Identifying Person Duplicates of Short Geographic Distance by Computer Matching

The Census Bureau conducted evaluations of person duplication in Census 2000. Duplicates of short geographic distances were identified by both clerical and computer matching. The evaluations showed that for these short distance duplicates that the computer matching algorithms were not able to find all of the duplicates identified by the clerks. However, the computer matching algorithms in the previous evaluations were primarily developed to identify duplicates of longer distances. This report analyzes the potential of computer matching when the focus is on short distance duplicates. I used the Bureau's record linkage software to do the computer matching. Using SAS, I was able to compare the computer matching results to the clerical results. First, I attempted to identify groups of links with high concentrations of true duplicates. I used Enterprise Miner to generate decision trees for several approaches and compared their results. Second, I analyzed clerical duplicates that were not identified by the computer matching to try to identify any patterns in these cases.