Census 2000 and Demography

Statistical Adjustment of United States Census 2000

David Freedman and Kenneth Wachter, research associates of the Fresh Pond Research Institute.

The U.S. Census has long suffered from net discounts that differ by race and ethnicity, sex, and age. These differentials may distort the allocation of political representation and public funds. Statistical adjustment of Census numbers has been forward as a solution. Despite deterrents and constraints from Congress and the courts, the Census Bureau is pressing forward with plans to include in Census 2000 a set of counts incorporating such adjustments. Freedman and Wachter compare strengths and weaknesses of the proposed sampling methodology as compared to more traditional methodologies, and analyzing the differences between Integrated Coverage Measurement for 2000 and the Post-Enumeration Survey of 1990. The strengths and weaknesses of the Bureau's statistical methods and the efficacy of the ICM will be analyzed in terms of four main sets of issues: (1) Changes in plans between 1990 and 2000 will be analyzed in order to assess their likely impact on the magnitude of processing error, (2) Some kinds of people are especially likely to be missed by both the Census and the ICM, introducing a kind of error called "correlation bias". The Bureau's stance toward this problem in Census 2000 is assessed, (3) Under the Bureau's sampling plans for Census 2000, counts for blocks and local areas will be based on the assumption that undercount rates are the same for all members of certain specified demographic groups within states. As more details of the Census Bureau's specification of groups are made public, estimates of the extent of errors likely to result from heterogeneity are made, (4) Bureau plans call for an effort to reduce the level of sampling error by a process called "raking" which is a form of statistical "smoothing". The method is different from that tried in 1990, but it has many features in common. Performance characteristics of raking need to be appraised, and it must be considered whether statistical biases can be quantified and controlled.

This project is partially funded by a grant from the William H. Donner Foundation.