BRENNEN

Industries / Government, education and community / Campaigns and polling

Voter files

Voter files are the official lists of people registered to vote, kept by election offices, and the larger lists that parties and data firms build from them to plan campaigns.

7

Milestones achieved, from 1951 to 2023.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1951

    Women enrolled by name

    For independent India's first general election, in 1951–52, officials compiled the rolls by a house-to-house survey and enrolled about 173 million voters. In several states many women had been entered only as someone's wife or mother, because they were "averse to disclosing their proper names to strangers." Chief Election Commissioner Sukumar Sen ruled that "no elector should be enrolled unless sufficient particulars, including the name, were given," and by his official account about 2.8 million of nearly 80 million women were deleted. After a campaign by officials, parties and women's groups, 92,141,597 women were on the rolls for the 1957 election, about 94% of adult women by Sen's count.

    LessonA rule on how names are recorded can keep millions off a roll unless officials work to enroll the people it affects.

    deccanherald.comorfonline.org

    Credit

    Sukumar Sen, Election Commission of India

  2. 2002

    One computer list per state

    After the disputed count of the 2000 presidential election, the Help America Vote Act (HAVA), signed on 29 October 2002, directed every state to keep "a single, uniform, official, centralized, interactive computerized statewide voter registration list." Pew Research Center notes that these digital lists let parties and companies build national voter files for the first time by combining the file from each state and the District of Columbia.

    justice.govpewresearch.org

    Credit

    the US Congress; signed by President George W. Bush

  3. 2008

    Fingerprints on the roll

    Brazil's electoral court, the Tribunal Superior Eleitoral (TSE), began taking voters' fingerprints, photo and signature in a 2008 pilot in three towns: São João Batista (Santa Catarina), Fátima do Sul (Mato Grosso do Sul) and Colorado do Oeste (Rondônia). On election day a fingerprint reader attached to the electronic voting machine confirms the voter's identity before the machine accepts a vote. By the count reported by the newspaper Correio Braziliense in May 2026, about 135 million voters, about 87% of the electorate, are now registered this way.

    diariodepernambuco.com.br

    Credit

    Tribunal Superior Eleitoral

  4. 2012

    States share their lists

    Research for the Pew Center on the States found that about 24 million US voter registrations, one in eight, were no longer valid or were significantly inaccurate: more than 1.8 million dead people were listed as voters and about 2.75 million people were registered in more than one state, while at least 51 million eligible citizens were not registered. The same year seven states founded the Electronic Registration Information Center (ERIC), a nonprofit run by member states that compares their records to find voters who have moved or died and to reach eligible people who have not registered. From 2022 Louisiana, then Alabama, Florida, Missouri, West Virginia, Iowa, Ohio and Texas left or announced they would leave, citing data privacy, election integrity or partisanship; ERIC now lists 27 states and the District of Columbia.

    LessonA shared list is only as complete as its membership, and each exit removes data the remaining states relied on, as Matt Germer of the R Street Institute told Route Fifty.

    pew.orgericstates.orgroute-fifty.com

    Credit

    the seven founding states; Pew

  5. 2018

    Commercial files put to test

    Pew Research Center tried to match about 5,000 members of its American Trends Panel, a nationally representative survey panel, to five commercial voter files: two from nonpartisan vendors, two from vendors working mainly for Democratic and progressive clients and one from a vendor working mainly for Republican and conservative clients. The vendors found between 50% and 79% of panelists, 69% on average, and found 77% of those who said they were registered but only 38% of those who said they were not. Among panelists all five files found, 85% had identical turnout records, and modeled party matched self-reported party fairly well, but education was missing or wrong 48% of the time and household income 63%.

    LessonThe files describe registered, settled adults well and young people and frequent movers poorly.

    pewresearch.org

    Credit

    Pew Research Center

  6. 2020

    Big-data claims deflated

    The UK Information Commissioner's Office (ICO) closed its investigation of SCL Elections and Cambridge Analytica with a letter of 2 October 2020 to Julian Knight MP. It found the firm had bought commercially available data, by one estimate more than 130 billion data points, mainly on millions of US voters, and combined it with data gathered through a Facebook app by the Cambridge University academic Aleksandr Kogan; Facebook estimated the app collected data on up to 87 million people. The ICO found the models were built mostly with standard open-source data science libraries such as "scikit", that SCL's claim of "over 5,000 data points per individual on 230 million adult Americans" "may have been an exaggeration," and that the predictions' accuracy on new people "was likely much lower" than in training.

    LessonClaims about the power of voter-file models need testing by someone outside the firm making them.

    cy.ico.org.uk

    Credit

    UK Information Commissioner's Office

  7. 2023

    Britain measures its registers

    The UK Electoral Commission's study of the December 2022 registers found the local government registers in Great Britain 86% complete and 88% accurate, and estimated that 7 to 8 million people across the UK were not correctly registered at their current address. It concluded there was little evidence that accuracy and completeness would improve much without major changes, and recommended letting public bodies share data so that people could be registered without applying.

    electoralcommission.org.uk

    Credit

    the Electoral Commission

Watch for

  1. Citizenship checks in bulk

    In 2025 US Citizenship and Immigration Services (USCIS) rebuilt SAVE, its system for checking a person's immigration or citizenship status: since May it has been linked to Social Security Administration data and can be queried with a name, a date of birth and the last four digits of a Social Security number, so states can run whole voter lists through it. By the agency's count more than 47 million voters had been checked by December 2025. A USCIS spokesman called SAVE "a secure, and reliable way" to verify the citizenship of people registering to vote. The Department of Homeland Security's own privacy assessment says that "due to misspellings of names, transposed numbers or incomplete information, the SAVE program may produce inaccurate results," and that the Social Security queries states use for voter lists do not allow its usual second and third steps of review. NPR found a naturalized citizen among the 2,724 Texas voters, about 0.015% of more than 18 million, whom the state flagged as potential noncitizens.

    Achieved whenan independent body, such as the DHS inspector general or the Government Accountability Office, publishes how often SAVE wrongly flags citizens on voter lists and how many flagged records were confirmed as noncitizens.

    kpbs.org

    Claimed by

    USCIS, December 2025

  2. Bihar's rolls rechecked

    The Election Commission of India (ECI) ran a Special Intensive Revision (SIR), a fresh revision of the whole electoral roll, in the state of Bihar from 24 June to 30 September 2025. The roll went from 7.89 crore electors (78.9 million; a crore is 10 million) to a draft of 7.24 crore after 65 lakh names were removed (6.5 million; a lakh is 100,000), then to about 7.42 crore after 21.53 lakh were added and 3.66 lakh more removed. The ECI called the exercise a "grand success" and plans to repeat it nationwide; opposition parties, some of which went to the Supreme Court, say it wrongly deleted eligible voters, and leaders of the governing BJP defend it.

    Achieved whenan independent check, such as a court-ordered or academic field survey of a random sample of deleted names, reports how many deleted electors were in fact eligible.

    pib.gov.inmillenniumpost.in

    Claimed by

    Election Commission of India, September 2025

  3. Registering without a form

    The UK government's Representation of the People Bill would let electoral registration officers add people to the register, or update their details, "without an application," using data held by other public bodies. Pilots in Wales added just over 14,500 new voters "with high accuracy," by the Electoral Commission's account, and the Welsh Government plans to roll the system out before the 2027 local elections.

    Achieved whendirect registration is in force for a UK or Welsh election and the Commission's independent evaluation reports higher completeness with no loss of accuracy.

    electoralcommission.org.uk

    Claimed by

    the UK Government and the Welsh Government; Electoral Commission briefing of 3 October 2026

Open

  1. 01

    Roll-matching tested on invented people

    A public test bench of invented registrants with a known truth table (twins, very common names, name changes, moves, deaths, typing errors and names written in more than one script), and code that runs published list-maintenance matching rules on them: first name, last name and date of birth alone; fuzzy matching on many fields; and Social Security number lookups of the kind SAVE now runs. For each rule it reports how many eligible invented people would be flagged, and how many true duplicates would be missed, per 100,000 records. No real person's record is used.

    The brief
    Credit · unclaimed
  2. 02

    Death removals against actual deaths

    For every US county and every two-year cycle since 2016, compare the number of registrations removed because the voter died, which election offices report to the US Election Assistance Commission (EAC) as item A12c of its Election Administration and Voting Survey (EAVS), with the number of adult deaths the Centers for Disease Control and Prevention (CDC) recorded in the same county and period, adjusted by the county's registration rate. Publish the ratios with notes on gaps. Only published totals are used; no individual is looked up.

    The brief
    Credit · unclaimed

Open 01 · the brief

Roll-matching tested on invented people

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test bench of invented registrants with a known truth table (twins, very common names, name changes, moves, deaths, typing errors and names written in more than one script), and code that runs published list-maintenance matching rules on them: first name, last name and date of birth alone; fuzzy matching on many fields; and Social Security number lookups of the kind SAVE now runs. For each rule it reports how many eligible invented people would be flagged, and how many true duplicates would be missed, per 100,000 records. No real person's record is used.

Why it is still open

DHS's own privacy assessment says SAVE "may produce inaccurate results," and NPR reports that the overhaul came without public data on its accuracy. The best-known measurement, by researchers from Harvard, Stanford, the University of Pennsylvania and Microsoft Research, used real Iowa records to test a single rule, name plus date of birth, and found it would remove about 200 valid registrations for every registration used to vote twice. Record-linkage researchers at Leipzig University publish a benchmark of up to 10 million person records built from North Carolina's public voter file, with duplicates made by the GeCo corruption tool, but it scores matching methods in general, uses real people's records, and has no labels for twins, deaths or moves.

kpbs.orgpropublica.orgdbs.uni-leipzig.de

Built with what exists

Splink, the open-source record-linkage library from the UK Ministry of Justice; Python's Faker library and public name-frequency tables, such as the US Census Bureau's surname file and the Social Security Administration's baby-name lists, to generate people; DuckDB for the runs.

  1. First step

    Write down five matching rules from public documents, including the name-plus-birth-date rule and SAVE's documented match fields, and generate a first population of 100,000 invented people with set shares of twins, name changes and movers.

  2. Done looks like

    At least 1 million invented records with a published truth table; at least five matching rules scored, each with false-flag and missed-duplicate rates per 100,000 and 95% confidence intervals; data and code released; a second team reproduces every rate within 5%.

Open 02 · the brief

Death removals against actual deaths

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

For every US county and every two-year cycle since 2016, compare the number of registrations removed because the voter died, which election offices report to the US Election Assistance Commission (EAC) as item A12c of its Election Administration and Voting Survey (EAVS), with the number of adult deaths the Centers for Disease Control and Prevention (CDC) recorded in the same county and period, adjusted by the county's registration rate. Publish the ratios with notes on gaps. Only published totals are used; no individual is looked up.

Why it is still open

The EAC reports that more than 21 million registrations were removed between the 2022 and 2024 registration deadlines and that 78.2% of states share data with bodies that hold death records, but its report gives no comparison between death removals and actual deaths. The last national estimate seen in this research, Pew's 2012 figure of 1.8 million dead people on the rolls, is fourteen years old.

eac.govpew.org

Built with what exists

The EAC's EAVS public data files for 2016 to 2024; CDC WONDER mortality data by county and age; Census Bureau citizen voting-age population estimates; Python and pandas.

  1. First step

    Download the 2024 EAVS file and CDC WONDER deaths of people aged 18 and over for October 2022 to October 2024, and compute the ratio for the counties of one state.

  2. Done looks like

    Ratios for at least 3,000 counties over at least four cycles, with every county that reported "Data Not Available" listed; a written method; and 50 counties checked by hand against their state's own list-maintenance reports.