes:
[1] R² is computed without centering (uncentered) since the model does not contain a constant.
[2] Standard Errors assume that the covariance matrix of the errors is correctly specified.