Alexander Brian Chow
Doctorate – Heterogeneous Effects and Machine Learning: na application to mineral resources in Brazilian municipalities
Advisor: Prof. Dr. Fernando Antonio Slaibe Postali
Comission: Profs. Drs. Maria Dolores Montoya Diaz, Sérgio Naruhiko Sakurai and Joelson Oliveira Sampaio
This thesis aims to contribute to the growing literature of applied machine learning techniques in causal inference problems. To reach this goal, a wide literature review is done, including the main methods for causal inference and the possible use of machine learning techniques to overcome some of its limitations. In addition, simulations are carried out to measure the performance of machine learning models for heterogeneous effects, under different data generating processes. Furthermore, some machine learning models are implemented, more specifically the S-Learner, T-Learner, X-Learner and R-Learner, with different base algorithms, to evaluate whether municipalities that receive mineral windfalls are more propense to an increase in overall and personnel expenditures, a decreased fiscal effort, and an increased probability of violating the Fiscal Responsibility Law (LRF). It is important to highlight that there are only a few machine learning techniques for causal inference with real data since most of the applications are carried out on simulated data. Therefore, this thesis contributes to the literature in two ways. Firstly, to the literature of machine learning applications on causal inference. Secondly, there is also an empirical contribution, considering that it also performs tests not only on simulated data, but also real data. The results show evidence of heterogeneity of in the effect of receiving mineral windfalls on the fiscal behavior of Brazilian municipalities regarding an increase in local expenditures and an increment on the probability of violating the LRF.
*Abstract provided by the author