Introduction to Special Issues on Historical Record Linking.

Historical record linkage has responded to two large opportunities in recent years. The growth of computational power and the emergence of full count historical census data are both revolutionizing the analysis of historical population change. The increased availability of full count census data has expanded the comparative terrain for addressing multigenerational or cross-population change. The exponential increase in the resolution of analysis invites scholars to revisit many assumptions about populations of interest, sample weighting, validation or ground-truthing, and measurement. As Ruggles, Fitch, and Roberts (2018) suggest the systematic effort to link repeated observations for social and economic research reaches back to work in the 1930s. But in the last two decades, the focus has shifted onto a larger geographic stage, moving from intensive studies of local and regional settings, to national and international studies of migration, mobility, and population change. The projects in this special issue, and the one that preceded it in volume 51(4) in 2018, are representative of this combination of computational power and geographic reach. As the authors argue, the richness of full count data allows for comparative and rigorously validated matches between historical individuals. But there is still great uncertainty. There are false matches and there are individuals who are missing over time. Bailey, Cole and Massey argue in “Simple strategies for improving inference with linked data: a case study of the 1850–1930 IPUMS linked representative historical samples” for closer attention to systematic bias introduced by machine linking algorithms in working with longitudinal or intergenerational data. They recommend that researchers adjust for the nonrandom false-matches (Type I errors) and missed matches (Type II errors) by incorporating validation variables in linking inference methods, and employing regression-based weighting procedure to customize research samples. Both approaches are illustrated in relation to the 1850–1930 Integrated Public Use Microdata Series Linked Representative Samples (IPUMS-IRS). Custom weights are developed in relation to a training data set (hand-linked) in order to document the performance of the linking algorithm. Validation variables are used to reduce the level of low quality links in a sample (conditioning on information like the commonness of a last name or disagreement about birthplace over time). This smaller and less biased sample is then evaluated for its representativeness of the reference population. A simple linear regression method and a heteroscedasticity-robust Wald test of joint significance test the null hypothesis of no relationship between the covariates and the likelihood of a linked observation. Abramitzky, Mill and Perez also argue for linking methods that customize large historical data sets to arrive at longitudinal samples that represent the populations of interest as closely as possible. In their paper “Linking individuals across historical sources: A fully automated approach”, they advocate that researchers move toward methods that are highly replicable and even provide STATA based code for reproducing their linking algorithm. Rather than relying on any kind of ground-truthing for validation, as was done in the original IPUMS-IRS (Goeken et al. 2011) and Bailey, Cole, and Massey (2019), Abramitzky et al are arguing for matching based on probability scores derived from the Expectation-Maximization (EM) algorithm. After blocking on subsets of the larger populations and using Jaro-Winkler to measure string distances between names, Abramitzky et al leverage the iterative nature of the EM algorithm to derive a local maximum likelihood function that describes the probability of a match. Once estimates of a match are derived, researchers can assess how the linked records are suited to the research question and the reference populations. In comparing the representativeness of linked IPUM-IRS samples to this automated approach,