CausalMatch is a Python package that implements two classic matching methods, propensity score matching (PSM) and coarsened exact matching (CEM), to estimate average treatment effects from observational data. This package was designed and built as part of the ByteDance data science research program with the goal of combining state-of-the-art machine learning techniques with econometrics to bring automation to complex causal inference problems. Our toolkit possess the following features:
- Implement classic matching techniques in the literature at the intersection of econometrics and machine learning
- Maintain flexibility in modeling the propensity score model (via various machine learning classification models), while preserving the causal interpretation of the learned model and often offering valid confidence intervals
- Use a unified API
- Build on standard Python packages for Machine Learning and Data Analysis
Table of Contents
If you'd like to contribute to this project, please contact xiaoyuzhou@bytedance.com. If you have any questions, feel free to raise them in the issues section.
August 20, 2024: Release v0.0.2, see release notes here
Previous releases
August 2, 2024: Release 0.0.1.
Install the latest release from [PyPI]:
pip install causalmatch==0.0.2
Propensity Score Matching (aka PSM) (click to expand)
- Simple PSM
from causalmatch import matching,gen_test_data
from sklearn.ensemble import GradientBoostingClassifier
import statsmodels.api as sm
df = gen_test_data(n = 10000, c_ratio=0.5)
df.head()
X = ['c_1', 'c_2', 'c_3', 'd_1', 'gender']
y = ['y', 'y2']
id = 'user_id'
T = 'treatment' # treatment variable must be binary with 0/1 values
# STEP 1: initialize object
match_obj = matching(data = df,
T = T,
X = X,
id = id)
# STEP 2: propensity score matching
match_obj.psm(n_neighbors = 1, # number of neighbors
model = GradientBoostingClassifier(), # p-score model
trim_percentage = 0.1, # trim x percent of data based on propensity score
caliper = 0.1) # caliper for p-score diff
# STEP 3: balance check after propensity score matching
match_obj.balance_check(include_discrete = True)
# STEP 4: obtain output dataframe, and merge X and y
df_out = match_obj.df_out_final_post_trim.merge(df[y + X + [id]], how='left', on = id)
# STEP 5: calculate average treatment effect on treated
X_mat = df_out[T]
y_mat = df_out[y]
X_mat = sm.add_constant(X_mat)
model = sm.OLS(y_mat,X_mat)
results = model.fit()
print(results.params)- PSM with multiple p-score models and select the best one based on f1 score
# STEP 0: define all classification model you need
from causalmatch import matching
import pandas as pd
import numpy as np
import statsmodels.api as sm
from sklearn.linear_model import LogisticRegression
from sklearn.svm import SVC
from sklearn.naive_bayes import GaussianNB
from sklearn.naive_bayes import MultinomialNB
from sklearn.neighbors import KNeighborsClassifier
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import RandomForestClassifier
from sklearn.ensemble import GradientBoostingClassifier
from lightgbm import LGBMClassifier
from xgboost import XGBClassifier
ps_model1 = LogisticRegression(C=1e6)
ps_model2 = SVC(probability=True)
ps_model3 = GaussianNB()
ps_model4 = KNeighborsClassifier()
ps_model5 = DecisionTreeClassifier()
ps_model6 = RandomForestClassifier()
ps_model7 = GradientBoostingClassifier()
ps_model8 = LGBMClassifier()
ps_model9 = XGBClassifier()
model_list = [ps_model1, ps_model2, ps_model3, ps_model4, ps_model5, ps_model6, ps_model7, ps_model8, ps_model9]
match_obj = matching(data = df, T = T, X = X, id = id)
match_obj.psm(n_neighbors = 1,
model_list = model_list, # input list of models you want to try
trim_percentage = 0,
caliper = 1,
test_size = 0.2) # train-test split, what portion does test sample takes
print(match_obj.balance_check(include_discrete = True))
df_out = match_obj.df_out_final_post_trim.merge(df[y + X + [id]], how='left', on = id)Coarsened Exact Matching (click to expand)
- Simple CEM
match_obj_cem = matching(data = df, y = ['y'], T = 'treatment', X = ['c_1','d_1','d_3'], id = 'user_id')
# coarsened exact matching
match_obj_cem.cem(n_bins = 10, # number of bins for continuous x variables, cut by percentile
k2k = True) # k2k: trim exp/base to have same observation numbers
print(match_obj_cem.balance_check(include_discrete=True))
print(match_obj_cem.ate())- CEM with customized bin cut
match_obj_cem = matching(data = df, y = ['y'], T = 'treatment', X = ['c_1','d_1','d_3'], id = 'user_id')
match_obj_cem.cem(n_bins = 10,
break_points = {'c_1': [-1, 0.3, 0.6, 2]}, # cut point for continuous variable
cluster_criteria = {'d_1': [['apple','pear'],['cat','dog'],['bee']],
'd_3': [['0.0','1.0','2.0'], ['3.0','4.0','5.0'], ['6.0','7.0','8.0','9.0']]}, # group values for discrete variables
k2k = True) See the References section for more details.
S. Athey, J. Tibshirani, S. Wager. Generalized random forests. Annals of Statistics, 47, no. 2, 1148--1178, 2019.
V. Chernozhukov, D. Nekipelov, V. Semenova, V. Syrgkanis. Plug-in Regularized Estimation of High-Dimensional Parameters in Nonlinear Semiparametric Models. Arxiv preprint arxiv:1806.04823, 2018.
S. Wager, S. Athey. Estimation and Inference of Heterogeneous Treatment Effects using Random Forests. Journal of the American Statistical Association, 113:523, 1228-1242, 2018.
V. Chernozhukov, D. Chetverikov, M. Demirer, E. Duflo, C. Hansen, and a. W. Newey. Double Machine Learning for Treatment and Causal Parameters. ArXiv preprint arXiv:1608.00060, 2016.