Missing data are ubiquitous in practical data analysis due to the study design or a subject's refusal to participate. An important related problem is the evaluation of non-randomized interventions, which can also be framed as a missing data problem. Together, missing data methods have important applications in biomedical sciences, economics, social sciences, and business. With practical usage in mind, the PIs will develop statistical methodologies that are simple to implement, easy to understand, and yet versatile enough to be applicable to a wide range of practically-motivated problems.
A general covariate balancing framework will be studied, which directly mimics missing completely at random and randomized trials in the case of observational studies, and has a better nonparametric flavor than existing model-based methods. The investigators will extend the scope of covariate balancing to various estimands and data structures, propose a unified framework for attaining uniform approximate balance in a reproducing kernel Hilbert space, and disseminate the findings to applied statisticians and non-statisticians.
