Variance Estimation for the 1995 Census Test: Methodology and Findings

The 1995 Census Test evaluated two fundamental changes in decennial census design: sampling for Nonresponse Followup (NRFU) and integrated coverage measurement (ICM). Sampling of nonresponding households after the mail phase of the census will reduce the workload and cost of the census. Housing units not returning their census forms by mail are classified by the Postal Service as occupied or vacant; a sample of occupied nonresponding households is then selected and used to estimate for the universe of nonresponding households. ICM attempts to mitigate differential undercoverage experienced in past censuses by incorporating estimates of omitted persons directly into the final census products. A subsample will be drawn to estimate the residual undercoverage of the census, and estimates of the undercoverage will be integrated into the final count. Although the use of sampling for characteristics has been a feature of modern censuses, the 2000 Census will be the first to use sampling to determine the number of persons. Consequently, estimation of variance for census data products has become a matter of increased importance. The paper will summarize the methodology implemented for the 1995 Census Test, aspects of which were previously presented (Town and Fay 1995). The paper elaborates features of the application, such as adaptations to the analysis of experimental panels assessing the effect of block vs. unit sampling in NRFU estimation. The paper then reports the findings from the production variance estimation for the 1995 Census Test. These findings will include comparison of estimated variances with the target variances assumed in the initial design. The test will provide important evidence on the relative levels of variance for a "Census Plus" strategy in the ICM vs. the dual-system approach incorporated in a number of previous attempts to measure census undercount. Finally, the analysis identifies features of sample design likely to affect the quality of variance estimation in 2000.