Jonas Helgertz

During the past decade, scholars have produced a vast amount of research using linked historical individual-level data, shaping and changing our understanding of the past. This linked data revolution has been powered by methodological and computational advances, partly focused on supervised machine-learning methods that rely on training data. The importance of obtaining high-quality training data for the performance of the record linkage algorithm largely, however, remains unknown. This paper comprehensively examines the role of training data, and—by extension—improves our understanding of best practices in supervised methods of probabilistic record linkage. First, we compare the speed and costs of building training data using different methods. Second, we document high rates of conditional accuracy across the training data sets, rates that are especially high when built with access to more information. Third, we show that data constructed by record linking algorithms learning from different training-data-generation methods do not substantially differ in their accuracy, either overall or across demographic groups, though algorithms tend to perform best when their feature space aligns with the features used to build the training data. Lastly, we introduce errors in the training data and find that the examined record linking algorithms are remarkably capable of making accurate links even working with flawed training data.