Limitations

MEASURING “TRUE” POPULATION.

To measure relative undercount and adjustment, a standard of true population must be set for each block cluster. For the purposes of this analysis, we accept the Census Bureau’s direct dual system estimate (direct DSE) for each block cluster as this standard.

Statisticians and demographers have provided ample evidence that the DSE is seriously flawed (Appendix E). Therefore, the validity of using the direct DSE as a standard may be debated.

However, the statistical adjustment proposed by the Bureau is based on a fundamental assumption that the dual system estimate methodology is valid. Therefore, this analysis examines an “even if” question: Even if the direct DSE accurately measures local undercounts, would the proposed statistical adjustment using synthetic estimation correct those undercounts?

RESULTS ARE UNWEIGHTED.
This analysis examines all but 10 of the 5,180 block clusters surveyed in the 1990 PES: an excellent basis for conclusions about the effect of statistical adjustment in those block clusters.¹⁷ It is also appropriate to draw some conclusions about the efffects of statistical adjustment in general. For example, the results in these clusters clearly indicate statistical adjustment fails to correct large undercounts in local areas.

However, without weighting, these data should not be used to make inferences about larger areas such as states and the nation as a whole. Heavily undercounted areas appear in the 1990 PES block clusters with greater frequency than if the sample were drawn randomly. For example, 13 percent of the block clusters in the sample have undercounts higher than 10 percent. In the nation as a whole, the percent of heavily undercounted clusters would be somewhat lower. Therefore, without the proper weighting of the sample, it is impossible to compare the proportion of heavily undercounted areas in the 1990 PES block clusters to the proportion of heavily undercounted areas in the nation.

This over-sampling of heavily undercounted areas results in a similar over-sampling of racial and ethnic minorities. As noted by the Bureau,¹⁸ heavily undercounted areas tend to have high minority populations. African Americans, Latinos, Asians and American Indians also appear in the 1990 PES block clusters with greater frequency than if the sample were a simple random sample.

While over-sampling limits the use of this sample as a proportionate reflection of the nation, the relatively large sample of undercounted areas allows for a better picture of undercounted neighborhoods with predominantly minority residents. At publication, the Board did not have the weights for the 1990 PES block clusters. We have requested these data, and will continue analysis.

DIFFERENCES DUE TO IMPUTATION.
At publication, the whole-person imputations reported in the official census counts in the block clusters surveyed in the 1990 PES were not available. Although the Bureau provided whole-person imputations for every block included in the PES, some of those blocks were subsampled. That is, portions of some blocks were not included in the survey. At publication, the Bureau was unable to separate imputations made in block portions included in the PES, from imputations made in block portions not included in the PES. Therefore, data provided to the Board overstate the number of imputations in the areas surveyed in the PES.

The Bureau does not include imputations in the calculation of the block-level direct DSE, because imputations in the census cannot be matched to responses from the PES. This analysis uses the exact value for the direct DSE provided by the Bureau.

However, we added imputations to the E-Sample counts for the purposes of this analysis. Adding imputations to the E-Sample counts approximates the official census counts in those areas. We believe the approximations will be very close. The error from this source is small, and does not affect the conclusions of this paper.

DISCREPANCIES IN BUREAU DATA.
In the course of this analysis, the Bureau provided the Board with two data files. One file, with block cluster-level data for those areas surveyed in the PES, is the XYZ file. The XYZ file was delivered on April 6, 1999, in response to a request submitted by the Congressional members of the Board in December 1998. The second file, the PES Block file, provides block-level data, including counts by race and Hispanic origin, for those same areas. It was provided upon request in July 1999.

When analysis uncovered discrepancies between the population counts in the XYZ file and the PES Block file, the Congressional members of the Board asked for an explanation. In August 1999, Bureau staff reported that an error in creating the PES Block file had resulted in inaccurate data, and undertook a revision. On September 10, 1999, the revised PES Block file was delivered to the Board, with assurances that the data had been corrected, and accompanying documentation. Again, the counts did not match.

The XYZ file reported that, in the 5,170 block clusters included in this analysis, the E-Sample count (not including imputations) was 392,543 persons. The revised PES Block file reported that, in the 5,170 block clusters included in this analysis, the E-Sample count (not including imputations) was 446,099 persons. From discussions with Bureau staff, and pertinent scientific literature, we surmise the XYZ file is accurate.

Rather than wait indefinitely on a corrected PES Block file, we simply noted the discrepancy. One set of data from the PES Block file was used in this report: the number of imputations in each cluster (a total of 5,905, which did not change in the Bureau’s revision). We have requested the corrected file.

In addition, the Bureau previously has reported that 5,290 block clusters were included in the 1990 PES. Through discussions with Bureau staff, we surmise that clusters with no recorded population were not included in the XYZ file or the PES Block file.

Our analysis of the effects of adjustment on people of various race and Hispanic origin has been seriously compromised by these data discrepancies. We will complete this analysis when the Bureau provides correct data on race and Hispanic origin, and PES weights for the block clusters.