Combining information from multiple surveys for small area estimation: Bayesian approaches.

Abstract

Cancer surveillance research requires accurate estimates of cancer risk prevalence for small areas such as counties. Two popular data sources are the Behavioral Risk Factor Surveillance System (BRFSS), a telephone survey conducted by State agencies, and the National Health Interview Survey (NHIS), a personal survey. Both surveys have advantages and disadvantages. The BRFSS is a fairly large survey and almost every county is represented, but it has poor response rates and excludes the non-telephone households. The NHIS is a smaller survey and not all counties are represented, but includes households with or without telephones and has a higher response rate. After a brief small area estimation literature review, the dissertation examines non-response and non-coverage errors in the BRFSS and NHIS. The distributions of demographic variables from two survey samples in 2000 are compared to those in the 2000 Census at both national and large area levels. The socio-demographic variables include gender, age, race/ethnicity, education, marital status, employment status, and household or family income. The BRFSS sample is found to be further from the target population than the NHIS. Hence the BRFSS design-based estimates are potentially subject to higher non-coverage and non-response biases. Next, a hierarchical Bayesian approach is used to obtain county-level estimates by combining information from both surveys. The model incorporates potential non-coverage and non-response bias in the BRFSS and complex sample design features in both surveys. A Markov Chain Monte Carlo (MCMC) method simulates draws from the joint posterior distributions for the model based on the county-level design-based direct estimates. Due to confidentiality concerns, the application of the model in Chapter III is limited since the design-based county-level direct estimates are only available from the in-house NHIS and BRFSS data. Therefore, in Chapter IV we explore a large area level model for publicly available data employing the same county level model as in Chapter III. A MCMC method combining Gibbs sampling and the Metropolis-Hastings algorithms is used in model inference. The estimates are compared to those in III. In Chapters III and IV, simulations and model validations evaluate the inference and model estimates.