Compares three causal machine learning methods: propensity score matching with machine learning, double machine learning, and causal forest across various treatment effect, model complexities, data dimensions and sample sizes.

Simulates different data structures based on DGP. The baseline DGP is structured as following:

$y_i=\theta d_i+x_i' \beta + u_i$

$d_i=x_i' \beta + v_i$

The treatment effect $\theta$ is set to $\theta=1$. The $d_i$ represents the binary treatment variable (approximately 50% of the observations receive treatment); $x_i$ represents a vector of k covariates, generated from a multivariate normal distribution; $\beta$ is a vector of k parameters.

The baseline DGP is defined as $n=150$, $k=10$, $\theta=1$ (homogenous treatment). We can then systematically vary these parameters to progressively more challenging estimation problems. In particular, we vary the following parameters, while keeping the others constant:

• Increase sample size $(n=150, 500, 5000, 15000)$
• Increase number of covariates $(k=10, 100)$
• Impose treatment heterogeneity $(\theta=1; \theta \tilde Normal(1,1))$
• Change structure of data $(y_i = \theta d_i + x_i' \beta + u_i)$ and $(y_i = \theta d_i + sin(x_i' \beta) + u_i)$

Runs the following causal estimation methods:

• PSM with Probit regression (PSM-Probit)
• PSM with Neural Net (PSM-NN)
• PSM with Random Forest (PSM-RF)
• PSM with eXtreme Gradient Boosting (PSM-XGB)
• Double machine learning with (E[̂D|Z]) ̂estimated using Logistic Lasso and (E[Y|Z]) ̂estimated using Lasso regression (DML-Lasso/Lasso)
• Double machine learning with (E[̂D|Z]) ̂estimated using logistic regression and (E[Y|Z]) ̂estimated using random forest (DML-Logit/RF)
• Double machine learning with (E[̂D|Z]) ̂estimated using classification neural network and (E[Y|Z]) ̂estimated using Random Forest (DML-NN/RF)
• Double machine learning with (E[̂D|Z]) ̂estimated using Logistic Lasso and (E[Y|Z]) ̂estimated using Random Forest (DML-Lasso/RF)
• Double machine learning with both (E[̂D|Z]) ̂and (E[Y|Z]) ̂estimated using random forest (DML-RF/RF)
• Double machine learning with (E[̂D|Z]) ̂estimated using classification neural network and (E[Y|Z]) ̂estimated using Lasso regression (DML-NN/Lasso)
• Double machine learning with both (E[̂D|Z]) ̂and (E[Y|Z]) ̂estimated using eXtreme Gradient Boosting (DML-XGB/XGB)
• Double machine learning with (E[̂D|Z]) ̂estimated using Logistic regression and (E[Y|Z]) ̂estimated using OLS (DML-Logit/OLS)
• Causal forest (CF)

Causal ML replication of Dhurandhar et al. (2014)

Dhurandhar, E. J., Dawson, J., Alcorn, A., Larsen, L. H., Thomas, E. A., Cardel, M., ... & Allison, D. B. (2014). The effectiveness of breakfast recommendations on weight loss: a randomized controlled trial. The American Journal of Clinical Nutrition, 100(2), 507-513.

Causal ML replication of Bryan et al. (2014)

Bryan, G., Chowdhury, S., & Mobarak, A. M. (2014). Underinvestment in a profitable technology: The case of seasonal migration in Bangladesh. Econometrica, 82(5), 1671-1748.

Dhurandhar et al. (2014) paper data.

Bryan et al. (2014) paper data.