The recount of the 2008 Minnesota Senate race gives us an opportunity to evaluate the accuracy of precinct-count optical-scan voting. Though there have been contentious disputes over which absentee ballot envelopes to open, the core technology for scanning ballots has proved to be extremely accurate.
The votes were counted by machine (except for part of one county that counts votes by hand), then every single ballot was examined by hand in the recount.
The “net” accuracy of optical-scan voting was 99.99% (see below).
The “gross” accuracy was 99.91% (see below).
The rate of ambiguous ballots was low, 99.99% unambiguous (see below).
My analysis is based on the official spreadsheet from the Minnesota Secretary of State. I commend the Secretary of State for his commitment to transparency in the form of making the data available in such an easy-to-analyze format. The vast majority of the counties use the ES&S M100 precinct-count optical-scanners; a few use other in-precinct scanners.
I exclude from this analysis all disputes over which absentee ballots to open. Approximately 10% of the ballots included in this analysis are optically scanned absentee ballots that were not subject to dispute over eligibility.
There were 2,423,851 votes counted for Coleman and Franken. The “net” error rate is the net change in the vote margin from the machine-scan to the hand recount (not including change related to qualification of absentee ballot envelopes). This was 264 votes, for an accuracy of 99.99% (error, one part in ten thousand).
The “gross” error rate is the total number of individual ballots either added to one candidate, or subtracted from one candidate, by the recount. A ballot that was changed from one candidate to the other will count twice, but such ballots are rare. In the precinct-by-precinct data, the vast majority of precincts have no change; many precincts have exactly one vote added to one candidate; few precincts have votes subtracted, or more than one vote added, or both.
The recount added a total of 1,528 votes to the candidates, and subtracted a total of 642 votes, for a gross change of 2170 (again, not including absentee ballot qualification). Thus, the “gross” error rate is about 1 in 1000, or a gross accuracy of 99.91%.
Ambiguous ballots: During the recount, the Coleman and Franken campaigns initially challenged a total of 6,655 ballot-interpretation decisions made by the human recounters. The State Canvassing Board asked the campaigns to voluntarily withdraw all but their most serious challenges, and in the end approximately 1,325 challenges remained. That is, approximately 5 ballots in 10,000 were ambiguous enough that one side or the other felt like arguing about it. The State Canvassing Board, in the end, classified all but 248 of these ballots as votes for one candidate or another. That is, approximately 1 ballot in 10,000 was ambiguous enough that the bipartisan recount board could not determine an intent to vote. (This analysis is based on the assumption that if the voter made an ambiguous mark, then this ballot was likely to be challenged either by one campaign or the other.)
Caveat: As with all voting systems, including optical-scan, DREs, and plain old paper ballots, there is also a source of error from voters incorrectly translating their intent into the marked ballot. Such error is likely to be greater than 0.1%, but the analysis I have done here does not measure this error.
Hand counting: Saint Louis County, which uses a mix of optical-scan and hand-counting, had a higher error rate: net accuracy 99.95%, gross accuracy 99.81%.