Replicate Weights for Variance Estimation of Subnational Areas

• Activity Number: 475 - SPEED: Predictive Analytics with Social/Behavioral Science Applications: Spatial Modeling, Education Assessment, Population Behavior, and the Use of Multiple Data Sources
• Type: Contributed
• Date/Time: Wednesday, August 1, 2018 : 8:30 AM to 10:20 AM
• Sponsor: Social Statistics Section
• Author(s): Stephanie Zimmer* and Marcus Berzofsky and Andrew Moore
• Companies: RTI International and RTI International and RTI International
• Keywords: subnational; crime; variance; sampling; small area estimation

Abstract

Criminologists have a desire to estimate local crime. This poses a challenge since sources for local crime, such as the UCR, do not include unreported crime while surveys require masking of local areas to protect respondent confidentiality. The National Crime Victimization Survey provides detailed victimization data but is designed for national estimation. We develop weights post-stratified to several large metropolitan areas. Direct variance estimation, rather than model-based small area models, are preferred as they can be used by analysts for any outcome. Design-based variance estimation through Taylor-series linearization is not possible because some clusters and strata are no longer included. We construct replicate weights for the national file and then post-stratify within each metropolitan area. We discuss the limitations of this methodology such as disclosure risk and instability of variance estimates for rare events. We discuss the pros and cons to different replicate weight approaches when developing weights for more than one area of interest. We compare the direct variance estimate using the national file to the variance estimates using the constructed replicate weights.