Does the US have a duplicate voter problem? We analysed 50 million records to find out.

Published January 2024, based on voter registration data obtained September–November 2023. Figures below refer to that 2023 snapshot.

Voter fraud is a politically charged topic in the United States. But strip away the politics and what remains is a data quality problem — one that entity resolution is well-suited to diagnose.

We gathered publicly available voter registration data from seven US states where the data is freely obtainable — Arkansas, Florida, Georgia, Michigan, North Carolina, Ohio, and Pennsylvania — representing over 49.5 million registered voters. We then ran Tilores entity resolution across the full dataset, comparing every registered voter against every other using name, address, date of birth, and other available fields. Here's what we found.

The methodology

Our analysis uses three matching techniques in combination:

Text similarity matching — fuzzy name matching to catch variations like "James" vs "Jim", hyphenated surnames, and transcription errors
Geographic proximity matching — linking records where addresses are similar or represent the same location under different formats
Temporal range matching — using date of birth ranges to account for data entry inconsistencies

Text similarity catches words with minor character differences — "Philip" and "Phillip" — as well as those that differ in spelling but are phonetically the same, such as "Steven" and "Stephen". Geographic proximity matching uses the coordinates of an address, so that two addresses within a defined distance of each other can be treated as a candidate match. Temporal range matching treats dates falling within a defined window, such as a date of birth out by a few days, as a potential match.

It is the combination of these techniques across the available fields that forms the deduplication rules. Multiple rules are generated to cover different duplication scenarios, and if any one of them is triggered when comparing two records, those records are treated as duplicates. Matched records are then connected into a single identity — an identity graph containing every voter record relating to that person — together with an explanation of which rule fired.

All source data is publicly available under state open records laws, purchased between September and November 2023. Costs varied by state: Michigan $23, Pennsylvania $20, Georgia $250. The National Conference of State Legislatures maintains a full list of US voter data accessibility by state.

One limitation shapes everything that follows: a duplicate registration is not a duplicate vote. Registration files tell you who is on the roll. They only tell you who actually voted if the state chooses to publish that field — and of our seven states, only Ohio and Pennsylvania do. Everything we say about actual double voting is therefore limited to those two states. In the other five the data is not publicly available — the state holds it, but does not publish it, so it was out of reach for an analysis run from public sources.

The headline numbers

Across all seven states, we identified 394,396 voters with duplicate registrations — representing 0.8% of the total voting population in our sample. A duplicate here means an active registered voter holding two or more active registrations, whether in the same county, in different counties within the same state, or in two different states.

That figure demands context. Almost none of it represents fraud. The overwhelming majority represents a genuine data management problem: voters who moved state or county, voters whose names appear in multiple formats, and voters who died but whose registrations were not removed in time. The system is producing duplicates faster than it can clean them up.

When voters move

The most common source of duplicates is interstate movement. When a voter moves from Georgia to Florida and registers in their new state, their old Georgia registration is not automatically cancelled. Both records remain active.

Name variations compound the problem. "James R. Wilson" in Georgia may appear as "Jim Wilson" in Florida — different enough that simple exact-match rules fail to link them, but identical enough that fuzzy matching catches the connection immediately.

This is not a detection failure so much as a structural one. US voter registration is managed at the state level, and a voter moving between states is not an event that automatically reaches the state they left. Tilores identified thousands of these cross-state pairs in the data.

The state breakdown

Every figure below is a duplicate registration rate — the share of a state's roll that resolves to a person already registered elsewhere on that roll, or in another state in our sample.

Florida — 148,516 duplicates, a rate of 1.10%. The highest count in the sample, both absolute and proportional.
Pennsylvania — 80,142 duplicates, a rate of 1.01%. The only other state above 1%.
Georgia — 51,876 duplicates, a rate of 0.73%.
Ohio — 47,000 duplicates, a rate of 0.79%.
Michigan — 35,723 duplicates, a rate of 0.47%.
North Carolina — 20,323 duplicates, a rate of 0.32%. The lowest rate in the sample.
Arkansas — 10,816 duplicates, a rate of 0.92%.

No state in our sample was a dramatic outlier, which is itself a useful finding: the problem is consistent and structural rather than concentrated in a few badly run jurisdictions. The spread between the best and worst performers — North Carolina at 0.32% and Florida at 1.10% — is real, but it is a spread within a common problem rather than evidence that some states have solved it.

We should be explicit about what these numbers do not tell you. They are a snapshot of the rolls as they stood in late 2023. They cannot be compared against the results of any past election — the rolls we analysed are not the rolls that existed on any previous election day, and a duplicate registration is not a vote in any case. Any comparison between these figures and a historical margin of victory would be meaningless, and we would ask that they not be used that way.

County-level variation

Duplicate rates vary significantly below the state level. In Arkansas the statewide average is 0.92%, but Searcy County — with a voting population of 5,805 — reaches 2.06%, or 121 individuals. In Florida, Hardee County has 13,654 registered voters and 314 duplicates (2.3%), while Osceola County has 214,588 registered voters and 5,376 duplicates (1.94%).

That last pair is the instructive one. Hardee and Osceola differ in size by a factor of fifteen, yet both sit around twice the national average we measured. County-level variation of this kind suggests the problem is partly driven by local data management practice rather than by scale or by systemic factors alone.

Registration rates: a separate signal

Something else emerged alongside the duplicate analysis: significant differences between the registration rate of certain counties and their general population. This is a registration rate, not a duplicate rate, and it needs reading separately.

Michigan stood out at one end. Several counties showed voter registration levels above 90% of the population — not feasible when roughly 22% of the US population is under 18, and therefore ineligible. The most likely explanation is that Michigan is not reconciling its voter lists against death notifications as effectively as other states, leaving deceased registrants on the rolls.

Arkansas sat at the other end, with some counties showing unusually low registration — Lincoln County at 26.02%. That points to room for improvement in voter engagement rather than to a data quality fault, and it is worth separating the two: a roll can be inaccurate because it holds people it should not, or because it is missing people it should hold. Entity resolution addresses the first. The second is a participation question.

A voter who dies while registered creates an orphaned record. Without automatic matching against death registries, that record persists. It cannot vote on its own — but it inflates registration counts in ways that distort analysis, and it is the kind of gap that erodes public confidence in the rolls even when nothing improper has happened.

Party affiliation in duplicate registrations

In every sampled state except Michigan, which does not disclose it, we were able to break duplicate registrations down by registered party affiliation. Looking only at voters registered as Democrats or Republicans, Democrats showed a duplicate rate of 0.96% and Republicans 0.89% — 121,985 and 120,384 records respectively.

By state, as a percentage of voters registered with each party:

Florida — Democrat 1.30% (58,455), Republican 1.29% (65,905)
Pennsylvania — Democrat 1.12% (39,797), Republican 0.96% (30,928)
Arkansas — Democrat 1.10% (632), Republican 1.23% (1,149)
Ohio — Democrat 0.68% (6,686), Republican 0.67% (9,061)
Georgia — Democrat 0.58% (8,713), Republican 0.43% (7,449)
North Carolina — Democrat 0.37% (7,702), Republican 0.29% (5,792)

The overall gap between the two parties is 0.07 percentage points, and no individual state gap exceeds 0.16 points. The direction is not uniform either: Republican-affiliated registrations show the higher rate in Arkansas, Democrat-affiliated registrations in most other states, and Florida and Ohio are effectively level at a hundredth of a point apart.

This is worth stating plainly, because it is the most misused kind of number in this subject area. Duplicate registrations in this dataset are not a partisan phenomenon. They are a function of how long someone has been registered, how often they have moved, and how their name has been recorded — none of which correlates meaningfully with party.

Double voting: what Ohio and Pennsylvania showed

Duplicate registration data is one thing, but on its own it means nothing more sinister than poor processes and untidy records. The larger question is whether anyone actually voted twice.

We only hold actual voting data for Pennsylvania and Ohio, so we examined the duplicates within that subset. There we found approximately 1,000 potential cases of an individual voting twice. Wanting to be stringent about a claim of this kind, we then applied additional deduplication rules and reviewed the remaining cases manually, one by one, narrowing them to 61 cases we were confident about.

Of those 61: 31 carried a Democratic Party affiliation, 21 a Republican Party affiliation, 6 involved both parties across the two records, and 3 had no affiliation. Two of the 61 span Pennsylvania and Ohio, both involving absentee votes; the rest sit within a single state.

Two examples give a sense of what these look like in practice. One individual registered to vote in West Chester, Pennsylvania in October 2020, shortly before the presidential election. Nine days later, the same individual registered in Philadelphia. There was a minor difference in the name; the date of birth and phone number were identical. Both registrations voted.

Another individual in Normalville, Pennsylvania registered twice at the same address, sixteen years apart almost to the day. The only difference between the two records was a small discrepancy in the date of birth. Both registrations voted. That is the most common pattern we saw — a date of birth out by a few days or months, with the name and address otherwise unchanged.

We want to be careful about what this is. It is a candidate set for review by election officials, not a legal finding. Entity resolution followed by manual review establishes that two records are very probably the same person; establishing that an offence occurred requires investigation by the relevant authorities, and that is not something a matching engine can or should do. The rate is vanishingly small relative to the registrations in those two states — but it was detectable, and it was only detectable because the vote history was published. Had the other five states published theirs, we would expect to find further cross-state cases.

What already exists: ERIC

We were not the first to run this analysis, and it would be misleading to present it as though we were.

A non-profit organisation called ERIC — the Electronic Registration Information Center — already performs cross-state list comparison for its member states using entity resolution, and supplements voter rolls with additional data sources such as state DMV records, in a privacy-preserving manner. We compared methodologies with them after publishing our first results. They were finding broadly the same duplicates we were, including the double-vote cases.

Two things follow from that. The first is that detection is not the bottleneck. The technology to find these records exists, is in use, and works. What happens next — actually reviewing and removing the records — sits with each individual state, and is outside ERIC's control.

The second is that coverage is uneven. A number of states have withdrawn from the consortium in recent years. A state that leaves loses the cross-state view of its own roll, and its departure also removes its records from the pool that remaining member states match against — so the effect is not contained to the state that left.

Where a state-level engine fits alongside ERIC

These are complementary rather than competing approaches, and the strongest position for a state is to do both.

ERIC membership gives a state the cross-state picture that no single state can assemble on its own, together with the data-sharing agreements and supplementary sources that make it work. Nothing a state runs internally can replicate that.

A state's own entity resolution instance answers a different question: what is the quality of my roll, right now? Running Tilores inside the state's own AWS account gives continuous resolution of its own records, under its own control, with results available between reporting cycles rather than after them. The same engine then applies to every other identity dataset the state holds — benefits, licensing, vehicle records, procurement — which is usually where the larger operational return sits.

A state would get better results than we did

This is the most important caveat in the whole analysis, and it cuts in the opposite direction to most caveats.

Everything above was produced from the fields that appear in public voter files: name, address, date of birth, registration status, party affiliation. That is a deliberately limited view, because a public showcase can only use public data. We were matching with one hand tied behind our back.

A state election office holds considerably more about each registrant — driver's licence numbers, full or partial Social Security numbers, prior addresses, application and signature history, and links into DMV and vital records systems. Those additional identifiers improve matching in both directions at once. They confirm genuine duplicates that public data alone cannot resolve, and they rule out false positives where two different people happen to share a name and an approximate date of birth. Precision and recall both go up.

Vote history is the clearest illustration. Five of the seven states we analysed record whether a registrant voted but do not publish it — which is the sole reason we could not test those five for double voting. Each of them already holds everything needed to run that test on its own data.

The practical implication is that the figures on this page should be read as a floor, not a ceiling. A state running the same engine against its own complete records would find duplicates we could not see from the outside, and would do so with materially higher confidence in every individual match. This is also why a state-level deployment is not merely a smaller version of what we did — it is a strictly better version of it.

What would actually reduce the problem

Wider participation in cross-state matching. The mechanism already exists. The constraint is how many states take part, and whether they act on what comes back.

Automatic death registry synchronisation. Social Security Administration death data should be matched against voter rolls continuously, rather than on an annual or ad hoc basis. The matching must use fuzzy logic, not exact match — death records and voter records use different name and address formats.

Fuzzy matching as the standard, not exact-match rules. Exact matching misses the large majority of genuine duplicates, because real-world names and addresses are never entered consistently. A rule that only catches perfectly identical records will report a clean roll while the duplicates remain on it.

The same technique protects eligible voters

It is worth being clear that entity resolution is not only a tool for removing records. It is equally a tool for not removing the wrong ones — and that matters just as much to the accuracy of a roll.

In 2018 the State of Georgia attempted to implement an "exact-match" requirement, under which a voter registration had to match the individual's driver's licence or Social Security record exactly. Any small variance placed their registration in a "pending" status until corrected. Around 51,000 registrations were pending in Georgia that year, disproportionately affecting voters with names more likely to be misspelled or transliterated inconsistently in official records. Several other states, including Wisconsin and Virginia, had already cancelled or scaled back similar requirements.

The underlying failure is the same one that lets duplicates survive: exact matching cannot cope with the ordinary variance of real-world data. Would it be reasonable for José Rodríguez to be unable to vote because his driving licence reads Jose Rodrigues, when the date of birth and address match? To a human reviewer these are obviously the same person. Entity resolution lets a computer reach the same conclusion.

Any state cross-validating its electoral register against another database — social security, driving licences, death records — should be doing so with a system that accommodates that variance. The alternative produces two kinds of error at once: duplicates that survive, and eligible voters who are wrongly flagged.

Identity resolution beyond the electoral roll

The deeper issue this analysis surfaces is structural: any process that depends on accurate records about people will be vulnerable when those records are managed without entity resolution.

This is not unique to voter registration, and voter registration is not even the largest instance of it. The same problem — records of the same person fragmenting across systems, formats, and decades — drives duplicate enrolment and improper payments in benefits programmes, fragmented case files across agencies, duplicate licence and permit records, tax records that split one person across years and filing types, and vendor duplication in procurement and grant systems. Those datasets are larger, messier, and carry more money.

Education data is a particularly close parallel. A student moves between schools, districts, and states, and re-enrols each time under a slightly different name, address, or identifier — the same mechanism that produces a duplicate voter registration when someone moves. The consequences differ, though: where a duplicate voter registration inflates a roll, a fragmented student record breaks longitudinal outcome tracking, distorts funding allocations that are calculated per pupil, and can mean a learner who needs support is invisible to the system meant to identify them, because no single record shows their full history.

We used voter data because it is one of the few large-scale, publicly available identity datasets in the United States — which makes it a good way to show what entity resolution finds, and what is missed without it.

Why we did this

Tilores was built for situations like this: large datasets, real-world name and address variation, high stakes if the resolution is wrong. If your organisation works with data where duplicate records carry real-world consequences — in benefits integrity, fraud, compliance, healthcare, or public records — entity resolution is the right place to start.

For US government agencies: Tilores is available to federal, state and local agencies through our partnership with Carahsoft, and deploys into your own AWS account including AWS GovCloud, on-premise, or into air-gapped networks. Book a demo or see our government use cases.

Download this analysis as a PDF →
View the interactive showcase →