Primary Submission Category: Machine Learning and Causal Inference
Comprehensive Causal Machine Learning
Authors: Jana Mareckova, Michael Lechner,
Presenting Author: Jana Mareckova*
Uncovering causal effect heterogeneity across various levels of granularity is invaluable for decision-makers. Comprehensive causal estimation approaches enable the use of a single machine learning method to estimate effects at all granularity levels, making them attractive for applied studies due to their computational tractability and unified empirical framework.
In this paper, we compare three comprehensive approaches: double machine learning (DML), generalized random forest (GRF), and modified causal forest (MCF). DML provides a generic framework for estimating (conditional) average causal effects using ML methods. MCF estimates conditional average causal effects and utilizes weighted representations for higher aggregation, while GRF employes doubly robust estimators for aggregates. The paper provides theoretical results for MCF including weight-based inference for causal effects and their asymptotic normality.
A simulation study examines these approaches, considering selection into treatment, effect heterogeneity and other properties of the data generating process, and provides valuable finite sample insights. Metrics include bias, standard deviation, root MSE and coverage probabilities. DML excels in estimating effects at higher aggregation levels. GRF exhibits higher bias with moderate to high treatment selection, while MCF shows robust performance across treatment selection scenarios, exhibiting lower bias than GRF and better or similar coverage probabilities.