Cast Vote Record Data

CVRs are electronic records of every choice made on every ballot (Wack Reference Wack2019). They are by-products of voting scanners and tabulators, that scan paper ballots and translate each marked vote into an official vote. Election officials also often use CVRs to conduct post-election audits. CVRs are anonymous. While CVRs are not usually made public following an election, officials sometimes release them voluntarily to promote the transparency of their administration, and more commonly release them in response to public records requests. As noted above, CVR data have been used in previous studies of voting behavior (Bafumi et al. Reference Bafumi, Herron, Hill and Lewis2012; Frisina et al. Reference Frisina, Herron, Honaker and Lewis2008; Gerber and Lewis Reference Gerber and Lewis2004; Herron and Lewis Reference Herron and Lewis2007; Herron and Sekhon Reference Herron and Sekhon2003; Kuriwaki Reference Kuriwaki2025; Lewis Reference Lewis2001; Morse Reference Morse2021; Wand et al. Reference Wand, Shotts, Sekhon and Mebane2001). CVRs are unique in the ability to measure the joint relationships between different contests exactly, avoiding the limitations of aggregated election returns and surveys (Kuriwaki et al. Reference Kuriwaki, Reece, Baltz, Conevska, Loffredo, Samarth and Mutlu2024).

Until recently, CVRs covering large portions of the US were not available for academic research. After the 2020 election, a group of election skeptics, data scientists, and other interested parties engaged in a large-scale, crowd-sourced effort to collect CVR data from county officials, often flooding them with public records requests.Footnote ¹⁰ We use the CVR database that resulted from this activity, which Kuriwaki et al. (Reference Kuriwaki, Reece, Baltz, Conevska, Loffredo, Samarth and Mutlu2024) describe in detail. We build on the data in Kuriwaki et al. (Reference Kuriwaki, Reece, Baltz, Conevska, Loffredo, Samarth and Mutlu2024) with records collected or cleaned later, and we also include cases where the CVRs cover only a part of a county.

Our initial data includes over 50 million voters. Our entire dataset covers 24 states and 480 counties. In this sample, Biden’s share of the two-party vote is 56.2%, which is 4 percentage points more than his vote-share nationwide. Thus, our subset is slightly more Democratic than the nation as a whole. Administered during the pandemic, this election featured roughly 43% of the national electorate voting absentee or by mail, compared to 21% in 2016 and 29% in 2024.Footnote ¹¹ In the collection of counties included in our data, 53% of the voting age population is white, 12% are Black, and 27% are Hispanic.Footnote ¹²

Categorizations of Local Offices and Ballot Measures

A time-consuming component of our data processing is the collection of information on the candidates and ballot measures. The raw CVRs do not include the wording of the local ballot measures. We used existing election databases, voter guides, and newspaper coverage to categorize as many offices and measures as possible. Details are given in subsequent sections.

For this article, we divide contests into six categories: a contest is either a partisan office, a nonpartisan office, or a ballot measure, and it is also either at the state or local level. Partisan offices are simply those where candidates run with party labels next to their names on the ballot. Almost always, those candidates must win their own party’s primaries to get to the general election. Nonpartisan offices are those without party labels. A given office, such as mayor or county council, or even state judicial offices, can be partisan in some places but nonpartisan in others. We define state offices to include elected officials the state government—statewide executive officers, commissioners, state legislators, and state judges. Local offices encompass everything else, from municipal and county districts to school districts.

Ballot measures are quite common in the US at the state and local level. All states except Delaware require voter approval of state constitutional amendments. Many states allow initiatives and referendums, allowing voters to pass laws or constitutional amendments directly, or to strike down a law passed by the legislature. Some states require voter approval of certain types of laws and policies, such as general obligation bonds or tax increases. In many localities, voter approval is required for bonds, tax increases, and to prevent some taxes from expiring. The state ballot measures in our sample are a mix of legislatively approved measures and initiatives. Almost all of the local measures were placed on the ballot by a local government, although a few are initiatives. Most measures require a simple majority to pass, although some require supermajorities.

Not all of the votes for President and other federal offices can be linked to subnational offices or ballot measures. Some voters do not face contested state or local races, or do not have the opportunity to vote on ballot measures. In certain cases, the CVR data split each voter’s complete ballot across two or more rows representing separate pages of their physical paper ballots, and we are unable to link voting across these split records (for details, see Kuriwaki et al. Reference Kuriwaki, Reece, Baltz, Conevska, Loffredo, Samarth and Mutlu2024). Therefore, we study various subsets of this data given the availability of offices on the ballot. Figure 1 shows our sample sizes and their locations after we limit them to those linked to a federal office. State-level offices are in the top row, while local offices are in the bottom. The columns are organized by partisan offices (left), nonpartisan offices (middle), and ballot measures (right).

Figure 1. Coverage of Cast Vote Records

Note: Total voters linked to their federal ballot, by state and category. Section A of the Supplementary Material shows exact counts by state.

Many local elections are held off-cycle, and our data are all from a presidential election year, 2020. Nonetheless, the data contain many local contests—for example, 544 school board races, 602 municipal legislative races, and 1,768 local ballot measures (see Tables 1 and 4). For each of these local contests—and also for state contests—the data contain a record of how each individual voted in that contest and how that voter voted for other offices such as the U.S. President and the U.S. House.

Table 1. Voting Patterns by Partisanship

Note: $ {D}^D $ is the average percentage of the two-party vote cast for the Democratic candidate by National Democrats, and $ {D}^R $ is the percentage of the two-party vote cast for the Democratic candidate by National Republicans. The Absolute Partisan Gap shows the average across contests.

We further limit the sample analyzed in the following ways. First, we drop uncontested races, and minimally contested races where the winning candidate or choice received more than 95% of the votes. Second, for nonpartisan races, we drop cases where the top-two candidates received less than 75% of the total votes cast. Third, we drop contests with less than 100 votes. These small contests often occur in special districts, for example, for water or sanitation. Fourth, we drop contests where voters could cast votes for more than one candidate—for example, multi-member district and ranked choice elections. Just over 15% of the contests in our sample were of this type. For our replication data, see Conevska et al. (Reference Conevska, Hirano, Kuriwaki, B. Lewis, Mutlu and M. Snyder2025).

Measures of Partisan Voting

We want to capture the differences or similarities in how Democrats and Republicans vote as simply and intuitively as possible. We therefore focus on a measure we call the Absolute Partisan Gap. Computed at the contest level, it ranges from 0 to 100, and it can be applied to a wide variety of first-past-the-post contests, including partisan races, nonpartisan races, and ballot measures.

To construct this measure, we first divide voters according to their National Partisanship based on their vote choices in the elections for the President, the U.S. House, and U.S. Senate. Specifically, we call a voter Democratic if they voted only for Democratic candidates for federal offices, and we call a voter Republican if they voted only for Republican candidates for these offices. All others are split-ticket voters, who we set aside in most of our subsequent analyses. By this measure, only 6% of our sample are split-ticket voters.Footnote ¹³ Although our vote-based measure of partisanship differs from the conventional survey-based measure of partisan identity, election surveys indicate that over 90% of voters that would be classified as national partisans using our definition in fact identify with that party.Footnote ¹⁴

To define the $ APG $ , let $ {V}_{1j}^D $ be the total votes cast for the winning candidate (or alternative) in contest j among Democratic voters, let $ {V}_{2j}^D $ be the total votes cast for the second place candidate (or alternative) among Democratic voters, and let $ {W}_j^D=100\hskip0.3em {V}_{1j}^D/({V}_{1j}^D+{V}_{2j}^D) $ be the percentage of votes cast for the winner among Democratic voters. Let $ {V}_{1j}^R $ , $ {V}_{2j}^R $ , and $ {W}_j^R $ be the analogs for Republican voters. The $ APG $ is simply:

⁽¹⁾ $$ \begin{array}{rl}{APG}_j\equiv |{W}_j^D-{W}_j^R|.& \end{array} $$

For example, consider a race where the Democratic candidate wins. If all Democrats vote for the Democratic candidate and all Republicans vote for the Republican candidate, then $ APG\hskip0.3em =\hskip0.3em 100 $ (i.e., 100 percentage points). If 75% of Democrats vote for the Democratic candidate and 75% of Republicans vote for the Republican candidate, then $ APG=50 $ . If 50% of Democrats vote for the Democratic candidate and 50% of Republicans vote for the Republican candidate, then $ APG=0 $ . As another example, consider a nonpartisan race in which candidate A wins, followed by second place loser candidate B, and third place loser C. Suppose that among Democrats who voted for either candidate A or B, 65% voted for candidate A, and among Republicans who voted for either candidate A or B, 40% voted for candidate A. Then $ APG=25 $ .Footnote ¹⁵

Note that we can compute $ APG $ in other ways. Consider a partisan office, and let $ {D}_j^D $ be the percentage of the two-party vote cast for the Democratic candidate among Democratic voters, and let $ {D}_j^R $ be the analogous percentage among Republican voters (that is, $ {D}_j^R $ is the percent of Republicans who split their ticket in the contest). Then

(2) $$ \begin{array}{rl}{APG}_j=|{D}_j^D\hskip0.3em -\hskip0.3em {D}_j^R|.& \end{array} $$

As a practical matter, the absolute value is unnecessary, because $ {D}_j^D>{D}_j^R $ in all cases. Note also that if there are only two alternatives in the contest and the absolute partisan advantage and the absolute vote share margin in the contest are equal, then the $ APG $ is equal to the individual level correlation between partisanship and voting. More generally, if $ {y}_i $ is a binary variable for voter i voting Democrat in a local contest j, and $ {d}_i $ is a binary variable for being a Democrat, then $ APG $ _j = |Corr(y, d)| · SD(y) / SD(d).Footnote ¹⁶

For ballot measure contests, similarly let $ {Y}_j^D $ be the percentage of votes cast for the Yes alternative among Democratic voters and let $ {Y}_j^R $ be the analogous percentage among Republican voters. Then $ {APG}_j=|{Y}_j^D\hskip0.3em -\hskip0.3em {Y}_j^R| $ . Finally, consider the following. We manually collect metadata about each state and local ballot measure to determine which of the two alternatives can be considered more “Liberal” or more “Conservative” (more details on this below). For each ballot measure j where we make this determination, let $ {L}_j^D $ be the percentage of votes cast for the Liberal alternative among Democratic voters, and let $ {L}_j^R $ be the analogous percentage among Republican voters. Then $ {APG}_j=|{L}_j^D\hskip0.3em -\hskip0.3em {L}_j^R| $ . The absolute value is almost unnecessary for these contests, because $ {L}_j^D>{L}_j^R $ in almost all cases.

One limitation of the $ APG $ is that it focuses on the top two candidates or alternatives, ignoring other choices and roll-off. Our measure using percentages may be sensitive to the small baseline prevalence of Democrats and Republicans in a locality. However, when we consider more complicated and general measures that incorporate these other choices, the quantitative estimates change only slightly, and the qualitative conclusions do not change at all.Footnote ¹⁷

Main Results

Table 1 presents basic summary statistics for three types of contests—partisan races in the top panel, nonpartisan races in the middle panel, and ballot measures in the bottom panel. Within each panel, the first row covers state contests and the second covers local contests.

The first two columns of the top panel show, for each type of voter (Democratic or Republican), the average percentage voting for the Democratic candidate in state and local partisan races. Evidently, voting in these races is highly partisan, with 92% to 97% of choices matching voters’ National Partisanship. As a result, the average values of $ APG $ are quite high—91 points for state partisan races and 85 points for local partisan races.

For nonpartisan races and ballot measures, the average $ APG $ is much lower. Also, within each panel, the average $ APG $ is noticeably lower in local contests than in state contests.

Table 1 shows that the average $ APG $ for partisan offices differs dramatically from those for nonpartisan offices and ballot measures. But what about the variation? Figure 2 below shows histograms that give a sense of the spread around the average values. The panels are ordered in the same way as in Figure 1. We see that among partisan offices, the $ APG $ s are all centered at the high end around 90 points, while the local ballot measures and nonpartisan offices are centered at the opposite end of the spectrum.

Figure 2. The Distribution of the Absolute Partisan Gap, by Type of Office

Note: Each histogram shows the distribution of the Absolute Partisan Gap (horizontal axis). The vertical axis shows proportions within each type of office.

How many nonpartisan races for state and local offices exhibit highly partisan behavior? How many state and local ballot measures are highly partisan? It depends on what threshold we use to define “highly partisan.” For partisan offices, the average value of the $ APG $ is more than 80 percentage points. Using 80 points as the threshold, almost no nonpartisan elections or ballot measures qualify. Relaxing the threshold to 60 points changes the story only slightly. Relaxing it further to 40 points—produced by partisan differences of [70% – 30%] or more—we see that 41% of state contests are classified as highly partisan, as are 16% of local contests.

Most of the nonpartisan state races with $ APG>40 $ are judicial contests in Ohio. Although these elections are nominally nonpartisan, they have a unique feature: party labels do not appear on the general election ballot, but the candidates are chosen in partisan primaries.Footnote ¹⁸ It is not surprising that these races are relatively partisan, since party organizations and politicians often campaign heavily for their nominees (Cheek and Champagne Reference Cheek and Champagne2003; Schotland Reference Schotland2006). If we set these aside, then the average value of $ APG $ for state offices is just 28 points, and only 25% of the contests have $ APG>40 $ .Footnote ¹⁹

Theoretically, the $ APG $ s for local nonpartisan offices and local ballot measures might be small simply because the contests are extremely uncompetitive. If the winning side receives 80% of the votes, for example, then the gap between Democratic and Republican voters is unlikely to be large. Indeed, the relationship between $ APG $ and the margin of victory is negative. However, as shown in Section C of the Supplementary Material: (i) most of the local contests in our sample, especially for nonpartisan offices, are actually fairly competitive; (ii) the relationship between the margin of victory and the $ APG $ is weak, and (iii) the relationship only becomes noticeable when the margin of victory is more than about 25 to 30 percentage points, where a small fraction of the contests lie. So, this factor cannot account for the small $ APG $ s.

One reason that the $ APG $ tends to be small in local nonpartisan contests is that the top two candidates in many of these races probably have similar ideologies, party affiliations, and positions on national policies. But even when that is not the case, voters might care more about candidates’ positions on local issues or their performance on local matters, rather than candidates’ ideologies, party affiliations, and positions on national policies. The latter are likely irrelevant for the day-to-day work of a county treasurer, town clerk, or school board member. They are probably not often relevant for county commissioners or mayors either, except perhaps in major cities or the largest counties. Based on our own reading of hundreds of newspaper articles, campaign advertisements, and editorials, it appears that local issues dominate the coverage and debates (see also Oliver, Ha, and Callen Reference Oliver, Ha and Callen2012).

Moreover, if the top two candidates in many nonpartisan local races have similar ideologies, party affiliations, and positions on national policies, it is important to acknowledge that this is equilibrium behavior. That is, there is probably no strong demand for candidates who espouse strongly opposing ideologies, party affiliations, and positions on national policies—otherwise, such candidates would have an incentive to run. Figure 3 shows that although there is some tendency for $ APG $ to be higher in areas that are more competitive, the differences between competitive and uncompetitive areas are small.Footnote ²⁰ Thus, even in areas where there are roughly equal numbers of Democrats and Republicans, most of the nominally nonpartisan local races do not divide the electorate along partisan lines.

Figure 3. The Absolute Partisan Gap vs. Partisan Competitiveness in Local Area

Note: Red lines indicate local averages. A contest is considered competitive (right panel) if the winner receives less than two-thirds of the total votes cast for the top two candidates.

We also investigated the degree to which $ APG $ varies depending upon constituency size and racial segregation, focusing on local contests.Footnote ²¹ In large counties and cities, the local government might not seem very “local.” Fewer voters may be personally acquainted with the candidates running, or have firsthand knowledge of the overall quality of services in the polity. In these cases, partisan or ideological cues might have a larger role. We find that the $ APG $ is modestly higher in larger constituencies. The total number of voters explains between 4% and 17% of variation in $ APG $ , depending on the type of contest. Second, given the claims regarding the significance of race in urban politics and the overlap between racial, ideological, and partisan cleavages in recent years, we might expect voting to be more partisan or ideological in racially segregated areas (e.g., Trounstine Reference Trounstine2018). An index of racial segregation explains between 1% and 7% of variation in the $ APG $ in local contests. Evidently, these variables can only account for a small percentage of the variation in the $ APG $ .

Defining Partisanship by Party Lever Use in Michigan

Above, we measure national partisanship using votes for presidential and congressional candidates. Voters in some states have the option to pull the party lever. This automatically votes a straight party ticket for all nominally partisan contests, which voters can then override on a race-by-race basis. For Michigan, we have information about whether voters exercised this straight-party option or not, and whether they deviated from this option anywhere on the ballot. This is arguably a more direct expression of partisan preference, and can also be used as a measure of partisanship. Using this measure produces $ APG $ s for partisan state and partisan local offices that are about 4–9 points higher than the estimates produced using our main measure (see Section C of the Supplementary Material). Part of this might be due to a type of mechanical effect—because voters know that their choices in partisan races are all filled in automatically, they might not check these races carefully down the ballot (Thornburg, Davis, and Buell Reference Thornburg, Davis and Buell2025). For nonpartisan contests and ballot measures, which are still left blank even after pulling the party lever, the $ APG $ s are similar using either definition of partisanship.

Roll-Off and Minor Candidates

The results above focus on the two leading candidates or alternatives. They therefore exclude roll-off—voters who turn out to vote but skip certain contests—as well as votes cast for candidates other than the winner and first runner-up, which we refer to as minor candidates. Here we show that including these other choices does not substantially change the results.

The first panel of Table 2 shows that 4.4% of voters in contested state partisan offices roll off, as do 4.5% in contested local partisan offices. Here, contested races include races with a Democrat and Republican candidate and both candidates receive at least 5% of the vote. The second panel shows a significant increase in roll-off in contested races for nonpartisan state and local offices, 22.5 and 16.2, respectively. In the last panel, we see that roll-off for both state and local ballot measures is only slightly higher than roll-off for the contested partisan offices, and less than half as large as the roll-off for contested nonpartisan offices. In many cases, this means voters are likely skipping nonpartisan offices but voting on local measures further down the ballot, or in columns further to the right.

Table 2. Roll-Off and Votes for Minor Candidates, in Percentages

Note: Average percentage of roll-off and votes cast for minor candidates, across contests.

Regarding minor candidates, Table 2 reports the vote percentage for all candidates other than the top two that received more than 1% of the vote. Evidently, few voters voted for minor candidates in 2020. In some states and localities, no such candidates are allowed on the ballot (e.g., the top-two systems in California and Washington), and in many races no such candidates ran even though permitted. In any case, the number of votes cast for minor candidates is too small to alter our main conclusions.

Explanatory Power of National Partisanship

An alternative way to summarize our data involves calculating the amount of variation that national partisanship can explain at the individual-level. Such analyses are impossible with aggregated election results, but straightforward with CVRs.

We estimate regressions at the individual voter level with voting on a down-ballot contest as the dependent variable and National Partisanship as the independent variable. The regression model is of the form:

(3) $$ \begin{array}{rl}{v}_{ij}={\alpha}_j+\beta \hskip0.3em {d}_i+{\epsilon}_{ij},& \end{array} $$

where $ {v}_{ij} $ is a binary variable that denotes voter i’s choice in contest j, $ {d}_i $ is voter i’s National Partisanship (1 if Democratic and 0 if Republican), and $ {\alpha}_j $ is shorthand for the contest-specific fixed-effects. If a single voter votes on three different contests, they enter the data three times with a fixed effect for each contest. For partisan races, we code $ {v}_{ij} $ = 1 whenever a voter votes for the Democrat. For nonpartisan races and ballot measures, we code $ {v}_{ij} $ = 1 whenever a voter votes for the candidate or ballot measure choice that is preferred more by national Democrats than by national Republicans.Footnote ²² This coding scheme allows us to use the same specification for nonpartisan contests and ballot measures, where the Democratic-preferred candidate is not readily obvious. This data-driven coding scheme likely overestimates the explanatory power of national partisanship on the outcome.

The within-group R-squared of the fixed-effects regression in Equation 3 captures the within-contest explanatory power of National Partisanship. This seems like the most natural measure to assess how well National Partisanship can account for voters’ choices.Footnote ²³

Table 3 shows the within-contest R-squared estimates for each subset of offices. Each row represents a regression for a different outcome. The first column indicates the coefficient $ \widehat{\beta} $ on the National Partisanship variable. The patterns are the same as in Table 1. National Partisanship explains over 80% of the variance in voting for local partisan offices, but less than 8% of the variance in voting for local ballot measures and local nonpartisan offices.Footnote ²⁴

Table 3. Voting Patterns by Partisanship, Regression Estimates

Note: Each row is a separate regression. Coefficient estimates in column labeled National Partisanship. Standard errors, clustered by contests, are in parentheses. Fixed effects for contests included in all cases.

Table 4. Voting Patterns by Partisanship, Selected Local Offices

Note: $ {D}^D $ is the average percentage of the two-party vote cast for the Democratic candidate by National Democrats, and $ {D}^R $ is the percentage of the two-party vote cast for the Democratic candidate by National Republicans.

When we limit our attention to contests where we can identify the Democratic and Republican candidates, or the Liberal and Conservative alternatives, the explanatory power of National Partisanship is several percentage points larger. We manually identified ballot measures that could be considered as some form of spending increase, or if not a spending increase, a proposal that would move policy in a conventional liberal or left ideological direction (we explain this process in more detail when discussing ballot measures). We then define $ {v}_{ij} $ in Equation 3 so that in nominally partisan contests, 1 indicates voting for the Democratic candidate, and in spending-related or ideological ballot measures, 1 indicates voting in the liberal direction. With these specifications, state ballot measures that are either spending-related or ideological exhibit an R-squared of 0.206 (N = 170,412,065), and local ballot measures that are either spending-related or ideological exhibit an R-squared of 0.104 (N = 42,966,899).