Post-election audits reduce the risk of confirming an incorrect outcome. Audits designed explicitly to limit such risk (risk-limiting audits) have advantages over fixed-percentage or tiered audits, which often count fewer or more ballots than necessary to confirm the outcome.
a. Risk-limiting audits have a large, pre- determined minimum chance of leading to a full recount whenever a full recount would show a different outcome.[9][10] (Correct preliminary outcomes are never overturned.[11]) After any audit, this chance should be calculated and published as part of the audit results to promote continuous improvement.
b. Audit units (precincts, machines, batches of paper records) should be selected using appropriate random sampling methods.[12] In a risk-limiting audit, the sample size will depend on the margin of victory and other factors; these other factors may include the number of ballots in each precinct and the overall number of ballots in the contests.[13] In general, smaller margins of victory and smaller contests require auditing a larger percentage of the audit units.
c. To reduce the burden of counting ballots while still auditing a variety of contests, it may be appropriate to use different rules for auditing some contests than others. For example, it may be appropriate to allow more risk for non-statewide contests.[14] Jurisdictions may require audits in some contests and randomly select others to be audited, so that every contest has some possibility of being audited. For smaller contests, it may also be appropriate to use alternative audit methods such as targeted sampling (see Best Practice 8) or random sampling based on a fixed number or percentage of audit units.
d. For each contest to be audited, votes cast in that contest on the ballots in the selected audit units must be fully and manually counted.[15] For each selected audit unit, the audit must compare vote count subtotals from the preliminary reported election results for the contest, with hand-to-eye counts of the corresponding paper records.
e. For efficiency, large groups of ballots can be divided into batches, each comprising an audit unit. In this case, the subtotals for each batch must be reported prior to the audit as part of the election results. For instance, absentee ballots (if not sorted and counted by precinct) can be divided into batches.
[9] “Outcome†refers to which candidates or ballot propositions won or lost, not necessarily a specific vote tally. Here we refer to the outcome as “correct†or “incorrect†depending on whether it corresponds with what would be the outcome from a complete manual recount. Note that the outcome from a complete manual recount may not always match the will of the voters. To ensure that outcomes reflect the will of the voters, additional conditions must be met including rigorous ballot accounting, accurate registration data, elimination of unreasonable delays at the polls, good ballot design, and controls on chain of custody for all election equipment and materials.
[10] Fixed-percentage samples are inadequate for risk-limiting audits, because the audit size needed to verify an election outcome depends on the apparent margin of victory, as well as the number of audit units and the amount of error each audit unit can harbor.
However, auditing some minimum percentage of votes or audit units regardless of jurisdictional size or election margin may be useful to monitor election accuracy. Generally, requiring a smaller chance of error (e.g. 1% versus 5%) will entail auditing more ballots.
[11] If audit results indicate that the initial outcome is incorrect, ultimately a full recount would be required to determine the final outcome. Preliminary outcomes cannot be overturned based on audit samples alone.
[12] In the selection, some units may be weighted more than others based on their size and the amount of error they could harbor. Random sampling is unnecessary if all audit units will be manually counted, or if so many audit units are counted that the remaining units cannot change the outcome.
[13] Discrepancies found during the audit can also affect the sample size, as discussed in 6a.
[14] All else being equal, contests spanning fewer audit units – for instance, local contests as opposed to statewide contests – re quire proportionally larger audits to ensure that the chance of confirming an outcome that is incorrect is low.
[15] “Manual counting†or “hand counting†refers to human visual inspection of paper records to interpret voter intent, followed by a tabulation of the individual vote interpretations. Only the tabulation portion is sometimes assisted by independent and well trusted equipment such as calculators and spreadsheets. All hand counts should be done blind to the expected result.