Department of Statistics Seminar
North Carolina State University

presents

Dr. Oliver Schabenberger

SAS Institute

"Local Estimation: A Parametric View of Nonparametric Modeling"

ABSTRACT

In situations where models that fit well throughout the design space are complex and/or require a large number of parameters, flexibility can be gained by fitting simpler models locally. The process of localizing a statistical model entails choosing a local model and a rule by which local contributions of the observations are gauged. We consider a model formulation in which the parametric nature of the local model is emphasized and estimation proceeds through quasi-likelihood methods. Based on the simple idea of equating an observations' impact on the analysis with its precision, standard cases such as kernel regression and local likelihood estimation are covered and extensions to the multivariate setting are immediate. Through applications we highlight local versions of generalized linear models and linear mixed models and also discuss local generalized estimating equations. A semiparametric method is introduced to combine parametric and nonparametric fits in a convex combination. The talk is delivered at a level accessible to students with a course in linear models; the emphasis will be on the concepts.

Friday, November, 01, 2002

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.