Advanced Methods: Missing Data Problems in Statistical Inference: Item and Unit Nonresponse, Ecological and Causal Inferences
advanced topics course (methods)
The seemingly unrelated statistical topics listed in the course title do share one common characteristic. They all can be formulated as missing data problems. When we think of missing data, we think of that person who did not answer that survey questions. For this reason we have to throw away all other information we have on them although this is usually not the best course of action. But there are other missing data problems out there worth considering. Sometimes a person is missing from the dataset because they refused the entire interview, a case or country is missing due to a complete lack of reporting information (think North Korea). Sometimes information is missing because we are working with aggregate level data, though want to draw inferences on a lower inferential level. Finally, Don Rubin reformulated one of the foundations of statistical inference: experiments, into a missing data problem. This reformulation lead to a revolution in how we view causality, leading to reconsiderations of where experiments fail in correctly inferring causality and, more importantly, where we can conduct true causal inferences even in absence of experimental data when only observational information is available. The unifying theme for this course is that all of the topics covered are formulated as missing data problems. For this reason the solutions to these problems are closely related to each other despite the problems seeming wholly independent of each other. This course wonders the wonders of statistical inference in the presence of missing data and leads us to 1, better think about our inferential strategies and 2, arms us with better tools to derive more unbiased estimates and 3, ask and answer more meaningful research questions using available statistical tools.