Linked Samples and Measurement Error in Historical Us Census Data

The quality of historical US census data is critical to the performance of linking algorithms. We use genealogical profiles to correct measurement errors in census names and ages. Our findings suggest that one in every two records has an error in name or age, and human capital is correlated with lower error rates. While errors in age decline across subsequent census rounds from 1850 to 1930, errors in names do not exhibit such trends. Fixing all transcription errors, hence leaving only those errors made at the time of enumeration, would reduce error rates in names by about a third. Correcting all names and ages using genealogical profiles leads to 17-33% more links and fewer false positives. Reassuringly, we find that reducing such errors has a negligible effect on estimates of intergenerational mobility