I’m not sure I trust this number either, because it implies a shift of 11 votes in a statewide race, and we know the shift in Franken-Coleman was more like a few hundred votes.

As I said, the Poisson distribution was a spherical cow approximation–I didn’t know enough about Minnesota’s process (I have never lived there) to get a better estimate than that. The reason for assuming a Poisson distribution is that it’s what you use when you are counting events that don’t have fixed recurrence times, and it’s easy to calculate the standard deviation: if you counted N events than σ = sqrt(N). An election may not satisfy the second part of that, but you are certainly counting events (votes) if you think the election means anything. The random error in polls is due to Poisson noise. Systematics, such as bad poll design or tendency to undercount votes, are the next level of sophistication.

Now that you have clarified that the machines will tend to undercount, Poisson statistics would apply to how you expect the additional votes to be distributed. Again, you are counting events (undervotes), and this time you know that the recurrence interval isn’t fixed (you know that a certain fraction of ballots will be undercounted, but you don’t know in advance which ones they are). The Poisson distribution looks like a normal distribution if the number of events is high enough, but for low N there is a noticeable difference.

I’m not sure what the justification is for calling this a poison distribution, nor am I comfortable attributing variance to this number based on a theoretical distribution of any type. Almost all the errors are undervotes, there are not many of them, and the question is not what is the probability of an undervote, but rather, what is the distribution of undervotes across candidates. That is probably a normal distribution.

Experience tells us that the actual errors in Coleman-Franken and Emmer-Dayton were substantially smaller than Poisson level, so that assumption gives us an additional safety factor. But I think I’d still want the 0.5% threshold if I were a candidate in a local race.

There are recounts every year, here and there in the state. Stwiching from 0.05 to half that would suddenly save a pile of money.