[EM] IRV vs Plurality

[Abd ul-Rahman Lomax]

At 01:57 PM 1/10/2010, Stéphane Rouillon wrote:
>Abd Ul,
>
>from the data you produce, I agree that for the 
>Burlington election, IRV did produce the same result
>FPTP would have produced.
>However, nobody can generalize this perticular case to any election.

Stephane, we now have much data on nonpartisan 
elections using IRV in the U.S. I think you've 
misunderstood my position. I'm not generalizing, 
I'm reporting what actually happens, and only 
from that, generalizing as a prediction, not as 
some kind of rule. There is no rule. It is 
possible for a comeback election to exist with 
IRV even in nonpartisan elections, but it seems to be, at least, very rare.

>I agree that in non-partisan election the 
>rallying pattern of defeated voters does not fit 
>only one typical set of preferences.
>However again, a statistical analysis of general 
>preferences shows an unbalance of preferences, even for non-partisan elections.

There is *some* "imbalance." I.e., those who 
prefer A are *somewhat* different from those who 
do not, as to B>C preference ratios. But that 
imbalance isn't strong, it is quite weak.

>Is this unbalance major to the point that IRV 
>could allow a come-back from another candidate than the plurality winner, or
>is this unbalance minor so IRV does not change 
>anything, it depends of each election.

Of course. Except that across many elections, we 
simply aren't seeing the "simulated comeback elections."

>On a last aspect, I do agree that Condorcet is 
>better than both. And I admit that , even if I believe that the distribution
>of a generalized set of preferences is 
>unbalanced, I have not yet been able to evaluate or quantify  this unbalance.

I think you are trying to swim upstream.... the 
general rule with nonpartisan elections seems to 
be what I've said, that is, *more often than 
not*, or *as a rough estimate*, the voters are a 
sample of a relatively homogenous population that 
happens to differ in a particular characteristic: 
their first preference. The second and lower 
preferences are *relatively* independent of the first preference.

Really, this is a quite remarkable finding!

>You can argue that as long that I was not able 
>to quantify by how much this unbalance occurs 
>(amplitude distribution), it is not
>acceptable to claim that this unbalance should 
>allow IRV to find a "better" winner.

My strong suspicion is that Plurality *usually* 
works well in nonpartisan elections, because of 
the effect I'm reporting. There are exceptions, 
for sure. By no means do I consider this avenue 
of approach to be fully explored! The election 
which was the basis for Brown v. Smallwood, the 
Minnesota Supreme Court case that ruled Bucklin 
voting to be unlawful there, had Brown with a 
plurality of first-choice preferences. But 
Smallwood won, by a good margin, after additional 
preferences were added in. Without actual ballot 
data, we don't know what would have happened with 
IRV, though, that reminds me, I should look at 
the data again. What if it had been batch 
elimination IRV? That kind of analysis might be 
applicable to the Bucklin results....

When there are *many* candidates, most methods 
tend to break down, unless a true majority is 
required. That, in fact, is the basis of the most 
advanced voting system in common use, to my 
knowledge: repeated voting, vote for one, no 
eliminations but only voluntary withdrawals or 
new nominations. Until a majority winner appears. 
If this weren't in common use, I can imagine 
voting systems experts complaining that it could 
take forever. But it does not actually take 
forever, the voters finally figure out that they 
have to make some compromises or else they aren't 
going to get to go home. They used to, in some 
organizations (like the Catholic church for papal 
elections, perhaps, or the Doge election in 
Venice, force the electors to stay in the room 
until they emerged with a decision. I think the 
papal elections required a two-thirds vote, and 
approval voting was used. The Venetians also used 
approval voting; approval is quite an efficient 
method if used for repeated balloting, and 
Bucklin simulates it to some extent.)

>  But, we do have data of previous elections and
>because we both agree that a Condorcet winner is 
>a "better" winner for this purpose, we can use this reference to evaluate the
>combined impact of IRV and the unbalance of the 
>preference sets. Thus, even if I do dot know the 
>general unbalance distribution,
>I can observe that IRV allows more often to 
>obtain a Condoret winner when plurality fails, 
>than plurality finds a Condorcet winner
>when IRV fails. So I claim IRV is more reliable than plurality.

What election data are you referring to?

The serious complication here, in comparing 
Plurality with a ranked ballot method, is that 
candidate behavior and voter behavior are 
seriously different if there is a single-ballot 
plurality election, than with the other methods. 
When I suggest that plurality might be a better 
method than we often think from theoretical 
considerations and simulations, it is due to this 
other behavior, which causes many voters to make 
the necessary compromises to gain the result 
they, after all consideration, prefer from among 
those they consider possible outcomes.

We cannot predict Plurality votes from IRV or 
other ranked system votes. I've claimed that 
there is some reasonable level of predictive 
power between IRV, Bucklin, and Condorcet method 
voting, because these methods do encourage 
sincere expression of preferences. In the case of 
IRV, in particular, that encouragement is based 
on a false promise, but I don't expect that 
knowledge of this has spread sufficiently to 
seriously compromise voting behavior. It will, if the use of IRV continues.

But, I suspect, Burlington may drop IRV. And I 
wouldn't expect to find much voting pathology in 
nonpartisan elections, not that is visible from 
the ballots. The damage IRV does in replacing top 
two runoff, in nonpartisan elections, is that the 
increased scrutiny of a runoff is missing. Thus 
Ed Jew was elected with less than 40% support in 
San Francisco (was that 2004?). And it then 
turned out that he didn't live in the district, 
he was an illegal candidate. In a runoff election, this would have been seen.

Unfortunately, San Francisco, just before it 
implemented IRV, outlawed write-in votes in 
runoff elections, a serious blow to democracy, 
supported by a poor decision by the California 
Supreme Court. Voting systems activists seem to 
have been completely asleep when this took place, 
in 2002, I think it was. That's because of the 
serious lack of attention to top two runoff as a 
voting reform among voting systems theorists, who 
have mostly ignored the method, which is, except 
for the center squeeze problem which is easily 
fixed, quite an advanced method, much better than 
simple-minded simulations show.

It's the effect of voter turnout on overall voter 
satisfaction with results! -- i.e., true Bayesian 
regret, considering the entire electorate, not 
just those who actually cast ballots. The voters 
are a biased sample of the entire electorate, 
biased toward, in a special election, stronger 
preference. Which then causes the results to 
shift toward minimized Bayesian regret.

To fix the Center Squeeze problem, simply use 
Bucklin for the first round; continue to require 
a majority, don't accept a plurality Bucklin 
winner, but hold a runoff. Perhaps the top two 
Bucklin vote-getters would be on the runoff 
ballot. If write-ins are allowed, and Bucklin is 
used in the runoff, there is no spoiler effect in 
the runoff, a write-in is at worst harmless.

And to get better than this, we'd need to use 
Range methods, to allow voters to directly 
express preference strength. My suspicion is that 
the most truly ideal method we could use would 
involve a Range ballot with explicit approval 
cutoff, and a runoff when needed. I'll leave 
"when needed" not clearly defined yet, it could get quite sophisticated.

And to move even further beyond that: Asset 
voting, which allows purely deliberative 
approaches to single-winner elections, and could 
allow rapid recall and other desirable features. 
(Note that recall has its own problems, but it 
should be up to the chosen representatives of the 
electorate to determine when it's truly 
appropriate. If we are considering elections of 
legislative representatives, they should be 
*representative*. They aren't judges or the kind 
of leaders or officers who should be insulated 
from a need for public support. They should be 
actual leaders and actually responsive to their 
constituents. Asset allows the creation of a 
representative assembly that is truly and 
completely representative, to the maximum degree 
possible, and even beyond that, to *total 
representation* in some respects. No wasted votes. Not one.