|
Abstract:
|
In this talk I will describe a semi-parametric Bayesian regression model for estimating heterogenous treatment effects from observational data. Standard nonlinear regression models, which may work quite well for prediction, can yield badly biased estimates of treatment effects when fitted to data with strong confounding. The new Bayesian causal forest model is able to eliminate this adverse bias by jointly modeling the treatment and the response conditional on control variables. Two empirical illustrations are given, analyzing the impact of smoking on medical expenditures and the impact of abortion laws on future crime rates.
|
