[EM] EMAV?

[Abd ul-Rahman Lomax]

For some unknown reason, Jameson responded with a new subject header, 
instead of to my original EMAV proposal, so I'm copying that at the 
end of this post.

At 10:13 PM 6/30/2013, Jameson Quinn wrote:

>Abd proposed Bucklin//Score, which he dubbed "evaluative majority 
>approval voting". My first, and still my principal, response was: 
>that's not bad, and if you can build were a consensus behind that, 
>I'll sign on. I'd still like to see Abd respond to that ( and 
>ideally commit to first mentioning a consensus Bucklin proposal once 
>per thread before discussing the relative merits between Bucklin 
>systems) before engaging with the specific discussion below.

EMAV is Bucklin. As proposed, it's three-rank Bucklin. An additional 
rating is added for full utility expression.
Hence, under certain conditions, it is Range. The minimum rating is 
the default No, which could be explicit.

Under those conditions, the rating values could be 2, 1, 0, -1, -2.

The ballot is a range ballot. The Bucklin evaluation is median Range 
except with ties, which are resolved with the social utility 
informoation from the ballots.

>My basic response to EMAV specifically is: Bucklin is good and Score 
>is good, so it's not bad, but it's still in some sense the worst of 
>both worlds.

It addresses the most significant problem of Range, majority 
criterion failure. The method is MC compatible. Yet it collects full 
Range data (and with interpolation and the explicit min rating, it 
would be a Range 8 ballot, very close to the commonly-proposed Range 
9. Range 9 *could* be used. But the interpolation idea -- not 
essential -- keeps the ballot simpler.

>  Not as responsive or simple as Score,

I'd disagree. It is more responsive than Jameson's proposal, MAV, 
which dumps and does not use the utility information available, if 
there is a multiple majority. EMAV decides a multiple majority based 
on the utility information. It is *very* easy to understand, and 
intuitive understanding is quite likely to be correct. It's Bucklin 
(which was easy to understand) and Range (also easy to understand.)

It will collect full Range data; the *basic voting strategy* is to 
vote a sincere Range ballot, same as Range, i.e, Range with the usual 
"max-rate favorite frontrunner, min-rate worst frontrunner" strategy. 
It allows clear distinction of the Favorite from the preferred frontrunner.

(Remember, multiple majorities in first rank will be extraordinarily 
rare, we should be so lucky. Anyone concerned about Favorite 
Wimpiness (equal ranking is not a betrayal, as such), can easily 
defer that additional approval to second or third rank.)

>  not as exaggeration or chicken resistant as most Bucklin methods, 
> it's just a compromise that will make nobody really happy.

1. "Exaggeration" is a made-up story. People extreme-rate when they 
have strong preference. That preference may be *simple,* such as "I 
don't know any other candidates and I don't want to give any voting 
strength to an unknown," or "I only vote for Democrats," but it *must 
be* a strong preference, or they would care about risk of the second 
preference beating the first. They'd be pleased to see the second 
preference win. "Exaggeration" is also used to describe *normal 
social choice.* I.e., we choose among realistic alternatives, and our 
"rating" of them depends on what we see as realistic. It's 
"strategic." And voters who vote absolute ratings will mostly be 
fooling themselves or disempowering themselves.

2. "Chicken resistant" is also a made-up story based on our 
imagination of how a voter faces a strategic choice. There is no 
speeding car heading toward us. There are options and assessments of 
probabilities. There is no other driver we are competing with. And 
it's just one vote. There are simply different methods of determining 
the winner if there is a multiple majority. EMAV is practically 
identical to MAV except with multiple majorities and the use of range 
rather than pure reduction to approval, losing all that utility 
information, when majority approval has not appeared. What are the 
two alternatives that supposedly create a "dilemma"?

1. Add B as an additional approval to our favorite A. Bucklin and 
EMAV allow two lower preference levels for this, which would 
rationally correspond to preference strength, with the minimum 
preference level indicating approximation to election expectation. 
Remember, if there is another candidate, C, whom we fear will win, 
our election expectation has gone down, so a mid-rating will be more 
appropriate for B than otherwise.
2. Do not add B, risk the election of C. EMAV provides multiple 
levels of response to this situation, which are then balanced with 
(1). How the voter actually votes will depend on preference strengths 
in the three pairs and the election expectation.

>  Bucklin hybrids can work well; Bucklin//Condorcet, for instance, 
> is in many ways best of both, though its counting complexity is a 
> bit high for me. But if EMAV is better than MAV, then Score is even 
> better; while if EMAV is better than Score, then MAV is even better.

The problem, Jameson, is that ordinary Score misses something, called 
Majority Approval, a basic element of democracy. It's a complex 
issue, and it may, in fact, make little practical difference, in many 
elections, but the principle of majority consent to decisions is 
quite basic and is a widespread idea; therefore getting Range 
implemented, just like that, is likely to be difficult, because MC 
failure *will* be asserted as a counter-argument.

EMAV begins collecting Range data, and uses Range evaluation for 
median tie-breaking (and plurality resoluttion), so it would break the ice.

I'm going to guess that SU study of this method, with proper 
application of voting strategy, would take this down to 
*almost-Score* Bayesian Regret. I've also argued that Range should 
include an explicit approval cutoff, which this method effectively 
incorporates and uses.

As you know, I would greatly prefer to complete with majority 
approval, and thus I'd see this method, or Range, used with a runoff 
if there is no majority approval shown. Ultimately, if it's a 
two-round system intrinsically (the first election is really a 
nomination primary), this could be the second round. The first round 
would also work, but I'd want to see Condorcet testing, both with 
EMAV and MAV, in the first round. Not the second.

original post on EMAV
From: Abd ul-Rahman Lomax
Date: Fri, 28 Jun 2013 10:06:10 -0500


>The other day I sent this post to the mailing list for the Center 
>for Election Science (electionscience at googlegroups.com). Comments 
>are very welcome.
>------------------------
>Okay, from discussions on the EM list, I've come to a method that 
>seems practically ideal to me. We can call it EMAV, Evaluative 
>Majority Approval Voting.
>
>It uses the fact that a Bucklin-ER ballot, sanely voted, represents 
>a descending approval cutoff; therefore the ballot
>
>1. Expresses approval explicitly for all candidates so categorized.
>2. Expresses preference *strength*, tied to approval cutoff, which 
>is, from a common suggested approval strategy, rooted to election expectation.
>
>The method:
>
>1. Ranked ballot, with three approved ranks, one explicit 
>disapproved rank, and default maximally-disapproved rating for blanks.
>2. The ranks are ratings, with values of 4, 3, 2, 1, with 0 being 
>default. A rating of 2 or higher is approval. (so rank number is the 
>inverse of rating.)
>3. Voters categorize candidates into categories, corresponding to 
>the ranks (ratings), and may categorize as many candidates as they 
>choose into each category. Categories may be empty.
>4. The count of all legal ballots is the basis for "majority."
>5. MAJORITY SEEKING. Canvassing begins with the first rank. If a 
>majority of voters have so categorized a candidate, and no other 
>candidate, the election is complete and that candidate wins. If more 
>than one candidate has a majority, see the Evaluation process.
>6. If no candidate has a majority in the first rank, the number of 
>second rank votes is added in to the previous sums, and majority 
>support is again tested as with the first rank.
>7. If no candidate has a majority in the first and second ranks 
>combined, the number of third rank votes is added in to the previous 
>sums, and majority support is again tested, as with the first rank.
>8. EVALUATION. Vote evaluation is performed if there is no majority, 
>or if the above process finds more than one candidate with a 
>majority at a rank amalgamation. For each candidate, all the cast 
>votes are summed, using the rating values. If there are multiple 
>approved majorities, from the prior process, the winner is the 
>majority-approved candidate with the highest sum of ratings. If 
>there is no majority approval, the winner is the candidate with the 
>highest sum of ratings.
>
>Notes:
>
>This method is "almost Range." In a runoff system, with runoff 
>conditional on lack of majority approval, it could send top approval 
>and top range candidates to the runoff. It could also be pairwise 
>analyzed to send a Condorcet winner to the runoff.
>
>As a runoff method, the voters will now presumably be much better 
>informed about the candidates. The candidate set will likely be 
>reduced. This method is safe from ordinary Favorite Betrayal in a 
>runoff. (There are very unusual situations that could require 
>strategic equal-rating, they involve multiple majorities. In the 
>primary, there are other rare situations where some level of 
>strategic voting is required; intrinsically, the way that many 
>define "strategic voting," any use of "election expectation" is 
>"strategic." In a runoff system, voters would presumably vote 
>conservatively as to adding additional approvals, those with strong 
>preferences. That's rational and *does express real perference 
>strength.* That includes "bullet voters," based on only knowing the 
>Favorite. A runoff gives these voters a new look at the candidates. 
>If someone the best candidate is not nominated for the runoff, a 
>write-in campaign would be based on actual knowledge of voter 
>preferences. A Condorcet winner, for example, if Condorcet testing 
>is not done, could still be visible from the votes if they are all 
>reported, and if those preferences are maintained, and voted with 
>knowledge, this candidate should win the runoff.)
>
>If there is Condorcet failure with this method, it is likely to be a 
>case where the Condorcet winner is *not* optimal.
>
>The canvassing is simple to understand, I suspect. This is Bucklin, 
>purely, unless there is majority failure, or a multiple majority, in 
>which case it becomes Range (completely or within the majority-approved set).
>
>This method could, of course, be trivially adapted to use more 
>rating categories. I suggest that they always be balanced between 
>approvaed and disapproved ratings, with midrange being considered 
>"barely" approved, if it is Range N, N being even. If N is odd, then 
>there is no midrange rating.
>
>Unless the previously suggested overrating method is used to allow 
>half-ratings. (This method would allow a voter to express 
>half-ratings by voting two adjacent ratings. It handles what would 
>otherwise be overvotes by counting the vote at the average of the 
>top and bottom rating expressed.)
>
>So if the method used a Range 9 ballot, voting the ratings of 4 and 
>5 would be considered a vote of 4.5, thus midrange, and a 
>'stand-aside' approval. This is a trick for almost-doubling the 
>resolution of a range method. Explicit zero might be added to the 
>ballot if this is used, that would then allow doubled disapproved rating.