who has this Nifty Blog, that should definitely be one of the Most Popular sites on the Internet ;)
http://e-voter.blogspot.com/Here was Howard's excellent response to my query.
Hi Melissa,
I agree that 3% is an arbitrary figure, but see below.
Parallel testing and post-election audits are apples and oranges. If you don't have anything to audit, such as voter-verified paper ballots, then there is no recourse other than to randomly select machines to be tested during an election to see if they count the votes correctly. This is known as parallel testing but it would still require some kind of voter-verified paper records of the votes cast to test the machines to see if they are counted correctly. These would either have to be kept secret if they are produced before the tests begin (so that no one will be able to program the machine to count these particular combinations of votes correctly), or generated randomly at the time the testing is done (so they cannot be predicted). If the machines do not count these votes correctly, there may be no recourse anyway since the testing is done during the election and any flaws in the remaining untested machines would still be present. DREs are of course especially vulnerable to this sort of thing, and they are the only e-voting systems that would not be auditable in the first place (assuming they do not produce VVPATs).
The other thing about parallel testing is that for the tests to really be statistically accurate, they would have to follow the same rules as sampling machines or precincts in an audit after the election. The difference is, with the audit, we know the amount of discrepancies we are looking for: it's the number of switched or uncounted votes that would change the outcome of the election reported by the voting system. With parallel testing, we don't even know this number because the election is still underway. So it's hard to come up with the right number of machine to test, although we could probably test enough to rule out some percentage of faulty machines such as 1%, 5%, etc. The problem is, there could be trouble with some ballot styles that might not affect others, and it's impossible to test all vote combinations. It might be possible to test every ballot style though, so that might be a better idea. The best thing that can be said about this kind of testing is that it might serve as a deterrent to deliberate tampering with the machines, but it doesn't do much to confirm the true winners of elections in my opinion.
As far as a 3% audit, or any percentage audit, it's not the percentage that matters, but rather the number of machines audited. So I agree that it's arbitrary. 3% may be very effective for a statewide race because it's a lot of machines, whereas 3% may be inadequate for smaller contests such as US House or members of a State Legislature where 3% would NOT involve many machines, or when the victory margin in any contest is narrow.
If you are interested in learning more about this, I'd suggest you read the paper on this subject available on Verified Voting's website:
http://www.verifiedvotingfoundation.org/auditcomparison There are 2 versions, one of which was recently published in the peer-reviewed journal, "The American Statistician."The longer one, entitled "Percentage-based versus SAFE Vote Tabulation Auditing: A Graphic Comparison" is written for a more general audience. I'd be happy to answer any questions about this work.
I would say that generally speaking, a properly conducted 3% audit of a large state would be quite effective, but this depends on the victory margin of the statewide contest. It would be better to use a smaller minimum such as 2%, or even 1% in the largest state, with the ability to automatically increase the initial sample size based on the closeness of the race without any cap on that sample size, and with additional requirements for escalating the audit if errors are found. In the Verified Voting paper, we show that an audit of about 3.5% of precincts nationwide would have been enough to confirm the outcomes of ALL federal elections from 2002-2006, assuming no discrepancies were found and assuming there were no caps on the size of the initial sample for close races, a few of which would have required full automatic recounts.
The audit law we wrote for the State of NJ requires all this. You can read about that on my blog if you scroll down past the NY lawsuit article:
http://e-voter.blogspot.com/ Hope this helps. Let me know if you need any other information. There are some newer, but slightly more complicated methods in the works that audit larger precincts more often than smaller ones, while still maintaining a truly random sample, that are very attractive.
Regards,
Howard (Stanislevic)