Discussion:
[EM] Automatic Primary + Pairwise Runoff (APPR), a class of cloneproof top-two-style methods
Ted Stern
2017-09-28 20:46:10 UTC
Permalink
I am interested single-winner methods that find the variance-minimizing
candidate, with resistance to strategic voting.

Top two approval, STAR (top two score), and 3-2-1 voting, while all very
good at resisting strategic voting, all fail clone resistance.

When I raised the topic of top two approval on the EM list last November (
http://election-methods.electorama.narkive.com/Vlwq75Zy/em-top-two-approval-pairwise-runoff-ttapr),
it was suggested that using the ballots to approximate a two seat
Proportional Representation style "parliament" would avoid the crowding
effect of cloned candidates.

There are several problems with this idea ... to start with, in a 3 person
election, it fails the Condorcet criterion, which would be a minimal
threshold for centrist approximating methods. Another problem is that
while picking the top two approved candidates is vulnerable to crowding,
replacing the second-place winner with the second-seat parliament member
means that there is no incentive for factions to cooperate, because doing
so would lead to elimination from the second round.

After playing around with this idea for a while, I think I've come up with
a fairly straightforward modification. I'm calling APPR a *class* of
methods, since the initial candidate ranking can be based on any of several
FBC-satisfying voter alignment metrics, such as Approval, Score, or
Majority Judgment.

We can start with APPR-Approval, as in the cited thread above, since that
is the easiest place to start.

Voters use a ratings ballot that is interpreted with ranking during
tabulation. I prefer a zero through 5 score rating, with scores 5, 4, 3,
approved and 2, 1, 0, disapproved, but the actual implementation could vary
as desired.


- *Round 1*: Find the top two approved candidates, A (top score) and B
(second-highest score).
- Then drop *every* ballot that approves of A, and determine the new
approval ratings for each candidate.
- *Round 2*: The top two approved candidates among these reweighted
ballots are C (top reweighted approval) and D (second-highest reweighted
approval). NB: the reweighted approval totals can be accumulated summably
during the round 1 count.
- *Candidates A and C are the Automatic Primary winners. They are the
candidates to beat.*
- From the original, non-reweighted, ballots, determine the pairwise
votes between candidates A, B, C, and D. NB: the pairwise totals can be
accumulated summably during the round 1 count.
- To win, one of the four top-two winners from both rounds must defeat
all other Automatic Primary winners (i.e., A & C) pairwise. If more than
one candidate satisfies this property, the pairwise preferred candidate is
the winner.

As an example, assume B > A and B > C, but one of A or C defeats D. Then B
wins. If *both* B and D defeat both A and C, the pairwise winner of B vs.
D is the APPR.

If B is defeated by either A or C, and D is defeated by either A or C, the
APPR winner is the winner of A vs. C.

To win, the APPR is either the beats-all candidate among the 4, or has a
beatpath through an Automatic Primary winner.

For 3 candidate elections, this is Condorcet compliant: either B or D must
be a repeat of one of A or C. The winner is either the pairwise winner
between A and C, and one of them defeats B/D, or B/D defeats both A and C
pairwise. In case of a Condorcet cycle, one of A or C must defeat B/D, so
B/D is dropped, and the APPR winner is one of the two Automatic Primary
winners, the victor of A vs. C.

For 4 candidate elections, APPR is not strictly Condorcet, because it might
be possible for B or D to overlap with A or C as before, and the fourth
candidate left out might be preferred pairwise to the other three. But
this is extremely unlikely except in highly fragmented elections with
extremely low winning approvals. When A through D are all distinct
candidates, APPR is Condorcet compliant.

Properties:

- The automatic primary avoids the both the splitting and the
clone/crowding problem, since the second round winner is chosen from only
those ballots that do *not* approve of A. So the second round is
clearly from a different set of voters than those who would be crowded
around A. Therefore, there is no advantage to be gained from crowding, but
no disadvantage either.
- Pushover is avoided because the automatic primary is based solely on
the highest approval winner, and it is not possible to engineer second
round top two placement for your favorite by approving one's weakest
opponent.
- By including the second-place approved candidates for each round, we
avoid the problem of eliminating the best representatives of strongly
aligned factions. Consider a 2016-type situation: Clinton wins round 1,
but after eliminating all Clinton-approving ballots, Trump wins round 2.
This is not a great choice for voters. By including the runners-up, we get
to choose the most preferred of the candidates in each faction who defeat
both Clinton and Trump pairwise. That is, if the Greens and Independents
partially aligned with Democrats, they are not penalized for that
alignment, and may in fact be rewarded for cooperation.
- Including the second-place candidates in each round adds a bit of the
flavor of 3-2-1 voting --- more than 2 factions can thus be considered.
- Chicken-dilemma problems can be addressed via having rankings below
the approval cutoff (see thread cited above).
- Within each round, Favorite Betrayal Criterion (FBC) is satisfied
through use of an FBC-compliant ratings method.
- While not Condorcet compliant for 4 or more candidates, APPR tends to
find the most preferred representatives of the two most preferred disparate
factional groups, and therefore should find the variance-minimizing
candidate most of the time. I will be doing Yee-metric tests on APPR to
see just how well it performs in this respect.

As described above, the particular method is *APPR-Approval*. But the APPR
process could also be implemented with either Score or Majority Judgment in
each round.

After thinking about this for a while, I have come to prefer *APPR-Score*
due to its combination of expressiveness and its natural summable extension
to the Automatic Primary part of the process. I think that APPR-Score is
the simplest way and most natural extension of STAR voting, without losing
too much of STAR's simplicity. Score based on total scores, instead of
averages, also satisfies Participation and Immunity from irrelevant
alternatives, in each round.

I've described the Automatic Primary for score voting in other posts, but
for clarity, I'm adding again here. Assume a ratings ballot with range 0
to 5.

- Accumulate total scores (not averaged) for each candidate, counting
blanks as zero scores.
- Round 1: Find the score winner and runner up, A and B.
- For each ballot that scores A above 0, accumulate scores of 5 minus
the ballot's A-score times the non-A score, for every other candidate on
the ballot. So, for example, if the ballot scores A at 3, and candidate X
at 4, accumulate (5-3) * 4 = 8 points for X, and similarly for all other
non-zero scored candidates on the ballot. Computationally, this preserves
exact integer arithmetic in the totals. These totals are the Round 2
scores. They can be converted into averages for reporting, if desired, by
dividing by the maximum score squared and the total number of ballots.
- C and D are determined from the Round 2 totals.

*APPR-MJ* is similar to APPR-Approval in terms of dropping A-approving
ballots to find the round 2 scores, but Majority Judgment is used in each
round. In the second round, the 50% level is determined by the number of
remaining ballots instead of the original number of ballots. There are
some attractive aspects to this method, but they come at the cost of more
complexity and unpractical summability. Nevertheless, I would happily use
this method if summabilty were not desirable.
Ted Stern
2017-09-28 21:25:47 UTC
Permalink
As an example of APPR-Approval, consider the following ballots:

98: Abby > Cora > Erin >> Dave > Brad
64: Brad > Abby > Erin >> Cora > Dave
12: Brad > Abby > Erin >> Dave > Cora
98: Brad > Erin > Abby >> Cora > Dave
13: Brad > Erin > Abby >> Dave > Cora
125: Brad > Erin >> Dave > Abby > Cora
124: Cora > Abby > Erin >> Dave > Brad
76: Cora > Erin > Abby >> Dave > Brad
21: Dave > Abby >> Brad > Erin > Cora
30: Dave >> Brad > Abby > Erin > Cora
98: Dave > Brad > Erin >> Cora > Abby
139: Dave > Cora > Abby >> Brad > Erin
23: Dave > Cora >> Brad > Abby > Erin

(modified from an example due to Rob LeGrand that is cited here:
http://wiki.electorama.com/wiki/Definite_Majority_Choice)

In this case, the Round 1 winners are Erin and Abby, and the Round 2
winners are Dave and Abby. The Automatic Primary winners are Erin and Dave.

Though Abby is never a round 1 or round 2 approval winner, she pairwise
defeats both automatic primary winners Erin and Dave, and is therefore the
APPR-Approval winner. It can also be seen that Abby, though not the
highest approved candidate, is nevertheless rated highly by Erin and
non-Erin voters alike, and could be seen as the best compromise candidate
by both factions. This should yield a high degree of satisfaction with the
results.

Compare with http://wiki.electorama.com/wiki/Marginal_Ranked_Approval_Voting,
which also finds the same result.
Post by Ted Stern
I am interested single-winner methods that find the variance-minimizing
candidate, with resistance to strategic voting.
Top two approval, STAR (top two score), and 3-2-1 voting, while all very
good at resisting strategic voting, all fail clone resistance.
When I raised the topic of top two approval on the EM list last November (
http://election-methods.electorama.narkive.com/
Vlwq75Zy/em-top-two-approval-pairwise-runoff-ttapr), it was suggested
that using the ballots to approximate a two seat Proportional
Representation style "parliament" would avoid the crowding effect of cloned
candidates.
There are several problems with this idea ... to start with, in a 3 person
election, it fails the Condorcet criterion, which would be a minimal
threshold for centrist approximating methods. Another problem is that
while picking the top two approved candidates is vulnerable to crowding,
replacing the second-place winner with the second-seat parliament member
means that there is no incentive for factions to cooperate, because doing
so would lead to elimination from the second round.
After playing around with this idea for a while, I think I've come up with
a fairly straightforward modification. I'm calling APPR a *class* of
methods, since the initial candidate ranking can be based on any of several
FBC-satisfying voter alignment metrics, such as Approval, Score, or
Majority Judgment.
We can start with APPR-Approval, as in the cited thread above, since that
is the easiest place to start.
Voters use a ratings ballot that is interpreted with ranking during
tabulation. I prefer a zero through 5 score rating, with scores 5, 4, 3,
approved and 2, 1, 0, disapproved, but the actual implementation could vary
as desired.
- *Round 1*: Find the top two approved candidates, A (top score) and B
(second-highest score).
- Then drop *every* ballot that approves of A, and determine the new
approval ratings for each candidate.
- *Round 2*: The top two approved candidates among these reweighted
ballots are C (top reweighted approval) and D (second-highest reweighted
approval). NB: the reweighted approval totals can be accumulated summably
during the round 1 count.
- *Candidates A and C are the Automatic Primary winners. They are the
candidates to beat.*
- From the original, non-reweighted, ballots, determine the pairwise
votes between candidates A, B, C, and D. NB: the pairwise totals can be
accumulated summably during the round 1 count.
- To win, one of the four top-two winners from both rounds must defeat
all other Automatic Primary winners (i.e., A & C) pairwise. If more than
one candidate satisfies this property, the pairwise preferred candidate is
the winner.
As an example, assume B > A and B > C, but one of A or C defeats D. Then
B wins. If *both* B and D defeat both A and C, the pairwise winner of B
vs. D is the APPR.
If B is defeated by either A or C, and D is defeated by either A or C, the
APPR winner is the winner of A vs. C.
To win, the APPR is either the beats-all candidate among the 4, or has a
beatpath through an Automatic Primary winner.
For 3 candidate elections, this is Condorcet compliant: either B or D
must be a repeat of one of A or C. The winner is either the pairwise
winner between A and C, and one of them defeats B/D, or B/D defeats both A
and C pairwise. In case of a Condorcet cycle, one of A or C must defeat
B/D, so B/D is dropped, and the APPR winner is one of the two Automatic
Primary winners, the victor of A vs. C.
For 4 candidate elections, APPR is not strictly Condorcet, because it
might be possible for B or D to overlap with A or C as before, and the
fourth candidate left out might be preferred pairwise to the other three.
But this is extremely unlikely except in highly fragmented elections with
extremely low winning approvals. When A through D are all distinct
candidates, APPR is Condorcet compliant.
- The automatic primary avoids the both the splitting and the
clone/crowding problem, since the second round winner is chosen from only
those ballots that do *not* approve of A. So the second round is
clearly from a different set of voters than those who would be crowded
around A. Therefore, there is no advantage to be gained from crowding, but
no disadvantage either.
- Pushover is avoided because the automatic primary is based solely on
the highest approval winner, and it is not possible to engineer second
round top two placement for your favorite by approving one's weakest
opponent.
- By including the second-place approved candidates for each round, we
avoid the problem of eliminating the best representatives of strongly
aligned factions. Consider a 2016-type situation: Clinton wins round 1,
but after eliminating all Clinton-approving ballots, Trump wins round 2.
This is not a great choice for voters. By including the runners-up, we get
to choose the most preferred of the candidates in each faction who defeat
both Clinton and Trump pairwise. That is, if the Greens and Independents
partially aligned with Democrats, they are not penalized for that
alignment, and may in fact be rewarded for cooperation.
- Including the second-place candidates in each round adds a bit of
the flavor of 3-2-1 voting --- more than 2 factions can thus be considered.
- Chicken-dilemma problems can be addressed via having rankings below
the approval cutoff (see thread cited above).
- Within each round, Favorite Betrayal Criterion (FBC) is satisfied
through use of an FBC-compliant ratings method.
- While not Condorcet compliant for 4 or more candidates, APPR tends
to find the most preferred representatives of the two most preferred
disparate factional groups, and therefore should find the
variance-minimizing candidate most of the time. I will be doing Yee-metric
tests on APPR to see just how well it performs in this respect.
As described above, the particular method is *APPR-Approval*. But the
APPR process could also be implemented with either Score or Majority
Judgment in each round.
After thinking about this for a while, I have come to prefer *APPR-Score*
due to its combination of expressiveness and its natural summable extension
to the Automatic Primary part of the process. I think that APPR-Score is
the simplest way and most natural extension of STAR voting, without losing
too much of STAR's simplicity. Score based on total scores, instead of
averages, also satisfies Participation and Immunity from irrelevant
alternatives, in each round.
I've described the Automatic Primary for score voting in other posts, but
for clarity, I'm adding again here. Assume a ratings ballot with range 0
to 5.
- Accumulate total scores (not averaged) for each candidate, counting
blanks as zero scores.
- Round 1: Find the score winner and runner up, A and B.
- For each ballot that scores A above 0, accumulate scores of 5 minus
the ballot's A-score times the non-A score, for every other candidate on
the ballot. So, for example, if the ballot scores A at 3, and candidate X
at 4, accumulate (5-3) * 4 = 8 points for X, and similarly for all other
non-zero scored candidates on the ballot. Computationally, this preserves
exact integer arithmetic in the totals. These totals are the Round 2
scores. They can be converted into averages for reporting, if desired, by
dividing by the maximum score squared and the total number of ballots.
- C and D are determined from the Round 2 totals.
*APPR-MJ* is similar to APPR-Approval in terms of dropping A-approving
ballots to find the round 2 scores, but Majority Judgment is used in each
round. In the second round, the 50% level is determined by the number of
remaining ballots instead of the original number of ballots. There are
some attractive aspects to this method, but they come at the cost of more
complexity and unpractical summability. Nevertheless, I would happily use
this method if summabilty were not desirable.
Ted Stern
2017-10-12 18:36:05 UTC
Permalink
Toby, you bring up a valid point. There are indeed pathological cases in
which the Condorcet winner does not have broad support.

My intention in introducing Automatic Primary as an adjunct method was to
address the main problem possessed by top two or top N methods (such as
SRV/STAR or 3-2-1), that the results are distorted by crowding.

For example, in 3-2-1, there might be a strong minority faction united
behind a crowded slate of candidates, while the majority faction has a
contentious chicken dilemma, effectively splitting its first place votes
beyond the top three. The same could be true in top two methods using
approval, score or majority judgment.

I was hoping that some form of Condorcet compliance would be a side benefit
of my earlier proposed automatic primary technique, but you have pointed
out a situation in which that compliance could actually lead to a result
with lower social utility or higher variance.

So I have thought of a further simplification that focuses solely on the
crowding/splitting problem:

For any top two or top N method, run two different versions: one with the
original method, and another with automatic primary. Each method will have
a pairwise winner. If the original top N method winner differs from the
winner of the automatic primary pairwise runoff, the overall APPR winner is
the pairwise preferred between those two candidates.

Automatic primary means: for a particular ratings method, find the top
scoring candidate Then deweight ballots according to the degree that each
ballot supports the ratings winner, and calculate the total rating winner
using the remaining ballots or fraction of a ballot. If using a top-3
method, apply another round of deweighting on those remaining ballots and
find the next ratings winner.

In top two approval:

- Use a rating ballot with an approval cutoff, e.g. 5,4,3 = approved;
2,1,0 = disapproved; infer rankings from ratings. Equal rating implies no
pairwise preference vote.
- Find candidates A0 (highest approved) and A1 (approval runner-up).
- Drop all ballots that approve of A0
- Find the candidate with highest approval among ballots that don't
approve of A0. Call this candidate B.
- Do pairwise comparisons A0 vs. A1 (top two) and A0 vs. B (auto
primary). If the top two pairwise winner is different from the auto
primary pairwise winner, the APPR winner is the most preferred the two.

In top two score [Score Runoff Voting (SRV) AKA Score + Automatic Runoff
(STAR)]:

- Use a ratings ballot, infer rankings from ratings. Equal rating
implies no pairwise preference vote.
- Find candidates A0 (highest total score) and A1 (runner-up total
score).
- For each ballot that gives A0 a score less than max-score, give each
candidate their original score times (max_score - A0_score). With scaling,
this is the equivalent of removing the total score fraction for A0 from the
original ballots.
- The automatic primary runner-up is the candidate with highest total
score among these remaining candidates. Call this candidate B.
- Proceed as before for APPR-Approval

One can proceed similarly for APPR-Majority Judgment.

It is interesting that the same idea APPR idea could also be applied to
3-2-1:

- Use a 3-2-1 3-slot ballot: Good, OK, Reject.
- Find the total votes for each score level for each candidate.
- Find the top three candidates by total Good rating, A0, A1, A2.
- Drop the candidate with the most Reject scores.
- The 3-2-1 winner is the pairwise preferred between the remaining 2
candidates. Call that winner A*.
- Find the automatic primary runner-up and second runner-up by first
counting Good ratings on only those ballots that Reject A0. The Good vote
winner on those ballots is B1.
- Then count Good ratings on only those ballots that reject both A0 and
B1. The Good vote winner on those ballots is B2.
- Apply 3-2-1 on A0, B1, and B2. Call that winner B*.
- If A* is not the same candidate as B*, the APPR-3-2-1 winner is the
the pairwise preferred between A* and B*.

As with top-N methods, APPR applies another level of stabilization to the
original ratings method.


On Tue, Oct 10, 2017 at 5:36 AM, 'Toby Pereira' via The Center for Election
I'm not sure that passing the Condorcet criterion is a primary aim of
these methods, otherwise we might as well propose using one of the many
Condorcet methods on offer. I think the main aim of methods like STAR and
3-2-1 voting is to maximise utility, without so much concern about strictly
passing particular criteria in all cases.
In any case, there are situations where I think passing the Condorcet
criterion is not a good thing. The following example has 3 candidates, 100
voters and score ballots with a maximum score of 10.
49 voters: A=10, B=0, C=1
49 voters: A=0, B=10, C=1
2 voters: A=1, B=0, C=10
C is the Condorcet winner, but a very weak winner. There are two main
candidates (A and B) that polarise support, and a non-entity (C). C wins
because the 98 A/B supporters decide to put C (marginally) above the main
candidate they dislike. They might not even know anything about C.
This is a fairly extreme example, but it shows what can happen when you
have two main candidates and a non-entity.
I think using score voting and then finding the top two candidates
proportionally is a good idea. I would do this sequentially though, since
we are ultimately looking for a single winner, not a two-person committee,
and it would be strange to eliminate the top scoring candidate, which is
what could happen with non-sequential election.
This method would still not be cloneproof as it stands, but I would
solve this not by trying to eliminate the effect of clones, but by
embracing clones. I would effectively clone each candidate. If the top two
candidates in the sequential proportional election are a candidate and
their clone, then that candidate would automatically be elected without the
need for a top two head-to-head. If a candidate has that much of a lead
over the others in a score ballot, then I don't think the head-to-head
would be necessary, or even desirable. It would help eliminate the
possibility of very weak winners, as could happen in my example above under
some methods.
Toby
Post by Ted Stern
I am interested single-winner methods that find the variance-minimizing
candidate, with resistance to strategic voting.
Top two approval, STAR (top two score), and 3-2-1 voting, while all very
good at resisting strategic voting, all fail clone resistance.
When I raised the topic of top two approval on the EM list last November (
http://election-methods.electorama.narkive.com/Vlwq75Zy/em-
top-two-approval-pairwise-runoff-ttapr), it was suggested that using the
ballots to approximate a two seat Proportional Representation style
"parliament" would avoid the crowding effect of cloned candidates.
There are several problems with this idea ... to start with, in a 3
person election, it fails the Condorcet criterion, which would be a minimal
threshold for centrist approximating methods. Another problem is that
while picking the top two approved candidates is vulnerable to crowding,
replacing the second-place winner with the second-seat parliament member
means that there is no incentive for factions to cooperate, because doing
so would lead to elimination from the second round.
--
You received this message because you are subscribed to the Google Groups
"The Center for Election Science" group.
To unsubscribe from this group and stop receiving emails from it, send an
For more options, visit https://groups.google.com/d/optout.
Loading...