Measuring Census Response Variation Using Automated Record Linkage

Matthew Sobek, University of Minnesota
Matt A. Nelson, University of Minnesota
Cheyenne Lonobile, University of Minnesota
Nesile Ozder, University of Minnesota
Diana Magnuson, University of Minnesota

The advent of full count census databases has enabled record linkage of entire national populations, but such linking is vulnerable to noisy data. Perhaps the most vexing errors stem from the original interaction between the enumerator and the respondent, as recorded on the census form. There is no way to correct such errors, but having an accurate sense of their scope at the variable level would inform our understanding of response variability that linking methods must account for. We aim to measure the degree of census response variation by linking the full-count U.S. censuses to themselves to identify instances where households were enumerated twice. Most often, such people moved during the enumeration period and were interviewed in two places, but sometimes enumerators inadvertently strayed from their district. We will focus on multi-person households, where the combination of personal characteristics can ensure that these are truly the same group of people. The 1940 field that identifies the actual census respondent will allow an analysis of gender and relationship effects on enumeration differences. Finally, the 1880 census enumerated the entire city of St. Louis twice to address concerns about potential undercounting. Linking these two 1880 enumerations for the same population will offer cases at a scale amenable to robust statistical analysis.

No extended abstract or paper available

 Presented in Session 177. Building and Interpreting Censuses II