Loopydude, Little Nemo - I think you guys are talking about different things than Sam (for instance) is talking about.
In an election, there’s no MOE on the voters who go to the polls, and their candidate preferences: if their choices could be recorded 100% accurately, that would be the population, and MOE would equal zero.
But then there’s other errors. As the boilerplate language from Census Bureau source and accuracy statements goes, “For a given estimator, the difference between the estimate that would result if the sample were to include the entire population and the true population value being estimated is known as nonsampling error.” (Cite: bottom of p. 319 on this big-ass PDF.) It then goes on to list several such sources of nonsampling error that are relevant here, such as:
· Definitional difficulties
· Differences in the interpretation of questions
· Respondent inability to provide correct information
· Errors made in data collection such as recording and coding data
· Errors made in processing the data
The first couple of bullets might only apply to referenda, but the last two would apply universally, and that’s what Sam was mostly focusing on. And by taking a test deck of known ballots, as he suggested, you can turn the last two bullets from a nonsampling error situation (in terms of sampling the voters) to a sampling situation (in terms of measuring error rates in one’s ballot-reading and tabulating processes).
And as we recall from 2000, ballots can get confusing. The third bullet point above, if we rewrote that as “respondent difficulty in providing correct information,” would describe the situations in Palm Beach County, where voters had problems correctly recording their preferences due to the infamous “butterfly ballot,” and in Duval County, where the instructions told voters to vote on every page, even though the Presidential ballot went on for two pages - resulting in numerous disqualified ballots when voters recorded their preference for a major presidential candidate on the first page, then checked a minor candidate on the second page.
A great deal of statistical analysis was done with the butterfly-ballot results, of course, and it’s old news that the butterfly ballot decided the 2000 election and the subsequent course of our country. (That’s why the chaos theorists talk about the “butterfly ballot effect,” don’t’cha know? ;))
Much of this is testable. Ballot designs can and should be tested on groups of people to see how they perceive them, before putting them into practice, to find out what the error rates are between people’s intended and actual selections. These rates can be measured, and once measured, they can be minimized. (Systems that provide a means of self-checking, as my hypothetical system above does, are one technique to minimize that kind of error.)
Similarly (and much more easily) machine error in reading and processing ballots can be measured, via test decks as Sam proposes. The accuracy of OCR technology in different environments, for instance, has been measured in considerable detail.
So the expected number of these kinds of nonsampling errors can be sampled and measured in advance, and like Sam suggests, those measurements could be used to define what constitutes a ‘virtual tie’ in an election with a given number of voters. I’d quibble with Sam in just one place:
My system would have the ballot prep computer print out the voter’s selections for each race in, say, eighteen-point Courier bold, all caps, that pretty much anyone with eyesight could clearly read. (If the printer screws up, hopefully the voter will ask for a new ballot.) As long as the printer has enough ink, the result should be a ballot such that the OCR reader rarely misreads a character, with a misreading of ‘BUSH’ as ‘GORE’ being almost infinitely improbable. So maybe we agree after all about the machines being better, but disagree on the likelihood of vote-changing machine error in a system such as this.
