[EM] Report on the 2008 San Francisco RCV elections

[Kristofer Munsterhjelm]

Abd ul-Rahman Lomax wrote:
> I took a preliminary look at the November 2008 San Francisco RCV elections.
> 
> There were 7 Supervisors elected. 3 elections found a majority in the 
> first round.
> 
> In the remaining four, no majority was found after transfers, I haven't 
> looked at all of them in detail, but one was on the order of 40%.
> 
> In no case did the first preference leader fail to win.
> 
> This is seems to be typical of nonpartisan IRV. So far, no comeback 
> elections, excepting a partisan election in Pierce County, Washington, 
> out of what must be approaching forty elections.
> 
> This appears to be normal: most of the time, when a majority isn't found 
> in the first round, no majority is found after vote transfers (except 
> for the "last round majority," which isn't a majority of the votes and 
> is a mathematical certainty from the method, which essentially discards 
> all votes which aren't for the top two).
> 
> Further, in nonpartisan elections, the preference order in the first 
> round tends to be maintained after transfers. No reversals have been 
> seen. In Top Two Runoff elections, reversal takes place roughly 
> one-third of the time.
> 
> I have the ballot image files for the "instant runoffs" from San 
> Francisco. Rather irritatingly, they don't have the files available from 
> the districts where there was a first round majority; but those results 
> might be in the overall result file that I also have, I haven't looked 
> at it yet.
> 
> San Francisco could have replaced their prior TTR system with Plurality; 
> the results would have remained the same. However, previously, there 
> were comeback elections where the runner-up did win. IRV eliminates 
> that, in nonpartisan elections, almost entirely. I say "almost" not 
> because I've seen an exception, but because the vote gaps do reduce, 
> sometimes, and in a very close election, a "comeback" could occur. It's 
> clearly rare with IRV.

I've run that data through my program as well. In all of the districts, 
the Plurality winner would have won using Schulze or Minmax. In all but 
one of the district, the same candidate would have won no matter which 
of the methods I tried were used. I had some trouble with the RCV 
truncation as far as regards Vote for and Against and Antiplurality, so 
I left them out.

There are some interesting patterns in the ballot data and the official 
interpretation. For instance, some people effectively bullet voted by 
voting the same candidate for first, second, and third place. Others 
voted A > A > B in order to vote A > B (and nothing else beyond that).

Also, it seems that the official IRV count stops counting a ballot when 
an overvote has been detected at that rank, but simply skips the rank if 
there's an undervote (no vote at the rank). For instance, A > B = C > D 
would count as a bullet vote for A, whereas A > (empty) > B is counted 
as A > B.

In my opinion, the restriction to three ranks worsened the results 
considerably. For instance, in district 9, 2398 votes didn't get counted 
at all. That's nearly a tenth (8% to be precise).

Oddly, there seems to be just about the same level of bullet-voting as 
in Burlington (about 25%), at least from my quick check of District 1.

The results are at 
http://munsterhjelm.no/km/elections/SF2008_results.txt and the ballots, 
in [number]: A > B > C format, is at 
http://munsterhjelm.no/km/elections/sf2008.zip . Note that the ballot 
files have one entry of the form [number]: each, with no ranking 
following - that's the count of invalidated ballots.