Mitigating Ethnoracial-Based Linkage Bias with Feature and Frequency-Aware Methods

Abstract

Record linkage using names as identifiers is central to population data integration yet may systematically disadvantage minoritised ethnoracial groups. We investigate how ethnoracial-based differences in name characteristics and error distributions interact with probabilistic linkage models to produce unequal missed match rates, and whether feature-based methods can mitigate this bias. Using 8.7 million records from the North Carolina Voter Registry, we derive race-specific error profiles and simulate three data corruption scenarios, including a biased setting with higher corruption among minoritised groups. We compare three linkage strategies within a mixed, cluster-based framework: (i) conventional probabilistic linkage using Jaro–Winkler and Levenshtein similarities; (ii) rarity-aware variants with term-frequency (TF) adjustment; and (iii) a proposed name-feature model combining principal-component embeddings of first-name characteristics with TF-adjusted last-name similarity. Across all settings, we fix the overall false negative rate at 20% and quantify group-specific false match rates differences relative to Non-Hispanic White voters. Baseline probabilistic models exhibit clear ethnoracial bias mechanisms: Non-Hispanic White individuals consistently have the lowest false negative rates, with larger missed-match penalties for Non-Hispanic Black, Asian, and Hispanic groups under race-specific error patterns. Term-frequency adjustment reduces but does not eliminate ethno-racial disparities. The name-feature model yields the most equitable outputs, minimising false negative rate gaps across racial groups in the most realistic corruption scenario, with modest increases in false match rate. Distribution-aware name features and rarity adjustment can be embedded in cluster-based linkage to reduce ethnic bias in name-based linkage without substantial loss of accuracy.