In his New York Times column last week (Aug. 17), Paul Krugman raised the specter that the results of the 2004 presidential election could be "suspect." After the 2000 election debacle with "hanging chads" and "butterfly ballots," Florida, along with other states, installed electronic voting machines in many precincts. However, because the machines provide no hard copy of the vote, there is no way to conduct a vote recount in the case of a close or disputed election. In addition to the lack of a paper trail, "independent computer scientists who have examined some of these machines' programming code are appalled at the security flaws," Krugman wrote. Theoretically, it is possible for someone to tamper with the vote count electronically and never be discovered.
As one of several solutions to the problem, Krugman noted that "some voting activists" have urged that there be intensive exit polling, "parallel to, but independent of, polling by media groups …" Krugman likes the idea because it could 1) act as a deterrent to anyone contemplating election fraud, 2) help validate the results, and 3) give an early warning of possible vote tampering.
Despite this ringing endorsement of exit polling, Warren Mitofsky -- the inventor of exit polls and currently co-director of Edison/Mitofsky (the group that conducts exit polls for the national news media) -- has doubts about the columnist's prescription. "Paul Krugman's suggestion that independent exit polling be used to detect errors in electronic voting is probably not going to be useful in individual polling unless the size of the error in any single polling place is very large," Mitofsky responded in an e-mail to my question about the feasibility of Krugman's proposal. "Small errors will be undetectable," Mitofsky wrote.
There are a couple of ways in which intensive exit polling might be used to detect fraud. One way is to try to get all the voters in a select number of precincts around the state to participate in the exit poll, and compare the exit poll results with the results in each of the selected precincts. Another method is to survey a sample of voters (typically around 100 to 200) in each of a larger number of randomly selected precincts around the state, so that the sample of precincts can be used to project the overall vote in the state (or even in regions in the state).
With either approach, there are possible "biases" and/or "sampling errors" in the data that would make it virtually impossible to detect vote fraud of a magnitude of one or two percentage points. Take, for example, canvassing all voters in selected precincts. This is a difficult operation, and would require several interviewers in each precinct to ensure no voter was missed. Even then, many people will simply refuse to participate, so that the vote count of the exit polls is bound to be lower than the official count.
More importantly, there could be "non-response bias" -- which would occur if, say, Bush voters were disproportionately less likely to take part in the post-election poll than Kerry voters, or vice versa. That would mean that not only was the vote count lower, but the percentages for each candidate would not be completely accurate. As Mitofsky wrote, "The effect of the non-response could easily be confused with any vote rigging or any other systematic errors at a polling place due to the voting equipment."
With the second approach, interviewing samples in a larger number of precincts, the margins of error in each precinct would be quite large. Mitofsky would not speculate on the sampling error in a precinct, because it depends on the size of the sample. But given sample sizes of 100 to 200 voters, the margins of error could be in the 8% to 10% range. When the data are aggregated into regions and across the state as a whole, the margin of error would still be substantial -- at least in the 2% to 3% range.
In 2000, Bush won Florida by 537 voters out of almost 6 million cast. His margin of victory was less than one hundredth of 1% of the vote. No exit poll operation would ever be able to detect a fraudulent addition or subtraction of a few thousand votes across the state.
Only if there were vote fraud of a significant magnitude -- probably at least 10% of the vote (or some 60,000 votes in the case of Florida) -- might a very expensive and intensive exit poll be used to detect the chicanery. Smaller differences could be attributed to normal sampling error and possible non-response bias.
The only real solution to disputed elections is a complete recount based on hard copy ballots. And even that, as we saw in Florida in 2000, has its margin of error.