Using Isotonic Regression to Smooth State-Level Variance Estimators from a National Complex-Design Survey

• Activity Number: 278
• Type: Contributed
• Date/Time: Wednesday, August 14, 2002 : 8:30 AM to 10:20 AM
• Sponsor: Section on Survey Research Methods*
• Author(s): Van Parsons*+ and Thu Hoang
• Affiliation(s): National Center for Health Statistics and Université René Descartes

Abstract

The National Health Interview Survey (NHIS) is a source of information on the health of the U.S. population. This annual survey, covering about 40,000 interviewed households, implements a state-level stratification. In principle, state-level data could be released from the NHIS, but for confidentiality reasons, no geographical identifiers are provided on public-use databases. To satisfy a public demand for state statistics, basic statistics and standard errors can be tabulated and released. The robust methods typically implemented for large domain variance estimation often result in unstable and/or biased variance estimators at the state-level. For such situations some smoothing of the state-generated variances may be appropriate. The standard techniques of using average design effects or generalized variance functions may require stronger modeling assumptions than the statistician is willing to make. For variance estimators of state proportions, we consider somewhat weaker smoothing models that are based upon orderings on a two-dimensional grid. In the simplest case, the grid orders are modeled by effective state sample sizes and magnitude of proportion.