Subclassification Propensity Score Matching Using Python Package Causal Inference. Propensity score estimation, subclassification matching, and treatment effect estimation

Subclassification Propensity Score Matching Using Python Package Causal Inference

Subclassification matching in causal inference stratifies the propensity scores into bins, and the treatment and the control units within the bins are compared to get the treatment effects estimation.

In this tutorial, we will talk about how to do subclassification propensity score matching (PSM) using the Python CausalInference package.

To learn how to do subclassification matching using R, please check out my previous tutorial 8 Matching Methods for Causal Inference Using R

Resources for this post:

  • Click here for the Colab notebook.
  • More video tutorials on Causal Inference
  • More blog posts on Causal Inference
  • If you are not a Medium member and want to support me as a writer (😄 Buy me a cup of coffee ☕), join Medium membership through this link. You will get full access to posts on Medium for $5 per month, and I will receive a portion of it. Thank you for your support!
  • Give me a tip to show your appreciation
  • Video tutorial for this post on YouTube
Subclassification Propensity Score Matching –

Let’s get started!

Step 1: Install and Import Libraries

In step 1, we will install and import libraries.

Firstly, let’s install dowhy for dataset creation and causalinference for subclassification matching.

# Install dowhy
!pip install dowhy

# Install causal inference
!pip install causalinference

After the installation is completed, we can import the libraries.

  • The datasets is imported from dowhy for dataset creation.
  • pandas and numpy are imported for data processing.
  • CausalModel is imported from the causalinference package for subclassification matching.
# Package to create synthetic data for causal inference
from dowhy import datasets

# Data processing
import pandas as pd
import numpy as np

# Causal inference
from causalinference import CausalModel

Step 2:Create Dataset

In step 2, we will create a synthetic dataset for the causal inference.

  • Firstly, we set a random seed using np.random.seed to make the dataset reproducible.
  • Then a dataset with the true causal impact of 10, four confounders, 10,000 samples, a binary treatment variable, and a continuous outcome variable is created.
  • After that, we created a dataframe for the data. In the dataframe, the columns W0, W1, W2, and W3 are the four confounders, v0 is the treatment indicator, and y is the outcome.
# Set random seed

# Create a synthetic dataset
data = datasets.linear_dataset(

# Create Dataframe
df = data['df']

# Take a look at the data
Causal Inference Data —

Next, let’s rename v0 to treatment, rename y to outcome, and convert the boolean values to 0 and 1.

# Rename columns
df = df.rename({'v0': 'treatment', 'y': 'outcome'}, axis=1)

# Create the treatment variable, and change boolean values to 1 and 0
df['treatment'] = df['treatment'].apply(lambda x: 1 if x == True else 0)

# Take a look at the data
Causal Inference Data —

Step 3: Raw Difference

In step 3, we will initiate CausalModel and print the raw data summary statistics. CausalModel takes three arguments:

  • Y is the observed outcome.
  • D is the treatment indicator.
  • X is the covariates matrix.

CausalModel takes arrays as inputs, so .values are used when reading the data.

# Run causal model
causal = CausalModel(Y = df['outcome'].values, D = df['treatment'].values, X = df[['W0', 'W1', 'W2', 'W3']].values)

# Print summary statistics
Python CausalInference raw balance and difference —

causal.summary_stats prints out the raw summary statistics. The output shows that:

  • There are 2,269 units in the control group and 7,731 units in the treatment group.
  • The average outcome for the treatment group is 13.94, and the average outcome for the control group is -2.191. So the raw difference between the treatment and the control group is 16.132, which is much higher than the actual treatment effect of 10.
  • Nor-diff is the standardized mean difference (SMD) for covariates between the treatment group and the control group. Standardized Mean Differences(SMD) greater than 0.1 means that the data is imbalanced between the treatment and the control group. We can see that most of the covariates have SMD greater than 0.1.

Step 4: Propensity Score Estimation

In step 4, we will get the propensity score estimation. A propensity score is the predicted probability of getting treatment. It is calculated by running a logistic regression with the treatment variable as the target, and the covariates as the features.

There are two methods for propensity score estimation, est_propensity_s and est_propensity.

  • est_propensity allows users to add interaction or quadratic features.
  • est_propensity_s automatically choose the features based on a sequence of likelihood ratio tests.

In this step, we will use est_propensity_s to run the propensity score estimation.

# Automated propensity score estimation

# Propensity model results
Propensity Score Estimation –

From the model results, we can see that the feature selection algorithm decided to include only the raw features, and not include interaction or quadratic terms.

To get the propensity score, use causal.propensity['fitted'].

# Propensity scores


array([0.99295272, 0.99217314, 0.00156753, ..., 0.69143426, 0.99983862,

Step 5: Subclassification Matching by Propensity Score Stratification

In step 5, we will do the subclassification matching by stratifying the propensity scores.

The Python CausalInference package provides two methods for subclassification matching: stratify and stratify_s.

  • stratify split the units into five equal-sized bins by default. But we can change the hyperparameter value for blocks from 5 to other values to create different number of bins.
  • stratify_s uses a data-driven approach to set the number of bins and the bin boundaries automatically. It recursively divides the samples into two groups until there is no need for further splitting.

We will take the automatic approach for propensity score stratification.

# Propensity score stratification

# Print stratification summary
Subclassification Matching by Propensity Score Stratification

From the Stratification Summary, we can see that:

  • The samples are divided into 18 subclasses.
  • The minimum and maximum propensity scores for each subclass are listed.
  • There are more control units than treatment units in the subclasses with low propensity scores. Similarly, there are more treatment units than control units in the subclasses with high propensity scores.
  • The within-subclass average propensity scores are similar between the treated and the control group.
  • The average outcome raw differences for subclasses are mostly between 9 and 11, which is much closer to the true causal impact of 10 than the raw difference of 16.

Step 6: Subclassification Treatment Effect Estimation

In step 6, we will talk about how to use the subclassification propensity score matching results for treatment effects estimation.

The treatment effects estimation takes the following steps:

  • Step 1: After stratifying the samples into 18 subclasses, estimate the within-subclass treatment effects using ordinary least squares (OLS). To learn more about the ordinary least squares (OLS) treatment effect estimation, check out my previous tutorial OLS Treatment Effects Estimation Using Python Package Causal Inference
  • Step 2: Take the sample-weighted average of the within-subclass treatment effects from step 1 to get an overall treatment effects estimation.

Instead of manually calculating the treatment effects, the function est_via_blocking performs the two steps above and outputs the final causal impact estimations.

causal.estimates prints out the treatment effects estimation results.

  • ATE is the average treatment effect
  • ATC is the average treatment effect on the control
  • ATT is the average treatment effect on the treated

To learn more about the average treatment effect (ATE), the average treatment effect on the control (ATC), the average treatment effect on the treated (ATE), and how the values are calculated, please check out my previous tutorial ATE vs CATE vs ATT vs ATC for Causal Inference.

# Subclassification treatment effects estimation

# Print out the estimation results
Subclassification Treatment Effect Estimation –

From the treatment effect estimation results, we can see that the average treatment effect (ATE), the average treatment effect on the control (ATC), and the average treatment effect on the treated (ATE) are all the same as the true causal impact of 10, which is a much more accurate estimation than the raw difference of 16.

For more information about data science and machine learning, please check out my YouTube channel and Medium Page or follow me on LinkedIn.

Recommended Tutorials


[1] DoWhy Documentation

[2] CausalInference Documentation

Leave a Comment

Your email address will not be published. Required fields are marked *