[EM] [ESF #1080] Re: Meta-criteria 3 of 9: Value: expressiveness

[Abd ul-Rahman Lomax]

At 05:00 AM 5/10/2010, Raph Frank wrote:
>On Thu, May 6, 2010 at 10:19 PM, Jameson Quinn 
><jameson.quinn at gmail.com> wrote:
> > Only when the vote is an
> > election, not a poll, does dishonesty come into the picture, as utility
> > conflicts with expressivity.
>
>However, as you say, choosing to vote approval style under range is a
>statement too.

What, precisely, is "expressivity" if it is not the ability to 
express utility? As we have been emphasizing for a long time, voters 
will naturally include probabilities in the energy they put into a 
choice. That "energy" has various aspects: the effort involved in 
learning about the choices involved in an election, showing up to 
vote, and how they use their voting power. They may use their voting 
power to "make a statement," entirely apart from what the expected 
results are. Or they may restrict their expressed choices to those 
they think might actually affect the outcome.

I've pointed out that, where write-in votes are allowed, it's quite 
possible that a majority of voters are suppressing their personal 
favorite. How many voters did not write in, say, Dennis Kucinich on 
their presidential election ballot, but did, in fact, prefer him? Or 
someone else.

We do not put effort, most of us, into useless gestures, mere "expression."

Now, it seems that some imagine that the ideal voting system will 
extract from voters some kind of "sincere expression" of utility, in 
order for utility to be useful; otherwise they may tend to think that 
utility is completely bogus, that the only thing that matters is pure 
preference. But we know that there is no preference-based voting 
system that is ideal; all of them will violate some intuitive 
criterion or other. As a result, we get phenomena that encourage 
voters to actually reverse preference, or, alternatively, even if the 
scenario is too complicated to reverse preference intelligently -- 
unless they have comprehensive knowledge! -- they may see results 
that show they would have obtained a better outcome if they had voted 
differently.

And from simulations, where we can create an abstract world of voters 
with known commensurable utilities, we can see that, in fact, these 
preferential voting systems do cause loss of utility under some 
conditions. One of the worst of these is instant runoff voting. Most 
of us, starting out in our study of voting systems, came quickly to 
the position that IRV was inferior to other voting systems, but, as 
Kathy Dopp is pointing out, it is even, arguably, inferior to 
Plurality, and the reason is that if the voting system is Plurality, 
voters know the ropes. They know what the consequences of their votes 
are likely to be; as well, candidates and political parties know the 
same. If a candidacy is likely to spoil an election, this can bring a 
backlash; my sense is that the Green Party of the U.S. was badly 
damaged by the events of 2000 and the candidacy of Ralph Nader.

IRV brings an illusion of safety; the propaganda has claimed that IRV 
allows voters to vote their sincere first preference, and then lower 
preferences, without harm. It's absolutely not true, under certain 
conditions, and those conditions can arise when there are three major 
parties, as is the case in Burlington. If the parties are split so 
that the middle-of-the-road party is the lowest in numbers -- or 
energy -- IRV will tend to award the election to one of the extremes. 
This is exactly the criticism of the single-winner STV method found 
in Robert's Rules of Order. Note that Robert's Rules is criticizing 
any sort of elimination, and they insist that, with their version of 
IRV, if a true majority is not found, the election must be repeated. 
Not a "runoff," where candidates are automatically eliminated.

They note that there are other forms of preferential voting, and they 
do not examine them or criticize them specially. My understanding of 
this is that Robert's Rules is a manual of actual practice, and there 
was too little actual practice for them to present and discuss the 
other methods.

The voting system that most closely simulates a series of runoff 
elections *without eliminations* is Bucklin, which was at one time, 
in the U.S., being adopted enthusiastically, with over ninety cities 
and towns implementing the method. What happened? FairVote has 
claimed that the Later-No-Harm violation of Bucklin was responsible 
for it being dropped. That's quite unlikely, in fact. Bucklin was 
dropped for political reasons; one of the most prominent was there as 
a seed from the beginning. In the very first Bucklin election, in 
Grand Junction, Colorado -- it was commonly called the "Grand 
Junction method" after thsi -- Bucklin elected, when the rounds were 
counted, a candidate who was not one of the top two in the first 
round, but who, in fact, from the votes, was clearly more broadly 
accepted. This was not a partisan election, and the rules prohibited 
the candidate's party affiliation from being on the ballot. But ... 
he was a Socialist, and this did not escape notice; it was an 
argument presented against the method in San Francisco, in debate 
before the Commonwealth Club.

Bucklin does not suffer from center squeeze like IRV. It has better 
performance in simulations (and it hasn't been accurately simulated, 
the simulations failed to recognize that Bucklin can be voted more 
flexibily; in particular, the Bucklin ballot was a kind of Range 
ballot, and you could leave the middle rank blank, thus indicating 
stronger preference for your favorite, and you could, with Duluth and 
original Bucklin, rank as many candidates as you liked in third rank.

But the strongest use of Bucklin was never tried. FairVote has been 
going around proposing IRV as a replacement for Top Two Runoff, when 
Bucklin as an *improvement* over Top Two Runoff, avoiding some but 
not necessarily all runoffs, and getting better candidates into the 
runoff, is far more appropriate, allowing voters much more flexibility.

>The ideal system would be something like declared strategy voting,
>where you just tell the system your utilities and let it worry about
>the back-end.

This is actually how Bucklin works. I've shown how the three-rank 
Bucklin ballot, if equal ranking is allowed in all ranks, is a Range 
ballot, specifically Range 4, with the two unapproved ratings of 0 
and 1 collapsed into one (i.e, no-vote). Bucklin could use and 
function well with a full Range ballot. When enough voters fail to 
approve the leading candidate, the election fails and would be 
repeated. In an ideal sense, it should be completely repeated, but 
given the Range data from the original ballots, it would be possible 
to present the voters with a decent reduced set of candidates. I have 
argued that the most effective strategy for this would be to present 
the method with a sincere expression of level of approval, with the 
minimum level that the method is looking for to declare a victory by 
a majority being the level at which the voter becomes indifferent to 
the election or to a runoff being held, or would actively prefer a 
runoff. That anchors the range ratings to a common level, making the 
votes commensurable.

But it would be easy to vote, easy to instruct voters how to 
effectively manage their ratings on the range ballot.

>Abd has made the point that actually turning out to vote is also part
>of the system.  If someone doesn't care about the election, then they
>won't show up.  Thus even a simple election with 2 candidates and
>majority rule becomes slightly (true) range-like.  You can vote A, B
>or don't care.

The effect is more than slight, I'm quite sure. But there is another 
effect from runoffs, and Robert's Rules mentions it as well. The 
voters get a better view of the candidates, for starters, and then 
they have the benefit of the original poll as a survey likely to 
express true preference. If the primary were an even better poll, 
i.e., a range poll, the data from that would be even more valuable, 
for it would show preference strength, to a degree, as well as raw preference.

I expect that a Range ballot analyzed Bucklin style might outperform 
Range itself, in reality, because the Bucklin approach, I suspect, 
incentivizes sincere Range rating. Most critics don't understand that 
bullet voting is an expression of strong preference, they imagine 
that it is some kind of insincerity motivated by partisan feeling. 
Sure, it's motivated by partisan feeling! But that's sincere!

> > This leads to a certain paradox: systems which seek to increase
> > expressiveness by increasing voter freedom - for instance, Range 
> as compared
> > to a Condorcet system - could increase strategic opportunities, and thus in
> > the end reduce expressiveness - for instance, if Range were to end up as
> > pure Approval in practice.

That might occur in a Range used as a plurality method, without 
regard for majority approval. Approval voting, in fact, used in 
repeated ballot, is almost as good as Range, failing only to find the 
rare cases where there is a majority preference that is not as 
socially useful as the Range winner, i.e., repeated approval, 
especially if an exclusive majority is required (or, in this case, 
there would be a runoff between the majority-approved candidates, to 
guarantee an exclusive choice). Bucklin simulates a series of 
repeated approval elections, where the voters, wanting to complete 
the election, gradually lower their approval cutoff. With a full-on 
range ballot used to control the voting in each round, the cut-off 
would slide down one rating level at a time, so voters could fully 
specify preference order. But there is a trade-off: if they want to 
give total voting power to their favorite, they run the risk of 
losing choice in the election. If their preference for the favorite 
is that strong, strong enough to risk that, fine. But there is no 
such thing as a free lunch, or there shouldn't be. This trade-off of 
risk is exactly how to encourage sincere ratings. There is a cost to 
assigning all your voting power to one candidate. However, if it's a 
situation where there are only two reasonably possible winners, you 
can do it with reasonable safety, with about any voting system.

>That isn't a paradox.  Range is more expressive than approval even if
>99% of the voters vote approval style.

I showed, I suspect, that if even one voter votes an intermediate 
vote, the overall utility is improved. It dithers the result. The 
chance of one voter actually affecting the outcome is low; but if we 
realize that voters are more like each other, generally, than they 
are different, what one will do may well be the same as what others 
will do, so we can imagine an election as being reduced to a sample 
of voters, with one voter possibly being quite significant.

>Expressiveness isn't reduced, it just isn't increased as much as it could be.

That's right. Reduction in expressiveness from bullet voting is 
actually not a reduction, it's an *expression.* It expresses strong preference.

>Maybe you could have 2 announcements after the election, the range
>winner and also the winner after strategy is applied.

What I've suggested for Bucklin, as a tweak on Top Two runoff, is 
just about exactly that, only more comprehensive. The method would 
have a device for determining a winner, if a majority or other 
specific level of approval can be found. But it would also analyze 
the ballots looking for a Condorcet winner, and it would report the 
sum-of-votes and average range ratings. If Bucklin is also used in a 
runoff, it would be possible to have three candidates in the runoff 
and still find an optimal result. Voters would be entering the runoff 
with good data from the primary, much better than they now have.

>For example, the method might be:
>
>- Determine the top-2 using the full range
>- Have everyone's vote recomputed using the strandard approval strategy of
>-- "vote for one of the top-2 and anyone preferred to the expected winner"
>- The approval winner is then elected, but the range info is available

Lots of alternatives are possible.