False Discovery estimation in Record Linkage

Abstract: Integrating data from multiple electronic sources allows researchers to expand studies without the cost of new data collection. However, such data are often collected for administrative or operational purposes rather than with specific future research questions in mind and, due to privacy constraints, unique identifiers are unavailable. This lack of direct identifiers requires the use of Record Linkage (RL) algorithms, which rely on partially identifying variables to probabilistically determine whether records belong to the same entity. Since these variables lack the strength to perfectly combine information, RL procedures typically yield an imperfect set of linked records. Therefore, assessing the false discovery rate (FDR) of RL is crucial for ensuring the reliability of subsequent analyses. In this paper, we introduce a novel method for estimating the FDR in RL by linking records from real and synthesised data. As synthetic records should never form links with real observations, they provide a means to estimate the FDR across different procedural settings. Notably, this method is applicable to all RL techniques. By identifying the FDR in RL results and selecting suitable model parameters, our approach enables to assess and improve the reliability of linked data. We evaluate the performance of our procedure using established RL algorithms and benchmark data sets before applying it to link siblings from the Netherlands Perinatal Registry, where the reliability of previous RL applications has never been confirmed. Through this application, we highlight the importance of accounting for linkage errors when studying mother-child dynamics in healthcare records.