Combining Information from Multiple Surveys for Small Area Estimation: A Bayesian Approach

• Activity Number: 188
• Type: Invited
• Date/Time: Tuesday, August 5, 2003 : 8:30 AM to 10:20 AM
• Sponsor: Section on Survey Research Methods
• Author(s): Trivellore E. Raghunathan*+ and Nathaniel Schenker and Dawei Xie and Van L. Parsons and William W. Davis
• Companies: Institute for Social Research and National Center for Health Statistics and University of Michigan and National Center for Health Statistics and National Institutes of Health
• Keywords: hierarchical model ; Gibbs sampling ; noncoverage ; nonresponse ; BRFSS ; NHIS

Abstract

Accurate small area estimates of cancer risk factors are required in the cancer surveillance research. There is no one survey that can provide data deriving accurate estimates. Typically, the Behavioral Risk Factors Surveillance Survey (BRFSS), a state-level telephone survey, has been used to derive such estimates. However, BRFSS excludes non-telephone households and has, generally, low response rate. Though BRFSS provides reasonable sample sizes in all states covers most counties, the estimates may be biased due to noncoverage and nonresponse. The other possibility is to derive small area estimates from the National Health Interview Survey (NHIS), a nationally representative, face-to-face survey with a high response rate. However, sample sizes are small and relatively the number of counties covered is small. This paper develops and evaluates methods that combine information from these two surveys to derive the county level estimates. A Bayesian hierarchical model is used that incorporates both noncoverage and nonresponse biases in BRFSS. Small area estimates for 11 outcomes useful in cancer surveillance research are developed using the data from 1997-2000 BRFSS and NHIS.