Home

Wednesday, 22 January, 2025

Master´s Thesis Defense Presentations

Room 3

Defense Committee: Jan Zápal (chair), Byeongju Jeong, Christian Ochsner

13:00 Bernhard Brunner: Predicting Policy Rate Changes from Central Bank Minutes Using Machine Learning: Evidence from the Czech Republic (1998-2024)

Chair: Sebastian Ottinger

Opponent: Paolo Zacchia

14:00 Klaus Hajdaraj: Causal Machine Learning for Heterogeneous Treatment Effects: An Empirical Application on Optimal Treatment Assignment

Chair: Paolo Zacchia

Opponent: Sebastian Ottinger

Klaus Hajdaraj

Abstract:

With the rising popularity of machine learning for uncovering complex patterns, there is growing interest in leveraging these techniques to understand how interventions affect individuals differently based on their characteristics, a concept known as heterogeneity (HTE). This paper compares two machine learning methods for predicting HTEs for optimal treatment assignment or so-called targeting: the causal forest (CF), a direct tree-based method, and the causal neural network (CNN), an indirect deep learning method. I use an empirical dataset from an online experiment on incentivizing manual labour to compare the methods. I show that CF outperforms CNN; assigning individual optimal treatments based on CF yields higher outcomes than assigning the overall best treatment to all individuals. Further, I address the winner’s curse in the optimal targeting context by introducing two shrinkage techniques: the James-Stein and the Variance shrinkers, which improve the performance of ML methods in assigning the optimal treatments. This study contributes to the literature by providing a detailed guideline for selecting and comparing ML methods for optimal targeting and introducing shrinkage techniques to adjust upward bias (overestimation). The findings highlight the importance of accurate HTE estimation in improving optimal targeting, and recommend development of personalized treatments. Personalized treatments can improve overall outcomes by tailoring policies to individuals’ characteristics.

Full Text: “Causal Machine Learning for Heterogeneous Treatment Effects"