CensusPlus: A sampling and prediction approach for the 2000 census of the United States

For a general audience, this paper offers details of a simple proposal for estimation of the population and housing in the year 2000 for the United States. Under CensusPlus, two surveys (mass enumeration and plus sample enumeration) are made of a universe with M blocks. The mass enumeration results in an initial preliminary count for each and every block in the country. The plus sample blocks undergo a second extra high quality count which when compared with the initial count leads to observed resolved counts for the sample blocks. Under a simple model, resolved counts are predicted for the nonsample blocks. Hence an optimal estimator of N, the universe size. is obtained by adding these observed (in sample) and predicted (not in sample) resolved block counts. In fact, this sum turns out to be the classical ratio estimator. This one number census collection is additive and consistent for all levels of geography. In addition, this paper presents sample sizes for the number of blocks required by the plus sample enumeration to support reliable state level estimates of population produced by CensusPlus. In particular and using data from the 1990 Census Files and the 1990 PES Block Data File, it is shown that a nationwide deeply stratified probability sample of 22,120 blocks is needed to ensure that the housing unit population of a given state is estimated with a standard error of 40,000 persons. The 1990 PES Block Data File also provides some early empirical evidence that the model is very likely to hold.