CEU Electronic Theses and Dissertations, 2021
| Author | Margad-Erdene, Nomin |
|---|---|
| Title | The Causal Tree Estimator for Heterogeneous Treatment Effects: Optimal Data Splitting Rules in Small Samples |
| Summary | Causal Trees leverage the supervised machine learning algorithm decision trees to estimate heterogeneous treatment effects across data-driven groups in a randomized treatment assignment setting. In my thesis, I modify the Causal Tree estimator by introducing a parameter theta that lets the user control allocation of data into training and estimation subsamples. The estimator implements "honest" sample splitting by default, which divides the sample into two equal parts: training and estimation subsamples. The new input parameter lets the user select the portion of data to be allocated to the estimation subsample. I test the performance of the estimator under various data allocations through Monte-Carlo simulations. My results suggest that in large samples theta between 0.3 and 0.7 can be an appropriate parameter value that minimizes the MSE of estimation. On the other hand, in small samples and in data sets with noise the recommended parameter range of theta is between 0.5 and 0.7 with optimal value of 0.6. |
| Supervisor | Lieli, Róbert Pal |
| Department | Economics MA |
| Full text | https://www.etd.ceu.edu/2021/margad-erdene_nomin.pdf |
Visit the CEU Library.
© 2007-2021, Central European University