Strategic Voting: Information Aggregation in Elections

(LSE) EC319 03/14 1 / 20

Strategic Voting: Information Aggregation in Elections

Why do we have jurors? Why do we have Democracy?

One explanation: Information aggregation.
Condorcet’s model ) when voters are non-strategic the more voters
the more acurate the decision.
Will strategic voting "ruin" information aggregation?

(LSE) EC319 03/14 2 / 20

Condorcet’s model

Condorcet’s model
n voters.
Two states of the world, L, R occur with equal probability.
Each voter gets a signal, l or r , correct with probability q > 12 .
Voters vote “sincerely”.

(LSE) EC319 03/14 3 / 20

Condorcet’s model

Result: As n goes large, the probability that the electorate takes the
correct action converges to one.
Proof: Law of Large numbers.
But what if voters are strategic?

(LSE) EC319 03/14 4 / 20

Strategic voting with information: Example.

Suppose we have 5 voters.

Suppose that for whatever reason two of them always vote L.
Is it an equilibrium that the other three voters vote sincerely?

(LSE) EC319 03/14 5 / 20

Strategic voting with information: Example.

The example has shown that sincere voting is often not an

equilibrium? Why?
Result (Feddersen and Pesendorfer): When voters are strategic, as the
electorate grows large, the electorate chooses the correct outcome
with probability one.

(LSE) EC319 03/14 6 / 20

Strategic Voting: Unanimity rule in Juries?

Why do we use unanamity rule?

We worry about convicting the innocent (intuitive).
Does this intution survive when voters are strategic?
It increases communcation (Deliberative democracy).
Information aggregation...

(LSE) EC319 03/14 7 / 20

Strategic Voting: Information Aggregation in Elections

Feddersen and Pesendorfer (American Political Science Review 1998)

Check whether the above two are correct,
Does unanimity minimize the probability of convincting the innocent?
Does unanimity induce e¢ cient aggragation of information?

(LSE) EC319 03/14 8 / 20

Feddersen and Pesendorfer (1998)

Example: A defendant faces 3 judges. Apriory there is a 50% chance

he commited the crime (G) and 50% he did not (I). Every judge hears
evidence and gets an impression of the defendant (the judges reach
an evaluation independently). The impression can be either “guilty”
(i) or “innocent” (n). The impression is correlated with the defendant
state: there is probability p > 12 that the impression is correct. Given
that the defendants true status (G or I), the impressions are
independent. In order to convict, a unanimous vote is needed. The
utility for each judge from a correct decision is 1; from a wrong
decision 0.

(LSE) EC319 03/14 9 / 20

Feddersen and Pesendorfer (1998)
The example as a Bayesian game

3 players – the judges, N = f1, 2, 3g.

2 actions: Ai = fC , Ag (Convict, Acquit)
States of the world: 16 – The defendant’s state fG , I g, and judges’
impressions fgj , ij g.

1 3
Pr(G \ (g , g , g )) = Pr(I \ (i, i, i )) = p
2
1
Pr(G \ (i, g , g )) = Pr(I \ (g , i, i )) = p 2 (1 p )
2
1
Pr(G \ (i, i, g )) = Pr(I \ (g , g , i )) = p (1 p )2
2
1
Pr(G \ (i, i, i )) = Pr(I \ (g , g , g )) = (1 p )3
2
(The judges’identity is not important)

(LSE) EC319 03/14 10 / 20

Feddersen and Pesendorfer (1998)
The example as a Bayesian game

Information function: Each judge knows her own impression but not
that of the others’.
Utilities:

Correct decision : uj ((C , C , C ), G ) = uj (Atleast one A, I ) = 1

Wrong decision : uj ((C , C , C ), I ) = uj (Atleast one A, G ) = 0

Note: We can also model this using types (the other defenition of
Bayesian Game).
For that we would need four players, the three judges and the
defendant which has two types, but is not an active player.

(LSE) EC319 03/14 11 / 20

Feddersen and Pesendorfer (1998)
Voting according to impression is not Bayesian eq.

Suppose that every judge votes sincerely: i.e., votes to Convict if and
only if his impression is g .
We now show that this is not an equilibrium.
Proof:
Every judge knows he is only pivotal when both his peers have decided
to convict.
Since they vote according to impression (in eq.), this means that she
would a¤ect the result only when both got the signal/impression g .

(LSE) EC319 03/14 12 / 20

Feddersen and Pesendorfer (1998)
Voting according to impression is not Bayesian eq.

Suppose this judge got the signal i. In the case in which he a¤ects the
outcome there are two g signals and one i.
Thus (since p > 12 ), it is more likely that the defendant is guilty:

Pr(G \ (g , g , i ))
Pr(G j(g , g , i )) =
Pr((g , g , i ))
Pr(G \ (g , g , i ))
=
Pr(G \ (g , g , i )) + Pr(I \ (g , g , i ))
1 2
= 2 p (1 p)
=p>
1
1 p 2 (1 p) + 1 p (1 p )2 2
2 2

Therefore he should vote C , a contradiction to his following his signal!

(LSE) EC319 03/14 13 / 20

Feddersen and Pesendorfer (1998)
Simple majority: voting according to impression is Bayesian eq.

Assume that now we use simple majority in the jury p[rocess.

We will show that voting sincerely is an equilibrium.
Proof:
This time a judge in‡uences the result when there is exactly one judge
who decide to convict (got signal g), and one that decided to acquit
(got the signal i).
Hence, if our judge got the signal g, it is more likely (with probability
p > 12 ) that the defendant is guilty, and if he got the signal i, that the
defendant is innocent.

(LSE) EC319 03/14 14 / 20

Feddersen and Pesendorfer (1998)
Finding Bayesian eq. in a more general setting (like in the paper)

There are n jurors.

The payo¤ in the case of a right decision is 0,
If an innocent man is found guilty the payo¤ is q,
If a guilty man walks free (1 q ).
This means that q is the threshold level of the jury’s “reasonable
doubt”.
They would want to convict if they believe that the defendant is guilty
with probability q 0 q:

Convict = q 0 0 + (1 q 0 )( q )
q 0 + (1 q )( q ) = q ( (1 q )) + (1 q) 0
q 0 ( (1 q )) + (1 q 0 ) 0 = Acquit

(LSE) EC319 03/14 15 / 20

Feddersen and Pesendorfer (1998)
Unanimous conviction rule

Assume that p, n, q are such that if at least n 1 jurors got the signal
g they want to convict, and if they got i they want to acquit.
Find a Bayesian eq. and check the probability of conviction as a
function of n.
Is the prob. going to 0 when the number of jurors goes to in…nity?
(Observe that under naïve voting, the prob. does indeed go to 0)

(LSE) EC319 03/14 16 / 20

Feddersen and Pesendorfer (1998)
How to …nd such equilibria?

First, we can see that there are some trivial eq. For example – all are
voting to acquit, regardless of their signal. In this case no one will
impact the result if she deviates, and that makes this an eq.
We look for a symmetric eq., in which the strategy (voting as a
function of the signal) is the same for all players.
We’ve seen that telling the truth (or lying) is not an eq.
) no pure eq.

(LSE) EC319 03/14 17 / 20

Feddersen and Pesendorfer (1998)
How to …nd such equilibria?

Look for mixed strategies eq.

A symmetric eq. is de…ned by two parameters: pg & pi – Prob. of
conviction if received the signal g /i.
Easy to see that pi > 0.
Otherwise, when someone is pivotal, he knows exactly that all the
other have received the signal g , and thus needs to convict no matter
what is her own signal.
pg = 1.
Since jurors who got the signal i are indi¤erent or want to convict,
then whoever has signal g really wants to convict.

(LSE) EC319 03/14 18 / 20

Feddersen and Pesendorfer (1998)
How to …nd such equilibria?

All that is left is to …nd pi .

Need to be such that whoever gets the signal i will be indi¤erent and
will agree to mix.
Meaning that when the other n 1 jurors have voted C , and I have the
signal i, then the probability I gives to the defendant being guilty is
exactly q:
(Denote with A the state in which I observed i and the rest have voted
C)
1
Pr[G ] 2
Pr[G jA] = Pr[AjG ] = Pr[AjG ] 1 1
Pr[A] 2 Pr [AjG ] + 2 Pr [AjG ]
(1 p )(p 1 + (1 p )pi )n 1
=
(1 p )(p 1 + (1 p )pi )n 1 + p ((1 p ) 1 + ppi )n 1
= q

We can …nd pi , and see that it converges to 1 when n goes to in…nity.

We can
(LSE) also calculate the various
EC319errors. Do they converge to 0? 19 / 20
03/14
 1
Pr[G ] 2
Pr[G jA] = Pr[AjG ] = Pr[AjG ] 1 1
Pr[A] 2 Pr [AjG ] + 2 Pr [AjG ]
(1 p )(p 1 + (1 p )pi )n 1
=
(1 p )(p 1 + (1 p )pi )n 1 + p ((1 p ) 1 + ppi )n 1
= q

We can …nd pi , and see that it converges to 1 when n goes to in…nity.

We can also calculate the various errors. Do they converge to 0?

(LSE) EC319 03/14 20 / 21

Feddersen and Pesendorfer (1998)
Other voting rules?

Feddersen and Pesendorfer check non-unanimous conviction rules.

Suppose that in order to convict, 0 < β < 1 percentage of the jurors
need to convict.
For n large enough there is an eq. in which the vote is a function of
the signal and when n goes to in…nity, to probability of an error goes
to zero.
Meaning – the unanimous vote is inferior to any role of a …xed
proportion!

(LSE) EC319 03/14 20 / 20

American Political Science Review Vol. 92, No. 1

FIGURE 1. The Probability an Innocent Defendant Is Convicted as a Function of Jury Size

0.25 -

0.2 -j(p=.7,q=.5,n)

a 0.15-

IL 0.1 -

0.05 -

N I ( cc 0 N q (0 cc 0 (,4 N q (0 co
o 0 (C4 q (0o 00 0 (.4 (0 00
N N Jury (
n) ) V V V IV I Iq I IV

Jury Size (n)

* /(1-~~~~q)
(1 -p) P/(2p - ) case of informative voting, in which the probability of
lrm l(p, q, n) = ( q -p ) conviction converges to zero as n -> 0, independent of
whether the defendant is guilty or innocent. Thus, a
1 -PI(2p -
second implication of strategic voting is that the prob-
* /(1-~~~~~q)
1 p) ability of a guilty verdict may be much larger than
lrM IG(P, q, n) = 1 - q p
under informative voting.
To provide an intuition for proposition 2, first ob-
If q ' 1 - p, then there is no responsiveequilibrium.In serve that equation 3 implies v(i) -> 1 as n - oo. As
this case u(i) = 1 is an equilibrium,and 1I(p, q, n) = a consequence, YG (the probability that a juror votes to
1, 1G(P, q, n) = 0. convict if the defendant is guilty) and Yi (the probabil-
ity that a juror votes to convict if the defendant is
Proof. For q > 1 - p we demonstrated in the text innocent) both converge to one. This is not enough to
above that the unique responsive voting equilibrium show that the probability of convicting an innocent
under the unanimity rule is given by equation 3. In defendant, (y)f)n, stays bounded away from zero. In
Appendix A we show that Appendix A we demonstrate that for large n, Yi can be
lim lI(p, q, n) approximated by
n-o

-
==lim
( (2 - 1)( - q)1 ~P))11(n )n 1+ n - 1 2p -1 lnf),
(2
() qp ) \
- 1)
where f = (1 - q)(1 - p)/qp, and hence (y)f)n
-q) (1-p ) )11(n
which is the bound given in
(o
n---V( converges to fP/(2P-1),
p -(1 p) - q)(1p proposition 2.12
The convergence to the bounds given in proposition
qpp 2 is fast, and hence the limit formula allows us to
q)P ) approximate the probabilities of each kind of error
even for small juries. Figure 1 illustrates the conver-
The proof that lim O 1G(P, q, n) = 1 - ((1 - gence of 1I(p, q, n) for the valuesp = 0.7, q = 0.5.
q)(1 - p)lqp)(1-P)F(2P-1) is analogous. The figure is startling for several reasons. First, the
If q ? 1 - p, then the argument given in the text limit probability of convicting an innocent defendant is
shows that there is no responsive equilibrium. Q.E.D. quite large-22%. Second, when there are only 12
jurors the probability of convicting an innocent is 21%.
Proposition 2 also implies that the probability of a
guilty defendant being convicted (1 - IG) is bounded
away from zero for all n. This is again in contrast to the 12 Recall that ex = limn?,(1 + x/n)n.

27

This content downloaded from 158.143.42.19 on Mon, 10 Mar 2014 05:12:20 AM

All use subject to JSTOR Terms and Conditions
The Inferiority of Unanimous Jury Verdicts under Strategic Voting March 1998

FIGURE2. The Probability a Guilty Defendant Is Acquitted as a Function of Jury Size

0.52-

0.51

0.5

0.49-

0.48

0.46-

0.45
14 IV (0 co 0 14 IV (0 co 0 (14 IV (0 00 0 (-4 q (0 c0 0 (-4 (000

JurySize (n)

There is a difference of only 1% between the probabil-

ity of convicting an innocent defendant with a jury size
of n ?-12 and the limit probability. Third, the proba-
bility of convicting an innocent defendant increases
with the size of the jury.
Figure 2 shows the probability of acquitting a guilty
defendant for the valuesp = 0.7, q = 0.5. The limit
probability of acquitting a guilty defendant is 47%.
Once again, the limit probability is a very good esti-
mate of the actual probability of this type of mistake,
even for small juries. Note that the probability of
acquitting a guilty defendant actually decreaseswith the
size of the jury.
Figure 3 shows the limit errors 1I(p, q) = limn,,
1I(p, q, n) and 1G(P, q) = limn,- 1G(P, q, n) for the
value p = 0.7 as a function of q. The figure demon-
strates that for large juries the probability of convicting
an innocent decreases in q, while the probability of
acquitting a guilty defendant increases in q. Thus, the
unanimity rule does a poor job of protecting innocent
defendants from unreasonable juries. Yet, if the jury is
responsible (e.g., q = .9), the innocent defendant is
protected at the cost of a high probability of acquitting
the guilty. In the next section we show that the una-
nimity rule is a uniquely bad voting rule in terms of the
probabilities of both kinds of error it induces.
Using the equilibrium strategies calculated in this
section, we may also compute the probability that a
convicted defendant is innocent and compare it to the
bound given in proposition 1. A straightforward calcu-
lim Pr(GIC)= -q >
n-->oo 2p
1
-1> 2p -l'
1 +q
where the right-hand side of the above inequality is the
bound given in proposition 1.
NONUNANIMOUSRULES
We now analyze the probability of making the wrong
decision under nonunanimous voting rules.13Suppose
k = an for some axwith 0 < ax< 1, and assume that an
is an integer. Thus, a defendant will be convicted if and
only if at least an a-fraction of the jury votes to convict.
For a fixed ax, consider a sequence of symmetric
responsive equilibria corresponding to an increasing
jury size. In the following proposition we show that as
n -> 0 the probability of making either of the two
kinds of error converges to zero.
PROPOSITION3. Fix any oawith 0 < oa < 1. (1) There is
a n' such that for n > n' there is a symmetric
responsive equilibrium. (2) For any sequence of sym-
metric responsive equilibria the probability of convict-
ing an innocent defendant and the probability of
acquitting a guilty defendant both converge to zero.
Proof. See Appendix B. Q.E.D.
13 See Appendix B for a computation of the (unique) symmetric

lation demonstrates that responsive equilibrium for general k.

28

This content downloaded from 158.143.42.19 on Mon, 10 Mar 2014 05:12:20 AM

All use subject to JSTOR Terms and Conditions
AmericanPoliticalScience Review Vol. 92, No. 1

FIGURE 3. Limit Error Probabilities as a Function of the Threshold of Reasonable Doubt

1.2 -

\ /~~~~~~~~~~~~~~G(p
z71q)
0.8-

2.0
06-

0.4-

0.2-

I1(p=.7,q)
0
C M (0 0 N 1 00 -) (0 (0 OW (N 0L CC ) ( 0)

Oodd . . . . ' ? o o dddddd O O O Oo ddd O

Threshold of Reasonable Doubt (q)

Proposition 3 shows that for any (x E (0, 1) the (,yI)a*(1 - y)l-a*

probability of convicting an innocent defendant and

(Y G)a*(1 G)Ia*
the probability of acquitting a guilty defendant both -

converge to zero as the jury grows large.14 This is in
sharp contrast to the results of propositions 1 and 2,
which showed that both types of mistakes stay bounded
away from zero for the unanimity rule.
To provide an intuition for proposition 3, recall that
YG is the probability a juror votes to convict if the
defendant is guilty and yI is the corresponding proba-
bility if the defendant is innocent. As we noted above,
in any responsive profile it must be the case that 1 >
YG > yi > 0. Suppose the actual fraction of guilty
votes is a. Then probability that the defendant is guilty
is given by
(YG)an(l - a)n
YG)
(YG) a(l - YG)(1a)n + (,y)an(l
It follows that if
(,YI)a(l - y&)la
1-
('YG)a(l - YG)
then for large n the defendant is either guilty with
probability close to one (if the above fraction is less
than one) or innocent with probability close to one (if
the above fraction is greater than one).
For any responsive profile there is a unique a *, such
that
14 Proposition 3 holds in much more general environments. Fed-
dersen and Pesendorfer (1997a) prove the analogous result for an
environment that includes preference diversity and a much broader
range of information environments. Myerson (n.d.) proves a similar
result for the case of simple majority rule.
and YG > a* > yI.15 If the actual fraction of guilty
votes is a < a *, then for large n the defendant is
innocent with probability close to one; if a > a *, then
the defendant is guilty with probability close to one.
In any responsive equilibrium it must be the case
that the event a vote is pivotal, that is, an a fraction
votes to convict, is not overwhelming evidence of either
guilt or innocence.16 This in turn implies that the a*
implied by a responsive equilibrium must be arbitrarily
close to Oxif n is sufficiently large.
Now recall that a responsive equilibrium may take
two possible forms. One is that jurors vote to acquit
when they observe signal i and randomize when they
observe signalg (i.e., v(i) = 0 and 0 ? v(g) ? 1). The
- y )(l-a)n other is that jurors vote to convict when they observe
signal g and randomize when they observe signal i (i.e.,
0 ? v(i) c 1 and v(g) = 1). Figure 4 depicts yI, YG,
and a * as a function of the strategy profile forp = 0.7.
It is convenient to represent the strategy by the variable
1 x E [0, 2]. Forx ? 1 the strategy is v(i) = 0, v(g) -
x; forx -1 the strategy is v(g) = 1, v(i) = x - 1.
Figure 4 allows us to find the unique symmetric
responsive equilibrium for large juries as a function of
the voting rule at. Suppose, for example, that at is as
indicated in the figure. Since a * in a large jury is close
to a, it must be that the equilibrium strategy profile is
15 It is easy to see that (-y)a*(1 - - 'YG)1-a =
y,)l-a /(YG)a*(1
1 impliesa* = (ln(1 - 'YG) - ln(1 - y1))/(lny, - In'YG + ln(1 -
YG) - ln(1 -y)).
16 Formally, it must be the case in any responsive equilibrium that
Pr(Gpiv, g) > q > Pr(Gtpiv, i).

29

This content downloaded from 158.143.42.19 on Mon, 10 Mar 2014 05:12:20 AM

All use subject to JSTOR Terms and Conditions