dragonwriter thinks the metric being used in the article linked here was cherry-picked because it favours particular voting systems that the author likes for other reasons.
(I do not believe this, nor do I believe the similar allegation made by vintermann in a top-level comment that's currently the highest rated. Neither offered any actual evidence, and the metric used here seems obviously reasonable to me. What I could believe is if someone thinks that VSE / Bayesian regret captures the essence of what matters in a voting system, then they are likely to prefer systems like approval voting or range voting or STAR. I don't see anything wrong with that.)
> while ignoring the cultural variation in honest ballot marking for these methods
dragonwriter thinks that different voting systems, implemented in different cultures, will exhibit strategic ("dishonest") voting to different extents and in different ways, and thinks this is a problem not addressed in the OP.
(The linked article does consider strategic voting, of various different kinds, but it's certainly possible that it isn't realistic enough about what real voters might do. It would be easier to tell whether dragonwriter has found a real problem here, if dragonwriter had given some examples of variation in strategic voting that are important but neglected in the OP.)
> resulting from the fact that the ballots for them ask a question that doesn't have an objective relationship to preferences.
dragonwriter thinks that what voters primarily have is preferences: candidate A is better than candidate B who is better than candidate C. This is what an IRV ballot asks for. STAR or approval-voting ballots ask different questions (please give a score for how much you like each candidate / please say for each candidate whether you would find them acceptable); dragonwriter finds that unsatisfactory on the basis that voters will have to translate their preferences into answers to those questions, and there is no One True Way to do that.
(I disagree with dragonwriter about what's in voters' brains. At any rate, looking within myself, it is very common that I have a better idea of some candidates' acceptability to me than I do of their relative ranking. E.g., maybe there's the Nice Party, which I like a lot, the Mean Party, which I don't like much but concede has some competence, the Evil Party, which I really don't like but again concede has some competence, and a bunch of Crazy Fringe Parties, none of which I know enough about to rate them relative to one another, but all of which I consider obviously unfit to rule. With a range/STAR system, I might rate Nice at 100, Mean at 30, Evil at 10, and the Crazies at 0. Or I might rate Nice at 100, Mean at 70, Evil at 10, and the Crazies at 0. These are importantly different opinions I might have. A preference-ordering system like IRV doesn't let me express the difference between those, but would really like me to say exactly what my order of preference between the Crazies is. For me, range/STAR does the best job of matching the actual kinds of opinions I have about candidates; 3-2-1 is also pretty good; approval and preference-ordering are similar to one another and distinctly worse than those two; and of course plurality/first-past-the-post is hopeless.)
No, that's not it at all. I mean, now that you mention it, that sounds like something I would agree with, but its not either what I was saying here or particulary relevant to this discussion.
What I said here is about honest not dishonest votes. For ballots that do not ask a question with an objective mapping to preferences, honest ballots for the same preferences are a cultural variable. (This article implictly recognizes that honest ballots have no consistent meaning for approval, when testing it with two different consistent mathematical models, though that doesn’t really address the issue and it ignores the same issue for other methods to which it applies.)
For typical public elections, this problem occurs mostly with: ballots that ask the user to assign a score or rating to each candidate which is either more detailed than an unforced ranked preference ballot (a ranked preference ballot allowing ties) or less detailed than a forced preference ballot (a ranked preference ballot not allowing ties). Phrased another way, it applies to ballots that ask the user to assign scores or categories to candidates, where either the maximum number of distinct scores or categories is either more or less than the number of candidates.
There's fairly extensive literature on scoring systems (mostly outside of voting research, because they aren't widely implemented as voting systems), whether with a wide number (e.g., 0-100) or narrow range (5 stars) showing that without an objective touchstone, these are highly variable between individuals, in a way which strongly correlates with culture, in how they relate to what people actually think, on an absolute or relative scale, about the things rated, with the only consistent thing being that they tend to be consistent with ordered preferences, but the additional information (for score systems) or compression (for limited-categories) is inconsistent.
For that reason, approval/score/etc. ballots are, at best (that is, considering only personal variation and discounting any significance of the cultural correlations) just ranked preference ballots with either randomized noise or randomized compression, into which the associated tallying methods read false signficance.
EDIT: MORE that this applies to the use of these ballots in typical public elections of candidates. For other types of voting scenarios, the ballot markings involved may have concrete meanings that make these methods better, even ideal in some cases.
For instance, when planning a group activity for which you want maximum participation from the voter pool, approval where an “approve” mark is a binding commitment to participate if that option wins is an ideal method. Similarly, score voting where the assigned score is a binding commitment of that many units some scarce resource to the common project in the event of victory for the chosen option makes perfect sense.
(But: Whatever the cause, I misunderstood your position and consequently misrepresented it, and I greatly regret that.)
It sounds as if you are arguing that it must be true because if you ask people to score things on a scale, people from different cultures tend to give different patterns of scores for a given preference ordering. I'll take your word for it that the "fairly extensive literature" does in fact show that they do, but I don't see how you get from that to the idea that preferences come first and scores are some sort of fundamentally unreal construct on top of them.
A person's culture may well influence their values. That doesn't make those values unreal or illegitimate or mean that when we're trying to aggregate people's values (in order to choose a candidate, policy, etc.) we should ignore that influence. That's obvious (or at least I think it is) in the case of specific values -- e.g., if a country has in it a lot of people whose culturally-influenced values favour peaceful foreign policy, they will tend to elect governments that favour peaceful foreign policy, and that's how democracy is supposed to work.
I think it applies equally if what's being influenced by culture is some other more abstract feature in the pattern of values. If members of one culture are more easygoing in a way that makes them find a wider range of political candidates "acceptable" than members of another, then candidates liked less by the latter culture will do worse than otherwise-similar candidates liked less by the former under approval voting, because being disliked by culture B will lead to more "not acceptable" votes than being disliked by culture A. I claim that's the right thing: being represented by the candidate B dislikes will cause more distress than being represented by the candidate A dislikes.
Now, it's potentially a different matter if what's happening is that people in different cultures have the same actual values -- both think X is a bit better than Y who is hugely better than Z -- but some cultural factor makes people from culture A who feel that way mark Y as "acceptable" and people from culture B who feel that way mark Y as "unacceptable". If you're saying that that "fairly extensive literature" demonstrates that that happens, as opposed to the situation where people in culture B really do feel a bigger gap between X and Y, then I agree that that's an argument against scoring-style or approval-style voting systems. I would be interested in some pointers to the literature in question, if so.
But to whatever extent there are actual differences in values that go beyond mere ordinal preferences, it is simply not true that "approval/score/etc. ballots are at best just ranked preference ballots with either randomized noise or randomized compression".
It seems obvious to me that real people, whatever their culture, can have values that genuinely and importantly don't reduce to ordinal preferences. Would you rather be given $30k, be given $25k, or have $30k taken from you? Would you rather eat a carrot, a radish, or a bowl of mud? Would you rather have your country led by Joe Biden, Adolf Hitler, or Josef Stalin? It's possible that people from different cultures would express their opinions about these options in different ways, but I flatly do not believe that there would be no information in what they say beyond "$30k > $25k > -$30k" or "Biden > Stalin > Hitler" or whatever.
Do you disagree with that? Or is it just that you are confident that the noise greatly exceeds the signal in typical cases? (Or some other option I haven't thought of?)
I think that ordinal preferences are the objectively comparable thing that voters fundamentally have, relevant to the question of normal elections of candidates in public elections. That’s not all they have, and if we had an objective way to measure and an objectively correct way to aggregate expected experienced (dis)utility you could probably build a better public choice system around such a measure and aggregate than anything using ranked preference ballots. But we know from the research done in scoring/rating systems that the kind of things done in score/range voting or compressed-preference ballots like approval aren’t anything like a consistent measure of that, even before considering aggregation issues.
For lots of group choice situations that aren’t elections of candidates for public office, either approval or score style voting (and probably in some cases other compressed-preference systems besides approval) have concrete comparable meaning across different ballots and make sense, as I discussed with some examples upthread.[0]
> But to whatever extent there are actual differences in values that go beyond mere ordinal preferences, it is simply not true that "approval/score/etc. ballots are at best just ranked preference ballots with either randomized noise or randomized compression".
It is, across a population, in normal secret ballot, no personal consequence for the relation of ballot markings to actual winners, someone must win, elections for public office [1], because there is no consistent mapping to what people have beyond ranked preferences and how they do scoring of the type those ballots call for. There’s been fairly extensive research on this where it comes to rating systems. It’s one of the reasons rating systems have trended away from generic forms (though there is a lot of inertia!) to context-specific forms that focus on what has consistent meaning across respondents for the particular use being made of the rating. E.g., Facebook a while ago dropping 5-star ratings for “Would you recommend?”, media platforms like Netflix and others that had star ratings switching to just Up/Down, etc.
[0] including some public elections, e.g., modified approval makes sense — and is in effect used in California (though it isn’t called that) — for conflicting ballot measures, where in the event of fundamental conflict of measures in the same election, the “winning” measure if multiple get at least the minimum required to pass is the one with the most yes votes.
[1] But even for public office elections, if you weaken any of those constraints you can find situations where this changes. E.g., plain approval would make sense in candidate elections if you eliminate the “someone must win” condition, and instead it is an optional position, and if no candidate (even the one with the most approval ratings) got more than some threshold (say, approval by a majority of voters), then the seat would be unfilled. Then, the approval/disapproval choice would have a concrete meaning rather than just being an arbitrary compression of preferences. (Actually, that’s effectively the same system as the one discussed above used by California for competing ballot initiatives, which works because in any such set of initiatives, “None are psssed” is a valid final outcome.)
A ratings ballot, on the other hand, involves massive amounts of thinking. First of all, the concept of what might represent an "honest" approach to intensity of preferences is open to debate. It's US culture to give their Uber driver max marks and tip the serving staff 20%; is this an honest reflection of the intensity of the average American's enthusiasm for service workers relative to people from non-American cultures or simply what they've been taught to do? And real world electoral situations give you all kinds of incentives to do other things (like give your preferred candidate max marks and all the other candidates 0, or all the candidates you don't hate at least 9 out of 10). Is the person who bullet votes max marks for his preferred candidate because that's what the campaign literature suggested really worth more votes than the person who carefully evaluated all the candidates on their merits but doesn't think any of the candidates are worth more than 6 out of 10? With approval voting it's arguably even worse because there's a massive dilemma over whether to vote positively for the candidate(s) you want or negatively to eliminate the risk of candidates you want to lose. Obviously this applies unevenly to different demographics: put someone downright nasty towards a particular demographic on the ballot paper, and suddenly that demographic is under pressure to Approve every other candidate, whilst others continue to happily only vote for candidates they actually want.
If you design an efficiency metric with the arguably unjustifiable assumption that votes always represent either a perfect expression of the intensity of their preferences or a simple uniform strategic vote, and you assume that the proportion of people voting strategically is unaffected by voting system, it is unsurprising that voting systems whose main shortcomings are strategic complexity perform well. Even looking at that data, IRV actually outperforms nearly all the systems if you assume IRV voters are honest and voters using the other system are very strategic which isn't unreasonable given the incentive structures of the different systems.
Framing is important, and I think STAR would be most optimal with good ol' Likert scale (five or seven point) rather than 0-N stars. But I think even simpler is the "3-point Likert" behavior of V321: approve, disapprove, and "meh, ok."
I don't like how IRV forces me to rank candidates I may know little to nothing about (e.g. downballot randos not on "my team"). I don't really care if Kick Puppies Party is above or below Trip Grandparents Party.
V321 is certainly more intuitive in the ratings (it's basically an upvote/downvote system) if not the counting, but it's still quite strategic. On the plus side, the most basic strategy of simply upvoting who you want to win and downvoting the candidates you hate ensures the most divisive candidates are removed from the pool and the obscure fringe candidates are unlikely to benefit, but there's a fair bit of juggling how many candidates to upvote or downvote depending on what you think of the candidates seen as most likely to win the final round
That is an implementation issue. IRV can be implemented as either requiring complete ordering or allow partial ordering (you have n candidates, you can assign each one level 1..n, but you may use the same level for more candidates to describe partial ordering).
Technically, that’s a complete ordering with unforced preferences, not a partial ordering.
A ballot with a partial ordering would be won for which for some pair of candidates A, B, none of A> B, A<B, nor A=B would be true. Equal preference and absence of preference are different.
A ballot that worked this way would be a meaningful extension in existing ranked preference methods that aggregate pairwise preferences from individual ballots, lile ranked pairs. (Whether the effect would be desirable is not something I’ve really analyzed, but that kind of method would naturally make distinct use of the information.)
Yes, it is weak ordering, not partial ordering.
> Equal preference and absence of preference are different.
Well, it may be understood as absence of preference, provided the absence of preference is transitive.