[EM] [ESF #1100] Re: [RangeVoting] Scenario where IRV and Asset outperform Condorcet, Range, Bucklin, Approval.

[Abd ul-Rahman Lomax]

At 10:37 PM 5/12/2010, Jameson Quinn wrote:
>Look. It's a scenario. It's not totally implausible. We are blowing 
>hot air if we argue too much about how plausible it is.

No, I showed some aspects of the scenario that were highly 
implausible as supposed "sincere" ratings. First, what was 
implausible was apparent discontinuity in the political spectrum, so 
that any voter who preferred one of the two centrists had no 
preference for the extremists. That indicated three basic groupings, 
with a large gap between them. Real voters are not distributed like 
that, when an election is large-scale.

The second problem was a problem, but not so implausible. Each of the 
two extremist groups were almost evenly divided on which of the two 
centrist candidates was preferred. Yet they all had a preference. The 
centrists were shown as equal-ranking the left and right candidates, 
equal bottom.

That indicates, from the centrist view, that the left and right are 
far away from them. The only way we identify this as a centrist party 
is the uniformity of left in ranking them above right, and right in 
ranking them above left.

>  A system that deals with it well is better than a system which 
> doesn't, all other things being equal. All other things are never 
> equal, so it's just one factor among many.

I'm not sure what "it" is, that's my point. From the preferences, 
there is no clear winner that is intrinsic. It seems that the best 
winner, though, would be one of the centrists, because a centrist is 
everyone's first or second choice. The question devolves, then, on 
which of the two centrists is the best choice. If they totally bullet 
vote, high partisan feeling is implied between them. Thus they are 
centrists only by relation to the other two parties. That's odd in itself.

There is really only one clear problem shown by the scenario. If the 
centrists bullet vote -- which would be very odd, since neither one 
of them is a frontrunner -- one of the extremists could win. If it's 
a top two runoff (FPTP primary), the runoff could be the two 
extremists. That's a known (and real) flaw of TTR with sincere voting 
in the primary. But that flaw seems highly unlikely to surface in 
Bucklin, where everyone gets to express their sincere first 
preference and then add approvals queued for voting in the second or 
third round. The centrists, given their equal bottom rating for the 
left and right, would very much dislike seeing the election go to the 
extremists, so pure bullet voting seems very, very unlikely. And 
since left and right get no support from anyone else, but will 
clearly support a center candidate over the other extreme, one or the 
other centrist will likely get a majority (or both), and I showed the 
range of votes consistent with the setup. If it were top-two runoff, 
the most-feared outcome for the centrists would be for neither 
centrist to make it to the runoff, so pure bullet voting becomes 
very, very unlikely.

>I am planning to one day make a program which explores how likely 
>and how severe strategic opportunities are in each system. Typifying 
>such strategic opportunities is useful prior work.

You could tweak Warren's simulator, IEVS is it called? It was 
designed to be heavily configurable. What "strategic opportunity" was 
there in this election that is not depending upon high knowledge? The 
election setup is extremely close. Left and right are in balance. The 
center candidates are close to each other, and every move toward 
bullet voting risks a very bad outcome.

A "strategic opportunity" is not one where it is merely *possible* 
that it improves the outcome, truly, but where it will *likely* 
improve the outcome. In order to use such an opportunity, one must 
know that the election can be moved in a desired direction without 
risking a larger loss.

Here is my analysis: Given any election method that is remotely 
practical, there can exist an electorate that is divided such that a 
voter can cast a deciding vote, if the voter votes last and has full 
knowledge of how others voted. In practice, this could be a very good 
knowledge of how others would vote, and this "final" vote could 
merely be reasonably likely as a deciding vote.

What the voter needs to know is where the near-tie is, and assuming 
that this is a two-way tie, three-way ties being extraordinarily 
rare, the voter then can cast an effective vote to improve utility. 
Is this a sincere vote?

Suppose that to cause the desired outcome, the voter must vote a full 
vote for the better of the two, and a zero vote for the other (if 
it's Range or Approval). If the voter does have another preferred 
candidate, then, the voter must rate the other preferred candidate 
equally. A Bucklin system does ordinarily allow casting a sincere 
vote, because Bucklin, if needed, converts a partial vote to a full 
vote, if it is above approval cutoff. However, if the situation is 
that the vote must be cast in first rank, or the less preferred 
candidate will win, so you can easily construct a scenario where a 
Bucklin voter must equal-rank or fail to exert strategic voting power.

Range doesn't do this as pure Range. IRV will do it, if there are 
sufficient ranks, because the favorite will be eliminated if other 
than one of the top two, but IRV seriously breaks down in other ways 
(remember, this is full-knowledge, whereas voters don't have that 
kind of knowledge, they have something much vaguer). Approval does 
it, because only full votes are allowed, but sacrifices the important 
expression of the favorite as such, and if there is a significant 
preference there, this is a form of strategic voting. Condorcet 
methods will do it, because all that is needed is for that pairwise 
preference to be sincerely expressed (and these are intrinsically 
full-vote comparisons).

"Strategic voting" in Range is meaningless unless it means equal 
ranking when there is a preference. There is never any preference 
reversal However, Range as a ballot form for Bucklin does allow 
ranking but equal voting power, for sets of candidates (one set per rating).

This is the paradox of the "strategic" analysis of Approval Voting: 
it is often asserted that a voter "really" approves of both A and B 
but only votes for one of them for "selfish" strategic gain. However, 
if there is a gain, that must mean that the voter *really* has a 
preference, so, *of course* the voter will vote to elect one over the 
other! The reverse is sometimes asserted, that Approval fails the 
majority criterion because a voter may approve of more than one 
candidate while, supposedly, preferring one to another.

Basically, fans of preferential voting, when Brams published his 
paper claiming that Approval Voting was "strategy-free," meaning that 
preference reversal was never an advantage, redefined strategic 
voting to include equal ranking in the presence of a preference, and, 
at the same time, to assert that there was some absolute quality of 
candidates called "approval," and that therefore bullet voting when a 
voter "really" approves of more than one candidate was "strategic."

If we step back and look at democratic traditions in peer 
organizations, we'll see that repeated ballot is a method of seeking 
majority approval; if it's vote-for-one, the adjustments take place 
outside of the ballot process, as candidates withdraw and voters 
shift their votes. Approval voting allows this to collapse, but the 
first round would normally start with voters voting for their 
favorites. If a Range ballot were used for the first round, with 
explicit approval cutoff, the voters can more efficiently adjust 
their votes, understanding what compromises are more likely to fly. 
Bucklin simulates this, quite well, actually. Make it two-round, and 
the simulation becomes even closer, with a deliberative stage (the 
runoff process including new campaigning).