The development of mortality models is important in order to reconstruct historical processes, understand current patterns and predict future trends. Mortality models are particularly useful when the available data are sparse, unreliable or incomplete. Traditionally, mortality patterns in data-rich populations were used to observe mathematical or empirical regularities, which could be applied to data-sparse populations. However, as a wider variety of data have become available, the focus of model building has shifted to developing flexible models that perform well in a variety of contexts.

This dissertation introduces Bayesian methods of mortality estimation in three contexts where the available data are imperfect. The first paper develops a method to estimate subnational mortality in situations with small populations and highly-variable data. The second paper develops a unified modeling framework to estimate and project neonatal mortality in all countries worldwide, including those with limited and poor-quality data. The third paper introduces a new dataset to study mortality inequalities in the United States, and develops methods to deal with the truncated and censored mortality information that is available. In all three contexts, the modeling approaches combine strengths from traditional demographic models, which capture mortality regularities across age, with the flexibility of Bayesian frameworks, which allow for multiple data sources to be incorporated, information to be shared across time and space, and uncertainty to be assessed.