Estimation in the Presence of Missing data: An Application of Bayesian Methods to the Analysis of North Dakota Death Certificate Data

• Activity Number: 404
• Type: Contributed
• Date/Time: Wednesday, August 10, 2005 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Betsy L. Cadwell*+ and Edward F. Tierney and Theodore J. Thompson and James P. Boyle
• Companies: Centers for Disease Control and Prevention and Centers for Disease Control and Prevention and Centers for Disease Control and Prevention and Centers for Disease Control and Prevention
• Keywords: missing data ; Bayesian methods ; random walk Metropolis-Hastings ; logistic regression

Abstract

To estimate the number of diabetic deaths, North Dakota death certificates include a check box to identify decedents with diabetes. The Behavioral Risk Factor Surveillance System (BRFSS) estimates the total number of diabetics in North Dakota and provides a denominator for calculating death rates among persons with diabetes. However, approximately 20% of the death certificate check boxes are not completed. To obtain estimates for the number of diabetic deaths using information from all death certificates, we generate the posterior predicted distribution of the incomplete check boxes using logistic regression. From the posterior distribution, we draw estimates for the number of diabetics among the incomplete check boxes and add the information from the completed check boxes to obtain estimates for the number of diabetic deaths. We then make draws from a normal distribution with mean and variance for the number of diabetics in North Dakota taken from the BRFSS. Dividing the estimates for the diabetic deaths by the estimates for diabetic population provides the distribution of diabetic death rates. The estimated death rate is 50.48 deaths per 1,000.