Department of Statistics Seminar
North Carolina State University

presents

Dr. Thomas A. Louis

University of Minnesota

"BAYES/EB RANKING, ENSEMBLE AND PARAMETER ESTIMATION: CAN ONE SET OF ESTIMATES SERVE THESE GOALS?"

ABSTRACT

The beauty of the Bayes/EB formalism is its ability to structure complicated models, inferential goals and analyses. Consider ranking school performance, physician complication rates or region-specific health risk. A valid analysis must deal directly with ranks; ranking posterior means can have very poor performance. Similarly, the empirical distribution function (EDF) of the posterior means is under-dispersed and always an inappropriate estimate of the EDF of the parameter ensemble. So, optimal coordinate-specific estimates produce poor estimates of the parameter ranks and EDF. Conversely, estimates tuned to these inferential goals may not have good coordinate-specific performance.

Though no single set of estimates will be optimal for all goals, communication and credibility will be enhanced by a single set of values that serve as good estimates of ranks, the parameter EDF and individual, coordinate-specific parameters. I will present and compare performance of candidate "triple-goal" estimates and show their operation in a variety of applications

Friday, November 19, 1999

3:35 - 4:35 pm

206 Cox Hall

Refreshments will be served on the second floor of Dabney Hall (left of Room 222) at 3:00 pm.