he machines that count our paper ballots every election in New Hampshire are so old they run on out-of-date Windows XP from 20 years ago, which is why the state is looking to replace them. It’s also why some people doubt their accuracy.
A lot of hot air is generated about voting accuracy these days, but at Granite Geek, we prefer numbers to noise, so we sought any objective information about how many errors these AccuVote machines actually make.
I’m happy to say there is such information, thanks to a natural experiment that has been running ever since they plugged in the first AccuVote machine in the late 1990s.
It comes from hand recounts that are done by election folks in a stuffy office in Concord, who gather the paper ballots that were run through AccuVote machines and check the results one by one.
Recounts can be requested by candidates in close elections (usually less than 1 percentage point between a winner and loser). Every year sees a half-dozen or more of these recounts, both for party primaries and general elections.
The tallies for all recounts are listed in the biannual Manual for the General Court, a.k.a. the Red Book. To test the AccuVote accuracy, all I had to do was compare before-and-after results to see how many machine errors were found by hand recounts.
I looked at 37 recounts of state representative races or primaries in 2020, 2018 and 2016, covering a total of 150 seats. (Why so many seats? Three recounts involved Derry, which has 10 seats all by itself, and most recounts involved districts that had three or four seats.)
I didn’t look at the 17 other recounts that occurred during this period because they involved at least one town that hand-counts ballots on election night rather than using AccuVote machines.
By my calculation there were a total of 476,203 votes cast in these races; that is, 476,203 times somebody filled in one of those little circles next to a candidate’s name and the AccuVote optical reader should have tallied it.
How many of those votes were changed in the recounts? I count 2,849 changes, meaning 0.6% of the circles – a little over one-half of one percent – changed either up or down. That’s a success rate of 99.4%.
Very few of the changes made any difference: Recounts changed winners in just 5 of the 150 seats I examined.
By the way, almost half of the ballot changes came from a 2020 Windham race that led to a much-publicized audit. The audit found that hundreds of ballots had been folded in a way that changed how the machine counted them. (None of the Windham election results were changed by that district’s recount.)
Without that one election-night screw up and its 1,368 changes, you’ve got a ballot change rate of 0.33% or one-third of one percent over 36 multi-candidate elections.
All in all, I’d say the AccuVote machines passed with flying colors. If somebody casts doubt on our ballot tabulation accuracy, tell them to find something else to worry about. There are certainly enough possibilities out there.
Note, by the way, that there are two complications in calculating percentage.
First, I didn’t count number of ballots but rather total number of votes for all candidates. Since most state representative seats have multiple members where voters are asked to vote for more than one candidate, my tally for any recount is larger than the number of ballots cast, sometimes much larger. Depending on how you look at it, this might be said to underplay the error rate.
Second, I double-counted some mistakes, which seems to exaggerate the error rate.
Let’s say the machine erroneously took a vote for Candidate A and gave it to Candidate B. When this is fixed in the recount it will show up in the Red Book as one vote less for A and one vote more for B. I have no way of knowing that both of these are from a single ballot mistake so I count them as two separate mistakes.
In the Windham audit, multiple votes were missed entirely on some ballots because they were along the fold in the paper ballot. The problem affected around 400 paper ballots but those multiple vote switches mean it produced a total of 1,368 changes in tallies of different candidates, more than tripling the apparent error rate.
The question is how these two inexact approaches interact. Did they result in an exaggerated error rate or did they downplay the error rate? I don’t know, although either way the rate remains minuscule.
You could get a more accurate tally by looking at every separate town ballot for every recount but that’s not practical – not for me, anyway. Maybe some poly-sci grad student looking for a thesis topic wants to take a crack at it.