Model card summary
Model: BISG v2, a deterministic Bayesian surname-and-geography estimator. Status: implemented but not generally activated; separately approved modeled-output lane only. Owner: VoterFile. Review date: July 11, 2026.
Intended use is aggregate language-access, field-coverage, and community-outreach planning. Prohibited interpretations include treating the output as self-identified race or ethnicity, using it as a candidate-support score, or presenting it as certainty about a person.
- Output: normalized percentages from 0 to 100
- Displayed groups: Hispanic, White, Black, Asian, and Other
- No outcome: the model does not estimate candidate preference
- No training target: this is a transparent formula, not a fitted classifier
What BISG means
Bayesian Improved Surname Geocoding combines surname distributions with geography to estimate broad race and ethnicity probabilities. The method is common in political science, voting rights, public health, and disparity research when individual-level race or ethnicity is not directly observed.
For campaign work, the useful output is not a hard label. It is an explainable signal that can help teams plan language access, field coverage, persuasion universes, and community-specific outreach with more care than surname matching alone.
- Probabilistic scoring rather than a personal identity claim
- Surname signal from public Census surname tabulations
- Geographic priors when reliable local demographic data is available
- Thresholds that campaigns can review before using a segment
Sources: Imai and Khanna BISG methodology · Science Advances fully Bayesian BISG research
How VoterFile uses the signal
When separately approved and activated, the implementation uses a surname prior from the U.S. Census 2010 surname tabulations and a geography prior from locally loaded American Community Survey five-year ZIP Code Tabulation Area shares. For each group, the two inputs are multiplied and normalized across all groups.
When a ZIP-code prior is missing, the implementation falls back to surname-only percentages. When the surname is not in the Census table, the model returns no positive signal rather than guessing from geography alone.
Sources: U.S. Census Bureau 2010 surname tabulations · U.S. Census Bureau American Community Survey data
Validation status and performance claims
VoterFile has not published an independent accuracy, calibration, or subgroup-fairness evaluation for this implementation. The current model card therefore makes no individual-level accuracy claim and publishes no unsupported precision, recall, or error rate.
Before the model is used for a consequential workflow, validation should compare aggregate estimates with an appropriate self-reported or authoritative benchmark, report calibration by geography and group, and document missing-surname and missing-ZIP coverage.
Known limitations
Surname and residential ZIP are incomplete proxies. Multiracial identity, name changes, transliteration, household composition, ZIP-level heterogeneity, stale addresses, and demographic change can all make the estimate wrong or misleading for an individual.
The underlying surname source is from 2010, and ZIP-level demographic priors are broader than neighborhoods or households. Small posterior differences should not be treated as meaningful distinctions.
- Not self-reported identity
- Not appropriate for eligibility, exclusion, or adverse decisions
- Less informative for rare or unmatched surnames
- ZIP-level priors can hide local variation
- Aggregate estimates can still carry uncertainty and bias
Responsible campaign use
BISG is most useful when it improves outreach quality: bilingual planning, volunteer assignment, local message testing, and resource allocation. It should not be used as a substitute for self-reported identity, local knowledge, or compliance review.
VoterFile keeps methodology language visible because modeled demographic signals are sensitive. Campaigns should understand what the score means before building a voter universe around it.
Complete California access review before any file upload.
Submit the campaign, requester, issuing source, viewer list, geography, and intended use. VoterFile does not open upload, import, browse, export, or voter-level modeled output until the California activation checklist is complete.
Questions
Does BISG identify a voter's ethnicity?
No. BISG produces probabilities from surname and geography signals. It should be treated as an outreach planning estimate, not as self-reported identity.
Why use BISG instead of surname matching only?
Surname matching ignores local context. BISG can combine surname information with geography, which usually gives campaigns a more useful and transparent planning signal.
Has VoterFile independently validated its BISG estimates?
Not yet. VoterFile has not published an independent accuracy, calibration, or subgroup-fairness study for this implementation, so the model card makes no individual-level accuracy claim.
Can BISG be used to infer candidate support?
No. The output is a demographic outreach-planning estimate. It does not measure ideology, issue preference, or candidate support.