Can we use customers' names to infer ethnicity?

Say we want to check whether our models treat ethnic groups fairly. We’d need to know which customers belong to which group. But, with some exceptions, banks and insurers don't record ethnicity.

A workaround is to guess it from the data we do have. Names can be proxies for ethnic origin, and we have our customers’ names.

Several methods have been built on this idea. But does it work?

Two example approaches

Surname lists. A simple approach: match surnames against a reference list, such as census tables showing how people with each surname described their own ethnicity. If most people named X described themselves as Y, assume this customer is Y.

Surname plus geography. A method called BISG (Bayesian Improved Surname Geocoding) combines surnames with the demographics of where the person lives. It has been used by U.S. regulators, including in fair lending analysis. The CFPB (Consumer Financial Protection Bureau) published its proxy methodology.

What the research says

Individual accuracy is poor. Studies testing BISG against self-reported ethnicity found high error rates for some groups. For example, a study in the American Political Science Review found that people from minority groups who live in wealthier, more educated areas were typically misclassified as belonging to a majority group.

The categories are muddled, and the mapping is ambiguous. Some name classifiers have category labels like "Muslim" alongside "British" and "East Asian". One is a religion, one is a nationality, one is a broad region. Names can carry a mix of language, religion, migration history and naming custom; they don't always map neatly to ethnic groups.

Take the most common names on Earth. One of the most common first names is Maria, and one of the most common surnames is Wang. Maria suggests a Christian family (a religion, not an ethnicity) and spans Europe, Latin America, the Philippines and beyond. Wang points to Chinese heritage. Put them together and a real person named Maria Wang, of whom there are many, can’t be reliably classified.

Rare names are also a challenge. The census tables only include names above a minimum count, so an uncommon surname simply can’t be classified.

Names move around. People change names at marriage, sometimes across ethnic lines. Names get anglicised, or transliterated into English in several different ways from the same original. Reference lists built in one country don't describe another: the U.S. census surname tables won’t work elsewhere.

The aggregate question is contested. A defence is that individual mistakes wash out when we only look at group-level statistics. A recent paper (yet to be peer reviewed) argues something worse: because the misclassification is systematic, group-level disparity estimates get pulled toward zero. So, testing with inferred ethnicity would tend to understate unfair treatment.

In short: individual-level inference is widely agreed to be unreliable; group-level use is defended by some institutions and questioned by serious researchers, and the question isn't settled.

So should we use it?

Collecting ethnicity properly, where it's permitted, is a separate conversation.

In the absence of that, banks and insurers face a difficult trade-off. Names and other data points carry enough signal to create bias risk, and inferred ethnicity may be the only measurement available.

Whether an unreliable measure is better than none is a judgement call.

If someone proposes to infer ethnicity from names, we first have to ask: whose ethnicity is likely to be guessed wrong, which way would those errors push the result, and would we rather have that imperfect answer, or none? Importantly, what level of inaccuracy are we willing to work with?

Disclaimer: The info in this article is not legal advice. It may not be relevant to your circumstances. It was written for specific contexts within banks and insurers, may not apply to other contexts, and may not be relevant to other types of organisations.