[E-voting] The VVAT Debate - Technical or Not ?
michaelmcmahon at vodafone.ie
Tue Mar 9 14:07:48 GMT 2004
Patrick O'Beirne wrote:
> At 10:10 09/03/2004, Michael McMahon wrote:
>> The key point is the "assumptions about distribution of errors" and a
>> (alone) can't determine that.
> True, they can't determine that, but they need that information to
> devise a sampling scheme to tell you how much you need to sample, where
> and when, to assure yourself that with a given degree of probability
> that the full data set (population in stat terms) does not have more
> than a chosen % of defectives
> Are all errors random. binomial distribution, and independent? Then use
> random sampling.
> Are they likely to be dependent .. i.e. if a machine has one error, may
> it have more? Then stratify sampling by machine.
> Are they clustered by time .. e.g. more likely to occur with higher
> volumes than lower? Then weight the sampling toward the end of the day.
These models can certainly be applied to parts of the problem, eg. hardware
errors. But for detecting software errors, and particularly malicious tampering
kinds of errors, I'm not so sure. I'm sure research has been done, and I'd like
to look at it, but the topic is bound to be controversial and I'm not convinced (yet).
There is also the issue of effective sample sizes. You can't take a sample
from an individual (constituency) election. At this level, auditing has to be all or nothing.
>> If an accountant audits the books of a company, would the shareholders
>> be happy if the accountant only checked a sample of the records?
> They are, because that's all auditors do. Auditors have many sampling
> schemes, and some involve using any prior knowledge they have to point
> them in the right direction; see above.
That's interesting. I'm not an accountant so it's news to me.
> I used to know this stuff when I worked in quality control thirty years
> ago but it's faded now. A few years ago, when I worked in TCD Statistics
> dept ( http://www.tcd.ie/Statistics ) my colleagues could do this kind
> of thing very easily. As I understand now that most on this list are CS
> academics, maybe it's time to talk to your colleagues?
I'm not an academic, and I think you know more about statistics
than I do, so I'm happy to read your views on the subject.
More information about the E-voting