Residual Votes and Ballot Design, 2000–2024

The test

The measure needs two numbers per place: ballots cast and presidential votes counted. No agency publishes both, which is why nobody publishes the residual vote. The ballots come from the Election Administration and Voting Survey, the questionnaire the U.S. Election Assistance Commission sends to every election office in the country after each federal election, asking each office to report on the election it just ran; the 2024 answers, Version 2, are downloadable from the Commission, and item F1a is the box marked total ballots cast. The votes come from the fifty-one state canvasses — the official count each state certifies after election day — gathered county by county into one public file by Tony McGovern, which this book’s data-sources chapter describes.

Why these sources, when the question is about individual ballots? The record that answers it properly is the cast vote record, which shows what each ballot actually carried — and it exists for a few dozen jurisdictions, which is not a country. Certified returns cover everywhere but are totals by construction, silent about any single ballot; the returns chapter sets out that limit. The subtraction is the one measure that works everywhere, and it buys that reach by giving up the reason a ballot came up empty.

The subtraction also has a property most measures lack. You cannot count more presidential votes than there were ballots to carry them, so the residual vote has a hard floor at zero. A negative residual vote is not a low rate; it is proof that one of the two sources is wrong. That turns the measure into a test of the record as well as a measurement of the ballots.

The data

The survey is one table, one row per election jurisdiction — the office that runs an election, 6,461 of them. Jurisdictions are counties in most states, but Wisconsin, Michigan and most of New England run elections by township. Those township rows are summed into counties using the first five digits of the FIPS code, the number the federal government assigns every place, whose first two digits name the state and first five the county. The returns come one row per county. The two are joined on the FIPS code, carried as text throughout, because 268 of the 2,890 county codes begin with a zero that a numeric column would silently drop. One joined row looks like this:

Allegheny County took in 722,248 ballots and counted 720,504 presidential votes. Subtract the second from the first and 1,744 ballots are left over that recorded no president: that is the residual vote. Divide it by the ballots cast — 1,744 out of 722,248 — and the residual-vote rate is 0.24% of everything cast. Every county in the file gets the same two steps.

Two cleaning decisions matter enough to state. First, the survey writes “did not answer” and “does not apply” as negative codes in the same columns as the counts. Every negative value is therefore set to missing before anything is added up; summed as they stand, those codes subtract real ballots from real counties. Second, where the join produced an impossible answer, more presidential votes than ballots, the row was not dropped. A disagreement between two official sources is a finding, and tidying it away would destroy it.

The result is four tables. counties.csv holds one row per county: fips, ballots (item F1a, summed to the county), total_votes (the presidential total), residual (the subtraction), residual_rate (the same thing as a percentage of ballots), and state_usable. That last column is a judgment this brief attaches, not a fact from either source: it flags three whole states whose answers imply they read “ballots cast” to mean something else. anomalies.csv holds the 103 impossible rows, with the same columns. unusable_states.csv names the three set-aside states and gives each a reason. states.csv rolls the usable counties up to state level.

Before you read on, commit to a number. Two official sources describe the same 2,993 counties, and the subtraction between them cannot come out below zero unless one of them is wrong. In how many counties does it? Write down a count out of 2,993 before you look.

What it says about democracy

The median county — the one in the middle when every county is ranked by its rate, with as many above it as below — has a residual vote rate of 0.93%: about one ballot in a hundred arriving without a usable presidential vote. The middle half of counties, setting aside the quarter with the lowest rates and the quarter with the highest, sit between 0.64% and 1.39%. One percent sounds small. In a state decided by twenty thousand votes it is not. And it is the size of number that design moves: one bad butterfly pushed one county’s spoiled ballots into the tens of thousands.

Most readers write down zero for the prompt above; an official count is not the sort of thing one expects to come out impossible. 103 counties report more presidential votes than ballots cast. Figure 2 puts them on the same axis as everybody else.

Figure 2. The arithmetic floor, and 103 counties on the wrong side of it. A histogram — bars counting how many counties fall in each narrow band of rate — with the floor drawn in, because this measure’s defining property is a boundary: nothing can sit left of the heavy black line and still be a measurement. The blue bars are the 2,803 counties whose two sources are at least consistent, with the median marked in red at 0.93%. Each red dot is a county reporting more presidential votes than ballots, spread vertically so overlapping cases stay countable. 23 fall left of the frame entirely, the worst at -269.8% in Potter County, South Dakota.

The largest gap is Cook County, Illinois, and it is diagnosable. The survey row for Cook County reports 1,112,978 ballots cast; the returns give the county 2,056,800 presidential votes, which is 943,822 more votes than there were ballots to carry them. The reason is that the City of Chicago runs its own elections, through a board separate from the Cook County Clerk, who runs them for the suburbs.⁶ The survey therefore has two rows where the returns have one: its county row is the suburbs alone, while the canvass counts the whole county, city included. The summing of smaller units into counties assumed that every smaller unit sits inside a county code, and here one does not, in the second most populous county in America. Which source is wrong cannot be determined from here; that one of them is wrong can be, and that is more than either source discloses about itself.

3,143 counties, from Census TIGER 2020 boundaries. Hover any county. County outlines are simplified to about ten vertices each; nothing in the argument rests on the shape of a county line.

Figure 3. Every county in the United States, colored by residual vote rate. A map, because the audit’s last question is where: 3,143 counties drawn from the Census Bureau’s published 2020 county boundaries (its TIGER map files), the blue ones passing the arithmetic test and shaded by rate. The 103 in red report more presidential votes than ballots; they fall in 19 states, most heavily in Texas with 14, a clustering that looks like reporting practice rather than like voters. The 87 in mid gray belong to the three states set aside below. The 150 in pale gray never entered the join at all, including every county in Alaska, Connecticut, and Wisconsin and 32 of Virginia’s independent cities.

The exceptions have their own stories, and each was found by the arithmetic failing rather than by any documentation.

Mississippi returns exactly 0.000% in every county — ballots cast equal to presidential votes, to the vote, statewide. That is not a state with perfect ballots; it is a state that answered the ballots-cast question with its presidential total. The District of Columbia’s 87.7% comes from a survey that reports by ward joined to returns that do not. Vermont’s 5.8% in every county at once is a signal about reporting, not about Vermont’s voters.

The floor also catches errors in one direction only. Wabash County, Illinois survives the test at 62.7%, nearly two thirds of its ballots recording no president, which no plausible mechanism produces. An error that inflates the rate passes the test and looks like data.

What survives is still most of the country: 2,803 of the 2,993 counties joined, or 94%. On those, about one ballot in a hundred is cast and records no president. And the audit is the other half of the finding. Two official datasets describing the same election disagree in 103 counties and three whole states, and nobody reconciles them, because no agency owns the comparison. The number that detected the Palm Beach ballot’s kind of failure still has to be assembled, by hand, from two files that were never built to meet.

What this brief cannot tell you

It cannot say why any ballot is blank. A deliberate abstention, a spoiled mark and a scanner failure all count the same in the subtraction. That is why the classic studies compare equipment types across otherwise similar places rather than reading levels directly: the level in one county alone attributes nothing. No equipment data is joined here, so the comparison that made this measure famous is not run in these pages.

It cannot say which source is wrong in any of the 103 impossible counties — only that one of them is. Settling it would take the underlying certified canvasses, county by county.

It cannot catch errors on the survivable side of zero. Wabash County is almost certainly as broken as the red counties, and it passes. A floor is a one-way test.

And it is one election. A single year is a photograph. Whether rates still vary by equipment as they did in 2000 needs more years and the equipment join, and both are extensions below.

What you should have learned

• A residual vote is ballots cast minus votes counted in the top race — a countable trace of ballot failure. In 2024 the median county’s rate was 0.93% of ballots cast.
• Ballot design moves that number measurably: after the 2000 election, residual rates sorted by equipment type, punch cards worst, and the finding helped end punch-card voting by law.
• The measure has a floor at zero, and 103 of 2,993 counties fall below it — proof that one of two official sources is wrong in each, though not which.
• Three whole states, one of them reporting exactly 0.000% everywhere, answered the ballots-cast question with something else and had to be set aside.
• No agency publishes both halves of the subtraction, so the number capable of detecting the next bad ballot design is published by nobody.

Extensions

None of these is answered above, and every one can be answered from the tables the Sources section hands over.

• counties.csv has residual_rate for every county, with state_usable beside it. Sort the usable counties by rate. Which places lose the most votes between the ballot and the count, and do they share anything?
• anomalies.csv holds the impossible rows. Sort by residual and read the worst ten. How many look like Cook County — a city counted separately from its county?
• unusable_states.csv gives each excluded state a reason. Decide whether you agree with each exclusion, and what including the state would do to the national median.
• states.csv has a residual_rate per state. Compare the highest and lowest usable states. Is a high state rate more plausibly voters, ballot design, or reporting?

Two stretch ideas point past the spreadsheet. The survey this brief takes its ballots from also asks what voting equipment each jurisdiction uses. Download the release from the Commission’s address below, join the equipment items onto counties.csv, and rerun the 2000-era comparison on 2024 data. And for one jurisdiction that publishes cast vote records, count its blank and overvoted presidential ballots directly and check the subtraction’s answer against the ballots’ own.

Sources

Data

• U.S. Election Assistance Commission, Election Administration and Voting Survey 2024, Version 2, item F1a. Version 2 rather than Version 1: the first release was corrected, and the Commission’s errata note says what changed. https://www.eac.gov/sites/default/files/2026-02/2024_EAVS_for_Public_Release_nolabel_V2_csv.zip
• County-level presidential returns for 2024, from Tony McGovern, United States General Election Presidential Results by District, Ward, or County from 2008 to 2024, by way of pres2024_counties.csv in the data-sources chapter, which documents the download and what the file is and is not. https://github.com/tonmcg/US_County_Level_Election_Results_08-24
• U.S. Census Bureau, TIGER/Line 2020 county and state boundaries, for the map. https://www2.census.gov/geo/tiger/TIGER2020/COUNTY/tl_2020_us_county.zip and https://www2.census.gov/geo/tiger/TIGER2020/STATE/tl_2020_us_state.zip

Academic literature

• Ansolabehere, S. and Stewart, C. III (2005). “Residual Votes Attributable to Technology.” Journal of Politics 67(2): 365–389. https://www.jstor.org/stable/10.1111/j.1468-2508.2005.00321.x
• Caltech/MIT Voting Technology Project (2001). Voting: What Is, What Could Be. https://vote.caltech.edu/reports/1
• U.S. Commission on Civil Rights (2001). Voting Irregularities in Florida During the 2000 Presidential Election. The Palm Beach ballot’s arrangement in Figure 1 is drawn from its account. https://www.usccr.gov/files/pubs/vote2000/report/main.htm
• Help America Vote Act of 2002, 52 U.S.C. § 20901 and following. https://www.law.cornell.edu/uscode/text/52/20901

The data itself

Every figure above is drawn from these tables. They are plain CSV, they open in a spreadsheet, and they are yours to take further.

anomalies.csv, counties.csv, fig_map_absent.csv, fig_map_attr.csv, fig_map_counties.csv, fig_map_frame.csv, fig_map_meta.csv, fig_map_state_lines.csv, fig_map_states.csv, states.csv, unusable_states.csv

I want to rebuild a public dataset from its original sources, without
any starting files. Please do the following and show your work.

THE FIRST SOURCE. U.S. Election Assistance Commission, 2024 Election
Administration and Voting Survey, Version 2, the no-labels CSV release:
https://www.eac.gov/sites/default/files/2026-02/2024_EAVS_for_Public_Release_nolabel_V2_csv.zip
A zip of about 2 MB holding one wide CSV. Each row is a local election
jurisdiction — 6,461 of them. Item F1a is total ballots cast. Negative
values like -99 and -88 are codes for "not reported" and "not
applicable", not counts. No account or key is needed.

THE SECOND SOURCE. County-level 2024 presidential returns. No federal
agency publishes these; use a public compilation such as Tony
McGovern's on GitHub (US_County_Level_Election_Results_08-24), which
gives votes for each candidate by county FIPS code.

WHAT TO BUILD. A residual vote is a ballot that came out of the
machine without a valid presidential vote on it: ballots cast minus
presidential votes counted.
1. One row per county: ballots cast, presidential votes, and the
   residual rate. EAVS jurisdictions are not counties everywhere — in
   New England, Wisconsin and Michigan they are townships, so sum
   jurisdiction rows to the first five digits of the FIPS code.
2. A separate list of the counties where the arithmetic is impossible
   — more presidential votes than ballots cast. Two federal sources
   disagreeing about the same election is a finding, not a nuisance;
   do not drop those rows quietly.
3. State rollups, setting aside states whose numbers cannot be right
   (every county exactly zero, or a median rate too high to believe).

CHECK YOUR WORK. The copies this book used were downloaded in August
2026. If your rebuild matches them, you should find:
- 6,461 jurisdiction rows in the survey file
- 2,890 counties in the finished table, and 103 more where ballots
  fall short of presidential votes — the impossible ones
- 48 state rollups, 3 of them then flagged as unusable
- a median county residual rate near 0.9 percent
The Commission reissues this file in numbered versions — the address
above already says V2. A later version shifts the counts because the
survey was revised, not because you made an error.