Missing Data Problems

Abstract

Missing data problems are often best tackled by taking into consideration specificities of the data structure and data generating process. In this doctoral dissertation, I present a thorough study of two specific problems. The first problem is one of regression analysis with misaligned data; that is, when the geographic location of the dependent variable and that of some independent variable do not coincide. The misaligned independent variable is rainfall, and it can be successfully modeled as a Gaussian random field, which makes identification possible. In the second problem, the missing independent variable a categorical. In that case, I am able to train a machine learning algorithm which predicts the missing variable. A common theme throughout is the tension between efficiency and robustness. Both missing data problems studied herein arise from the merging of separate sources of data.