The Athena Class of Risk-Limiting Ballot Polling Audits
share
The main risk-limiting ballot polling audit in use today, BRAVO, is designed for use when single ballots are drawn at random and a decision regarding whether to stop the audit or draw another ballot is taken after each ballot draw (ballot-by-ballot (B2) audits). On the other hand, real ballot polling audits draw many ballots in a single round before determining whether to stop (round-by-round (R2) audits). We show that BRAVO results in significant inefficiency when directly applied to real R2 audits. We present the ATHENA class of R2 risk-limiting stopping rules, and prove that each rule is at least as efficient as the corresponding BRAVO stopping rule applied at the end of the round. We have software libraries implementing most of our results in both python and MATLAB. We show that ATHENA halves the number of ballots required, for all state margins in the 2016 US Presidential election and a first round with 90 stopping probability, when compared to BRAVO (stopping rule applied at the end of the round). We present simulation results supporting the 90 probability claims and our claims for the risk accrued in the first round. Further, ATHENA reduces the number of ballots by more than a quarter for low margins, when compared to the BRAVO stopping rule applied on ballots in selection order. This implies that keeping track of the order when drawing ballots R2 is not beneficial, because ATHENA is more efficient even without information on selection order. These results are significant because current approaches to real ballot polling election audits use the B2 BRAVO rules, requiring about twice as much work on the part of election officials. Many states performing risk-limiting audits for the first time plan to use ballot polling audits in November 2020 for the US Presidential election and could substantially benefit from improvements.
READ FULL TEXT