|Efficient Analysis of Missing Data|
Ursula U. Müller Texas A&M Universityy
Monday, October 6th
The fastest and simplest method of dealing with missing data is "complete
case analysis", i.e. using only cases that are completely observed. It is
well known, however, that statistical analyses based only on those data do
not always perform well. Approaches which impute missing values often give
better results. Recent work has shown that -- perhaps surprisingly --
there are situations where complete case analysis is appropriate and even
situations where it is optimal.
In this talk I will review some of my results on efficient estimation in
semiparametric regression. My only structural assumptions on the
statistical model are (1) responses are missing at random, and (2) the
regression function has a semiparametric form. I will explain when an
imputation estimator should be used and when complete case analysis is the
preferred (efficient) method.
I will also present a general method, a transfer principle, for obtaining
limiting distributions of complete case statistics (for general missing
data models) from corresponding results in the complete data model. This
provides a convenient method of adapting established methods without
This talk is based on joint work with Anton Schick and Hira Koul.