Big data and machine learning for behavioral analytics and inference: Cases in sports and education

This thesis focuses on the use of big data and machine learning methods in behavioral analytics and causal inference. The main motivation of the thesis is to illustrate how the researchers working with traditional econometric methods can benefit from big data and causal ML methods. In the absence of well-established literature, finding the right regression specification is a challenging task, especially when working with high dimensional data set. In this study, I have combined causal ML techniques with explainable AI methods and provided guidelines on how to measure heterogeneous treatment effects with the right regression specification (i.e. which main effects and interactions to be used, what control variables to be included). To empirically test these guidelines, I have curated a large data set in football including detailed variables about interim feedback, match-specific conditions, team features, and most importantly manager characteristics. Empirical evidence contributes to the sports analytics literature suggesting when and how risk-taking behavior of football managers pays off in light of interim and ex-ante information revealed to the manager (i.e. the decision maker). Moreover, this thesis contributes to the causal ML literature by evaluating the performances of two well-known causal ML techniques (a recently popular matching algorithm focusing on finding average treatment effects (FLAME) and Causal Forest that directly aims to estimate heterogeneous treatment effects) are evaluated by using synthetic data generated with known heterogeneous treatment effects. In addition to sports analytics, I have also worked with education data and demonstrated how grit, a non-cognitive skill, predicts academic achievement for students. I used a unique dataset from a digital learning platform to construct a behavioral measure of grit and showed that behavioral grit is a better predictor of student performance compared to survey grit that has been traditionally used by the researchers. I have also found that machine learning algorithms perform well in predicting academic resilience even without constructing any structural model or regression specification, thanks to the power of big data. I believe that my findings from cases in sports and education put forward the benefits of using Machine Learning and big data for researchers working with traditional and theory-based models for causal inference.