Confidence Intervals for Estimates of Rare Events Based on Complex Survey Data

• Activity Number: 419
• Type: Contributed
• Date/Time: Wednesday, August 6, 2003 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Computing
• Author(s): Margaret D. Carroll*+
• Companies: Centers for Disease Control and Prevention
• Keywords: rare event ; extreme percentile

Abstract

One of the precision requirements for The National Health and Nutrition Examination Survey, 1999-2000 (NHANES 1999-2000) is that an estimated prevalence of order 10 should have a relative standard error not exceeding 30%. Estimated prevalences of rare events (< 10%) are more likely to violate this criteria. Furthermore, when confidence intervals are constructed using the normal approximation, it is possible to obtain a negative lower limit. The purpose of this talk is to make recommendations concerning the analysis of these estimated small proportions as well as extreme percentiles. We will present examaples based on the NHANES 1999-2000 data including extreme obesity (BMI>=40 kg/m square), very high triglyceride(>=500 mg/dl) and extreme percentiles (1st and 99th) for selected health-related variables. We will first run simulations of standard distributions (e.g., the beta distribution with selected location and shape parameters). We will then compare these results to the empirical distributions of the underlying variables upon which these estimates are based.