Increasingly, government agencies, businesses, and researchers have to combine information about the same individual from multiple databases, a process known as data or record linkage. Crucially, linking records often relies on calculating how similar a name of “John Snow” in one database is to “Jon Snowy” in another database.

A new study published in the International Journal of Population Data Science (IJPDS) reveals that, due to ethnic characteristics of names, the choice of method used to calculate similarities between names can quietly determine whose records get matched and whose get missed.

During data entry, name variations can occur for various reasons, including spelling mistakes, differences in pronunciation, or even name changes over time. The name of a person can, therefore, be recorded differently across databases.

As a result, the decision about whether two records refer to the same individual can be crucially affected by how names are compared. If the methods used to compare names respond differently to name variations, some names may more likely be identified as a match than others.

This raises an important yet unanswered question: does the way names are compared in data linkage work equally well for names from different ethnic backgrounds?

To answer this question, the researchers studied over 300 million publicly available US voter records from 2016 and 2026. They captured individuals’ name changes over time and analysed around 67,000 pairs of first names and 305,000 pairs of surnames that belonged to the same individual but contained name variations.

This study revealed notable differences in the similarities calculated between records depending on which method is selected for name comparison. The same method produced different similarity distributions depending on the ethnic group associated with names, while the same name pair could result in very different similarities depending on the comparison method used.

As a result, linkage bias and fairness might be affected across ethnic groups.

As an example, if a linkage method identifies 95% of matches for the largest ethnic group and only 65% of matches for a minority group, an overall linkage quality of close to 90% might still be reported. While this might be an acceptable performance, in a real-world scenario such a result would mean that one third of records of an entire ethnic group have not been linked.

The researchers show that simply checking overall linkage performance is therefore not suitable when handling real-world databases containing records from diverse ethnic groups. This study emphasises how seemingly invisible design decisions can cause unexpected biases that can have an impact on real lives.

Every time researchers analyse linked health data, or a government checks whether a relief programme reached the families it was meant to help, the answer may depend on a technical decision deep within a data linkage method that is often overlooked.

The study provides researchers and practitioners with recommendations for checking whether  their linkage methods are quietly missing some ethnic groups, and, more importantly, how to avoid such unfair outcomes.

 

Click here to read the full article

Yashithi Dharmawimala,  School of Computing, Australian National University

Dharmawimala Y, Christen P, Ziyad S, Schnell R, Lam J, Vidanage A, et al. How Ethnic Name Characteristics Influence String Similarities and Cause Bias when Linking Large Population Databases. IJPDS [Internet]. 2026 Nov. 9;11(1). Available from: https://doi.org/10.23889/ijpds.v11i1.3767