[EM] some July 4 comments

[John B. Hodges]

>Kevin Venzke wrote:
>Message: 5
>Date: Fri, 4 Jul 2003 00:25:30 +0200 (CEST)
>Subject: [EM] A few short replies
>
>First, I want to clarify that I don't claim Condorcet's "turkey problem" is a
>serious issue.  If plain Condorcet were as easy to explain and 
>hand-count as Approval,
>I would easily prefer it to plain Approval.

(JBH) It occurs to me that the advocates of Condorcet/Approval etc. 
are in the same position as advocates of socialism in the years 
before 1917. (See, for example, LOOKING BACKWARD by Edward Bellamy.) 
They are comparing a theoretical system as they imagine it would 
work, to a real system as it has been found to work in practice. 
Naturally the imaginary system comes off best.

>John Hodges wrote:
>>Whether IRV is the best method possible is
>>open to debate; but it ain't all THAT bad. C'mon, people.
>
>I think it is comparably bad to plurality, because it still 
>encourages two-party
>dominance.  An electoral method that can't even avoid that, in my view, isn't
>worth the effort to implement.

(JBH) Oh, really? I'd love to hear your reasoning on that. Even 
better would be some evidence. How many significant parties are there 
in Australia and Ireland? In ELECTORAL SYSTEMS, A Comparative 
Introduction, by David M. Farrell (2001) I find mention of seven 
different parties in Australia.

IRV is a system designed to fill a single seat; the winner must get 
50%+ of the vote. It is a majoritarian system, not a form of 
proportional representation. Two-party dominance depends on the 
strength of the "spoiler effect". If (as under Plurality) small 
parties reliably spoil the chances of the major party nearest them, 
while also reliably failing to win the seat themselves, few will vote 
for small parties, even if those parties are their most-preferred. 
Since all multicandidate elections are open to manipulation, the 
"spoiler effect" under IRV is nonzero; hypothetical cases can be 
constructed where voters would find it to their advantage to falsely 
rank a major party first. But I think any honest person would grant 
that IRV drastically reduces the "spoiler effect" relative to 
Plurality.

Another point: Assume Condorcet-plus-some-completion-method; instead 
of two-party dominance, should we fear one-party dominance? Condorcet 
methods strongly tend to pick centrists. Could one party stake out 
the center and sit on it, unassailable? Or perhaps, might we see a 
competition of several "center" parties each vying to be 
indistinguishable from the others, so we routinely have a choice 
between Coke and Pepsi? What incentive do "center" parties have to 
listen to/coopt the proposals of "wing" parties? Under IRV the 
"center" can be squeezed from both sides, can be eliminated if a 
"wing" accumulates more second-choice votes from the "far wing", so 
we can be confident of at least two viable and meaningfully distinct 
choices. If a "center" party exists, it will be working hard to lure 
voters away from the wings, fighting the squeeze. Every party will 
have a reason to campaign for the second-choice votes of other 
parties near them. Can this be said for Condorcet-based systems? And 
on what grounds?

Since neither Condorcet-plus-whatever or Approval has ever been tried 
on any large scale, we do not know how they would work in practice. 
In CHAOTIC ELECTIONS! A Mathematician Looks at Voting, Donald G. 
Saari reports on the only serious experimental test of Approval, Page 
53: in 1999 the Society for Social Choice and Welfare elected their 
president by AV, and ALSO asked each voter for a complete preference 
ranking of the candidates, as a "test of voting theory". There were 
three candidates, all highly regarded, so it was a competitive 
election. It turned out that the AV winner was not the Condorcet 
winner. It also turned out that only one voter listed a second choice 
in the AV election; since everyone knew the election was close, most 
chose to "bullet-vote" for their first choice. Saari speculates that 
AV will reduce to Plurality in close races.

Saari makes the point that, to get a definite predicted result for 
AV, you not only have to specify a profile of voter preferences, you 
ALSO have to specify some assumptions about voter behavior in 
choosing how many candidates to "approve". I.E in an election with 
five candidates, in AV the voter may vote for one, two, three, or 
four; which will they choose to do and why? With ONE assumed profile 
of voter preferences, AV may produce just about ANY result, depending 
on how the voters choose to set their individual thresholds of 
approval. One implication of this is that ANY quirky (undesirable) 
result possible under Plurality, Runoff, Borda, or any other 
"positional" method may theoretically be reproduced under AV, and in 
addition many OTHER quirky results are possible. So, in predicting 
the performance of AV, your behavioral assumptions are crucial. And, 
really, you should justify them, try to make them realistic, 
encompassing all the ways we actually see humans behave.


>You also said:
>>Oddly enough, even Condorcet-completion methods cannot be relied on
>>to select the Condorcet winner, once you take into account the
>>incentives for and effects of insincere voting.
>
>Can you clarify this?  As I understand those terms, Condorcet 
>completion methods
>are only used when a Condorcet winner doesn't exist.
>
>Kevin Venzke
>stepjak at yahoo.fr

(JBH) Plain Condorcet is not an option, because it does not reliably 
pick a winner. There may not BE a candidate or party that beats all 
others in pairwise comparisons. All Condorcet methods are 
Condorcet-Plus, some additional procedure for picking a winner when 
plain Condorcet does not.

Adam Tarr wrote:
>Can you really put the Condorcet "nighmare" scenario I wrote in the same
>ballpark as the IRV nightmare scenario?  My Condorcet scenario ends with
>the compromise candidate winning, and nobody regretting their vote.  In the
>IRV scenario I showed, the clearly, indisputably preeminent candidate
>loses, and a third of the electorate is kicking themselves on election night.

(JBH) For reference, I will copy: From Adam Tarr:
>the nightmare scenario of IRV:
>
>10% FarRight>Right>Centrist>Left>FarLeft   [call this group "A""- JBH]
>10% Right>FarRight>Centrist>Left>FarLeft   [B]
>15% Right>Centrist>FarRight>Left>FarLeft   [C]
>16% Centrist>Right>Left>FarRight>FarLeft   [D]
>15% Centrist>Left>Right>FarLeft>FarRight   [E]
>13% Left>Centrist>FarLeft>Right>FarRight   [F]
>11% Left>FarLeft>Centrist>Right>FarRight   [G]
>10% FarLeft>Left>Centrist>Right>FarRight   [H]
>
>Centrist has the most first place votes, the most second place votes, and
>the most third place votes.  Centrist is the only candidate who does not
>appear fourth or fifth on any ballot.  Centrist would win in a landslide
>over any other candidate in a two-way race.  Centrist is quite obviously
>the popular choice by ANY reasonable measure.
>
>In Condorcet, plurality, top two runoff, or really any reasonable method,
>Centrist wins.  But in IRV, Centrist is eliminated before the final runoff,
>and Right wins in a squeaker.

(JBH) This scenario is worth careful study. There are eight distinct 
groups of voters, and five parties; the "50-yard line" is occupied by 
group D, center-right. The Center party is appealing to both D and E, 
but the process of elimination in IRV results in two "wing" 
coalitions ABC and FGH, both of which are larger than "center"'s DE, 
so the Center party is squeezed a bit too much and loses out.

Right-of-center has slightly more voters than left-of-center. The 
question is how many parties will colonize this spectrum, and how 
they will split it. If we had two parties, ABCD and  EFGH, the 
Center-Right party would win with 51%. If four parties, appealing to 
AB, CD, EF, and GH respectively, again Center-right would win, 
getting votes from ABCD. If we had eight parties, corresponding to 
the eight groups of voters, and voter preferences among the parties 
followed the same linear logic as in the example above, the process 
of elimination would again assemble a center-right coalition of ABCD 
winning with 51% of the vote.

So, basically what you have shown with this example is that pure 
center parties can get squeezed out under IRV. To remain viable, a 
pure center party has to stay larger than at least one wing, taken as 
a coalition. It has to poach voters from at least one wing; it cannot 
allow itself passively to be poached upon. In trying to poach from 
either wing it may lose some voters to the opposite wing; probably 
the equilibrium solution involves two parties glaring at each other 
over the 50-yard line, each fighting also to guard their rear from 
additional parties. "Wing" parties may win if Center-wing and Wing 
together make a majority but Center-wing allows itself to get too 
centrist. There will always be at least two distinctly different 
parties.

So we return again to the question, so far hypothetical, of how many 
parties would form if elections were Condorcet-plus, and where on the 
spectrum they would settle. Would the equilibrium solution be a 
one-party state, like Mexico was for so many decades? (And like the 
U.S. may be becoming?)

My own conclusion is that one-seat elections should be avoided 
wherever possible. Legislatures should be elected by proportional 
representation, with a "district magnitude" of at least five seats. 
Single-seat methods should be reserved for electing executives, such 
as Mayor, Governor, President. If a "mixed-member" system is 
established, where some legislators are elected in single-seat 
districts, then we need to find a single-winner voting system that 
avoids, as much as possible, the spoiler efect. IRV is a contender; 
I'm open to arguments for other systems.
-- 
------------------------------------
John B. Hodges, jbhodges@   @usit.net
The two-party system is obsolete and dysfunctional.
Better forms of democracy: www.fairvote.org
REAL CHOICES, NEW VOICES, by Douglas J. Amy