White Paper: Election Audit Strategy and BRAWL -- Balanced Risk Audit with Workload Limitation
Citizens Oversight (2018-09-10) Ray Lutz
This Page: https://copswiki.org/Common/M1879 More Info: Election Audits, Election Integrity
A key element of defending our democracy is ensuring elections that are fair and free of scams. There are many dimensions of this problem.
You may have heard a lot of about Risk Limiting Audits of various kinds, as well as "batch comparison audits." The general concept is easy enough, we sample some of the paper ballots and review them by hand, and compare with the computer report. If this is done properly, it can detect -- to a very high degree of confidence -- most types of hacking that may occur.
When you review these methods, however, you will soon find that the papers on the subject include fairly sophisicated statistics and equations that are not at all intuitive. We want to believe the prestigeous mathematicians that are promoting their particular approach, but we need a way to fully understand how these approaches will pan out in practice.
This document is an attempt to understand these approaches using a fairly simple simulation, known as Monte Carlo methods, which simply try things over and over and then find out their performance as compare with each other. In doing this, we began to understand the underlying "physics" of the problem, and an alternative method arose we are calling the Balance Risk Audit with Workload Limitation (BRAWL) method.
Finally, we compare the statisical sampling methods along with the more traditional batch comparison method and the ballot image audit, where validated and secured ballot images are used to conduct a slightly different type of audit which has a different set of risks.
We found that statistical audits can be very good for almost all elections that have margins greater than 10% and in some cases for those with margins greater than 2%. But really close elections will require increasing sample sizes to a point where it may be better to start with a sequential hand count from the get go.
Finally, we attempt to compare the economic costs for each type of audit. This part of the white paper is still being refined as we get additional data from recent audit pilots, particularly the one in Rhode Island at the beginning of 2019. At this writing, the final sections 12 & 13 of Part 2 does still need to be refined based on new data from RLA pilots, particular the Rhode Island RLA pilot which provided critical timing data. That data has been reduced to suggest that a Ballot-Polling audit and Ballot-comparison audit require 20% and 25% more time respectively. This is a bit less costly than the very rough guesstimates in the document.