Seminar Notice
 
                                    Statistical Laboratory
                                    Iowa State University
 
DATE AND TIME:       Tuesday, July 19    1:10 p.m.
 
PLACE:                          1219 Coover Hall
 
SPEAKER:                Lihua Chen
 
                                                                  
TITLE: Adaptive Regression by Mixing: an Alternative to Model Selection Demonstrated on a Capture-Recapture Data Set
      
 

ABSTRACT

 
Traditional data analysis techniques that depend on the selection of a model are vulnerable to model uncertainty. Model uncertainty arises when a small change in the problem makes a model selection criterion lead to selection of a different model. When the model choice is unstable, the results of the analysis may be unreliable. The present talk reports on an alternative to model selection, Adaptive Regression by Mixing (ARM).
 
The method is derived from the insight that, for a family of functions, a centroid can be constructed which is close to every member of the family, even if those members are far from each other. This is possible due to properties of the Kullback-Leibler divergence. This leads to the derivation of a model combining method that has weights based on the performance of the candidate models and whose cumulative risk is bounded by the risk of the best model plus a penalty term of order 1/n.
 
This presentation demonstrates the properties of ARM in the context of loglinear models. Here ARM is shown to outperform the commonly used model selection criteria, Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), and a popular model combining method, Bayesian Model Averaging (BMA).
 
A data example is taken from the literature on capture-recapture studies where the object of interest is the hidden population size. In the context of loglinear models, this problem can be formulated as estimating a missing cell mean based on an incomplete contingency table. In this data set, ARM leads to a better estimate than AIC and BIC as verified by a parametric bootstrap method.
 
COFFEE:  12:45 p.m., 104 Snedecor Hall
 
Seminar schedules and abstracts are available via WWW:  http://www.stat.iastate.edu/