Selection Bias Encountered in the Systematic Linking of Historical Census Records

Abstract

<jats:title>ABSTRACT</jats:title> <jats:p>Linked historical records typically are unrepresentative of the population from which they are drawn even if the method of linking is restricted to time-invariant matching criteria. An example drawn from Canadian census records illustrates the nature of bias that may afflict even a carefully linked sample. The use of potentially time-varying match criteria doubles the size of a linked sample at a modest cost in terms of additional bias. This trade-off is attractive for some research purposes if care is taken in the uses to which the data are put. Reweighting to mitigate the effects of bias in visible characteristics is desirable.</jats:p>

Intergenerational Mobility in a Mid-Atlantic Economy: Canada, 1871–1901

“…A trade-off exists between increasing sample size and introducing linked records more selective on less representative characteristics when using family information to disambiguate. Our view is that this process is worthwhile if it allows researchers to address questions where large sample size is essential (Antonie et. al., 2020).…”

Business Entry and Exit: Career Changes of Proprietors in England and Wales (1851–81) Using Record-Linkage

“…The average linkage achieved was 22.7 percent of the starting population of tagged employers and masters. 7 This linkage rate is fairly typical for census linkage lacking accurate training data, and where transcription errors and variations in name recording occur (Winkler 2014;Massey 2017;Goeken et al 2017;Antonie 2020); it is almost identical to that achieved by Long (2013: Table 1). 8 It also reflects drawing match thresholds high to minimise false positives to maximise accuracy.…”

“…7 Coding and cleaning of proprietors summarized in (van Lieshout et al 2017); birthplace standardization follows Schürer et al (2015) and Schürer and Day (2019). linkage rate is fairly typical for census linkage lacking accurate training data, and where transcription errors and variations in name recording occur (Antonie 2020;Goeken et al 2017;Massey 2017;Winkler 2014); it is almost identical to that achieved by Long (2013: table 1). 9 It also reflects drawing match thresholds high to minimize false positives to maximize accuracy.…”

“…This meant rejecting those with multiple links and there is no alternative means of "disambiguation", e.g. as used by Antonie et al (2020). As a result of this and high thresholds we will have lost people who were truly linked: we have high accuracy, but low "efficiency" (Abramitzky et al 2020).…”

“…Because no training data or alternative fully accurate "truth" data are available that can be used at the scale of the whole population it is essential to have strong confidence in the accuracy of the links. This meant rejecting those with multiple links as there is no alternative means of "disambiguation," for example as used by Antonie et al (2020). As a result of this and high thresholds we will have lost people who were truly linked: we have high accuracy, but low "efficiency" Note: Shaded area is the main focus of subsequent analysis because the unshaded area is not a full sample of switching for workers and others.…”