Estimating treatment effects becomes harder when the outcome is semicontinuous — that is, when a large share of observations are exactly zero while the positive values are skewed. According to a new paper on arXiv, this structure is common enough to matter, yet standard causal forest methods are not designed for it.
The authors propose a Bayesian causal forest framework specifically for zero-inflated continuous outcomes. The work aims to estimate both average and heterogeneous treatment effects while accounting for the mixture of a discrete zero mass and a skewed positive distribution.
The significance lies in extending a popular nonparametric causal inference tool to a class of outcomes that often appears in health economics, medical research, and other fields. The abstract does not provide simulation or application results, so the practical performance of the method remains to be assessed in the full paper.