[EM] Part 3--What's wrong with IRV

[MIKE OSSIPOFF]

I've now talked about Monotonicity and the lesser-of-2-evils problem.

Now I want to briefly mention a few other things:

Were you here when we were discussing social utility (SU)?

Say we rate the candidates according to how much we'd like them to win.
Those are our "sincere ratings" of the candidates. They measure how
much good the candidates would be expected to do for us if they won.

Say, when someone wins, we add up everyone's sincere rating of that
winning candidate. That's his "social utility", abbreviated "SU".

SU matters, because: Say there's going to be an election, years down
the line, by a method that hasn't been chosen yet. We hope it's a
good method. A method's average SU is the average of the SU of its
winners over a long series of elections.

Now, say there's reason to expect Approval to have a higher average
SU than IRV. That means that _your_ expectation, estimated now, for
that future election, is better if the election will by by Approval
than it would be if the election will be by IRV.

Well, guess what? In all the simulations I've heard of, IRV does
significantly worse than Approval. These are "zero-info" simulations,
in which it's assumed that voters don't have polling data or winnability
information about the candidates. But don't let that make you think
that that invalidates the results. IRV advocates always claim that
people will vote sincerely anyway, just as they do in our zero-info
simulations. And, when winnability information is added, Approval's
average SU score can only improve.

Samuel Merrill describes his simulation study in his book
_Making Multicandidate Elections More Democratic_. Approval reliably
& robustly did much better than IRV.

As we add more candidates, IRV does even more dramatically worse.
If we have the candidates distributed more tightly about the voter
median than the voters are--as candidates tend to try to do--that
makes IRV do worse still.

Pairwise-count methods, such as Condorcet's method get the top scores
in average SU in zero-info elections. With Condorcet's method we
can expect those high scores to continue even when information is
available.

Another thing: Recently on this list you may have noticed some discussion
about how _bad_ an SU score the winner can have, with various methods.

It's a new subject here, but so far it's turned out that IRV can do
a lot worse than Approval, especially as we add more candidates.
In regards to how low can be the SU of the winner by that method.

I hope I've clarified why we don't want IRV. But by all means if you
have any questions or comments, or agreements or disagreements, be
sure to let me know.

Mike Ossipoff


_________________________________________________________________________
Get Your Private, Free E-mail from MSN Hotmail at http://www.hotmail.com.

Share information about yourself, create your own public profile at 
http://profiles.msn.com.