Using Machine Learning Methods to study research questions in health, labor and family economics

Abstract

Over the last decades, machine learning became increasingly popular as a toolbox of methods for making precise predictions on a wide spectrum of different tasks. Despite their success, economists only slowly started to incorporate them in their research. As of now, the literature combining conventional econometric approaches with machine learning methods is growing fast and new methods to answer economic questions are developed and applied by practitioners. My doctoral thesis contributes to applied machine learning research by exploring and discussing novel methods to a number of relevant research questions. I specifically look into the question of how and when machine learning methods can be useful to answer economic questions. To this end, each chapter focuses on one specific area in which recent methodological advances have been made that are of particular interest for economists. Chapter 2 applies post-double-selection to estimate average effects. Chapter 3 uses the generalized random forest framework to work out the case of a Two-Stage Least Squares random forest aimed at estimating heterogeneous effects. Chapter 4 applies latent dirichlet analysis for survey data to study the role of latent variables in a family economics application. In summary, I conclude that machine learning methods contribute to economic research in many ways. First, they allow to flexibly model the relationship between variables and to account for high-level interactions. Second, the methods are designed to handle a large number of variables. Third, most of the machine learning methods limit the freedom of the researcher in making rather arbitrary decisions. This makes empirical research more traceable and increases the trust in empirical work. Finally, new tools to analyze data entail new perspectives and new questions which can be answered. The ability to estimate personalized effects is the key to efficiently assign policies on an individual level. Moreover, the machine learning literature provides methods for dimensionality reduction which lead to well-interpretable results despite their complexity.