Guessing race from a name and an address

Bayesian Improved Surname Geocoding (BISG) estimates the probability that a person belongs to each racial and ethnic group using only their surname and where they live. Courts, regulators, and researchers rely on it when records leave race out. Type a name below, add more information, and watch each piece of evidence change the prediction.

Your inputs

Neighborhood size used Smaller areas are more informative, but noisier.
Try an example

Everything runs in your browser. Names are never sent anywhere. If you click Locate, the address alone is sent to the U.S. Census Bureau Geocoder to find its census block. Nothing is saved.

Prediction

Fill in any field to begin.

Step by step: how the evidence adds up

BISG is Bayes’ rule applied over and over. Each piece of evidence multiplies every group’s current probability by a factor: how much more (or less) common that evidence is among members of the group than among people overall. The results are then rescaled to add up to 100%. Switch a step off to see what it contributes.

Steps will appear here once you fill in a field.

How BISG works

BISG (Elliott et al. 2009) combines two facts that are each only partly informative. Some surnames are held mostly by members of one group: the Census reports that about 96% of people named Nguyen identify as Asian, for example. And neighborhoods in the United States remain racially segregated, so where someone lives also carries information. Bayes’ rule combines the two, assuming that, within a racial group, surname and location are unrelated.

P(race | surname, place) ∝ P(race | surname) × P(place | race)
P(place | race) = (residents of the place in that race) ÷ (all U.S. residents of that race)

The BIFSG extension (Voicu 2018) multiplies in a first-name term the same way, P(first name | race). This tool also lets you add a middle name and sex, following the fuller model in Imai & Khanna (2016). With every input switched on, the full calculation is:

P(race | all evidence) ∝ P(race | surname)
  × P(first | race) × P(middle | race)
  × P(sex | race) × P(place | race, sex)

Each multiplier shown in the step-by-step panel is a group’s new probability divided by its old one. A multiplier of ×3 means the new evidence is three times as common among that group as among people overall, given everything learned so far.

Where the numbers come from

  • Surnames: U.S. Census Bureau, Frequently Occurring Surnames from the 2010 Census (names.zip). It covers 162,253 surnames held by 100 or more people. Where the Bureau suppressed a small cell for privacy, marked “(S)”, the leftover percentage is split evenly among the suppressed groups. Surnames not on the list get the Bureau’s “All other names” distribution.
  • First and middle names: Tzioumis (2018), “Demographic aspects of first names,” Scientific Data. It covers 4,250 first names from mortgage applications (Harvard Dataverse; CC0 copy used here from Rosenman, Olivella & Imai 2023). P(name | race) is computed by weighting each name’s race shares by its count. No public list exists for middle names alone, so middle names are scored against the same first-name list. That is a simplification.
  • Neighborhoods and sex: 2020 Census Demographic and Housing Characteristics File, tables P12H–P12O (sex by race and Hispanic origin), retrieved from the Census API for every state, county, tract, and block group in the 50 states and D.C. Puerto Rico is not included.
  • Address lookup: the Census Geocoder, using the current address benchmark and 2020 census geography.

The six groups match the Census surname file: non-Hispanic White, non-Hispanic Black, non-Hispanic Asian or Pacific Islander, non-Hispanic American Indian or Alaska Native, non-Hispanic multiracial, and Hispanic of any race. People who chose only “Some other race” (and are not Hispanic) are dropped, because the surname file has no matching category. The starting point (step 0) is each group’s share of the 2020 U.S. population on these definitions.

Things to notice and cautions

  • It predicts probabilities, not identities. A 70% White prediction means that among many people with the same inputs, roughly 7 in 10 would identify as White. It does not tell you how any one of them identifies. Applications either add up the probabilities across many people (the more defensible use) or assign each person to the most likely group (simple, but it throws information away).
  • It works best for groups that are segregated or have distinctive names. It is accurate for Hispanic and Asian surnames and for Black residents of segregated neighborhoods. It is poor at identifying American Indian and multiracial people, whose names and neighborhoods are rarely distinctive.
  • Zeros are absolute. If a block group recorded no residents of a group, or a surname shows 0% for a group, Bayes’ rule rules that group out entirely, however strong the other evidence is.
  • Smaller areas are noisier. Starting in 2020 the Census Bureau adds deliberate statistical noise (differential privacy) to protect respondents, and that noise matters most for small areas.
  • The data are from different years. Names come from 2010, neighborhoods from 2020, and people move, marry, and change their names. Women who took a spouse’s surname are a well-known source of error.
  • The independence assumption is wrong at the margins. A Black Garcia and a Hispanic Garcia may not live in the same kinds of places, but BISG assumes they do.

Questions for discussion

  1. Enter your own name. Is the prediction right? If not, which input misled it, and why?
  2. Try Lee, Smith, or Jackson with addresses in different cities. Which input moves the answer more for these names, the surname or the neighborhood?
  3. Find a case where switching from tract to block group changes the most likely group. Should we trust the smaller area more?
  4. A regulator uses BISG to check whether a lender charges minority borrowers more. Should it add up probabilities or assign each borrower to one group? What could go wrong with each?
  5. Voting rights cases use BISG to estimate the racial makeup of a voter file. Whose votes are most likely to be miscounted, and in which direction?

Further reading

  • Elliott, M. N., et al. (2009). Using the Census Bureau’s surname list to improve estimates of race/ethnicity and associated disparities. Health Services and Outcomes Research Methodology 9: 69–83. doi
  • Consumer Financial Protection Bureau (2014). Using publicly available information to proxy for unidentified race and ethnicity.
  • Imai, K., & Khanna, K. (2016). Improving ecological inference by predicting individual ethnicity from voter registration records. Political Analysis 24(2): 263–272. doi (R package wru)
  • Voicu, I. (2018). Using first name information to improve race and ethnicity classification. Statistics and Public Policy 5(1): 1–13. doi
  • Tzioumis, K. (2018). Demographic aspects of first names. Scientific Data 5: 180025. doi
  • Kenny, C. T., et al. (2021). The use of differential privacy for census data and its impact on redistricting. Science Advances 7(41). doi