Estimating Mortality Risk in Populations and Individuals: Applications of Bayesian and Machine Learning Methods

Abstract

Understanding mortality risk, including its distribution and determinants, is fundamental to the health sciences and any effort to prevent needless deaths. This dissertation is an exploration of modern methods for estimating mortality risk in populations and individuals. While Bayesian statistics and machine learning have both benefited from advances in modern computing, they are polar opposites in terms of modeling strategy. Bayesian methods are rigid by design -- rigid in a particular way determined by the modeler -- in an effort to provide contextual information to the model. Machine learning methods are more flexible, seeking the information content of the data wherever it may exist. Chapter 1 of this dissertation describes a novel Bayesian method called a spline cascade that is capable of characterizing how a non-linear curve varies across hierarchical subsets of a dataset. We developed the method to model age patterns of COVID-19 mortality for global locations. Chapter 2 demonstrates how machine learning and variable attribution methods can and should be used in analytic epidemiology. We used XGBoost and SHAP values to investigate patterns in the relationship between anthropometric measurements and mortality risk, adjusting for age. Chapter 3 is a simulation study comparing Bayesian spline cascades, XGBoost and existing methods for estimating child mortality risk in 193 countries. We develop a theory of model validation and discuss the role of Bayesian statistics and machine learning in the health sciences.