Statistical Machine Learning Meets Social Science

Abstract

There has never been a more remarkable time for machine learning than the current epoch. However, when applied in a blanket manner, machine learning models can often be ill-fitted, misinterpreted, and even misused or abused. Without a nuanced approach that caters to the specific context of the problem at hand, the results may not be as accurate and meaningful. I dedicated my doctoral research to optimizing and tailoring the usage of statistics and machine learning models for three main applications: biology and healthcare, finance, and political science. The first application covers a critical look into the COVID-19 pandemic, where we characterize the growth trajectories of the virus in counties across the U.S. We combine a community detection methodology with semiparametric learning to make near-term case growth predictions. This study also incorporates demographic and behavioral variables, which in tandem produced convincing clustering and prediction results. Within the second case study, we improve existing option pricing models by introducing a two-step approach where given any parametric option pricing model, we add a nonparametric neural network to correct the error of the parametric model. We show that our machine-boosted model always outperforms the original parametric model and is also relatively indiscriminate in performance. The final application of this dissertation concerns the incumbency advantage in Mexico and whether term limits have an impact on a political party's electoral outcome. We apply a regression discontinuity (RD) design to Mexican mayoral election data and demonstrate a significant negative RD effect for states when strict term limits are still in place and a null effect when reelection is allowed. Moreover, the effect for the group of states that have adopted the reelection reform is much larger than the group of states that have yet to adopt the reform.