| DATE: | Thursday, November 7, 2002 |
| TIME: | 4:30-5:30 PM |
| PLACE: | LN 2205 |
| SPEAKER: | Wolfgang Wefelmeyer (University of Siegen, Germany, and Binghamton University) |
| TITLE: | Imputing responses that are not missing |
We compare two related models: A mixture model with additional observations on the mixing distribution (the Ibragimov-Hasminskii model); and a regression model with responses missing at random. In both models one has i.i.d. copies of bivariate random variables (X,Y), with a parametric model for the conditional density of Y given X. In the first one observes partly X, partly Y; in the second partly X, partly (X,Y). Efficient estimation is much harder in the first model because it requires inference about the reverse conditional distribution, namely of X given Y. In the second, for the missing responses Y one simply imputes conditional expectations given X, with an appropriate estimator plugged in for the parameter. The purpose of this talk is twofold: First we review and compare known results about the two models, then we make a small but useful point: In the second model, with missing responses, it is better to impute conditional expectations also for the responses that are not missing.