Why "Majority Judgment" failed in France 2007.

[Warren D Smith]

I now understand what went wrong to cause Balinski & Laraki's 2007 MJ
voting experiment to elect Sarkozy instead of the winner France
wanted, Bayrou.
That is now explained here:

http://rangevoting.org/MedianVrange.html#twopeak

But it seems rare that MJ and average-based range voting yield
different winners in the real world. At present, France 2007 is the
only clear example I know (and even then,
Balinski & Laraki did not find a very clear MJ win for Sarkozy, in the
sense it was close).
But at least in this one example mean-basedrange voting was superior,

The page
http://rangevoting.org/MedianVrange.html
has now also been updated in other ways, e.g. to mention
the "Burr dilemma" effect which might be important and which seems to favor
MJ over average.

[Jameson Quinn]

2013/5/25 Warren D Smith <warre...@gmail.com>

I now understand what went wrong to cause Balinski & Laraki's 2007 MJ
voting experiment to elect Sarkozy instead of the winner France
wanted, Bayrou.
That is now explained here:

http://rangevoting.org/MedianVrange.html#twopeak

Interesting. A few comments:

1. MJ didn't fail to get the right results in B+L's poll; it only failed to get the right results if you take their polling data and try correct it to agree with national results. This is a dodgy procedure, which involves assuming that the average voter for candidate X in Orsay is the same as the average voter who votes for candidate X in France as a whole. Despite this statistical dodginess, I do consider this study to be a valid argument against MJ, but it's an awfully weak one.

2. Later in the page you claim that MJ with all medians at bottom is equivalent to antiplurality. Of course, that's not true at all; if anything, it's equivalent to approval, or at worst, to plurality. In no cases does it even remotely resemble antiplurality.

But it seems rare that MJ and average-based range voting yield
different winners in the real world.  At present, France 2007 is the
only clear example I know (and even then,
Balinski & Laraki did not find a very clear MJ win for Sarkozy, in the
sense it was close).
But at least in this one example mean-basedrange voting was superior,

The page
   http://rangevoting.org/MedianVrange.html
has now also been updated in other ways, e.g. to mention
the "Burr dilemma" effect which might be important and which seems to favor
MJ over average.

Thank you.

Jameson

[Abd ul-Rahman Lomax]

At 02:46 PM 5/25/2013, Warren D Smith wrote:
>I now understand what went wrong to cause Balinski & Laraki's 2007 MJ
>voting experiment to elect Sarkozy instead of the winner France
>wanted, Bayrou.
>That is now explained here:
>
>http://rangevoting.org/MedianVrange.html#twopeak

That's an excellent page by Warren.

What I've noticed is how common it seems to be that a "poor outcome"
is accompanied by some kind of majority failure, and "majority
failure" can include multiple majorities. When there is majority
failure, an electorate has not made a clear choice, it's down in the
noise, including misinformed strategic voting.

(And "strategic voting" must include normalization. Unless an
election is zero knowledge, all voting will be heavily influenced by
"strategic considerations," and voters do *choose* to normalize, it
is what we do with *all choices*. We restrict applying weight to the
realistic candidate set. That is "strategic.')

Hence my suggestions:

1. A Range ballot should have a defined approval cutoff. I've
generally suggested 50% range. In a runoff environment, approval has
an *absolute meaning.* That is, it has an absolute effect, easy for
voters to understand. Approving a candidate does not mean, I like
this candidate. It does not mean that, on some absolute scale, I give
the candidate a passing grade. It means this: I prefer electing this
candidate to having a runoff be held. Or, alternatively, it might
mean, with some specific systems, I prefer that this candidate be
included in a runoff than not included. -- which is a more complex decision.

2. Whenever the majority criterion would be violated, or a majority
is not demonstrated, or there are multiple majorities, a runoff be
held. The weakest of these is the question of multiple majorities,
because tradition is clear on that. When two conflicting ballot
measures both obtain a majority, the one with the most Yes votes
prevailed. A compromise might be wherever the "most Yes votes" is in
conflict with a Range result or with a pairwise winner.

3. In Bucklin-type systems, use a complete range ballot. The lower
half is not used in the basic amalgamation (though I'd want to see
all the votes be counted), but it *is* used for pairwise comparison,
which can make a difference if a runoff is indicated.

4. Always include a pairwise winner in a runoff. This makes the
overall system Condorcet-compliant, at least is such substance that
any violation would be ridiculously rare. We can expect that rankings
within the lower half of a ballot will be relatively sincere.

5. I have a lot of ideas about these systems, but also would greatly
prefer to see more sophisticated simulations. In particular,
simulations should also analyze and study the effects of voter
ignorance and voter turnout. In a runoff, for example, voters have
much more idea of the actual strategic situation, and voters with
weak absolute preference will tend not to vote. A great deal depends
on details.

To my knowledge, all simulations of runoff voting have, with utter
unrealism, assumed the identical electorate with identical
preferences. That obviously does not hold for real runoff voting.