Post-election audits

Traditional audits

In a traditional audit, election administrators often manually compare the results tabulated by a set proportion of voting machines with the paper record to determine whether there is a discrepancy significant enough to affect the election outcome. If a notable discrepancy is found, states outline different processes for how administrators should proceed. In Massachusetts, for example, the secretary of state can order additional audits if a post-election audit raises doubts about the accuracy of the count. In Kentucky, if a post-election audit identifies a discrepancy in the tally of votes cast, the county board of elections can open an investigation into the discrepancy. During this investigation, county boards review election day irregularity reports and, if more discrepancies are found, potentially expand the investigation “until the reason for the discrepancy is discovered.”

Traditional audits generally require the same proportion of machines, precincts, or ballots to be audited regardless of how narrow an election is. Tiered traditional audits, such as those used in New Mexico and Oregon, use a different approach. These states will audit more ballots in close elections than in landslide contests, tying the proportion of precincts to be tested to different tiers of election outcomes. This balances the interest in election accuracy with the resource limitations election administrators face, as decisive margins are less likely to be swung by the kinds of discrepancies that audits pinpoint. 

Procedural audits

Few states conduct procedural audits, and an even smaller number require them, but they are still an important tool that administrators can use to ensure an election was run correctly. Like traditional audits, procedural audits include a ballot counting and reconciliation process, but instead of verifying that voting systems are functioning correctly, they ensure that the human processes and procedures throughout the election were conducted correctly.  This includes verifying that the proper election forms are signed, vote tabulation equipment is tested, ballot materials are securely sealed, and that custody procedures for certain election materials are followed. In states such as Florida, jurisdictions may even reconcile voter registration records and their associated voting credits with the total number of ballots cast. Procedural audits share similarities with the election canvass; therefore, some states conduct procedural audits as part of the canvass process as opposed to a standalone procedure.

Risk-limiting audits

Risk-limiting audits (RLAs) are a relatively new post-election auditing technique that use a pre-specified risk limit to provide statistical confidence that an election outcome is correct, rather than auditing a fixed proportion of a precinct’s results or ballots. A risk limit is the maximum probability that the audit will stop short of a full hand tally when the original observed outcome is incorrect. For example, if an RLA has a 10% risk limit, there is at least a 90% chance that the audit will correct an outcome, and at most a 10% chance that it won’t. Risk-limits are often misconstrued as the chance that the final results of an election are wrong. 

RLAs allow election administrators to audit fewer ballots than in traditional audits while also providing assurance that results are correct. This reduces the time and resources that jurisdictions must devote to the audit process. The size of the risk-limit does, however, influence the scale of the audit. A smaller risk-limit (e.g., 1% versus 10%) entails gathering stronger evidence that the outcome is correct given that there is less room for error. As such, more ballots must be examined for smaller risk-limits and for races with smaller margins between candidates.

Risk-Limiting Audit Methods

RLAs are conducted using one of three main methods: ballot-level comparison, batch comparison, and ballot polling. Researchers and election administrators have also developed an RLA methodology, referred to as SUITE, or Risk-Limiting Audits by Stratified Union-Intersection Tests of Elections, which allows for multiple audits methods to be combined in one valid audit. However, its use thus far has been confined to Colorado.

Ballot-level comparison

In a ballot-level comparison RLA, a random selection of paper ballots is examined manually and individually. Auditors’ interpretations of voters’ markings are compared to the voting system’s cast vote record for the same ballot. This is a highly efficient method given that it typically requires the fewest number of ballots to verify the observed outcome; however, it requires the voting system to export a machine-readable cast vote record for each paper ballot, which places significant strain on tabulation systems.

Batch comparison 

RLAs using batch comparison methods are similar to ballot-level comparison audits, except that they tally a physical batch of ballots for a contest instead of looking at each ballot individually. The tally of the batch of physical ballots is then compared to the voting system’s exported subtotal for that same batch, thereby placing less strain on the voting system. This method is generally less efficient than using ballot-level comparison methods and usually requires more ballots to be inspected to reach the same confidence level. It is often used when a voting system does not export a machine-readable cast vote record for each paper ballot, but does export batch-level subtotals in physically identifiable batches, as is the case with older voting systems. 

Ballot polling

An RLA using ballot polling methods is often less efficient for close-margin contests, but comparable to ballot-level comparison audits for wide-margin contests. Here, a random sample of individual ballots is counted, and the sample’s subtotal is compared to the contest outcome, as reported by the voting system. This method does not rely upon cast vote records or batch tallies; therefore, it is often used in cases where an older tabulation system does not produce these records. A ballot polling RLA becomes less efficient as the contest margin narrows because, with a smaller margin, a larger sample of individual ballots is needed before the hand-tally gives strong enough statistical evidence to confirm the reported outcome.

Statisticians, computer scientists, political scientists, and election administrators have contributed to innovations in ballot polling methods that promote more efficient RLAs by validating election results more quickly. In 2011, Mark Lindeman, Philip B. Stark, and Vincent S. Yates developed BRAVO, or the Ballot-polling Risk-limiting Audits to Verify Outcomes protocol. In its simplest form, auditors using BRAVO randomly select a ballot and examine it for a valid vote, noting which candidate the vote was cast for before selecting another ballot. This process continues until the evidence in favor of one candidate winning is strong enough—as determined by the risk limit—to conclude the audit or order a hand count. Often, jurisdictions using BRAVO will sample ballots in batches, rather than one at a time. Doing so is logistically easier, but can require auditing more ballots over all before the risk limit is reached. 

Additional innovations in ballot polling RLAs include the Audit that Learns from Previously Hand-Audited Ballots (ALPHA) method, and MINERVA. The ALPHA method strongly resembles BRAVO, but generalizes its methods to more audit settings and improves on its efficiency by learning from previous ballot results and deploying a modified betting strategy in the family of betting martingale tests. Conversely, MINERVA uses a different statistical model to present a more efficient ballot polling alternative to BRAVO. Under MINERVA, election officials pull a random sample of ballots, tally them by hand, and compare the subtotals to those produced by the voting machine—as is done in a ballot polling RLA. The difference lies in how the auditors engage with the sample. More specifically, MINERVA employs advanced probability techniques to identify the tail of possible outcomes based on the sample, not just the single number derived from the subtotal. In doing so, MINERVA techniques allow for smaller sample sizes, and thus more efficient audits.

SUITE methodology

The SUITE audit methodology was developed by Ottoboni, Stark, Lindeman, and McBurnett and improves on existing RLAs by allowing administrators to combine two different RLA audit methods into one valid audit. This approach is particularly beneficial in states such as Colorado where all counties or jurisdictions do not use the same voting systems, and thus have different processes for recording votes which cannot be audited using the same method. For example, one county’s voting system may allow for ballot-level comparisons by producing exportable cast vote records, while another does not. Prior to SUITE, auditing a statewide contest would have required election officials to conduct a ballot-polling audit throughout the state, rather than a more efficient ballot-level comparison audit. Using SUITE, counties can conduct their different audits that are then assessed according to a defined risk limit for the whole contest.

Risk-Limiting Audit Usage

Jurisdictions in many states have been increasingly piloting and implementing risk-limiting audits. In 2009, Colorado was the first state to authorize RLA pilots, with the first audit taking place in Arapahoe County during the 2013 general election. In 2017, the state implemented the first statewide RLA using SUITE methodology to support county variations in voting systems. In the years since, seven states have adopted RLAs, six have conducted RLA pilots, and three allow state or local election officials to opt into conducting an RLA. Figure 2 below displays the status of RLA usage throughout the U.S. in 2026.

Implementation of RLAs can be challenging for some states given that they depend  on several conditions. First, conducting an RLA requires that a jurisdiction’s voting system is software independent, meaning that if ballots are recorded by an electronic system such as a Direct Recording Electronic (DRE) device, that device must produce a paper audit trail (also called a voter-verified paper audit trail, or VVPAT). Most states meet this requirement; Louisiana is currently the only state to utilize DREs without a VVPAT. Second, many states dictate the specific type of post-election audit that must be used in the state via statute. Any changes to these procedures would require legislative action. This may pose a barrier given that the type of audit conducted can be seen as an administrative technicality; legislators are often concerned with whether an audit is conducted rather than the specific type of audit that is used. Third, in states where elections are administered primarily at a local level, variations in voting system usage throughout the state can make implementing RLAs complex, despite the statistical methods that have been developed to account for this. It took Colorado four years, for example, to progress from its first RLA pilot to full statewide implementation.