[EM] Approval Voting and Long-term effects of voting systems

[Kristofer Munsterhjelm]

On 11/29/2016 05:33 AM, Daniel LaLiberte wrote:
> Hi Kristofer,
> 
> Thanks for your message responding to the more neglected aspects of what
> I wanted to discuss, not that any of the other subjects are uninteresting.
> 
> On Sat, Nov 26, 2016 at 9:27 AM, Kristofer Munsterhjelm
> <km_elmet at t-online.de <mailto:km_elmet at t-online.de>> wrote:
> 
>     On 11/21/2016 07:08 AM, Daniel LaLiberte wrote:
>     > This message is about two related subjects:
>     >
>     > 1. Factoring in the long-term emergent effects of each voting system.
>     > 2. An example of how Approval Voting results in better long-term effects.
>     >
>     > Among the many criteria for evaluating voting systems
>     > (https://en.wikipedia.org/wiki/Voting_system#Evaluating_voting_systems_using_criteria
>     <https://en.wikipedia.org/wiki/Voting_system#Evaluating_voting_systems_using_criteria>)
>     > I don't see any that address the long-term effects of using each voting
>     > system.  In other words, the effect on one election is certainly
>     > important, regarding the satisfaction of the election results by voters
>     > and candidates.  But I would argue that it is even more important to
>     > consider the long-term effects that emerge when applying a voting system
>     > repeatedly over many elections.  A small bias one way or another may not
>     > be very apparent if you only look at one election, but over may
>     > elections, they can add up and perhaps compound the bias exponentially.
> 
>     There are basically two ways of doing this, as far as I can see, without
>     actually running experiments.
> 
>     The first one is to use game theory. Determine what the strong Nash
>     equilibrium is for the method in question, and if honesty is an
>     evolutionarily stable strategy.
> 
> 
> Do you mean to suggest that honesty (i.e. sincere voting?) is necessary
> in order to possibly achieve stability?

No, what I mean is that you can get a good idea whether voters will stay
honest by determining if an initially small group of dishonest voters
can make other honest voters have to turn to strategy in order to not
lose. That's what an evolutionarily stable strategy means: a strategy
(say, honesty in voting) is an ESS if an initially small group can't
grow and displace that strategy by the gains they get.

A strong Nash Equilibrium is stronger in that it requires that (say,
honesty) is stable in the sense that if everybody is initially using
that strategy, no group (no matter how large or small) can benefit by
using another strategy.

The point, in any case, is: suppose the ESS for some voting system is
honesty or some strategy that doesn't lead to two party domination. Then
you're not going to be pushed out of that state by any small group that
falls for the temptation to strategize; only a large group will do. And
if the strong Nash equilibrium for a voting method in a particular
situation is to elect some candidate X, then no coalition/group of
voters can make some other candidate Y win by strategizing, even if the
method as such isn't strategy-proof -- as long as the voters know enough
about each other.

So game theory can give you an idea of when a method will lead to
two-party dominance, and conversely, when it won't. The standard
Plurality argument "I can't vote for my party because then the greater
evil might win" is in a sense a game-theoretical one. But the problems
with game theory are:

- it only answers the precise question you ask (e.g. the questionable
realism of the very strong conditions on strong Nash)
- it's hard to figure out whether a system has a particular equilibrium,
particularly when there are many voters and parties.

The stability I'm thinking of is thus not stability of government, but
rather stability of the method. The method is stable in some state if it
doesn't trail off into another state. E.g. strategic Plurality (in
isolation) might be stable in the two-party state since it's hard for
third parties to get any support if everybody thinks "but if I vote for
them, the other guy might win". But strategic Plurality isn't stable in
the multiparty state; if you start with many parties, it'll veer
chaotically around until the parties coalesce into just a few.

(I'm being a little fuzzy here because the law isn't ironclad. Consider
the existence of the Liberal Democrats in the UK, for instance; and
regional variety may also matter, where you could have two-party rule in
different areas of the country, but which two parties they are depends
on which part of the country you're in. I've heard India is like that.)

If a method is stable in some multiparty state, that doesn't imply that
the governments it leads to are, in turn, stable or impervious to
change. You'd expect the voting method to change which parties are in
power as the people change their opinions -- if the voting method is good.

Whether the greater political system is in itself stable is a much more
difficult question, as well. We don't know whether parliamentary or
presidential democracy is stable. In some even wider sense, we know that
they aren't, as both have failed numerous times - consider Germany as an
example of the former, and various South American nations as examples of
the latter.

> But an election method might not result in stability and could still be
> a good method, depending on the nature of the instability.   It could
> oscillate between different states quickly, slowly, mildly or wildly, or
> perhaps it will grow in some direction possibly toward an inevitable
> catastrophic collapse, and maybe that pattern of gradual growth followed
> by sudden collapse will be repeated over a longer cycle. 

I suppose that is possible, but I find it hard to see how a method in
itself could have complex attractors. The complexity would arise from
outside interaction, I think. For instance, a winner-takes-it-all system
can rather severely swing from one side to another and then back again.
Some two-party states are like that: party X gets to rule, and then they
do so until the people becomes dissatisfied enough, then party Y gets to
rule, and rinse and repeat. You get something reminiscent of hunting
oscillation, and the overshoot to either side can be rather harmful.

> Stability or various kinds of instability is one class of long-term,
> emergent effects, concerned with how the state of things changes over
> time.  
> 
> Another important class of emergent effects should be about how well the
> voters are represented through party organizations or otherwise.  It
> might be, for example, that only half the voters are represented very
> well while the other half is represented very poorly.  Or it might be
> that most of the voters are represented very well to moderately well, as
> in homing in on the median, with a smaller number on the fringes being
> less well represented.
> 
> And another important class that I mentioned previously is how
> vulnerable a government is to a growing concentration of power leading
> to self-serving corruption.  This is somewhat in conflict with the
> stability effect since increased stability, which might otherwise be a
> good thing, allows a concentration of power to remain in power and find
> ways to gradually increase its power.

The increased stability has to be real, not illusory. A dictatorship is
pretty stable (until the revolution, at least) but it is so because it
manages to repress what would otherwise lead to change.

Or to take an analogy of two people who don't like each other, but
decide not to voice their dislike because they want to be nice -- and in
the end they get so angry at each other that the encounter explodes. A
better system can also be stable if it's responsive as well, which means
that it deals with changing demands placed upon it before these demands
become severe enough to cause instability. The system then smoothly
changes instead of with a jerk.

> What other emergent effects should we be aware of that we have a chance
> understanding and linking back to the election methods and other
> governmental structures that affected them?

I've mentioned some of them at the bottom, as regards political systems
in general. Another one we should be aware of is that political systems
are self-reinforcing. So suppose we find a way to implement a much
better system. This system should still leave open the possibility that
something better yet may appear at a later time, which means it should
be able to be improved upon in turn. But how to do that without also
becoming more vulnerable to say, authoritarianism... that's a tricky one.

>     The second is to assume the worst and use properties as defenses against
>     strategy. For instance, if a single-winner method passes monotonicity,
>     there's no point in attempting pushover strategy because it won't work
>     (although I'd say that's not the main point of the monotonicity
>     criterion). Similarly, if a multiwinner method is weakly invulnerable to
>     Hylland free riding, there's very little point in trying to tempt fate
>     by voting E>F when your honest preference is X>Y>E>F, even if you think
>     X and Y are sure to win.
> 
>     The logic is like that of an upper bound: if the method is completely
>     invulnerable to some kind of strategy, it'll also be invulnerable to
>     that strategy for the kind of ballots that appear in the real world.
> 
> 
> This seems closer to what I was hoping for in terms of finding way to
> determine the likely bounds of emergent behaviors based on the
> underlying properties of election methods.  How bad or good could it get?

Well, Arrow and Gibbard-Satterthwaite says that no ranked method can be
perfect and there will always be some need for strategy. In another
reply, I mentioned classes of strategy being compromise, burial and
pushover. It seems not too hard to make methods that are invulnerable to
one of these classes, but by Arrow and G-S you can't make methods that
are vulnerable to all of them.

Approval seems to do the impossible here, but as I said in the other
post, I think it's just hiding the problems by shifting the burden to
the mind of the voter.

Call the mechanism the voter uses to turn his preferences ("I like A
more than B, B more than C") into an Approval ballot, "manual DSV".

Then, as an example, if if "Approval with manual DSV" were to pass
independence from irrelevant alternatives (the Arrow criterion that
pretty much every ranked method fails), then that means there must exist
situations where a reasonable voter would approve of no candidate or of
all of them. But it's hard to imagine that ever happening in a real
election: perhaps a voter could submit a no-approvals ballot as a
none-of-the-above sentiment, but submitting one that approves of every
candidate?

>     The reason there's so little consideration of long term effect is that
>     they're hard to model, and getting any definite answers from game theory
>     is also hard. The closest you get are, I think, arguments of the form
>     that "method X passes criterion Y which is very important to deter the
>     kind of strategy that would render the method useless in practice, in
>     the long run". Different people disagree about which criteria are most
>     important, and there we are.
> 
> 
> I would expect people to disagree about which criteria are most
> important, but I suspect that is largely because each criteria is
> relevant to varying degrees and in different ways for each emergent
> effect, and we haven't agreed on what emergent effects we want in the
> first place.  I suspect it might actually be easier to agree on the
> emergent effects, however, and then we may have a better chance of
> determining which election methods could work, or which criteria are
> most relevant for affecting the emergent effects. 

Suppose we start from considering emergent effects. Then the reasoning
would go like this:

1. We want emergent effects X, Y, Z.
2. Criteria A, B, C lead to X, Y, Z.
3. So we want A, B, C.

I think that while step 1 is pretty easy - just declare the effects you
want - step 2 is *very* hard. Suppose we know that IRV leads to
"two-party plus change" rule. We know what criteria IRV passes and
fails. But which of them makes IRV lead to two-party-plus-change rule?
For that matter, how do we know that not every single-winner method used
with single-member districts leads to two-party rule?

You are right about criteria sometimes hiding what effects we want,
though. In a past discussion on this list with an IRV supporter, I
eventually found out that he wanted to keep a two-party system where the
third parties shouldn't interfere too much. In other words, he wanted
the third parties to pull the major parties in their direction but
otherwise stay very small so not as to interfere with the process. For
him, that IRV led to two-party rule was a feature, not a bug, and in his
opinion, the benefit of IRV was that it eliminates a fringe spoiler
effect. If the sum of support for the fringe parties is less than the
support for the weakest major party, then IRV satisfies IIA, but
Plurality doesn't.

I didn't figure this out until relatively late, and so I spent a lot of
time talking about bizarre outcomes from IRV, but those outcomes were of
no problem for him.

>     For instance, I think clone independence is important (to the degree it
>     generalizes to near clone situations) because parties could always run
>     multiple candidates instead of one, or third parties could be deterred
>     if there's a vote-splitting incentive. Others think the favorite
>     betrayal criterion is very important - possibly due to Plurality
>     elections forcing voters to choose between the lesser evil whereas a
>     method that passes the FBC would not.
> 
> 
> Whether a criterion is more or less important relative to an election
> method depends on the desired emergent effects. For Approval Voting,
> where the likely emergent effect is homing in on the median, I would
> think clone independence is going to be very low, but it won't matter at
> all because it is actually desirable (at least in my view) to have most
> winning candidates be fairly close to that median.  But for Plurality,
> where two party dominance should be an expected emergent effect, high
> clone independence is very desirable, especially in primary elections,
> to avoid vote-splitting.

I think that's kind of having it backwards. Part of the reason Approval
has a chance of getting to the median is because if voters either vote
for all the clones or none of them, there's no difference how many
clones there are, so Approval is cloneproof.

If say, Approval had instead been single-winner cumulative voting (where
each approval counts one point if you approve of one, each approval
counts half a point if you approve of two, etc), then that method,
despite having the same ballots, would have a pretty serious
vote-splitting problem and so would have a smaller chance of getting to
the median.

But speaking of emergent properties, what I'm thinking is a real
possibility for Approval is one of these:

Situation 1: two parties plus a growing third.
- Everything goes well until the third party becomes a real factor.
- Uncertainty in whether the third party supporters should support only
the third party or both the third party and the lesser evil causes the
election to go to the greater evil.
- Voters become pissed (like in Burlington) and repeal it.

Situation 2: noisy polls, many parties
- There are many parties and some are close.
- The polls are incorrect so some voters approve of too many or too few
candidates.
- The election goes to the wrong winner and the outcome suffers.

Neither of these are really desirable. Furthermore, at least if I were a
voter in an Approval situation, I wouldn't be unduly happy that I'd have
to run "manual DSV" in my head just because the ballot format isn't
expressive enough to let me say what I mean.

(I just thought of an unusual voting method: "Approval-Asset",
where you give a favorite elector and a ranked ballot. After all the
ballots have been gathered, the favorite candidates/electors get to set
the Approval cutoff of the ballots they're assigned to in an interactive
process, and then the Approval winner wins. Say you vote A>B>C>D and set
A as your elector. A can then turn this into the "A" approval ballot,
the "A B" ballot, the "A B C" ballot, or the "A B C D" ballot, but not
the "A C D" ballot. In essence, you make your favorite candidate/elector
run the manual DSV for you.)

> Multiple parties seems much closer to what I would like to see, at least
> compared to only two parties, but there are several other things going
> on that should be considered.  I think one thing that may be missed in
> the self-sorting by voters into parties is that there are really many
> issues, and while each voter has a mix of positions on the issues that
> make hir (gack) a relatively unique individual, each party tends to
> adopt a particular set of positions.  There may be some reasonable
> correlation between the positions, but more often than not, fit is
> uncomfortable.  And then we may get various awkward and temporary
> coalitions between parties (or factions) based on some arbitrary subset
> of common positions while other positions remain in conflict.
> 
> I think it would be much better if we could avoid the exclusive
> boundaries of hierarchical parties based on arbitrary bundles of
> positions, and instead actively support a much more natural collection
> (or web) of overlapping groups where each group focuses on just one
> issue or one position.  This would be a rather messy thing but it would
> correspond more directly to the true nature of collections of voters.
> 
> This organization, by the way, corresponds to what I am thinking of as a
> House of Delegates where each voter can select a different delegate for
> each issue.

Perhaps delegable proxy could work for this purpose. Let each person
have one person's worth of voting power. He can then either use his vote
on issues that come up from time to time, or delegate part or all of his
vote to someone else. You could then imagine more sophisticated programs
that let you delegate your vote to person X only if the upcoming issue
is a foreign policy issue (determined by some trusted independent
group), otherwise delegate your vote to person Y.

The problem with DP is that everybody knows who you're delegating your
vote to, so vote buying and coercion can potentially become a serious
problem. You could use crypto to hide this, but then the people at the
top ("movers and shakers" with a lot of voting power) would also be
unaccountable.

>     Right. You've identified the spoiler effect as a problem. But you seem
>     to have drawn the conclusion that *all* ranked methods suffer from the
>     spoiler effect, and that in every method it must be true that "if [the
>     voters] don't, then they [...] strengthen the chance of the less
>     preferred leading candidate".
> 
> 
> Yes, I have been saying that.  It reflects my intuition, but I am trying
> to provoke a clear counter-argument. So far, it sounds like it may be
> mostly true, but not so clear on the edges.

>     Consider Condorcet methods, for instance, where the winner is the
>     candidate who beats every other candidate one-on-one according to the
>     ballots. Suppose for the sake of the argument that such a winning
>     candidate exists. Then if some voters vote
> 
>     very good > good > bad
> 
>     or
> 
>     good > bad,
> 
>     there's no difference as far as the "good" vs "bad" contest goes. 
> 
> 
> But would you actually vote that way if only one of "good" or "bad" is
> likely to win. Wouldn't you want to insincerely vote "good" > "very
> good" to give "good" a slightly better chance, and also rank "good"
> better than any other candidates you might prefer as long as they are
> not too bad?   I honestly don't know enough details about any of the
> Condorcet methods to say for sure, but my intuition is that this is
> true.  I would tend to vote this insincere way, at least as long as my
> preferred "very good" candidate is not very likely to win.

Condorcet works like this: the candidate who beats everybody else one on
one wins. So if there's no cycle, what happens when you vote

very good > good > bad?

1. You give "very good" one point in the "very good vs good" contest
2. You give "very good" one point in the "very good vs bad" contest
3. You give "good" one point in the "good vs bad" contest.

This can lead to "very good" winning if "very good" now has a majority
in both "very good vs good" and "very good vs bad", or it can lead to
"good" winning if "good" now has a majority in "good vs bad" and used to
have a majority in "good vs very good".

If one of "good" or "bad" was going to win before your vote, then the
"good > bad" part of your vote makes sure you strengthen "good" rather
than "bad". The only thing "very good > good" and "very good > bad" can
do is pull the victory away from either "good" or "bad" over to "very good".

Now consider

good > very good > bad

The effects of this ballot are:

1. You give "good" one point in the "good vs very good" contest
2. You give "good" one point in the "good vs bad" contest
3. You give "very good" one point in the "very good vs bad" contest.

Number 1 is not something you want, because in the "good vs very good"
contest, you want "very good" to win. Numbers 2 and 3 are effects that
you also get by voting "very good > good > bad". So as long as you don't
establish a cycle, the "very good > good > bad" vote is better than
"good > very good > bad".

> If the only cases where this is not true are rather obscure and
> difficult to determine from the naive voter's perspective, then it
> hardly matters anyway.

What evidence we have seems to rather suggest the opposite: that
Condorcet cycles are rare. Of course, it's possible that it would change
if a Condorcet method were adopted, because the strategy incentives
would also change, but then again, it might not. In the absence of
information about what will happen in the future, all we've got to go on
is present data, and that data says Condorcet cycles are rare.

> The only benefit I can see for a ranking system is if, despite the
> drawbacks, when repeatedly applied overtime, it actually improves the
> degree of representation by homing in on the median, just as Approval
> Voting does.  But then why not just use Approval?

Because you shouldn't have to do strategy when you're a sincere voter.
There shouldn't be multiple sincere votes that you have to choose from,
where you get burned if you choose the wrong one. It's okay to have a
burden if you want to game the system (i.e. vote insincerely), but
ordinary voters shouldn't have to be burdened.

> That is, even though a ranking system may have a spoiler effect or other
> similar effects that tend to result in two parties (maybe plus some
> change), this can still work out well if subsequent elections attract
> more candidates who more voters would like to rank higher on the scale.  
> 
> But I think that if there is still a tendency to separate candidates and
> policies into a few parties, and if only one party wins, we will never
> satisfy more than about half the people at a time, at best.  We will
> never get very close to any kind of unity that crosses party boundaries
> that most people will be happy with.  And we will still tend to
> flip-flop between parties and deadlock in the meantime.

Not necessarily. Suppose you have 1D politics like this, where 0 is
left-wing and 1 is right-wing:

position candidate
0        A
0.2      B
0.4      C
0.6      D
0.8      E
1        F

Suppose that the public median oscillates with intermediate speed
between 0 and 1. In a two-party system, you might have two parties, one
with a candidate at 0.33 and one with a candidate at 0.67, and then
you'd get a hard shift once there's enough public support built up to
move the median closer to 0.67 than 0.33.

But in a multiparty system, as the public shifts from 0.2 to 0.4, the
elected candidate also shifts from B to C, and the same as you go from
0.4 to 0.6, the candidate shifts from C to D.

I suppose you could argue that if the public opinion oscillates slowly,
then while the voters spend time around the right-wing axis, the parties
on the left-wing end will stop being influential, lose support, and thus
shrink; and when the public opinion shifts back to the left, then they
aren't strong enough to field credible candidates for a single-winner
position like president.

That may be true. On the other hand, if the same method is used for
single-winner districts on a more local scale, the left-wing parties may
have solid bases in certain areas of the country, so that they're not
entirely wiped out even if public sentiment moves to the right. And if
you combine a reasonable single-winner method with proportional
representation, then as long as there's some left-wing support, the
left-wing parties will be represented in the PR assemblies.

> If many smaller parties can cooperate, then they could add up to more
> than half the voters, but how many parties and individuals outside of
> this cooperative would still be excluded entirely?  Is that really what
> we want, or is it just what we have ended up with because we couldn't
> anticipate it and we didn't have any better ideas?

In the 1D model above, I'm assuming a voter around 0 (at the very left)
would vote

A>B>C>D>E>F

and a voter around 1 (at the very right) would vote

F>E>D>C>B>A.

So you get a sort of "inherent cooperation effect" because a method like
Condorcet would find the median here. For single-peaked preferences like
the above (where voters just rank candidates from closest on the line to
most distant) there's always a Condorcet winner, so for such a method,
it doesn't matter if there are 11 candidates from 0 to 1 spaced 0.1
apart, or if there are six spaced 0.2 apart; if the median position if
0.6, it'll pick the candidate closest to 0.6.

>     Now, since no ranked method passes IIA, it is of course possible that
>     the long term effects of every ranked method would be to push it towards
>     the type of elections it handles badly.
> 
> 
> That would be an interesting and useful result, if it is true and could
> be proven. 
>  
> 
>     If that were to happen,
>     Duverger's law could be reestablished. But there are a lot of "coulds"
>     here, and there's no evidence that every ranked method would be pushed
>     towards its vulnerable region in that manner, or that even if they were
>     to be, that it would lead to two-party hegemony.
> 
> 
> The "vulnerable region" of an election method sounds like something
> unexpected could emerge that we don't like, whatever it is, though there
> is a chance it could be better.  If we can't anticipate where a voting
> system is going to evolve toward, there is at least a risk that we won't
> like where it takes us. What we should prefer instead is a voting system
> that tends to evolve to a known good place.

We can't know for sure what it's going to evolve towards because the
real world is messy. We can do game theoretical calculations (to some
extent), give guarantees in terms of criteria, and simulate utilitarian
efficiency, but these are only as strong as the assumptions.

>     As an analogy, two round runoff doesn't seem to lead to two-party rule.
>     See http://rangevoting.org/HonestRunoff.html
>     <http://rangevoting.org/HonestRunoff.html>. Yet delayed runoff
>     definitely fails IIA.
> 
> 
> Delayed runoff strikes me as very much like a two-vote Approval since
> you always get a chance to vote for one of the leading candidates in the
> second round, so you can safely vote for your preferred candidate in the
> first round.  If the first round used normal Approval, I think it would
> be even better in terms of not rewarding parties, and maybe it would
> help mitigate the underdog effect, and moreover, it would completely
> negate the main argument of approval-opponents that you would feel
> obligated to approve one of the likely winners.
> 
> So how about this as a compromise: The first round would be a primary
> election using Approval Voting with all candidates regardless of party
> on one ballot.  The second round, general election will be between the
> top two most approved candidates from the primary.   Each state could
> have a primary, but the approval votes would be totaled across all states.

That has the problem that you don't get much diversity. Suppose each
party fields two candidates. Then the supporters of each party will
approve both candidates, so both candidates get the same number of votes
and the winning party gets both candidates to the runoff.

In practice, something like this could happen: each party fields the
chosen presidential candidate and the chosen VP and tells party
supporters to approve of both. Then the second round simply becomes a
question of who gets to be president (where the loser becomes the VP).

It would be better to hold a runoff between the Approval winner, and
whoever wins when you count the ballots again but reduce the weight of
every ballot that approved of the first winner. But make this
deweighting too large, and you're back to the "I want to vote third
party but don't know if I should vote for the lesser evil too" problem.

>     > In any election, there will be two candidates who are the strongest in
>     > terms of popular support, and thus the most likely to win. Consequently
>     > (to grossly over-simplify the process) with any voting system that
>     > permits ranking, groups of voters will tend to coalesce around support
>     > for these two leading candidates to encourage everyone to support their
>     > preferred leading candidate. Eventually two major parties arise, and
>     > everyone who doesn't join one of these two major parties is excluded.
>     >
>     > So once voters and candidates figure this out, any such voting system
>     > ends up devolving into the dominance of two major parties that we get
>     > with simple plurality voting.  In fact, one might argue that plurality
>     > voting is better just because it is simpler.
> 
>     That happens if the voting method punishes not voting for the main two
>     to such an extent that the voters feel they're making the main party
>     they didn't vote for win. But it remains to be shown that all ranked
>     methods do punish voters that harshly.
> 
> 
> Would be interesting to see this investigation.  
> 
> I am tempted, as a matter of fact, to build an election simulator using
> neural nets to model voters who would learn how to vote to get their
> preferred outcome.  Has anyone done anything like that?

Not that I'm aware of. But go ahead if you'd like to try :-)

>     > Given that there are, as before, two leading candidates, how does
>     > Approval Voting affect whether one of those two leading candidates will
>     > win?  One of the two leading candidates is likely to win even with
>     > Approval Voting, so it would appear there is no benefit, but that would
>     > be a short-sighted way to judge a voting system.  In subsequent
>     > elections, it would seem likely that more candidates will run who have
>     > broader appeal to ALL voters, not just a majority or plurality. Because
>     > the winning candidates will be those who are most approved of by the
>     > most voters, there will be no value in parties that typically focus on
>     > appealing to no more than half of the voters.
> 
>     I don't see why, on the other hand, the above won't apply to voting
>     methods that don't punish voters harshly, and on the other, the above
>     necessarily applies to Approval.
> 
> 
> By "the above" I assume you mean that one of the leading candidates will
> likely win. And that is almost always the case for any half-way decent
> voting system, depending on how "leading candidates" are determined. 
> Any poll uses some voting system while sampling the population, so which
> voting system is used makes all the difference, assuming they do the
> sampling and reweighting in statistically reasonable ways.

What I mean is that if you assume that voters will put the lesser evil
first, then I can just as easily assume that voters will always approve
of the lesser evil as well. There's no obvious reason that cautious
voters would compromise in ranking but not in Approval voting.

If the voters rank the lesser evil first in ranked methods to protect
themselves from the greater evil winning, then you could just as well
argue that the voters would also approve of the lesser evil so that the
greater evil doesn't win if the polls were wrong. And if the voters keep
approving of the lesser evil, third parties can never win.

> And as I said, I don't think it is fair to judge Approval Voting or any
> voting system based on running candidates who would do badly. The simple
> reason is that candidates who would likely do badly will tend to not run
> at all. We should be looking at who the likely candidates are who
> believe they have a chance of winning.

Parties who will do well in the future have to start somewhere. That
said, I don't think any of the advanced voting methods have problems
with candidates who would do badly.

>     If you take a very long perspective, it might be the case that party
>     democracy itself can be improved upon, in which case what election
>     method we use is irrelevant. 
> 
> 
> I strongly doubt that we can rely on parties to improve democracy.  They
> tend to be anti-democratic if the power structure of the parties allows
> it, and there is no reason they would give up control unless required to
> do so. So if we have to live with parties, we have to impose democracy
> on them, in which case they effectively become proper institutionalized
> bodies of the government itself.
> 
> And even if they can be made to function democratically, then there are
> all the other problems with parties I mentioned above in this message
> and in a couple others.  I remain strongly anti-party.

I don't think changing election methods on their own can get rid of
parties, so in that case, a much more thorough change would be required.

>     If something like Gohlke's recursive
>     democracy is better, then that sidesteps direct elections altogether. Of
>     course, one can ask similar mechanism design questions of the system
>     that ends up being better. For instance, I discussed the need for
>     minority representation in Gohlke's method, where I thought requiring a
>     supermajority for the lower levels to support a higher level (and/or
>     having larger groups than three) could help with this, and where Gohlke
>     settled on a party declaration mechanism.
> 
> 
> I like the direct democracy aspect of Gohlke's recursive democracy.  But
> I think its structure is wrong for proportionally representing all the
> positions on issues.  It is a people organization first, and that may
> have some benefits for cross-fertilization of ideas, but I would suspect
> that widely dispersed minority opinions will tend to be slighted.

That's the impression I had as well. The decision process on each level
pares away the wings and elects the centrist (assuming the people in the
triad are reasonable). You end up with a very majoritarian result: if
every voter was either type A or type B, and 51% were type A, you'd most
likely get only type A representatives at the top.

I suggested, as a counter to this, that the triads could be larger and
use proportional representation. But that sort of weakens the intuitive
simplicity that "some people come together, discuss, and select a
candidate to go to the next level". So Fred suggested virtual parties
instead, where say, all type-A voters elect within their own "party",
and all type-B voters within theirs, and then you end up with a 51-49
result at the top instead of 100-0.

There's another thing to think of, as well. Some political structures
may become reversed (power goes from the top down instead of the bottom
up) or they may be incapable of defending themselves from outside attack.

Consider the idealized council structure of the USSR as an example of
the former: in theory, you have large local councils that elect
delegates to the next level, that elect delegates to the next and so on.
But what happened in practice was that the flow of power reversed: the
upper levels controlled the lower ones. "The party substituted itself
for the masses, the central committee substituted itself for the party,
and the leader substituted himself for the central committee", and all
that. In some respect, it could do that because the upper levels could
control the lower ones more effectively than the lower ones could make
the upper ones stop. As a result, the system didn't end up very democratic.

Gerrymandering might be another example of reversed power, where the
politicians select their constituents before the constituents elect the
politicians. It's even worse in Singapore: in 2011, the PAP managed to
get 93% of the seats with 60% of the vote.

I kind of like the idea of democracy by sortition because it sidesteps
the whole problem of imperfect election methods and is automatically
proportional given enough seats. Even proportional representation
elections fail absolute proportionality in the sense that every
candidate is a politician, and you have e.g. the Aristotlean observation
that elections are aristocratic rather than democratic. However,
sortition may fail to give the people a sense of participating in the
process, because there's nothing any given person can do to get himself
elected (or not). His chances are always exactly the same.

Furthermore, it might be vulnerable to attacks from the outside. A
non-democratic example of a structure being vulnerable to attacks from
the outside is the House of Lords. I'm not saying it is a good
structure, but since the people in general don't feel like they're part
of it, the elected politicians have been free to reduce its influence
without any repercussion. A chamber or house chosen by lot could also be
vulnerable for the same reason: it's hard for any given person to see
the influence of the sortition house diminish since their chances of
getting selected to it are very low to begin with.

So if we're to consider different types of democracy, and not just
different election methods within the current model, things get a lot
hairier. And it's very hard to conduct experiments.