SEMINARS
|
|
Spring 2008
STATISTICS
COLLOQUIUM
Wednesday, March 5, 2008
3:30-4:00—Refreshments, Yost 327
4:00-5:00—Talk,
Yost 101
Balgobin Nandram, PhD
Professor, Mathematical Sciences
Worcester Polytechnic Institute
Bayesian Nonresponse Models for Small Areas:
An Application to Body Mass Index
This presentation has two parts. The first part is a review of some work we have done on Bayesian nonignorable nonresponse models for small areas. Specifically, we discuss missing data mechanisms and nonignorable Bayesian models that we have constructed for both continuous and categorical response variables. The key idea is to center a nonignorable nonresponse model on an ignorable nonresponse model. The second part is a Bayesian analysis of data on body mass index (BMI) from the third National Health and Nutrition Examination survey for children and young adults. The objective is (Bayesian predictive) inference for finite population BMI percentiles (85th and 95th) for each age-race-sex domain within counties (small areas). Because there are a large number of outliers we modify our previous methods to make them more robust, and also incorporate differential selection probabilities. There are two robust procedures. The first, fully Bayesian, uses a Student’s t regression for the response indicators to obtain robust propensity scores. The second, coupled with the first, is a rank-based method to predict the nonsample BMI values, and it includes the selection probabilities for the children and young adults in the sample.
|
