Abstract: This paper proposes a non-parametric algorithm for estimating and inferring partially identified treatment effects. In particular, we introduce multivariate random forests that can be used to fit the bounds of treatment effects identified as the solution to a set of local moment equations. To detect heterogeneous subgroups, multivariate random forests adaptively search for subsets of data that exhibit the highest variation in the treatment effect bounds. We provide consistency guarantees for the estimators of the treatment effect bounds and derive their asymptotic normality under certain regularity conditions and sample splitting assumptions. Simulation experiments and applications to the National Longitudinal Survey of Youth reveal significant heterogeneity in the effect of the Head Start program on years of schooling.
- Feature Selection for Personalized Policy Analysis (with Nicholas Polson and Vadim Sokolov)[ArXiv]
- Deep Partial Least Squares for IV Regression (with Nicholas Polson and Vadim Sokolov) [arXiv]
- Deep Ensemble Transformers for Dimensionality Reduction (with Marius Geitle) [Abstract]
Work in Progress:
- Parameter-Tailored Loss Functions in Neural Networks (with Nicholas Polson and Vadim Sokolov)
- Overeducation and the Occupational Mismatch in the Labor Market (with Knut Røed and Simen Markussen) [Project]
- Oil Price Shock and Low-Carbon Transition (with Elisabeth Isaksen and Cloe Garnache) [Project]