[EM] bullet voting and strategy on Approval ballots.

[robert bristow-johnson]

okay, Warren, i do not want to get into a picky little argument about  
what i said and how it is represented by others.

On Aug 29, 2010, at 12:55 PM, Warren Smith wrote:

> look, what I did not say at any point in this thread so far:
>  * Robert Bristow-Johnson (RBJ) is associated with FairVote.

i hadn't read that anyone suggested you did.  why did you bring it  
up.  i think you can remember back in 2009 when we were both  
harraunging Rob Ritchie at FairVote.  i do *not* think they count me  
as an ally even though we share the same concerns.  i just think they  
bet all of their chips on the IRV solution when they should have  
considered backing other solutions.

>  * He thinks IRV is better or worse than approval.

i tried to make it clear that, although different methods *sometimes*  
use different ballots or ballot-marking rules, i decouple the issue of  
what ballot to use from the method in cases of a common ballot type  
from the tabulation method.  from that i've been consistent and clear:

    Ranked-Choice ballot   -->  "good"  (and better than either Score  
or Approval or FPTP ballots)
            IRV tabulation -->  "not so good"
      Condorcet tabulation -->  "much, much better"

so i'm saying that the ballot that IRV uses is better than Approval  
(it collects more relevant information from the voter), and am *not*  
taking a position on whether IRV as a well-defined method (of both  
ballot and tabulation) is better than Approval.  they both have their  
different and serious problems.

>  * The approval voting ballot is inherently better or worse than a  
> rank-order ballot.

i am clear about it: the Approval ballot is inferior to the Ranked- 
Order ballot.  it does not collect sufficient information from the  
voter to allow that voter to differentiate their favorite candidate  
from others that they merely (and marginally) approve of.  that  
creates an incentive to "bullet vote" that favorite candidate so as  
not to harm that favorite with votes for less-favored candidates that  
are indistinguishably marked from the vote for their favorite.  voters  
will be more willing to mark contingency candidates if they can  
express that they *are* indeed contingency candidates.  Approval  
voting does not allow a voter to do that.  if enough voters do not  
mark contingency candidates (they bullet their fav), it devolves to  
Plurality.

> What I did say:
>   * evidence from comparing IRV & approval elections indicates  
> "bullet voting" is NOT more prevalent with approval voting.  If  
> anything it is more prevalent with IRV.

i didn't dispute that.  but i *am* questioning that bullet voting was  
a *strategy* in IRV, because, unlike with Approval, the ballot used in  
IRV allows the voter to mark their favorite, yet approve of (in a  
sense) less-favored contingency candidates without equating their  
approval of the contingency candidates to that of their favorite.

you guys keep side-stepping that obvious observation regarding  
Approval.  it's not the case with Score, but Score has it's own  
problems.

> RBJ incorrectly stated my evidence was not based on real elections.
>
> That's untrue.

i am saying that you have no certain way to compare how an Approval  
election result would have been in comparison to another election  
method that was actually used, unless
   * that other election method used the very same ballot with the  
same rules (no limit or virtually no limit, like my state senate race  
in Vermont) as Approval, or
   * you have a plausible way to map the ballot used in the other  
method to an Approval ballot.

now, about the latter, i will make two obvious observations:
   1. we *are* able to compare an IRV election to Condorcet, Borda,  
and Bucklin, if the ballot data is public, because the ballots for  
those various methods are identical and it is completely reasonable to  
assume that the voters would mark them identically in either of these  
cases.
   2. we *are* able to compare an IRV election (or any other that uses  
the ranked ballot) to a FPTP plurality election (or even a TRR) by  
mapping the ballots that *assume* (but it's a good assumption, in my  
opinion), that the "single affirmative vote" on the Plurality ballot  
is the same as the top-ranked choice on the ranked-order ballot.

now, i am still saying that you do not have a direct mapping of IRV to  
Approval nor from FPTP to Approval nor from Approval to either a  
ranked-order ballot or a single-vote ballot.  now exit polling, if  
widely used in an election, can postulate such a mapping in a  
statistical sense.  it has some value, but the conclusions are not as  
solid as when you (or i) examined the ranked-order ballots coming from  
Burlington and concluding that Condorcet winner was different from the  
IRV winner and also more preferred by voters over the IRV winner.

that is the basis of my skepticism about the confidence you have in  
some of the conclusions you make on your various pages.  you can write  
off the "bullet voting and strategy on Approval ballots" as a bugaboo,  
but it's not so in estimation of other persons.  you *cannot* write it  
off theoretically, you can try to write it off from interpretation of  
other election data (of elections where Approval was not the law), but  
i consider such confidence ill-placed.


> It is based on real IRV and real approval elections, as well as exit
> poll studies (exit poll thus-simulated approval elections), but mostly
> real.  Just read the page:
>    http://rangevoting.org/BulletBugaboo.html
>
> There may be good arguments IRV is better than approval, but this
> bullet-voting business is not one of them.

the argument i am making is really about the ranked-choice ballot  
(that IRV happens to use) in comparison to the ballot that Approval  
uses.

>
> Further, the fact IRV voters in Burlington -- or any other voters in
> any IRV or approval or any election anywhere else on the planet --
> supposedly cast bullet votes out of "spite" is irrelevant.  (Actually,
> I can think of other reasons besides the ones RBJ listed as the "only"
> ones, why they may have done so, but I simply do not care.) I do not
> give a damn why they did it.  I do not know why they did it.  RBJ may
> think he knows, but he may be right or may be wrong and has no way to
> prove it
> or even define it (what exactly is "spite"? How do you kow for sure
> Jane Voter was "spiteful"?).

because Jane said so in public testimony in front of the Burlington  
City Council and other forums during the slugfest that ended last  
March.  Jane also told me the same thing.

AGAIN, the point i am making is that there is no need with a ranked- 
order ballot (now, it's a different story depending on how those  
ballots might be tabulated) to bullet vote for favorite candidate  
strategically in order to prevent a less-favored candidate from  
harming the fav.  but there *is* a need for such a strategy with  
Approval.  unless (from election polling) that you are confident that  
a less-favored candidate will not be competing with the favorite (one  
is far ahead of the other, like Nader and Gore), then there is a  
danger that approving the less-favored candidate will harm the  
favorite.  but what about the possibility of the less-favored  
candidate losing to Satan because you (and others) didn't vote for  
her?  "oh me, oh my, oh me, oh my, what to do, what to do..."  that is  
the burden of strategy inherent to Approval voting.

>  Even if RBJ is right, he cannot even
> pretend to know why the IRV and approval voters in all the other
> elections on that page, did what they did.  Maybe they too experienced
> spite or rabies. I do not know.  He does not know.  All I know and
> care about is *DID* they do it, and how much.  I do not know or care
> why.  I am simply gathering the data and reporting the conclusions.

the conclusions say that many IRV voters bullet voted.  they do not  
say that Approval voting does not suffer a burden of strategic voting  
for a particular voter due to the consideration of bullet voting.

> For whatever reasons -- spite, generosity, whatever fantasies RBJ may
> have, I do not care -- voters appear to bullet vote more often with
> IRV than approval.  Period.

no, not Period.

if people bulleted in IRV, it was not because the ballot force them to  
equate their support for the favorite candidate to their less-favored  
candidates they might approve of.  their reasons are different.  if  
people bullet vote in an Approval election, it may very well be  
because of a strategic concern to not harm their favorite candidate by  
lifting up less-favored candidates to the same level as their fav.   
but the strategy is not without risk of voter regret.  if they (and  
others of like political leanings) bulleted their fav, the less- 
favored approved candidate may lose to someone they completely hate.   
you have not refuted that observation, Warren.

>  Therefore, criticisms of approval voting
> based on its propensity to attract bullet votes, are misguided (at
> least, unless the same critic also claims IRV is unacceptable exactly
> because of this same reason... which would be a first).

Nope.  the ballots are different.  bullet voting on a ranked-choice  
ballot need not be (and is likely not, unless the voter is ignorant)  
for the same reason that one might bullet vote in Approval.  bullet  
voting in Approval might be deemed necessary to prevent harming your  
favorite candidate.  bullet voting on a ranked-choice ballot is not  
necessary and (depending on tabulation rules) is unlikely to help your  
favorite candidate.

>   Let's not overcomplicate matters.

i agree with this sentiment.


> -----------
>
> Now, RBJ has added one more claim: he thinks rank-order ballots make
> voters provide exactly the right amount of info.  In contrast
> (continues RBJ) approval-style ballots are too little info, while
> score-voting style ballots are too much.
> I riposte that I think score voting is the right amount of info. Why?
>
> * A voter may be ignorant about candidate X vs Y, or just about X.
> IRV forces that voter, against her will,  to provide a dishonest
> opinion ("X>Y") or (if its IRV with truncation allowed) forces that
> voter to rank X dishonestly last (which is what truncation does).

only if X and Y are both preferred over Satan (who is ranked the  
lowest by not marking at all).  in a ranked-order ballot, anyone not  
ranked is tied for last place.  unfortunately, IRV doesn't do too well  
with two candidates marked with equal rank, because if that vote is  
transferred, either they have to divide the vote in half or the  
voter's vote gets unfairly amplified if the tied ranking is promoted  
to the top level when another candidate is eliminated.

the ranked-ballot inherently does not mind having equal ranks and  
Condorcet would do just fine with that.

> * But with score voting, we can easily allow a voter to leave X  
> unscored
> and thus intentionally express ignorance about X, if she so chooses.

but then X is not just neutral, X is tied with Satan for last place.   
what score should the voter assign to X or Y assuming neither is the  
favorite candidate and neither is Satan, who happens also to be in the  
race?  should they get 50? 90? 40? 10?

the ranked-order ballot does not require the voter (who is not an  
Olympic gymnastics judge) to agonize over that quantitative decision.   
all the voter needs to do is decide who they like better.  if they  
like candidate A more than any other candidate, they mark A as #1.  if  
they like B better than C, but not as much as A, they mark B as #2, C  
is #3 and Satan is left unmarked.  they are not called upon to make  
the strategic decision of quantitatively scoring these candidates that  
they might approve of a little (they like better than Satan), but are  
not their favorite.

> Also, we can allow voter to score X&Y equal.  IRV does not permit  
> either.

but the ranked ballot does in general.

>
> * A voter may have a strong opinion that Hitler and Stalin are both a
> lot worse than Gandhi and Jimmy Carter.   IRV forces that voter to  
> pretend all her
> preferences have the same strength and refuses to allow her to express
> an opinion strength.

No.  Borda does that.  If you mark

     Gandhi > Carter > Stalin > Hitler

all you are saying is that if the election was held between Gandhi and  
Carter (or Gandhi and Stalin or Hitler), you would vote for Gandhi.   
if the election was between Carter and Stalin or Hitler, you pick  
Carter.  and between Stalin and Hitler, you pick Stalin.  that's all  
that the ballot says.

now, IRV is opaque to how you feel about Stalin vs. Hitler unless the  
election comes down to it, Gandhi and Carter are eliminated.   
otherwise it has no idea.  but Condorcet does not make that mistake  
(IMO).

Borda says that the amount that you dislike Carter in comparison to  
Gandhi is the same as the amount you dislike Stalin compared to Carter  
and that you dislike Stalin "twice as much" as you dislike Carter  
(assuming your dislike of Gandhi, your fav, is set to zero).

> Let me put it to you this way.  If H or S win the election. you die.
> If C or G win, you live.  There may also be some comparatively minor
> reasons you prefer, say, S>H and C>G.   IRV refuses to allow you to
> say some of your preferences are life-vs-death and others
> comparatively minor.   Why is this
> the "right amount" of info?

it (IRV) is not the right tabulation method.  it is the right amount  
of info in a real election where, instead of Hitler, you have Michele  
Bachmann or Ben Quayle.  no one is going to line up the liberals  
against the wall and shoot them.  but one knows they are not preferred  
to other candidates that are more preferable.  it's the right amount  
in that all it asks is "who do you prefer over whom else?"  it does  
not ask the voter to evaluate quantitatively how much you prefer one  
from the other (as Score does), like ("i really hate this guy worth 75  
points.").  and it does not voter to separate all candidates into two  
groups: white sheep and black sheep as does Approval.  some sheep are  
sorta gray and there is no group to segregate the grey sheep in.  once  
you start adding a third group (like you can approve, not approve, and  
mark one candidate as your favorite), you're already on the road to  
the ranked-order ballot.

> So I don't agree rank-order ballots are "exactly the right amount"  
> of info.

they differentiate between favorite candidates to lessor "approved-of"  
candidates.  and they don't require regular ol' voters to act like  
Olympic figure-skating judges and agonize whether some candidate gets  
a 9.1 or a 9.2 for their performance.  but the voters have to decide  
who they like better- who it is they would vote for in a contingency  
if their favorite was out of the running.  and they have to decide  
that by Election Day.

>  They can force the voter to provide too much (i.e. necessarily  
> fake) info,

That is UNtrue.  it's the Score and Approval ballots that forces them  
to provide fake information.  Score requires them to cook up more  
(quantitative) information than just the preference ranking (which is  
likely to be a guess) and Approval requires them to *ditch*  
information by quantizing their relative preference to a 1-bit  
number.  that is also a falsehood.  when the voter is asked to turn  
their true preference information to *more* information (like in  
Score), they have to make it up (or bring dice to the polls) and that  
is a strategic burden.  also when the voter is asked to truncate or  
quantize their true preference information to less information (as in  
Approval), they also have to strategize.

> while refusing to allow the voter to express important honest info.

when you say, "I would vote for Candidate A over Candidate B", how  
much more "honest info" is there?  legitimate honest info?

> In addition to that, there are also theoretical indications rank-order
> ballots are a bad idea.  They inherently yield "impossibility
> theorems" and contradictions
> (e.g. "Arrow's theorem") which simply never arise in the score- 
> voting world.
> See
>   http://rangevoting.org/ArrowThm.html
>   http://rangevoting.org/CondorcetCycles.html

never said that ranked-order ballots can beat Arrow or the Condorcet  
paradox.  but, without a cycle or credible threat of one, i cannot see  
how Condorcet misses in evaluating the will of the majority and, at  
the same time, weighting each voter's vote equally (Score doesn't  
really weight a voters preferences equally).

> These foundational problems strongly suggest that the entire area of  
> rank-
> order ballots, was a mistake, a road that should never have been  
> taken.


i know (and respect) that such is the product you're selling.   
similarly to how FairVote sells IRV.

--

r b-j                  rbj at audioimagination.com

"Imagination is more important than knowledge."