BCAUS

Reference: Belthangady et al, Minimizing bias in massive multi-arm observational studies with BCAUS: balancing covariates automatically using supervision, BMC Medical Research Methodology 2021

BCAUS on IHDP

from causalforge.model import Model , PROBLEM_TYPE
from causalforge.data_loader import DataLoader

# load IHDP dataset
r = DataLoader.get_loader('IHDP').load()
X_tr, T_tr, YF_tr, YCF_tr, mu_0_tr, mu_1_tr, X_te, T_te, YF_te, YCF_te, mu_0_te, mu_1_te = r

# model
params={}


bcaus_dr = Model.create_model("bcaus_dr",
                               params,
                               problem_type=PROBLEM_TYPE.CAUSAL_TREATMENT_EFFECT_ESTIMATION,
                               multiple_treatments=False)

from causalforge.metrics import eps_ATE_diff, PEHE_with_ite
import numpy as np

experiment_ids = [1,10,400]

eps_ATE_tr, eps_ATE_te = [], []
eps_PEHE_tr, eps_PEHE_te = [] , []



for idx in experiment_ids:
    t_tr, y_tr, x_tr, mu0tr, mu1tr = T_tr[:,idx] , YF_tr[:,idx], X_tr[:,:,idx], mu_0_tr[:,idx], mu_1_tr[:,idx]
    t_te, y_te, x_te, mu0te, mu1te = T_te[:,idx] , YF_te[:,idx], X_te[:,:,idx], mu_0_te[:,idx], mu_1_te[:,idx]


    # Train your causal method on train-set ...
    bcaus_dr.fit(x_tr,t_tr,y_tr)

    # Validate your method test-set ...
    ATE_truth_tr = (mu1tr - mu0tr).mean()
    ATE_truth_te = (mu1te - mu0te).mean()

    ITE_truth_tr = (mu1tr - mu0tr)
    ITE_truth_te = (mu1te - mu0te)

    eps_ATE_tr.append( eps_ATE_diff( bcaus_dr.predict_ate(x_tr,t_tr,y_tr), ATE_truth_tr) )
    eps_ATE_te.append( eps_ATE_diff( bcaus_dr.predict_ate(x_te,t_te,y_te), ATE_truth_te) )

Results

import pandas as pd

pd.DataFrame([[np.mean(eps_ATE_tr),np.mean(eps_ATE_te)]],
             columns=['eps_ATE_tr','eps_ATE_te'],
             index=['BCAUS_DR'])

	eps_ATE_tr	eps_ATE_te
BCAUS_DR	0.012877	0.260656