STRATA: A Name-and-Geography Race Inference Model for Fair Lending and Housing Equity Applications
This paper introduces STRATA, a novel race and ethnicity inference model that combines character-level name sequences with census tract geolocation using stacked Bidirectional LSTMs and XGBoost to significantly reduce socioeconomic bias and improve accuracy over existing methods like BISG for fair lending compliance and housing equity analysis.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Great Guessing Game: Who Lives Where?
Imagine you are trying to understand the story of a neighborhood, but the residents have left their names off the mailboxes. You can see the houses, the parks, and the schools, but you don't know who lives inside. This is a common problem in the world of data science, specifically in a field called "fair lending" and "housing equity." Banks and regulators need to know if people are being treated fairly based on their race or ethnicity, but often, that information is missing from loan applications or property records.
To fill in these blanks, scientists have developed "proxies"—smart guessing tools. The most famous one, called BISG, works like a simple recipe: it looks at a person's last name and their zip code. If the name sounds like it belongs to a certain group, and the neighborhood is mostly that group, the tool guesses that's who lives there. It's a bit like guessing someone's favorite ice cream flavor just because they live on a street where everyone else loves chocolate. But here's the catch: this method isn't perfect. It often makes mistakes, especially with people who live in mixed or wealthy neighborhoods, sometimes mislabeling them as "White" even when they aren't. This can hide real unfairness, making it look like there are fewer problems than there actually are. The big question is: Can we build a better, smarter guesser that doesn't get tricked by fancy neighborhoods?
Enter STRATA: The Super-Sleuth of Names and Neighborhoods
This paper introduces a new, high-tech detective named STRATA (Socioeconomic and Tract-Referenced Attribution for Algorithmic analysis). Think of STRATA as a super-smart robot that doesn't just look at a name and a zip code; it reads the entire story of a name, letter by letter, while also studying the neighborhood in deep detail.
While the old method (BISG) is like a person glancing at a name and a map and making a quick guess, STRATA is like a detective who reads the spelling of a name to understand its history and then cross-references it with 13 different facts about the neighborhood, like income levels and population density. It uses two powerful tools working together: a "neural network" (a type of computer brain that learns patterns in names) and a "decision tree" (a logic machine that double-checks the work).
The Big Discovery
The researchers tested STRATA against the old methods using a massive dataset of nearly one million voter records and a national set of small-business loan records. The results were a game-changer.
- The Old Way (BISG): When the old method tried to guess, it made a specific kind of mistake: it often misidentified non-White people as White. In the voter test, this happened 28.0% of the time. This is like a security guard at a club letting 28 out of 100 people who shouldn't be there slip in because they looked like they belonged.
- The New Way (STRATA): STRATA slashed this error rate down to 17.8%. It also got the overall right answer 88.8% of the time in the voter test and 88.5% of the time in the national loan test.
Why This Matters
The paper argues that the old method's mistakes aren't just random; they are "socioeconomically correlated." This means the old tool gets confused when people of color live in rich neighborhoods, often assuming they must be White because "rich neighborhoods are mostly White." STRATA learns that names and neighborhoods interact in complex ways, so it doesn't get fooled by a fancy address.
What STRATA is NOT
The authors are very clear about what this tool cannot do. They explicitly state that STRATA is not accurate enough to make decisions about a single person. You cannot use it to approve or deny a specific loan for a specific individual. It is a "population-level" tool, designed to look at the big picture of thousands of people to see if a bank or a city is treating groups fairly. Using it to judge a single person would be "reckless," according to the authors.
The Catch
STRATA is a powerful tool, but it has limits. It was trained on data from Florida, Georgia, and North Carolina, plus national loan data. While it works well in other places, its accuracy can drop in very specific areas, like Hawaii, where the mix of names and cultures is unique and doesn't fit the standard categories perfectly. Also, it struggles with the "Other" category (people who identify as mixed-race or Native American), often misclassifying them because the data it learned from didn't have enough clear examples of them.
The Bottom Line
STRATA isn't a magic wand that solves all discrimination, but it is a sharper lens. By reducing the number of times it accidentally labels non-White people as White, it helps regulators and banks see the true size of unfair gaps in lending and housing. It turns a blurry, biased guess into a much clearer, more honest picture of the data, ensuring that when we look for inequality, we aren't missing it because our tools were too simple to see it.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.