Probabilistic Models for Detecting Census Person Duplication

The net undercount of the population by the decennial census arises from the balance between: (1) omissions of persons the census should count but misses, and (2) erroneous enumerations the census incorrectly includes. Duplication is a form of erroneous enumeration; typically a duplicated person is counted correctly where they should be but also incorrectly elsewhere. Coverage measurement surveys, such as the 2000 Accuracy and Coverage Evaluation (A.C.E.), must account for the effect of census duplication in order to accurately measure the net undercount. Because Census 2000 captured both names and dates of birth for most respondents, computer matching can identify possible duplicate enumerations. Exact matches on name and date of birth appear to be likely evidence for duplication, but such matches include persons with the same name coincidentally sharing birthdays. The paper develops probabilistic expressions for exact matches for the relative effects of duplication and coincidental sharing of birthday.