|
Activity Number:
|
148
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, July 30, 2007 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Health Policy Statistics
|
| Abstract - #309172 |
|
Title:
|
A Bayesian Two-Part Model for Bounded Non-Negative Data: Estimating Extra Time Spent on Diabetes Self-Care
|
|
Author(s):
|
Betsy C. Gunnels*+ and Theodore J. Thompson and Louise B. Russell and Susan L. Ettner and James P. Boyle
|
|
Companies:
|
Centers for Disease Control and Prevention and Centers for Disease Control and Prevention and Rutgers University and University of California, Los Angeles and Centers for Disease Control and Prevention
|
|
Address:
|
3470 Buford Highway MS K10, Atlanta, GA, 30341,
|
|
Keywords:
|
Two-part models ; Bayesian modeling ; Diabetes Mellitus
|
|
Abstract:
|
Non-negative data with a large proportion of zeros is a candidate for two-part modeling. Two-part models consist of a binary regression distinguishing zeros from positive values and a linear regression for the (transformed) positive values. We extend this approach in two ways. We use a t-distribution for the errors in the linear regression instead of the normal distribution. The t-distribution is an option for modeling heteroscedastic normal data. We also consider a set of transformations that constrain the retransformed predictions to be bounded. Candidates include versions of the logistic, probit, complementary log-log, and log-log. Cross validation of posterior predictions is used to compare models and evaluate the fit of the final model. We demonstrate the methods using extra time spent per day on diabetes self care as our outcome with income and education as main exposures.
|
