NCJ Number
249474
Date Published
December 2015
Length
39 pages
Annotation
This guide provides documentation for calculating direct variance from the National Crime Victimization Survey (NCVS), which is sponsored by the U.S. Justice Department's Bureau of Justice Statistics (BJS) for the purpose of estimating the frequency and characteristics of criminal victimization in the United States.
Abstract
When calculating NCVS estimates, researchers must take into account the complex stratified, four-stage sample design and analysis weights. Stratification, clustering, and variation in analysis weights affect the variance of survey parameters, and inappropriately accounting for these factors during estimation can lead to invalid results. There are two broad methods for calculating variances of estimates from complex sample designs. One method is generalized variance functions (CVFs), which model the design-consistent variances for multiple survey estimates to obtain GVF parameters. This method has traditionally been used for NCVS estimation. The GVF equations and parameters that have historically been used to calculate NCVS standard errors are presented in an appendix of this guide. The second broad method for calculating variances of estimates from complex sample designs is to estimate variances directly, using software that accounts for complex sample designs. Two direct variance techniques are Taylor series linearization (TSL) and balanced repeated replication (BRR). Compared to BRR, TSL has proven to be a more straightforward method for NCVS estimation. The design variables required to calculate TSL estimates are already available on NCVS public use files (PUFs); whereas, the BRR weights must be computed for single and pooled years of NCVS data. The current guide provides documentation for calculating direct variance estimates from the NCVS using TSL. The guide includes the three key NCVS estimate types found in most BJS reports, i.e., victimization totals, proportions, and rates. 7 tables