A Comparison of Neural Network and Model-Based Methods for Small Area Estimates of Uninsurance Rates

• Activity Number: 16
• Type: Contributed
• Date/Time: Sunday, August 3, 2003 : 2:00 PM to 3:50 PM
• Sponsor: Section on Health Policy Statistics
• Author(s): Cynthia Garvan*+
• Companies: University of Florida
• Keywords: small area estimates ; survey ; neural networks ; health insurance

Abstract

A 1999 telephone survey interviewed 14,011 households in order to estimate the proportion of Florida residents without health insurance, providing estimates within a 1% margin of error statewide and 3% for demographic subpopulations. To generate estimates for counties, congressional districts, and zip codes, a consultant used a neural network approach, a nonlinear pattern recognition devices which involves geocoding respondent phone numbers and training and validating a neural net model. In 2003, Miami-Dade County conducted a similar survey to estimate the proportion of residents without health insurance; the sample size of 1,500 households was comparable to the 1,631 Miami-Dade households from the 1999 survey. Model-based small area methods were used to generate estimates for zip codes in the 2003 Miami-Dade project, due to concerns that a neural network approach might 1) lack the ability to provide estimates of precision; 2) be unable to utilize all available survey data; and 3) provide no means of explaining patterns of uninsurance. In our study, the zip code-level estimates using the two procedures are compared.