2012-02-20

What if we used Score Voting in San Francisco?


San Francisco currently uses Instant Runoff Voting, restricted to ranking up to three candidates. But imagine we used Score Voting (aka Range Voting), where you could rate as many candidates as you wanted to. Here's what that would look like if we used a simplified 0 to 2 point scale.


Bad candidates get a 0. Good candidates get a 2. Average candidates get a 1. Simple. Expressive (even more expressive than IRV, because you can vote for as many candidates as you want, not just three). And way less ballot space than the current Instant Runoff Voting ballot.

Other benefits:

Lower ballot spoilage
IRV results in about seven times as many spoiled ballots as Plurality Voting, on average. This is especially common among voters of lower socioeconomic status. That's one reason that many of IRV's detractors argue it's too complicated. But Score Voting experimentally results in fewer spoiled ballots. This is because there are actually fewer ways to spoil your ballot with Score Voting, since it's fine to give the same rating to multiple candidates (whereas you cannot give two candidates the same ranking with IRV, nor can you vote for more than one candidate with Plurality Voting). This is explained in much greater detail by Warren Smith (a Princeton math Ph.D. who studies voting systems) here.

People intuitively understand Score Voting
The most basic sanity check for a voting system is to have random individuals fill out a ballot, and then ask them how they intuitively think the result is tabulated. If their intuition tends to be right, then you know you have a voting system that will be easily understood, even without an expensive voter education campaign.

Instant Runoff Voting fails this sanity check, because it is not very intuitive. In my experience, polling many intelligent residents who claim to vote, the vast majority of them do not understand how IRV works. They tend to assume it's a "weighting" system, where candidates get more points for being ranked closer to first. Here's a conversation I had with a very intelligent software engineer I used to work with, who has voted in IRV elections in San Francisco.

San Francisco has tried to remedy this problem with a massive public education campaign. See this example, and another, and another. The problem is that this education not only costs money, but it isn't even very effective. Just poll some random friends or co-workers who vote in San Francisco elections. Show them the same example I showed my former co-worker, and see whether they get it right. Or even just ask them to correctly explain the tabulation procedure. You will likely find that many of them get it wrong.

Some people have suggested that it doesn't matter how well people understand the tabulation system, as long as they understand how to correctly fill out a ballot. But the problem is, if people assume IRV is a  weighting system, then they'll be more likely to strategically exaggerate. For instance, say a voter prefers Joanna Rees over John Avalos over Ed Lee over Tony Hall (ignoring the other candidates for simplicity). That voter may choose to strategically push Avalos and Lee apart, ranking Avalos in first place and Lee in last place, because he or she assumes that will help Avalos (the "lesser evil" frontrunner) defeat Lee (the "greater evil" frontrunner). It actually turns out that this tactic is partially correct, but that's not even very relevant. The relevant issue is that many voters will just do it "naively", based on their intuition of how the system works.

Score Voting simply doesn't have this problem. Virtually everyone is familiar with the concept of ratings. We've seen average ratings for movies, restaurants, and even products on Amazon. If you hand someone a Score Voting ballot like the example pictured above, and ask them to explain how they assume the tabulation works, they will just intuitively get it right most of the time. That is, people will better understand Score Voting the first time they use it than they will understand IRV after having already used it for years. I'm especially confident of this after having conducted polls, like this one in Beaumont, Texas, for the 2006 Texas gubernatorial election.

Voting machine simplicity
Unlike IRV, Score Voting can be counted on ordinary "dumb totaling" Plurality Voting machines, without any need for complicated (and potentially fraud-prone) electronic voting machines. We think the Department of Elections will really appreciate this.

Precinct subtotals
Score Voting can be subtotaled in precincts, unlike IRV. So you can just add up the subtotals at each precinct to get the final result. You don't have to transport the ballots to City Hall like you do with IRV. That's not only cheaper and easier, but it's better for election integrity. Moving ballots raises chain-of-custody concerns.

Summation in a single "round"
Score Voting doesn't have multiple "rounds of elimination" like IRV does. So the results are very clear. You just see a single sum of points for each candidate. You don't have this multiple rounds/columns business like you do with IRV.

Score Voting has none of the bizarre paradoxes of IRV
IRV has a crazy number of bizarre paradoxes. You can get a worse result by voting than not voting, or get a better result by staying at home instead of voting. A candidate can change from winner to loser by gaining support, or go from loser to winner by losing support. In some cases it's possible to reverse all the ballots (as if trying to elect the worst candidate) and still get the same result — meaning IRV thinks the best and worst candidates are the same person! IRV can also punish you (give you a worse result) for voting for your favorite candidate. Score Voting has none of these problems.

Better average voter satisfaction
This is perhaps the most noteworthy, albeit also the most esoteric, benefit of Score Voting. The following graph, from page 239 of William Poundstone's book Gaming the Vote, shows a graph of Bayesian Regret values for a host of different voting methods. Score Voting (called "Range Voting" in the graph) far exceeds Instant Runoff Voting.


The methodology behind Bayesian Regret is a bit complicated, but here it is if you care to understand it better. Bayesian Regret is sort of the opposite of "average voter satisfaction", so less is better (like a golf score). It's framed this way because a zero Bayesian Regret would be "perfection", and that's a convenient benchmark.

The further each bar is to the left, the better that system performs. The width of each bar represents its change in performance in response to a change in the ratio of strategic-to-honest voters. For example, Borda performs better than Condorcet or Approval Voting, if all voters are honest. But it behaves much worse than Approval Voting if there is a lot of strategic behavior.

While Bayesian Regret is indeed a very esoteric and mathematical concept, I believe it is the best (really the only) measure of ultimate voting method performance, or "democratic-ness". It precisely measures how well a voting system satisfies the will of the voters.

Here's an even more comprehensive list of benefits from Warren D. Smith, the Princeton math Ph.D. behind The Center for Range Voting.

What about getting a majority winner?

Score Voting occasionally faces criticism due to the (extremely unlikely) possibility of failing to elect a candidate who is the favorite of a majority of voters. We address that criticism in great detail here.

In a nutshell, IRV can do even more bizarre things. In the 2009 IRV mayoral race in Burlington, Vermont, the Progressive won, even though the Democrat was preferred to the Progressive by a 54% to 46% majority. And in the 2010 San Francisco District 10 supervisor race, Malia Cohen won in the 20th round with 4321 votes from 18308 voters who voted in that race. That's a 23.6% "majority".

IRV can even elect X in cases where Y was preferred to X by a majority and got more first-place votes than X. And to consider worst case scenarios for a moment, here's a hypothetical IRV election with over a million voters, in which the winner is the favorite of only two voters, and would lose in a head-to-head matchup against 19 of the other 20 candidates.

This is not meant as an indictment of IRV, but the moral here is to judge voting systems based on their average/typical behavior, rather than on the most bizarre and improbable worst case scenarios you can dream up.

What about having a runoff?

Some people like having a runoff in cases where no candidate got a majority in the first round. They see a runoff as a way to both ensure a majority winner, and allow voters to become better informed about their choices. Detractors of runoffs argue that they allow special interests to dictate election outcomes, and tend to attract older/whiter/wealthier voters, and decrease turnout in general.

As for the part about ensuring a majority winner even one no candidate was the favorite of a majority of voters, that actually turns out to be something of a mathematical impossibility.

With regard to the other pros and cons mentioned above, that's a very complex and controversial topic which we won't delve into here. The one case where there is pretty broad agreement from both the pro-IRV and anti-IRV advocates is that runoffs make sense in the mayoral race.

So we propose a simple way of adding a runoff requirement to our Score Voting proposal, for the mayoral race only. If no candidate receives at least as many points as there are voters, then a runoff is held between the top two finishers. So if 30,000 voters vote in the mayoral race, then a candidate must receive at least 30,000 points to win outright.

Why do we use as many points as there are voters? The derivation is pretty straightforward. We're just requiring 50% of the maximum number of points, which is the closest possible analog to "getting a majority". But it just conveniently/coincidentally happens that with a 0-2 scale, that number is the same as the number of voters. To demonstrate: number of voters × 2 (points per voter) ÷ 2 = number of voters.

No comments: