Matching Error in the Census 2000 Accuracy and Coverage Evaluation

1. BACKGROUND A potential source of error in the Accuracy and Coverage Evaluation (A.C.E.) coverage estimates is a matching operation which determined whether the respondents in the population sample (P-sample) were enumerated in the census and whether the enumerations in the enumeration sample (E-sample) were correct. In preparing for Census 2000, the A.C.E. planners put much effort into improving the person matching process from 1990. These improvements include: completing all matching in one location, utilizing a computer system in the clerical matching process, targeting the surrounding block search area, and automating the quality assurance process. To evaluate this source of nonsampling error, the Matching Error Study (MES) conducted an independent rematch in A.C.E. block clusters selected for the evaluation sample. The Census Bureau conducted a MES for the Census 2000 Dress Rehearsal Integrated Coverage Measurement (ICM) and for the 1990 Post Enumeration Survey (PES). The MES for the Dress Rehearsal was unable to measure significant matching error, presumably because there was a 100 percent quality assurance (QA) during the ICM. The match code discrepancy rates (which represent the magnitude of the difference between the person-level ICM and MES matching) for both the P-sample and E-sample were less than one percent in all sites. The 1990 M ES found that the PES generally tended to overestimate the P-sample nonmatches, especially when matching Central City, Minority persons. The magnitude of the biases in the population sizes due to matching error by evaluation poststratum (based on region, urbanicity, and minority status) ranged from approximately 0.7 percent to 1.3 percent. Of particular concern, nonmatches for Blacks were overestimated by about 4.5 percent (which equated to an approximately 0.7 percent positive bias in the total Black population). (Davis and Biemer, 1991a) The erroneous enumerations, on the other hand, were underestimated by about 5 percent for nonminorities (resulting in a positive bias in the overall population of about 0.25 percent). (Davis and Biemer, 1991b)