Comment by yyyk | Hacker News Reader

Comment by yyyk | Hacker News Reader

Parent Full thread

>Voters don't know which candidates they prefer?

They know which candidates they prefer, they don't have a score for it. One voter may prefer A over B and B over C. There's an argument to give A 10.0 and C 0.0 (it's not the only way to allocate, but I'll grant it for the sake of argument). What is B's score however?

Most voters do not have an answer. The only people who will have an answer are ultra-ideological voters who are willing to grade every legislative vote. Not sure this is what we want to encourage.

>This is a feature, not a bug. Cardinal systems can capture ordinal preference

This isn't entirely true. Let's give an American example.

Say the candidates in 2024 are Biden, Romney and Trump. For the sake of argument, your preference is Biden > Romney > Trump.

So you give Biden 10.0, Trump 0.0, and for Romney, say 1.0. It's possible your 1.0 vote gives Romney the election. Now, if only Biden and Romney ran, you'd have given 10.0 and 0.0 respectively. So Romney gains by having Trump run despite no change in the ordinal preference.

> So you give Biden 10.0, Trump 0.0, and for Romney, say 1.0. It's possible your 1.0 vote gives Romney the election. Now, if only Biden and Romney ran, you'd have given 10.0 and 0.0 respectively. So Romney gains by having Trump run despite no change in the ordinal preference.

In an ordinal perspective Romney would have a larger edge because you've given a higher weight to him. You have to think about the embedding process and how much information is lost through that embedding.

You're also only looking from the democratic viewpoint. Flipping this matters too because you have a potential favorite betrayer (Romney v Trump). Something Ordinal systems aren't good at handling. In a plurality voting those three could make an advantage for Biden because Romeny takes votes from Trump (and vise versa). In ordinal we have a similar but less severe problem because people will be placing their order strategically rather than preferentially. Even preferentially you are still embedding a decently high preference for Biden despite the fact you'd give him a 0 in Cardinal.

See it this way. You could vote Trump 10, Romney 9, Biden 0. (or any variation). This indicates a strong preference of republicans over democrats. In an ordinal system you'd have to preference Trump > Romney > Biden for the same ballot. This does not express this large preferential difference. You've lost a lot of information through this embedding process. You maintain the order but you lost the degree of preference. There's no reason to drop this information.

>In an ordinal perspective Romney would have a larger edge because you've given a higher weight to him.

Perhaps. Is that a necessity of ordinal voting? I'm not so sure. Trump could have been dropped in the initial round, and then you'd have gotten a straight Biden > Romney vote.

>You maintain the order but you lost the degree of preference. There's no reason to drop this information.

The reason to drop this information is because it's often garbage (there's no real score, at most an hidden ordinal preference), and its very existence gets voters to create even more garbage in order to get their actual preferences.

In our example, the strongly Republican voter probably actually wants something like this: Trump 10.0 > Romney 0.0, Trump 10.0 > Biden 0.0, Romney 10.0 > Biden 0.0.

However, Romney has to get only 9.0 because our Republican voter can only give a global score and wants Trump > Romney. Now Biden may win because Romney got 1.0 less. By forcing a global preference score, you got garbage info.

Another point is that we shouldn't conceive of candidates as passive observers. They can also create tactical manipulations, and that's much much worse than when voters 'lie'.

Say in 2024 the country has more Republican voters than Democrats. Biden now gets Warren and Sanders to run. Most Republicans have Biden > Sanders > Warren (or Biden > Warren > Sanders, it doesn't matter). So they may wish to give Biden 2, Sanders 1. Or Biden 2, Warren 1. Biden gains the more radical Democrats run, since Republicans have to give him a higher score to express their ordinal preference! (They could vote Biden 1 and give 0 to the others, but then their ordinal preference between the other democrats disappears).

Another thing to note is that in cardinal voting systems the candidates have a big interest in strong negative advertising against the opposition. If the Republican candidates want to avoid Biden getting 2 or 1 from Republicans expressing their ordinal preference between Democratic candidates, they need to loudly proclaim to Republican voters that Biden is equivalent to Sanders/Warren. You'll get an even more extremely polarized country that way.

In general, the effects on _candidates_ are hugely underlooked in the voting system discussion which tends to look only at the effects on voters.

> Another thing to note is that in cardinal voting systems the candidates have a big interest in strong negative advertising against the opposition.

This is stronger than plurality or ranked? Ranked encourages you to run opposition ads against your in-party competition (see favorite betrayer) because ordinal systems do not account for the strong spoiler effect (only the weak spoiler effect).

I'd also strongly encourage you to read up on the Monotonicity Criteria[0]. It seems your last two comments are acting under the assumption that IRV passes this criteria.

[0] https://en.wikipedia.org/wiki/Monotonicity_criterion

>This is stronger than plurality or ranked?

Much stronger and more damaging. In a non-score system, a candidate does not care if their supporter ranks the other side as a 1, 3, 5 or 9. They voted for the candidate and that's what's important. In score, it's critical for every candidate to demonize the other side completely to ensure they get a 0 from their supporters (yes, that happens anyway, but now there's a much greater incentive). Way more polarizing.

[EDIT: Note that ads against your in-party opponent can only be so negative, because you'll need the opponent's supporters to vote for you in the general. This is true in general for wavering voters - they dislike too negative ads against people they also like, and that provides a limit of sorts. In score systems you have an incentive to demonize your opponents, even when talking to people who already support you]

>It seems your last two comments are acting under the assumption that IRV passes this criteria.

Actually, I prefer AV, but my considerations are... atypical. It's just I am very much against score/range methods, unless they are a thin disguise for a ranked system*, in which case it's slightly less bad.

* e.g. STAR, highlighted in a different comment below, is a ranked system in disguise. It's clever but way too complicated for voters.

Approval voting is a cardinal system, not a ranked choice system. That's why it's able to satisfy all of the Arrow's Theorem criteria.

Technically yes, but it's doesn't force voters to get a nonexistent score out (I also don't mind the tactical voting implications). 0/1 is pretty close to a ranked choice which allows equality - which is a problem... but no perfect system exists.

My problem with the cardinal systems is score/range systems which believe voters have a super-exact utility function and we just need to get it out. I strongly disagree with that assumption.

> Perhaps. Is that a necessity of ordinal voting

Necessary? No. Likely? Yes. This depends on if the candidates are included in the ranking or not. So for example, if we had 10 candidates and you rank your top 5, then no this wouldn't happen. But if you rank your top 10, then yes. But if you limit your ranking then some votes don't get counted because their top 5 don't make it past the first 5 rounds.

> The reason to drop this information is because it's often garbage

This is a pretty strong claim that needs strong evidence to counter. Especially considering there's lots of works demonstrating the usefulness of this information.

> In our example, the strongly Republican voter probably actually wants something like this:

If the voter truly has no preference of Romney over Biden, then yeah, that makes sense. And you can't really express this in an ordinal manner. You can only do: Trump > Romney > Biden. Cardinal systems you can do any of the following: 1) Trump > Romney >> Biden, 2) Trump >> Romney > Biden, 3) Trump >> Romney = Biden, 4) Trump > Romney = Biden. There's a whole slew of permutations that are available at no additional cost (I'd actually argue the cost is lower because `argmax(reduce_sum(candidates))` is a hell of a lot easier than `while [candidate_votes < 0.5].all() pop(lowest_score)`). I'm not sure why you're convinced that these permutations are "garbage".

> Another point is that we shouldn't conceive of candidates as passive observers

This is a point that has not been ignored and I'm not sure why you think I am ignoring this. I've stated numerous times that no system is immune to strategies. I am specifically INCLUDING this parameter. I've made zero claims that cardinal systems are impervious to strategies and manipulation. I have made claims about the effectiveness of this manipulation being reduced but reduced is not dismissal.

> Most Republicans have Biden > Sanders > Warren

To be clear, it is easier to simulate a ranked system FROM a cardinal system than the other way around. In fact, ordering candidates in a cardinal system is perfectly valid and will become equivalent to an ordered system. You are also perfectly valid to vote {Trump >> Biden > Sanders = Warren, Trump >> Biden = Sanders = Waren, etc}. Those votes are all valid and none are considered "a bug."

To be clear, in an ordinal system you CANNOT given Biden, Sanders, and Warren an equal score (of 0 or any value) unless we limit the number of candidates a voter is able to rank (which is necessary for large elections, but comes with a lot of other problems and may be unconstitutional).

>This is a pretty strong claim that needs strong evidence to counter. Especially considering there's lots of works demonstrating the usefulness of this information.

My claim: "Data X does not exist". Your reply: "Data X would be very useful!". Note that your reply does not actually refute what I said. As for support, IMHO, I think the burden should be on score supporters to prove the data exists. All my experience talking to people is that almost nobody has a real score. They have candidates they'll never support, a few they may waver between, and eventually they settle on some ordinal preference.

>If the voter truly has no preference of Romney over Biden, then yeah, that makes sense.

Err.. I think you meant the voter has Romney > Biden, right? Or you have misread me here.

>Cardinal systems you can do any of the following: 1) Trump > Romney >> Biden, 2) Trump >> Romney > Biden, 3) Trump >> Romney = Biden, 4) Trump > Romney = Biden. There's a whole slew of permutations that are available at no additional cost...

Most of these preference either don't matter (Why should a >> vote matter more than >), or create more problems then they solve (= preferences create actual draws. I don't mind making voters choose an ordering).

>In fact, ordering candidates in a cardinal system is perfectly valid and will become equivalent to an ordered system.

I got that, my point is that it creates a lot of other effects, like giving Biden a reason to run Sibyls.

---

Your general assumption is that elections are a large public survey and we need to get max data and then average it out. I strongly disagree with that model. Most people have vague preferences. Scores don't really exist and we end up with GIGO.

Reply on news.ycombinator.com