←Back

564 – SPEED: Advances in Survey Research Methodology, Part 2B

Subsample Weights Studies - Alternate Methods for Constructing BRR Weights for NHANES Single Year Samples

at National Center for Health Statistics

Balance repeated replicate (BRR) weights are commonly used for variance estimation for multi-stage complex sample surveys. BRR requires two primary sample units (PSU) per stratum. Creating pseudo design variables, BRR stratum and/or BRR PSU, is needed when there is only one PSU per stratum. The National Health and Nutrition Examination Survey (NHANES) is a multi-stage sample survey with a sample design that has changed over time. For NHANES 2011-2014 stratification, five state groups were formed using external health information (e.g., infant mortality rates). Using these five state groups, major strata were created by geographical and urbanization characteristics. This design is used to create variance units on released data files for variance estimation; major strata and certainty PSU are both used to form variance strata. NHANES has two (sometimes three) variance PSUs per stratum for each two-year data file. For single year data files, there is generally one variance stratum and 15 variance PSU. Therefore, it is not straightforward to create BRR sample weights from single year data files. Although most health measures are available on two-year data files, some components are only available for one year. The purpose of this paper is to compare three methods for creating BRR sample weights with a one-year NHANES data file. First, two PSUs are randomly paired into one BRR stratum (Random). Second, the five state groups are used as BRR strata and PSUs within each state group are paired (State Group). Third, each PSU is defined as a BRR stratum and segments within the PSU are divided into two BRR PSU (Dividing PSU). Simulations using 16 years of NHANES data are used to examine variance estimates using these three methods. Relative bias and coefficient of variations (CV) are compared among the three proposed methods.