Premier AccuVote Machines Missed 0.4% of Ballots in Aspen Elections
Error’s discovery shows value of redundant ballot-counting methods
On May 5th 2009, Aspen (CO) held municipal elections for mayor, two city council seats and a ballot measure. Pitkin County's Premier (formerly Diebold) AccuVote optical scan voting machines failed to register 11 (0.4%) of 2,544 ballots, which was discovered due to the ballots also being counted on Election Day at a central location with a separate system. Premier is one of the three largest providers of voting equipment in the United States.
Aspen had more voters at its polls than ever before in its history. It was the first election with the city's system of instant runoff voting for electing the mayor and a variation of instant runoff voting (IRV) for city council. In both ranked voting methods, voters were allowed to rank candidates in order of choice. First choices were counted at the polls with Premier's AccuVote system provided by the county, then the full paper ballots were counted at a central location where there was an independent re-scanning of the ballots. This re-scanning uncovered errors in the results reported by the AccuVote machines, and most alarming, revealed the fact that 11 ballots were entirely missed by the AccuVote machines. The ballots were re-scanned using a commercial off-the-shelf scanner as part of an independent system TrueBallot, Inc., employed under its contract with the city to tally the instant runoff voting elections.
The TrueBallot system did not simply record votes and keep running totals for each candidate, but rather captured an actual graphic image of each individual ballot. This redundant record of each individual ballot allowed for a higher level of accuracy in determining voter intent, with apparent over-vote or under-vote ballots being able to be projected onto a screen allowing election judges to rule on voter intent. There is currently a legal battle in Aspen about whether these graphic images are public records that should be made available to the public. FairVote believes the public interest is best served by full election transparency and that the images should be made public.
To underscore the importance of the missing 11 ballots, it is not uncommon for manual recounts of optical scan elections to find new valid votes that were discounted by the optical scan voting machine, either because the machine detected a stray pen mark as an over-vote (voting for more candidates than allowed), or because a voter marked a choice too lightly or outside the designated spot on the ballot, such that the machine detected an under-vote, or skipped race. Such "found" votes are common in manual recounts where humans can recognize a voter's intent that the optical scan machine could not. In the recent Aspen election, the independent scanning of ballots did indeed allow election officials to find at least one such valid vote missed by the AccuVote voting machines.
However, this problem is unrelated to the discovery that nearly a half percent of ballots - 11 ballot cards in all - went entirely unrecorded by the AccuVote machines. According to Aspen City Clerk Kathryn Koch, both the poll book record of the number of voters who voted and the TrueBallot record of ballots processed agree that there were 2,544 ballots. The AccuVote machines, however only recorded 2,533 ballots. These numbers can be calculated from the first round mayoral vote counts from the AccuVote and TrueBallot data on the city web site. We have put them into the following chart.
|Mayoral candidates||AccuVote||TrueBallot||Missed Ballots|
|Residuals (e.g. skipped race)||17||16||-1|
According to city clerk Koch and the data from TrueBallot, Inc., the AccuVote voting machine errors came from two different AccuVote voting machine units, designated #4 and #5B. The errors were unrelated to the city's use of instant runoff voting. The voting machines were running in their customary plurality election mode, necessitating the independent re-scanning of ballots to record voters' alternate rankings, which uncovered the AccuVote errors. This means that the same sort of error could have occurred in previous elections in Pitkin County for elections for president, Congress and state offices and could occur again, unless the source of the errors is identified and fixed.
The only other known independent re-scanning of ballots cast using Premier AccuVote machines occurred in Humboldt County, California in conjunction with three recent elections. The independent scanning was done by the Humboldt County Election Transparency Project (http://humtp.com/page3.html) authorized by Registrar of Voters Carolyn Crnich. It is troubling to note that in November 2008 the re-scan uncovered a serious error. The independent re-scan discovered 197 ballots that went unreported by the version of the Premier GEMS ballot upload program in use, and were thus missing from the certified election results.
Fortunately for Aspen, the official results of the election were the TrueBallot results, not the Premier results that were missing 11 ballots. The post-election manual audit of a random selection of 10% of the ballots confirmed that the TrueBallot system accurately recorded the votes. (http://www.aspentimes.com/article/20090508/NEWS/905089992)
Recommendations for Aspen:
Aspen immediately should undertake a detailed investigation to identify the source of the errors. This investigation may include interviewing poll workers associated with machines #4 and #5B to learn if there were any unusual events or lapses of proper procedures. For example, are they certain all of the ballots actually went through the tabulators? Could some poll workers have put some kicked-back ballots (that were wrinkled, or mis-marked) directly into the "emergency" side slot of the ballot box (that exists in case of power failures, etc.) and thus did not pass through the tabulator? If this possible error was combined with a second error when the ballots were removed after the close of polls, such that nobody noticed these 11 ballots were in the special compartment of the ballot box for un-scanned ballots, this could explain the error.
If the scenario above is ruled out, it is recommended that representatives from Premier and the testing laboratory that passed this AccuVote model be asked to explain how such an error could occur. Machines #4 and #5B should be secured and computer forensic experts should be employed by the Colorado Secretary of State to attempt to identify the source of the errors.
As to the matter of public release of the ballot images, we believe the city of Aspen should release these records, just as it has previously released all of the ballot data derived from the images. This was the practice in Humboldt County (Humboldt County election ballots can be found online at http://earc.berkeley.edu/hosting.php). The remote risk that voters might in the future be solicited to sell their votes and make identifying marks on their ballots is not relevant to the already completed election and unconvincing as to future elections, as making any identifying mark on a ballot is already illegal under existing state law (1-13-712: disclosing or identifying vote). If there is still concern, as with other states, Colorado statutes should probably be amended to clearly state that ballots with identifying marks are defective and invalid.
Broader Policy Recommendations:
Since both U.S. elections that have been tallied or audited in a transparent manner by scanning with commercial off-the-shelf scanners have proven to be more accurate than results reported by proprietary voting equipment, the fundamental assumption that vote tabulations should be conducted by non-transparent proprietary means should be questioned. The value for election integrity of having redundant records of each individual ballot (both paper and machine) rather than merely running totals from proprietary voting machines is evident. Colorado should require that voting machines purchased after 2010 record individual ballot records showing each voter's choices, rather than merely keeping running totals and establish manual audits of a statistically appropriate random sample of all ballots.