Debiased machine learning causal inference for time-varying social variables

Thesis

This thesis reviews and develops efficient debiased machine-learning estimators for causal inference and applies the developed methods to empirical research in sociology and demography.

The thesis is composed of five chapters, organized in three parts. Part 1 is the first chapter, which gives a comprehensive methodological review of causal inference and efficient debiased estimators. Part 2 discusses doubly robust/debiased machine-learning techniques for causal inference with survival data; it includes one methodological chapter that develops a twice doubly robust estimator for left-truncated-right-censored survival data and one empirical chapter that addresses the causal effect of widowhood on mortality. Part 3 discusses doubly robust/debiased machine-learning techniques for causal mediation analysis; it includes one methodological chapter that derives debiased nonparametric estimators for both static and time-varying marginal structural models and one empirical chapter that applies these methods to analyze how labor market participation mediates the wage penalties and premiums associated with parenthood and marriage for both genders.

Authors: Pan, G

• DOI: 10.5287/ora-wrmngvenx
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: causal mediation analysis; debiased machine learning; survival analysis; causal inference; quantitative methodology
• Subjects: Econometrics; Statistics; Sociology