[EM] Why I think IRV isn't a serious alternative

[Abd ul-Rahman Lomax]

At 12:18 PM 11/25/2008, Markus Schulze wrote:

>If I remember correctly, Abd wrote that, in every
>IRV election for public office ever held in the
>USA, the IRV winner was identical to the plurality
>winner. Doesn't that mean that -- when we apply
>your logic -- plurality voting always elects the
>right winner?

Since I'm being mentioned, I'd better correct what was said.

In every *nonpartisan* election for public office held in the U.S., 
since San Francisco started RCV up, the winner was the first round 
leader. Something similar has been noted in Australia; vote transfers 
tend, strongly, to benefit the first round leader (and probably do so 
in proportion to the relative votes of the leader and the runner-up).

I thought, at first, that Pierce County -- the very recent election 
-- was an exception, but it turns out that this was a partisan election.

I think I understand what's going on. In nonpartisan elections, it 
turns out that, more or less, on average, voters who prefer C are a 
representative sample of the rest of the electorate with respect to 
the A/B pair, so that if C is eliminated, the votes from the C voters 
will tend to transfer in the same ratio as already exists with those 
who prefer A or B. Hence vote transfers don't radically shift the 
position of the frontrunners, normally. (It's also, apparently, rare 
for a third-place candidate in the first round to go on to win an 
election, I'm not sure it has ever happened in Australian, or if it 
has, it's been once or twice, in a huge number of elections.)

When an election is close, it may become an exception. Usually, 
however, the lead of the frontrunner is simply amplified with each 
round of eliminations. Exceptions may be, in fact, due to a false 
lead, i.e., one which was simply a reflection of statistical variations.

Yes, this means that plurality voting is a much better method *in 
practice* than we have thought, particularly for nonpartisan 
elections. It is roughly as good as IRV *in that context,* and there 
is no clear evidence that IRV produces better results there. However, 
Top Two Runoff is better than both of these, and in particular if 
write-in votes are allowed, for reasons that have been neglected by 
voting systems theorists, in general, who tend to assume fixed 
preferences, a single electorate, and both assumptions are far from the truth.

In a real runoff, voters have an opportunity to examine a reduced set 
of candidates, and will make a more informed choice between them, on 
average. Thus voters may reverse the preferences that they voted in 
the first round, i.e., even if they voted for A as their favorite, 
they might vote for B in the runoff.

Secondly, voter turnout may be different in the runoff, if it is a 
delayed runoff with a special election. (Not all are that way: Cary, 
NC, had a primary held in October, as I recall, with any necessary 
runoff being with the November general election. The city elections 
were held in off-years, not with the Presidential election, and what 
I saw for runoff turnout was that it was about the same as with the 
primary.) Top-Two Runoff is a method which effectively considers 
preference strength, and my expectation is that this would improve 
results from an overall voter satisfaction perspective, exerting a 
Range-like effect. Voters who, for example, don't have a strong 
preference between A and B, whether this is because they like both, 
or don't like both, will be less likely to turn out to vote in the 
runoff. I can say that I wouldn't bother, unless there was something 
important on the ballot besides that. Australia's practice of 
mandatory voting probably has a somewhat negative effect on the 
quality of results. None of this was likely to be noticed until we 
started realizing how important preference strength is. It's kind of 
amazing that it took so long. I attribute the delay to the disconnect 
between economists and political scientists.

(Borda was probably attractive because of an assumption that if a 
preference was "long," with many intermediate candidates, it was 
probably stronger than if it was "weak," with just a single candidate 
in between. But it is far more straightforward to simply allow the 
use of ratings, which automatically show preference strength. I.e., 
Range Voting and Borda Count are really the same method, except that 
Borda has this rigid assumption about preference strengths that makes 
it vulnerable to cloning, for example. Range is Borda with equal 
ranking allowed, and a fixed number of ranks such that ranks can be 
empty on a ballot. It's also been shown that Borda with a very large 
number of candidates, broadly distributed across an issue space, is 
equivalent to Range. And while we are mentioning large numbers of 
candidates, Range probably handles them better than any method, even 
Condorcet methods can have a lot of trouble finding a decent winner, 
because of the neglect of preference strength and the difficulty of 
defining majority approval of a result.)

Regardless, we know that, with the nonpartisan elections held in the 
U.S. recently, and FairVote has been successful, mostly, in gaining 
implementations only as top-two runoff replacements in nonpartisan 
elections, there aren't any comeback elections happening, whereas, 
from real runoff elections, as I recall the number, we should have 
expected about three, so far (before this year, I have not looked at 
this year's results yet; quite a few of them are not available yet!)

That's serious, if it is not a statistical fluke. Real runoff 
elections have a cost, to be sure, but they are necessary if one is 
to look for a true majority winner. IRV does not fix the problem; 
whenever runoffs are necessary, most of the time -- exceptions are 
uncommon, actually -- no majority is found. The same thing happens in 
Australia where they have optional preferential voting.

Consider a dark horse candidate, someone who is a good candidate, but 
who, for whatever reason, perhaps lower funding, hasn't attracted 
much attention. With IRV, that candidate may manage to get up to 
number two, but still loses the election. Getting up to number two in 
a Plurality election, with enough other votes cast that there is 
majority failure, the dark horse is now a serious candidate, and 
voters will make a far more informed choice. In addition, the 
supporters of the dark horse, who may have been relatively 
dispirited, are now highly motivated to work for the final election, 
and there will be high turnout among supporters. If supporters of the 
first-round winner only have a weak preference for their favorite 
over the dark horse candidate, they may not be highly motivated to 
turn out. Hence it's not surprising that we see roughly one out of 
three real runoff elections result in the runner-up in the first 
round winning the runoff.

What we don't know, because of poor data, is why this happens, in 
detail. I've given some speculations that may be reasonable. That the 
voters are more informed and focused is a common opinion. On the 
other hand, FairVote offers "evidence" that real runoffs are a 
problem because of the lower turnout, which is assumed to be a Bad 
Thing. (The number of voters for the winner in the runoff may be 
lower than the number of voters for the loser in the first round. 
Yes. So? Preferential voting methods don't collect preference 
strength information, and many of those first round voters may not 
really care which candidate wins the runoff, so they don't vote. Not 
a bad thing; in fact, it is normal in democratic process that those 
who care, vote. Those who don't, don't.)

Why are partisan elections different? Well, here, the candidates have 
clear affiliations, easily recognized by the electorate. In Ann 
Arbor, Michigan, in the single preferential voting mayoral election 
held there, there was a "comeback election," where the Democratic 
candidate was behind the Republican one in the first round, and, in 
fact, Democratic candidates had been losing to Republicans for some 
time due to vote-splitting with the Human Rights Party candidate. The 
Human Rights Party later morphed into the Green Party in Michigan. 
The Democratic candidate was the first African-American mayor of Ann 
Arbor. So IRV allowed those HRP members to vote for their party 
favorite, but then support the Democrat.

Good thing? Probably in that election. However, the problem there was 
a politically naive or careless HRP, that apparently didn't care 
about spoiling elections. We've seen this with Nader in Florida, he's 
still defending his candidacy. He claims that he had the right to 
run, and, of course, he did. We have the right to cause all kinds of 
political damage, we could vote for Adolf Hitler, if he were on our 
ballot. The issue, though, is the effect of the actions. If HRP 
members didn't care that they were "spoiling" elections, well, they 
were making a statement. The claim that Gore would have won in 
Florida if it had been IRV is actually quite questionable.

Nader's position then, and it is still the same, was that there was 
no important difference between the Democrat and the Republican. If 
those who supported him believed that, why would they have added a 
second preference for Gore? Maybe they would have. Maybe not.

No doubt about it. In partisan elections, in the presence of a 
minor-party spoiler, IRV is an improvement. But we probably get the 
same improvement simply by allowing multiple votes and counting all 
of them. We'd get the improvement with Bucklin, which is "instant 
runoff Approval." And both of these are simpler than IRV to 
implement, much simpler. Approval is really easy to vote; I'm 
recommending, as I have been for quite some time now that voting 
activists promote Approval immediately, because of its minimal cost 
(it might even be cheaper than Plurality; certainly there would be no 
overvotes to worry about). Approval need not be the end of the road; 
it opens the door for Range, or for Bucklin, or for other hybrid 
methods. Even IRV would be improved if voting more than one candidate 
at each rank were allowed. (I.e, the voter could vote for A and B in 
first rank, say. I won't address the technical complications here.... 
But you want to avoid "harming" your favorite, don't vote for anyone 
else in first rank. Don't care about that, and want to support two 
alternate candidates, equally, so that neither of them is eliminated 
early because of your concealing vote for the other, vote for both of 
them. Anybody but Joe? Vote in last votable rank (or above) rank for 
every other candidate besides Joe.

One more consequence to IRV. IRV makes the world safe for major 
parties; it becomes more difficult for a minor party to spoil the 
show; this may, in fact, harm minor parties; certainly the long-term 
effect in Australia has been the fading of independent minor parties. 
Warren Smith calls the support of minor parties in the U.S. for IRV 
"suicidal," and he's probably right. Because candidates are rated 
individually in both Approval and Range, and most accurately, of 
course, in Range, Range probably will have an "incubator effect," as 
Smith claims. Both Range and Approval fix the spoiler effect, and 
more deeply than IRV.

It should be realized that a Range ballot, with at least as many 
distinct ratings as there are candidates, is a more expressive ballot 
than that used by any other method. It also is obvious to me that 
preference strength matters. If a group of friends are trying to 
decide what pizza to buy, preference strength is crucial, if the 
group is to be satisfied overall. No business which neglected 
preference strength and only looked at raw preference in its market 
would be likely to survive, at least not on the basis of those 
analyses! So why do we use methods for *political* decision-making 
which don't even collect that information? And once you have it, how 
are you going to use it?

It's obvious: Count All the Votes. Voters will *naturally* vote von 
Neumann-Morganstern utilities, a big name for asking for and 
supporting with your voting power what you think you might succeed in 
getting, instead of merely to express moot preferences. The method is 
good if they can simply vote for a candidate or not, i.e., Approval. 
The method, from Smith's simulations, gets better if they can express 
those utilities with more accuracy. There is probably a point of 
diminishing returns where higher resolution in Range doesn't produce 
enough benefit to be worth the trouble.

But, politically, right now: Approval Voting. Just Count All the 
Votes. We need some bumper stickers.

The only concern I've seen about Range, expressed with any coherence, 
is that some kind of damage will be done by voters "exaggerating." 
This exaggeration, however, is what preferential systems *require*.

It is so ironic. A voter votes only for A, because the voter prefers 
A, even though the voter supposedly "also approves" of B. This is 
then considered strategic voting; the solution being supported 
through this criticism is to require voters to express exclusive 
preferences, which are tantamount to requiring all voters to vote 
this allegedly insincere way. But, of course, the critics never will 
explore these implications.

I have been looking for quite some time for the logic behind 
prohibiting overvotes. The only argument I've seen in print is that 
overvoted ballots are rejected because the "true intent of the voter 
cannot be discerned." But that is circular.