Electronic Voting, Spring 2020

Auditing Elections

Again, an apology for requiring reading something I wrote.

The big takeaway, as far as I can tell on reading it over 15 years after I wrote it, is how bad election auditing was in the 20^th century. The bare minimum information needed to reconcile the numbers coming out of election offices was not available in most states.

The ther thing to note is that auditing involves far more than just doig some kind of recount. In the 15 years after this paper, many people have focused so closely on the issue of various forms of post-election ballot audits that the big picture sometimes gets lost. If you don't audit the chain of custody, an audit of the ballots is not all that valuable. If you don't audit the pollbooks, how do you know that only legal votes were cast.

Stark

Stark invented the modern idea of risk-limiting audits, and most of that work traces back to this paper. Trigger Warning: This paper contains math. There are some simple cases, however, that help understand what's going on.

Suppose you have an election in which b ballots were cast, and the winner's victory margin was m votes. Stated another way, the victor won by a margin of 100(m/b) percent. If we randomly select just b/m ballots for manual inspection, comparing how those ballots were counted in the official count to how a human interprets the ballot, then we have a 50% chance of catching any miscounts. If we double our sample size, we roughly halve the chance that an error (or fraud) in ballot interpretation will go undetected.

This leads to the idea that, so long as we can link each ballot to how the machine tabulated that ballot, the sample size we need to check in order to verify the machine interpretations can be very small, largely independent of the number of ballots cast but depending closely on the apparent margin of victory of the winning candidate. The larger the margin, the fewer ballots we need to check to assure ourselves that the correct victor was declared.

Hall

This really works! The challenge is, how do you relate the actual paper ballot to the record of how that ballot was interpreted. This is easy for central-count paper ballot systems. How to apply this to precinct-count systems remains subject to argument.

Its worth noting that Colorado looked at this experience and enacted it into law. They've been doing risk limiting audits for several years now, and it works.