others), housing unit a match/not a match, and person had one or more characteristics imputed/no characteristics imputed (Cantwell et al., 2001:Table 10). After imputation, the percentage of correct enumerations dropped slightly, from 95.5 percent of resolved cases (correct and erroneous enumerations) to 95.3 percent of all cases because the imputation procedure assigned lower correct enumeration probabilities to unresolved cases (76.2 percent overall).
A sensitivity analysis determined that alternative procedures for imputing P-sample residence and match status probabilities and E-sample correct enumeration status probabilities could have a considerable effect on the estimated value of the DSE for the national household population, particularly when combined with alternative procedures for making P-sample household noninterview adjustments (see Section 6-A.5).6 One of the alternative imputation procedures substituted multivariate logistic regressions for the average cell values used in the original A.C.E. Another procedure, which assumed that unresolved cases differed significantly from resolved cases (what is termed nonignorable missingness), used 1990 PES data to develop alternative (lower) probabilities of residence, match, and correct enumeration status. These probabilities are illustrative; there is no evidence for their reasonableness compared with the probabilities used in the original A.C.E.
The results of the sensitivity analysis demonstrate the difference that alternative procedures could make. Thus, for all 128 combinations of noninterview adjustment and imputation procedures, about one-third of them differed from the average DSE population estimate by more than plus or minus 0.7 million household members; the remaining two-thirds differed by less than this amount (Keathley et al., 2001:2, as revised in Kearney, 2002:5).
For dual-systems estimation to produce highly accurate population estimates, it is critical not only for there to be very low household