A comparison of causal search and double machine learning using simulated data

Abstract

The paper discusses conditional average treatment effect estimation methods under nonlinear confounding. We use Monte-Carlo simulation with data generating processes based on convolutional neural networks with special adjustments, which allow comparison of counterfactual distributions. The generated data is subsequently analyzed through discretization and estimation of conditional distributions, as well as calculation of effects using double machine learning methods. We show that in this setup both discretized and double machine learning-based estimation perform poorly, showing very low correlation with true conditional effects