3. Applications to Census Undercount Estimation
The results of Section 2 provide guidance about the likely size of errors due to heterogeneity in the Census Bureau’s small-area estimates of undercounts from the 1990 PES. They provide such guidance to the extent that the P-12 variables provide meaningful analogues to undercounts with respect to place-to-place variability, and to the extent that P-12 resembles the PES in sample design and post-stratification. The P-12 variables were chosen specifically to provide such analogues. Like undercounts, they are Census coverage indicators, and the Census Bureau goes so far as to call them “proxies” or “surrogates” for undercount. The P-12 sample design was chosen to be essentially the same as that for the PES, and the post-stratifications are identical. These considerations all support the idea of taking P-12 as a guide to the effects of heterogeneity on 1990 undercount estimates.
On the other hand, there is no direct validation of the posited similarity between P-12 variables and undercounts. The main available comparisons are in terms of overall levels and indices of dispersion. These are presented in this section. It turns out that undercounts fall well within the range of alternatives spanned by the four P-12 variables, but no single P-12 variable is a close match in both level and dispersion.
Net undercounts can be negative (when there is an overcount) but the P-12 variables are always non-negative. This is an important difference which weakens the analogy. The net undercount is approximately equal to the difference between two non-negative variables, the rates of “gross omissions” (e.g., missed persons) and “erroneous enumerations” (e.g., duplicates or fabrications). The P-12 variables may be better analogues for these two components of undercount than for their difference, but the overall picture is complicated by the correlations between gross omissions and erroneous enumerations which extend within post-strata all the way down to Census blocks.
Information on levels and indices of dispersion for undercount variables are shown in Table 4. They are to be compared to the corresponding rows for P-12 variables in Table 1. In Table 4, following common Bureau practice, centered adjustment factors are used in place of undercount rates. The centered adjustment factor for any unit is calculated by taking the estimated true count, dividing by the Census count, and subtracting one. The centered adjustment factor is close to the undercount rate itself. The first column in Table 4 pertains to the Bureau’s “smoothed” adjustment factors, the factors actually used for the Bureau’s calculation of adjusted counts. The second column pertains to the “raw” adjustment factors. These are dual-system estimates from PES data, calculated post-stratum by post-stratum. The raw factors were transformed into the smoothed factors by an empirical Bayes smoothing algorithm [Freedman et al. 1993]. The final two columns pertain to the gross omission and erroneous enumeration rates. Neither Table 4 nor Table 1 is weighted for post-stratum size.
The level and dispersion of a variable undoubtedly affect the numerical values of for the variable, so the comparisons between Table 1 and Table 4 are important indicators of the relevance of P-12 to undercounts. With one exception, we see that all entries in Table 4 fall between the corresponding values for substitutions and for allocations in Table 1. The exception is the 5.9% sampling standard error for the raw factors, which falls above the standard error for allocations and just below the high estimate of standard error for multi-unit housing. Thus, in terms of the quantities shown in Table 4, the P-12 variables do span the relevant range, but none matches on all dimensions.
An important conclusion is suggested by comparing the figure of 2.0% in the lower left of Table 4 with the figures in the first row of Table 1. The 2.0% is the RMS of the Bureau’s estimates of sampling standard error for its smoothed adjustment factors, and it is lower than any of the RMS values of for local areas in Table 1. If the P-12 variables are at all valid analogues, then the estimated PES sampling variances are evidently dominated by the variance due to heterogeneity measured by 2. Sampling variance is the contribution to error which the Bureau did include in its error margins for adjusted local counts [U.S. Bureau of the Census 1991]. Variance due to heterogeneity is one of the contributions it did not include. The data here suggest that what was left out is more important than what was put in.
It is likely that some part of the true contribution from sampling variability was also left out. The 2.0% figure for sampling standard deviation is believed to be a considerable underestimate [Fay and Thompson 1993, Freedman et al. 1993]. In principle, sampling variance can be traded off against variance due to heterogeneity by adopting a coarser or finer post-stratification. But the variances due to heterogeneity implied by Table 1 are so large that the leeway for such tradeoffs appears rather slight.
The particular use we are making of P-12, with our concentration on heterogeneity alone and our direct calculation of within post-strata, avoids certain difficulties which would confront more ambitious uses. We are not calculating measures of overall error for local counts or shares. Thus we are not engaged in assessing the augmentations or cancellations of error that take place when the positive or negative estimated adjustments for different post-strata in the same local area are added together to yield the total estimate for the area. We cannot do so with P-12, because P-12 superblocks for different post-strata do not coincide. Heterogeneity implies error both in Census counts and in adjusted counts, and the balance between these errors appears to be a delicate function of patterns of cancellation when post-stratum contributions are summed. We are also not engaged in studying the interaction between errors in local counts due to heterogeneity and errors at all levels due to bias in post-stratum-wide adjustment factors. We are studying errors in an idealized, bias-free setting. This setting would correspond to a PES in which the post-stratum-wide adjustment factors were known perfectly. Our counterparts of post-stratum-wide factors, that is, our s, are unbiased.
The post-stratum-wide adjustment factors in the real PES are known to be biased. There is, of course, some ratio-estimator bias. That is a side-effect of heterogeneity, and should be distinguished from the heterogeneity studied in this report, which affects estimated rates for local areas within post-strata. There are other, more important, biases in the adjustment factors estimated by the PES. Attempts have been made to measure some of these by quality-control and followup studies, but only at the level of large aggregations of post-strata. Biases are quantified in [Breiman 1994] and in Table 15 of the Census Bureau’s P-16 Project Report. Unfortunately, this crucial table is omitted from the published version [Mulry and Spencer 1993]. There is also unmeasured “correlation bias” resulting from the tendency for people missed by the Census to be more likely to be missed by the PES estimates. Essentially nothing is known about how the measured biases are distributed among the post-strata, and even less about the size and distribution of correlation bias. Thus there is not yet a basis on which definitive assessments of the relative accuracy of adjusted and unadjusted counts for local areas could be made - unless some rather heroic assumptions are to be imposed on the data. For recent reviews, see [Brown et al. 1999, Wachter and Freedman 2000], but those findings seem to be disputed in [Prewitt 2000].
In short, at the local level, what can be made are assessments of components of error like heterogeneity, not assessments of relative accuracy. To strengthen the assessments, it would be valuable to relate P-12 more closely to the PES. The Census Bureau (as far as we can tell) has not released data sufficient to calculate place-to-place correlations between the variables studied here and undercounts. In principle, substitutions, allocations, non-mailback rates, and multi-unit housing rates exist along with undercount estimates for the 5392 PES block clusters. Even more relevant than such cross-correlations would be autocorrelation functions for the variables, calculated as functions of physical or notional distance between areas. The PES sample size is small for this purpose, but some insights could be gleaned. At present, the correlations that can be computed are those that are least relevant - across post-strata. Smoothed adjustment factors correlate 0.60 with non-mailback rates, 0.23 with multi-unit housing rates, 0.18 with substitution rates, and 0.07 with allocation rates, across post-strata. Substitution and allocation rates correlate 0.61 with each other.
The PES sample is too small to give estimates of heterogeneity of the precision obtained from P-12. At the local level, the data for an calculation are not available to us at all for most post-strata. At the state level, using weighted data by post-stratum and calculating as if the sampling weights were uniform within post-strata, we find RMS values for for state-to-state heterogeneity of 10% for gross omissions and 7% for erroneous enumerations. These figures fall near the upper end of the RMS state-level values in Table 1. The PES estimates for single post-strata are unstable to the extent that about 25% of post-strata come out with negative estimated values of 2. The RMS values over all 1380 post-strata are bound to be more stable, and the figures suggest that heterogeneity in components of undercount is at least as great as heterogeneity in the P-12 variables.
Measuring Local Heterogeneity with 1990 U.S. Census Data
Kenneth W. Wachter, David A. Freedman
© 2000 Max-Planck-Gesellschaft ISSN 1435-9871