New Sightline article by Kristin Eberhard
This is a new post by Kristin Eberhard.
I'll respond to some of it here.
> Political scientists and mathematicians have come up with many criteria by which to evaluate voting systems, resulting in complex tables like this one. But as Nobel prize winner Kenneth Arrow proved, no system can satisfy all criteria.
This is false. Arrow's Theorem says nothing about satisfying "all" criteria; it specifically refers to three specific criteria. And Arrow's Theorem only applies to ordinal (ranked) voting methods, so cardinal (rated) systems such as Score Voting and Approval Voting do in fact satisfy his criteria.
http://scorevoting.net/ArrowThm.html
> The Institute for Democracy and Electoral Assistance has a tool that translates criteria into priorities and selects the best voting systems for you based on your stated priorities.
This general way of thinking is fallacious. It's kind of like evaluating race cars by stating how much you value horsepower vs. drag vs. weight. What you care about is: How fast will this car complete a race? What voters really care about is: How satisfied will I be with the elected office holders with this system? It turns out we have an objective way to measure that, called Bayesian Regret. Thankfully this is briefly mentioned by the Wikipedia article Kristin linked to, but it is astonishing that she didn't mention it directly in a discussion about how to assess the quality of various systems.
Now that's not to say that "externalities" like cost or political viability (which aren't captured by Bayesian Regret) are irrelevant. But this kind of analysis needs to be grounded in science first and foremost. The same would be true of a complex topic like climate change. You start by understanding the science and the ideal policy, and then you incorporate practical considerations. If you don't get the ideal right, then the practical considerations are of much less importance. Who cares about the political viability of the wrong policy?
> Under Instant Runoff Voting, it is always safe to rank a weak third-party candidate like Nader.
This is simply false. Even a weak candidate can change the order of elimination, leading to a major change in the final outcome. E.g.
33% LePen > Macron
32% Macron > X
35% X > Macron
Macron is preferred to LePen by a huge 67% majority here, and preferred to X by a huge 65% majority. But thanks to vote splitting, Macron is the first eliminated.
LePen is the Condorcet loser—the weakest candidate. But if some LePen supporters insincerely rank Macron in first place, then he wins—which helps them get their 2nd choice instead of their 3rd.
And the bigger issue here is that you do not know ahead of time exactly what's going to happen. This is why Green supporters often vote Democrat under the present system, even if the Democrat ends up with a margin of victory that would have made it safe to vote sincerely. They did not know exactly what would happen. They just knew that a vote for Green was more likely to be a spoiler than to help them.
> For example, if you ranked Terry Tea Party first, Larry Libertarian second, and Ronald Republican third, your vote would count for the Tea Party candidate in the first round; if she was eliminated, your vote would transfer to the Libertarian;
Not if the Libertarian was eliminated before the Tea Party.
> In extremely rare cases—0.7 percent of Instant Runoff Elections in US cities—IRV creates a “center squeeze” situation
It doesn't matter that it's rare. You're failing to understand basic statistics here. It's also rare for a third party candidate to be a spoiler in our present system. But people vote strategically because of the relative probability of "Green is a spoiler" vs. "Green wins".
This is explained in great detail by a math PhD here if these relative probabilities aren't clear and obvious enough to you.
http://scorevoting.net/TarrIrv.html
> All of these systems suffer from a flaw voting experts call “Later-No-Harm”
Okay, this statement makes it clear Kristin is not acting as an objective researcher but more of a pro-IRV salesperson. Because there's a strong case that it's a flaw to satisfy Later-no-harm.
http://scorevoting.net/LNH.html
> When voters realize this, they often “bullet vote” (only score or vote for their favorite candidate among the perceived frontrunners).
"Among the perceived frontrunners"?! Are you telling me that a Green who votes Democrat and Green is "bullet voting" because the Green isn't one of the perceived frontrunners? Are you now suddenly redefining the term "bullet voting"?
In any case, this whole bullet voting argument is specious and deceptive based on empirical data.
http://scorevoting.net/BulletBugaboo.html
> Experience suggests that most voters using Approval and Score give their favorite candidate the maximum score or rank and all other candidates a minimal score or no vote.
That is an outright lie, as the previous link showed. Also...
http://scorevoting.net/Honesty.html
http://scorevoting.net/HonStrat.html
Great counterexample from a high stakes election. (A poll, but a heavily contested one.)
http://scorevoting.net/RLCstrawPoll2015.html
> Score Runoff Voting should, in theory, encourage voters to give a maximum score to their favorite and also a score to their second-favorite
Only your first and your second? This is simply false. You want to top-rate your favorite frontrunner, even if she's your 3rd, or 4th, or 5th (etc.) overall favorite. The runoff component of SRV makes this even more so, since distinguishing between the candidates is how you have influence in the second round.
> Different people have different ideas about who the “most right” winner is. The candidate whom a majority of voters support? The candidate whom most voters would choose over any other individual candidate in a head-to-head race? The candidate the fewest voters strongly object to (even if that also means that fewer voters strongly support him)? The candidate whom voters most strongly adore, even if many voters object?
This implies it's subjective, when in fact you can apply logic to this question and get an objectively correct answer.
http://scorevoting.net/UtilFoundns.html
> Score Voting would likely lead to even more negative campaigns than Plurality Voting.
Wow. Just, wow.
As far as I'm concerned, this goes beyond legitimate disagreement. This is outright anti-scientific Rovian FUD.
Kristin Eberhard Mon 15 May 2017 7:55PM
Yikes! Thanks for the catch Adam - I'll get that graphic fixed right away.
Adam Zielinski Sat 13 May 2017 3:58AM
@kristineb
Regarding Later No Harm, I think it might be more accurate to give Score Runoff a label of either "Resistant" or "Vulnerable" rather than a complete "Fail."
One of the reasons for or result of the runoff step is to minimize the Later no Harm criterion.
It incentivizes all voters to score candidates more honestly to ensure their preferences are taken into account in the runoff. So in this way discourages bullet voting.
So SRV is better on the later no harm criterion than plain score or approval voting. So it should not be given a total "fail" like those systems. There is a meaningful difference.
Kristin Eberhard Mon 15 May 2017 8:12PM
@adamzielinski Thanks for the suggestion about the Later No Harm criterion.
However, SRV does fail Later No Harm, it just also provides a countervailing motivation for voters not to bullet vote. Later No Harm asks: could you hurt your more preferred candidate by giving a score to your less-preferred candidate? For Score the answer is yes--if your more-preferred were in first place, and your score bumped your less-preferred into first place instead. For SRV the answer is yes--if your more-preferred were in second place and could have won the runoff and your score bumped your less-preferred into second place instead (botting your favorite out of the runoff).
I take your point that SRV provides a separate consideration (having a say in the runoff) that pushes back against bullet voting, whereas Score does not have that separate consideration. So although they both fail LNH, voters are more likely to give additional scores in SRV than in Score if they weigh the risk of having no say in the runoff more strongly than the risk of harming their favorite. In other words, the criterion tells us that technically they both fail, but other considerations tell us that voter behavior will likely be different between the two systems. I know I explained this in the section about SRV, but I'll look back at the language in the criteria section and try to make that distinction between pass/fail on the criteria and resulting voter behavior more clear.
Adam Zielinski Thu 18 May 2017 5:23PM
@kristineb
Thanks, I agree that technically SRV still fails LNH even though the runoff is a countervailing motivation to score more than one candidate.
What I find interesting is that in the marketing of RCV/IRV, there is hardly any focus or promotion of how IRV passes LNH or how this is such a wonderful feature of IRV. Instead, most of the marketing is focused on how IRV reduces or eliminates favorite betrayal / spoiler effect, which it doesn't really. So SRV performs better on the main criterion that most people care about and what most of the marketing is all about.
In order for a significant percentage of the electorate to resort to bullet voting with SRV, they would all have to be true partisans who really thought that all other candidates besides their favorite were equally bad.
I've known people like this who truly think there is no meaningful difference between the Democrats and Republicans and think they are both equally bad. They are usually Greens or Libertarians! However being in minor parties, they know that for most elections, their candidate isn't the favorite, so it would behoove them to give a score to at least their preferred major party candidate.
I have also met some Democrats and Republicans who think their party is best and all others, including Libertarians and Greens, and equally bad, for different reasons. But they were party operatives or campaign managers - a very small percentage of the electorate.
The vast majority of voters do have actual preferences and think there is a difference between the parties. They are not partisan enough to be motivated or persuaded by partisan campaigns to bullet vote only for one candidate and not score any others, especially with the runoff motivation that is part of SRV.
But even if there were large numbers of partisan bullet voters, they would all tend to cancel each other out. There would not be a scenario where only one candidates' supporters bullet voted and the other candidates' supporters didn't, resulting in a worst case scenario LNH result. In order for the bullet voting strategy to actually work in real life, only you and your supporters can bullet vote, while all other supporters for other candidates don't bullet vote. If a significant percentage of all voters bullet vote but for different candidates, then they all defeat each others' strategy and the bullet voting strategy doesn't work or have any meaningful effect.
So for this reason I think that LNH is a second tier criteria that is not as important as Favorite Betrayal or the Spoiler effect.
So I would put a premium on voting systems that perform better according to the Favorite Betrayal / Anti-Spoiler criteria, and not as much to the LNH criteria especially with the countervailing motivation that SRV has with the instant runoff feature.
Clay Shentrup Sat 13 May 2017 8:44AM
To Kristin,
Nader got <3% of the vote
I responded to this. Again, my example used only three candidates for simplicity. If you use more (e.g. think of the recent French election), you can indeed get a scenario where even an extremely weak candidate with 3% (or less) of the first-place votes acts a spoiler. This is a mathematical fact.
So when you say, "Engaging with Clay here didn't help me improve the article", I find that to be very disappointing. I showed you mathematical proof that your claim is inaccurate, and you're telling me that won't affect your article?
The article links to a wikipedia article that says strategic voting in Score is the same as Approval, which links to an article that says strategic voting in Approval is Bullet or Compromise. So... strategic score=strategic approval=bullet or compromise. Bullet voting is a strategy in score. How is that an egregious error?
Aside from touching on a few of your other comments briefly at the end, I'd like to primarily focus on this one comment for the remainder of this post, in order to go really deep instead of broad. You ask how this is an egregious error, so let me repeat precisely what you said before:
The best strategy is to give a maximum score to your favorite and minimum to everyone else, making Score Voting act like Plurality Voting
There are two things about this statement that are egregiously wrong.
First, imagine I linked to a page that said the best way to make a toddler happy is to draw his favorite shape, which for some toddlers will happen to be a circle. And then imagine I claimed that the best way to make a toddler happy is to "draw a circle".
Do you see how that is a drastic misrepresentation? There is a crucial difference between a rule and the particular outcome produced by that rule in a specific situation.
Now, the rule for optimal Score Voting strategy is known with precise mathematical certainty. It is to "approve" (give the maximum score to) all candidates you prefer to the expected value of the winner, and "disapprove" (give the minimum score to) the others. The voter in the example from that link approves: Dodd, Gravel, Obama, Richardson, Kucinich. That is his best strategy.
Of course we don't actually expect most voters to do such precise calculations. A simple intuitive approximation of the optimal strategy would be to just "vote for your favorite frontrunner, plus everyone you like better". In which case that voter would start with Obama (since he and Clinton were the obvious frontrunners by the time period that polling data is based on) and then also approve Gravel and Dodd.
But what if your actual favorite happens to be one of the frontrunners? Then of course you bullet vote. In that particular election. But that doesn't mean bullet voting is the rule.
That ties in to the second egregious error in your statement, about "making Score Voting act like Plurality Voting". You didn't say it, but what you mean there is honest Plurality Voting, aka "bullet voting". But everyone (including IRV proponents) says that strategic Plurality Voting users are those who do not vote for their favorite. These two views are contradictory.
Going back to your other similar claim:
Experience suggests that most voters using Approval and Score give their favorite candidate the maximum score or rank and all other candidates a minimal score or no vote.
You linked to a couple of things in that statement. One of them was a post about YouTube abandoning its rating system, because "five" was almost the only rating used. But as I wrote back in 2010, this doesn't support your argument. Because for one, people aren't only rating a single video and ignoring all the others. At worst, you could say that this is Approval-style voting, not Plurality-style.
Second, this is not like an election, where you have a finite list of candidates and there is a risk of someone you don't like winning. So voters aren't motivated to rate all the videos (actually that would be impossible). The voters are self-selecting, typically those who feel strongly about a video in the first place. People who think a video is just mediocre are much less likely to weigh in.
Your other two links are from FairVote. I'll try not to go too far into the hominem attacks here, but FairVote has an astonishingly bad record of basic logical/mathematical illiteracy and just getting basic facts wrong. We have a rebuttal to every argument offered at those links, but that would take pages and pages. Let me just pick out a couple of examples.
Speaking of the repeal of Approval Voting at Dartmouth, FairVote says:
Many backers of the petition candidates were more strategic, however. They were more likely to bullet vote, reflecting the passion that led to getting that petition candidate on the ballot in the first place.
Robert Z. Norman, a Dartmouth professor emeritus of mathematics, responded:
In 2007 there was a per voter average of voting for 1.81 candidates. Hence the proportion of bullet votes had to be fairly small (or else nearly everyone voted for one or all three candidates, but not two, which would seem crazy).
Warren Smith continued:
Specifically, if all ballots approved either 1 or 2 candidates, there must have been 19% approve-1 and 81% approve-2 ballots. Norman in later email later hypothesized that actually there may have been a strategy of "either voting for the petition candidate or voting for all [3 opposing] nominated candidates." If that was the only thing going on then 60% of the votes would have been approve-1 and the remaining 40.5% approve-3s, but in this case approval voting was clearly showing its immense value by preventing an enormous "vote-split" among the 3. In any case the fraction of "approve≥2" ballots presumably had to be somewhere between 40.5% and 81%.
I could go into depth on any of the arguments offered by FairVote, and they basically all suffer from the same class of errors.
maybe over beers is a good place to discuss philosophical differences of utilitarianism vs majoritarianism.
This statement implies that it's a subjective difference that two reasonable people can "agree to disagree" about. What I'm trying to tell you is that no, if you actually break it down into axiomatic logic, it can be objectively mathematically proved that majoritarianism is incorrect. A group can actually prefer X to Y even though a majority of its members prefers Y to X. It intuitively seems impossible but it isn't.
Here's another thought experiment, proposed by me. You can have a majority of voters in support of Proposition X, and a majority in favor of Proposition Y—while at the same time having a majority who would prefer that neither proposition pass to having both pass. This is literally possible in real life! In that case, what does majority rule even mean?
On a more practical level, let me get to your concern about a scenario where X wins even though Y is the favorite of a majority of voters. Even if you cannot accept axiomatic logic, you have to accept that all voting methods can produce wrong-seeming outcomes in some cases. For instance, here's an insanely wrong IRV worst case scenario.
Given that, it should simply be about the frequency of such anomalies, multiplied by their negative impact. That seems to me an unassailable position. But for some reason you seem to espouse a hard line position that Score Voting is a non-starter because this particular bad outcome is theoretically possible at all—even if extremely rare. I don't think that position is defensible.
Clay
Aaron Wolf Sat 13 May 2017 11:37PM
Regarding Clay's more recent points above: I don't think they are all described in a fashion everyone will like, but it's an improvement to the earlier style.
It's fair to conclude that Clay doesn't just disagree with FairVote et al but thinks their arguments are entirely specious. Without expressing my view, it's possible for this sort of perspective to be true. That's why we can't just do he-said-she-said reporting. Flat-earthers are wrong, and it's totally right to go beyond disagreement and actually say that the flat-earthers have zero credibility. We need to check whether this sort of thing applies in other cases too, although that may be hard. The main point is not to automatically reject Clay as being extreme in the same way you don't reject defenders of global-earth as being extreme.
But to weigh in with my views: I think Clay is guilty of sometimes mischaracterizing and exaggerating the views of others. I think he's right about his conclusions on the voting systems largely, but for example: majoritarianism can be a view people support for social and practical reasons rather than for mathematical ones (even though I personally disagree with it regardless). I can sympathize with Clay's reactions when he sees repetition of the same arguments that really don't hold up to scrutiny and aren't presented with careful qualifications…
Clay Shentrup Sun 14 May 2017 4:21AM
This is one area where it's probably important to weigh in on the strength of the evidence. Critics do make this claim, but they make it based on several assumptions and hypotheses. It would be fairer to word it as "worry about the potential for score to devolve to bullet voting" because there's really no compelling evidence for any claim that it will do this.
Exactly.
Regarding Clay's more recent points above: I don't think they are all described in a fashion everyone will like, but it's an improvement to the earlier style.
I dialed the "cool, calm and objective" knob up to an 11, so if there's still something that concerns you, I'd really appreciate it if you could go into some detail. We may just have different perspectives on this.
Aaron Wolf Sun 14 May 2017 3:33PM
if there's still something that concerns you, I'd really appreciate it if you could go into some detail
If all along, you'd worded things like you did recently, it would be basically fine. There's both skill and hard work / extra time that goes into making the ultimate expression for maximum good reception. I rarely ever achieve it myself (but welcome push and feedback that way).
Two things that would likely come across even better:
- avoid words like "egregious" (which isn't a personal attack and may accurately describe your judgment but just isn't likely to be taken well by anyone with any tendency toward defensiveness).
- literally follow Rapoport's Rules by taking the time to fairly express those core things about understanding the other person, agreeing with them, looking for ways you like or have learned from their points…
I'm not saying that these are requirements or that less than this is bad. I think your recent posts were all around positive and fair. I'm just saying this is the sort of thing that goes from good to exceptional. Thanks for working hard to be tactful, it does show, and I think it's only fair for @kristineb to recognize the effort and reply constructively.
Clay Shentrup Tue 16 May 2017 5:16AM
Regarding the Later-no-harm section:
All of these systems are flawed because they do not support what voting experts call “Later-No-Harm”
You assert that this is a flaw, but don't provide supporting evidence. I think that's a really big problem given that this is such an oft-touted claim with so much emphasis put on it. And given that rigorous mathematical analysis says LNH is not a good measure of quality.
If voters know which candidates are viable and which are not, they might vote for their favorite of the viable candidates and also any other candidates they like, so long as they are sure those other can’t beat their favorite.
There are two claims here, both in error.
First, it's incorrect that voters need to know which candidates are viable. If you don't know, then all can be considered equally viable in your expected value calculus. Say I feel Sanders=5, Clinton=4, Trump=0 and have no reliable polling data. Then my expected value is 3, so I want to "approve" Sanders and Clinton. If that causes Clinton to defeat Sanders, I'm only a little less happy. The math should be clear here: If not voting Clinton causes Trump to win, I'll be way less happy.
Second, you're saying they'll only vote for other "non-front-runners" if those can't beat their favorite. For example, if you feel Kucinich=5, Gravel=4, Obama=3, Clinton=0 (partially invoking the 2008 primary), you certainly don't want to approve only Kucinich and Obama (your favorite and your favorite frontrunner, respectively). If you believe Obama's a frontrunner, then you want to approve everyone you prefer to him, including Gravel. You aren't going to worry that it helps Gravel to defeat Kucinich, because it's already unlikely that Kucinich is going to perform well in the first place.
The underlying statistical calculations here are very straightforward and this shouldn't be controversial.
you can harm your favorite candidate by giving any other candidate a score or vote. When voters realize this, they often “bullet vote” (only score or vote for one candidate).
Presumably this is meant to be the evidence that LNH is a problem. But it's a non-sequitur.
First simply consider this equally valid logic: You can harm your 2nd favorite candidate by giving any lesser liked candidate a score or vote. When voters realize this, they often "double vote" (only score or vote for their top two candidates).
You can replace 2nd with 3rd or 4th or what have you and it's the same argument. Yet those who invoke LNH seem to always focus purely on the "1st" case.
To say that again a slightly different way, suppose I prefer X over Y over Z. If the argument is that I'm going to have an incentive to bullet vote for X because I'm afraid of Y beating X, then the same logic says I should approve Y because I'm afraid of Z beating Y.
This is particularly glaring if I have preferences like: X=5, Y=4, Z=0
So there is nothing special about "bullet voting". For some voters, the best strategy will be to bullet vote. For others, it will be to vote for their top two, or top three, etc.
And of course we know this, because we frequently see people betray their favorite and vote for someone "electable" with Plurality Voting. IRV proponents themselves make this argument all the time.
And beyond all this theory, we just have so much empirical data.
So in my view, this Later-no-harm boogey man has got to die. It's just not particularly useful, and there are numerous more important criteria that we should be focusing on, like Favorite Betrayal, precinct summability, transparent sum-based tallies, ballot spoilage rates, voting machine concerns, propensity to escape two-party domination, etc.
Adam Zielinski · Sat 13 May 2017 3:48AM
@kristineb
Thank you Kristin for all your hard work! I definitely appreciate your articles and perspective.
I think I have found an error in the graphic associated with the article:
In your article it says,
“Immune to spoilers, vote splitting, and clones: None
Resistant to spoilers, vote splitting, and clones: Approval, Score, Score Runoff, Top-Two Runoff, Instant Runoff
Vulnerable to spoilers, vote splitting, and clones: Plurality, Bucklin”
However then in the graphic, the icons indicate the opposite of the above.