4 May 2006
At the Midwest conference last week I saw Walter Mebane presenting his new paper entitled "Detecting Attempted Election Theft: Vote Counts, Voting Machines and Benford's Law." The paper is really fun to read and contains many cool ideas about how to statistically detect vote fraud in situations where only minimal information is available.
With the advent of voting machines that replace traditional paper ballots physically verifying vote counts becomes impossible. As Walter Mebane puts it: "To steal an election it is no longer necessary to toss boxes of ballots in the river, stuff the boxes with thousands of phony ballots, or hire vagrants to cast repeated illicit votes. All that may be needed nowadays is access to an input port and a few lines of computer code.��?
How does Mebane utilize statistical tools to detect voting irregularities? He relies on two sets of tests:
The first test relies on Benford’s Law. The idea here is that if individual votes originate from a mix of at least two statistical distributions there may be a rationale to expect that the distribution of the digits in reported vote counts should satisfy the second digit Benford's law. Walter provides simulations showing that the Benford's Law test is sensitive to some kinds of manipulation of vote counts but not others.
The second set of tests relies on randomization. The idea is based on the assumption that in each precinct (especially crowded ones) voters may be randomly and independently assigned to each machine used in the precinct. The test involves checking whether the split of the votes is the same on all the machines used in a precinct. If some of the machines were indeed hacked, the distribution of the votes among candidates would differ on the affected machines. Mebane tests these expectations against data from three Florida counties with very interesting findings.
In general, the paper was very well received by the audience. Some attendees raised concerns about the randomization test, arguing that voters may not be randomly assigned to voting machines (for example old voters may be more likely to go to the first machine in line etc.). The discussant, Jonathan Wand, raised the idea of actually using random assignment of voters to voting machines as an administrative tool to facilitate fraud detection ex post. He also proposed to use sampling techniques to make recounts more feasible (but that would require voting machines that do leave a paper trail). Another comment alluded to the fact that if somebody smart wants to steal an election, he or she might anticipate some of Walter's tests and design manipulations so that they satisfy the test.
Overall, my impression is that although his research is admittedly still at an early stage, Mebane is onto something very cool here and I am eager to see the redrafts and more results in the future. This is a very important topic given that more and more voting machines will be used in the future. Everybody interested in the vote fraud should read this paper.