American Politics Research

Volume 37 Number 4
July 2009 700-724
© 2009 SAGE Publications

A Bayesian Prediction 10.1177/1532673X08330670

http://apr.sagepub.com
hosted at
Model for the U.S. http://online.sagepub.com

Presidential Election
Steven E. Rigdon
Southern Illinois University, Edwardsville
Sheldon H. Jacobson
Wendy K. Tam Cho
University of Illinois, Urbana-Champaign
Edward C. Sewell
Southern Illinois University, Edwardsville
Christopher J. Rigdon
Arizona State University, Tempe

It has become a popular pastime for political pundits and scholars alike to
predict the winner of the U.S. presidential election. Although forecasting
has now quite a history, we argue that the closeness of recent presidential
elections and the wide accessibility of data should change how presidential
election forecasting is conducted. We present a Bayesian forecasting model
that concentrates on the Electoral College outcome and considers finer
details such as third-party candidates and self-proclaimed undecided voters.
We incorporate our estimators into a dynamic programming algorithm to
determine the probability that a candidate will win an election.

Keywords: presidential elections; election forecasting; operations

research; Bayesian prediction models

Election Forecasting and Recent Elections

Recent presidential elections indicate that the American voting popula-

tion is deeply divided. This was plainly obvious in 2000 when Al Gore
won more popular votes than George W. Bush but lost the presidency in
the Electoral College tally. In 2004, when Bush secured 50.7% of the pop-
ular vote to John Kerry’s 48.3%, claims of voting irregularities again sur-
faced. This time the disputes were in Ohio, where Bush bested Kerry by
700

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

 Rigdon et al. / Bayesian Model of Presidential Election 701

118,775 votes. Although Kerry ultimately decided not to challenge the

vote count, if he had somehow mounted a victorious challenge and Ohio’s
20 electoral votes had been placed in the Democratic camp, Kerry would
have won the election with 271 electoral votes—and for the second time
in two elections, the president would have won while losing the popular
vote! When Gore lost, pundits and journalists alike chanted refrains that
this phenomenon had not surfaced since 1888 when Benjamin Harrison
beat Grover Cleveland in the Electoral College while pulling in 0.8% less
of the popular vote. In fact, this oddity is slightly more common, with the
more recent manifestation in 1960 when young John F. Kennedy secured
the presidency over Richard Nixon by winning the Electoral College vote
but not the popular vote (Gaines, 2001). These disparities between the
winner of the popular vote and the Electoral College are rare, but far from
unheard of, and becoming common enough that there have been numerous
calls in the last decade for a reform of the Electoral College system
(Belenky, 2008; Edwards, 2004; Goux & Hopkins, 2008; Grofman &
Feld, 2005; Hansen, 2008; Hiltachk, 2008; Hirsch, 2008; Leib & Mark,
2008; Tokaji, 2008; Whitaker & Neale, 2004).
A reform of the Electoral College system seems unlikely given the
intentional difficulty involved in modifying the U.S. Constitution
(Bennett, 2006; Rathbun, 2008). Although Gore, especially, appeared to
be drumming up support for the idea that he was the legitimate winner of
the 2000 presidential election, all presidential candidates enter the race
knowing full well how votes are counted and, hence, tailor their campaign
strategies accordingly. Candidates spend far more of their campaign funds
and time in states that are close and pay scant attention to voters in states
where the outcome is basically foregone (Althaus, Nardulli, & Shaw,
2002; Doherty, 2007; Shaw, 2006). It seems a bit disingenuous after votes
are counted to then appeal to voters that a different vote counting scheme
conveys the true and legitimate winner. Instead, there is no ambiguity that
the division of the popular vote is interesting, but not decisive. Moreover,
it is unclear that the popular vote accurately conveys the preference of the
voters better than the Electoral College system. Instead, the weight of the
evidence supports the claim that majority will is difficult to measure and

Authors’ Note: We gratefully acknowledge support from the National Science

Foundation (Grant IIS 08-27540 SGER) and from the Undergraduate Research Academy at
SIUE. Please address correspondence to Wendy K. Tam Cho, University of Illinois at
Urbana-Champaign, 361 Lincoln Hall, 702 S. Wright St., Urbana, IL 61801-3696; e-mail:
[email protected].

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

702 American Politics Research

fundamentally ambiguous (Gelman, Katz, & Tuerlinckx, 2002; Grofman,

Brunell, & Campagna, 1997; Grofman & Feld, 2005; Longley & Dana,
1992; Rae, 1972).
Also interestingly, presidential vote forecasting models have been heav-
ily focused on the two-party popular vote (Abramowitz, 2004; Campbell,
2004; Holbrook, 2004; Lewis-Beck & Tien, 2004; Lockerbie, 2004; Nor-
poth, 2004; Wlezien & Erikson, 2004), with only a few exceptions (Kaplan
& Barnett, 2003). These models have had varying degrees of success by
Campbell’s (2005) accounting method, in which less than 2 percentage
points is ‘‘quite accurate’’ (Lewis-Beck & Tien, 2004; Wlezien & Erikson,
2004); between 2 and 3 points is ‘‘reasonably accurate,’’ and closer to the
accuracy of the final polls than the precampaign polls (Abramowitz, 2004;
Campbell, 2004); between 3 and 4 points is ‘‘fairly accurate’’ (Norpoth,
2004); between 4 and 5 points is ‘‘inaccurate’’ (Holbrook, 2004); and in
excess of 5 points is ‘‘very inaccurate’’ (Lockerbie, 2004). The Wlezien and
Erikson (2004) error was a mere 0.5%, which is undoubtedly an impressive
forecast of the popular vote. However, we know, theoretically through sim-
ple math, and empirically from recent elections, that 0.5% is far and away
more than enough votes to swing the election in the Electoral College if
they hail from particular geographic locations. It is an artifact of our elec-
tion system that even the most impressive forecasts of the popular vote and,
indeed, the precise popular vote could point to the loser rather than the vic-
tor of a U.S. presidential election.
In the past, a focus on the national two-party popular vote was under-
standable from a data perspective because it greatly reduces data demands.
One needs only a single representative national poll to forecast the popular
vote. If one wants to focus instead on the Electoral College, one would
need a sample for every state in the union. Today, the proliferation of the
Internet has greatly facilitated the transfer of large quantities of data and
made state-by-state polling data readily available. Because the data acqui-
sition obstacle has been dissipated, it seems clear, then, that the ease of
data acquisition and the evident dissimilarity between the popular vote
and the Electoral College outcome should push presidential forecasting
models to shift their attention from the popular vote to electoral votes. In
addition, the closeness of recent elections might also imply that forecast-
ing models should include factors such as third-party candidates and the
proportion of undecided voters. Both of these entities have the potential to
change the course of a close election. Even if elections are not close, a bit
of fine-tuning, if done well, is always a welcome addition.

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

 Rigdon et al. / Bayesian Model of Presidential Election 703

In this article, we propose to move election forecasting in both of these

directions with a model focused on the Electoral College that also pays
attention to finer and potentially election-determining details such as
third-party candidates and voters who proclaim to be undecided in election
polls. Our framework uses a Bayesian estimator that allows one to apply
both prior and current information for each state to determine each candi-
date’s probability of winning that state in the presidential election. The
estimators incorporate both the previous election’s results (to capture each
state’s party tendency) and current polling data (to capture each state’s
current candidate tendency). We utilize an informative prior based on
long-term voting trends within a state and different probability distribu-
tions. Once we estimate the probability that a candidate will win a state,
these values can be used to determine the probability that each candidate
will win the election by using a Monte Carlo simulation or a dynamic pro-
gramming algorithm introduced in Kaplan and Barnett (2003). The idea of
using Bayesian estimation rather than frequentist estimation techniques
has been explored by others (Jackman & Rivers, 2001; Kaplan & Barnett,
2003), though our formulation differs.
Our article is organized as follows. The next section provides a descrip-
tion of the Bayesian estimators for each candidate’s probability of winning
each state’s Electoral College votes. The following section uses these esti-
mators, coupled with polling data that were available, to retrospectively
provide a prediction and analysis of the 2004 U.S. presidential election.
Next, we discuss the results and provide concluding comments and direc-
tions for future research.

Bayesian Estimators

A Bayesian analysis differs from a frequentist analysis in that it uses

observations to update or newly infer some unknown quantity of interest,
here, the outcome of the presidential election. The term ‘‘Bayesian’’ arises
from the frequent use of Bayes’s theorem, which relates the conditional
and marginal probabilities of two events. Two fundamental elements of
Bayesian analyses are the prior distribution and the posterior distribution.
The prior distribution encompasses our beliefs prior to observing some
data. In our application, the prior distribution is based on historical voting
patterns that are used to construct each state’s normal vote (Converse,
1966; Nardulli, 2005; Stokes, 1962). We note that our formulated
Bayesian estimators could, of course, incorporate any prior, including a

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

704 American Politics Research

Table 1
Voting Behavior of Alabama and New York
Alabama New York

Year Republican Vote (%) Democratic Vote (%)

1984 60.5 45.8

1988 59.2 51.5
1992a 47.7 49.8
1996 50.1 59.5
2000 56.5 60.4

a. Ross Perot captured approximately 19% of the popular vote in 1992.

noninformative prior or a prior based on a different forecasting model run

at the state level (e.g., Wlezien & Erikson, 2004). Constructing a prior
based on a popular vote forecasting model is certainly reasonable and
would be one way of combining our proposed methods with existing
methods. Without other knowledge of the current election, in our applica-
tion here, our belief about the outcome of the election will be based on
historical trends. Consider Table 1, which shows the voting behavior of
Alabama and New York over 5 election years (1984-2000). Quite clearly,
a state’s partisan tendency is a useful and informative prior.
This ‘‘prior’’ information can be used along with current polling data to
formulate the posterior distribution (using Bayes’s theorem) for the
probability of each candidate winning a state. The posterior distribution is
a transformation of the prior distribution in which the transformation
embodies an updating of beliefs after observing some data. Because state-
by-state polling data are more reliable for predicting the winner of the
election than the previous year’s election results, the posterior distribution
is constructed such that the likelihood function dominates the prior distri-
bution. The posterior distribution for each state is then used to compute
the probability that a candidate wins the state’s Electoral College votes.
Our Bayesian estimators also take into account the impact from voters
who declare that they are undecided prior to the election, as well as third-
party candidates, to obtain the posterior distribution or an estimate of each
candidate’s probability of winning the election. Finally, these probabilities
are used as input into a dynamic programming algorithm to determine the
distribution of Electoral College votes for each candidate and, hence, the
probability that each candidate wins the election.

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

 Rigdon et al. / Bayesian Model of Presidential Election 705

Bayesian Formulation
More formally, define pi to be the true proportion of voters in a state
who intend to vote for candidate i in the election (for simplicity, let i = 1
correspond to the Republican candidate, i = 2 correspond to the Demo-
cratic candidate, i = 3 collectively correspond to all third-party candidates,
and i = 4 correspond to no candidate or voters who have declared that they
are still undecided). These proportions are assumed to be continuous
(between 0 and 1) and sum to 1. The joint prior distribution for p = (p1 ,
p2 , p3 , p4 ) is assumed to be a conjugate prior distribution (i.e., when com-
bined with a multinomial distribution, the same type of posterior distribu-
tion is obtained). To satisfy this requirement, assume that p follows a
Dirichlet distribution, p ~ DIRICHLET(b1 , b2 , b3 , b4 ), which is a multi-
variate generalization of the beta distribution and is often used as a prior
for the probability of a success in Bernoulli trials. Therefore, the joint
probability density function of p can be written as

b − 1 b4 − 1
X
4
f ðp1 , p2 , p3 , p4 Þ = c pb11 − 1 pb22 − 1 p33 p4 , pi ≥ 0 for i = 1, 2, 3, 4, pi = 1,
i=1

P Q
where c = ( 4i = 1 bi )/ 4i = 1 (bi ):
The probability that a candidate wins a given state can be computed
using the marginal probability densities. To obtain these marginals, we
sequentially integrated the remaining variables out of the joint Dirichlet
probability density function. We now illustrate this process by first rewrit-
ing the joint Dirichlet probability density function as

b −1
f1, 2, 3 ðp1 , p2 , p3 Þ = c pb11 − 1 pb22 − 1 p33 ð1 − p1 − p2 − p3 Þb4 − 1 ,
X
3 ð1Þ
p1 , p2 , p3 ≥ 0, pi ≤ 1:
i=1

Integrating over p3 leads to an expression for the joint probability density

of p1 and p2:
Z 1 − p1 − p2
b −1 b −1 b −1
f1,2 ðp1 , p2 Þ = c p11 p22 p33 ð1 − p1 − p2 − p3 Þb4 − 1 dp3 ð2Þ
0
Z 1 − p1 − p2
b −1
= c pb11 − 1 p2b2 − 1 p33 ð1 − p1 − p2 − p3 Þb4 − 1 dp3 : ð3Þ
0

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

706 American Politics Research

Using the result

Z 1−y
ðaÞðbÞ
xa − 1 ð1 − y − xÞb − 1 dx = ð1 − yÞa + b − 1 ,
0 ða + bÞ
R∞ R∞
tx − 1 e − t dt = 2 u2x − 1 e − u du,
2
where the gamma function is (x) = 0 0
leads to the expression
b − 1 b2 − 1
f1,2 ðp1 , p2 Þ = C1 p11 p2 ð 1 − p1 − p2 Þb3 + b4 − 1 , p1 + p2 ≤ 1, p1 , p2 ≥ 0, ð4Þ

where C1 = c · (b3)(b4) / (b3 + b4). Integrating Equation 4 over all pos-

sible values of p2 gives the marginal density of p1,
Z 1 − p1
b −1 b −1
f1 ðp1 Þ = C1 p11 p22 ð1 − p1 − p2 Þb3 + b4 − 1 dp2 ð5Þ
0
Z 1 − p1
b −1 b −1
= C1 p11 p22 ð1 − p1 − p2 Þb3 + b4 − 1 dp2 ð6Þ
0
b −1
= C2 p11 ð1 − p1 Þb2 + b3 + b4 − 1 , 0 ≤ p1 ≤ 1, ð7Þ

where C2 = C1 · (b2 )(b3 + b4 ) / (b2 + b3 + b4 ). Therefore, by the form

of f1(p1 ), p1 is distributed as a beta random variable with parameters b1
and b2 + b3 + b4 . Using the identical argument, p2, p3, and p4 are also dis-
tributed as beta random variables; hence,
!
X
4
pi ∼ BETA bi , bk − bi , i = 1, 2, 3, 4: ð8Þ
k=1

Calibration of Prior Parameters

The beta distribution is a family of continuous probability distributions
characterized by two shape parameters, here, b1 and b2 + b3 + b4 , which
must be chosen. Different choices can be incorporated, which would result
in quite varied substantive implications. For example, one possibility is to
set these values so that the expected value and the variance, or the first
and second moments, for pi , i = 1, 2, 3, 4, closely match observed elec-
tions. However, there are an infinite number of combinations of b1 , b2 , b3 ,
and b4 that result in the same values for the expectations and the variances
for b1 , b2 , b3 , and b4 . That is, if the expectation and variance of a beta ran-
dom variable are given by

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

 Rigdon et al. / Bayesian Model of Presidential Election 707

bi
E(pi ) = P4 ð9Þ
k=1 bk

and
P 
4
bi bk − bik=1
Var (pi ) = P  P 2 ð10Þ
4 4
k = 1 bk + 1 k = 1 bk

 
ρ
ðρ + τ Þ 1− ρ
ρ+t
= , ð11Þ
ρ+τ+1
P
where ρ = bi and t = 4k = 1 bk − bi , then if b1 = b2 = 4 and b3 = b4 = 1 or
b1 = b2 = 40 and b3 = b4 = 10, the prior means are the same, namely,
4
Eðp1 Þ = Eðp2 Þ = = 0:4
10

1
Eðp3 Þ = Eðp4 Þ = = 0:1:
10
The choice of the shape parameters is essentially arbitrary if we do not
issue any constraints or do not use any substantive guidance. Fortunately,
in presidential forecasting, we have a great deal of substantive knowledge
that can be integrated. One way to constrain our choices is to choose the
P
bs so that (a) bi / 4k = 1 bk equals what pi is expected to be, prior to obser-
ving the polling data, and (b) the spread of the prior distribution for pi
P
(determined by 4k = 1 bk , with larger values indicating less uncertainty)
reflects the perceived uncertainty in pi.
To illustrate this process, consider the marginal probability density of p1
in the candidates’ home states: Massachusetts and Texas. In 2000, the most
recent presidential election before 2004, Bush received 33% of the popular
vote
P4 in Massachusetts and 59% of the popular vote in Texas. If we let
k = 1 bk = 4 and set b1 such that E(p1) equals the percentage of the popular
vote that the Republican candidate won in the 2000 election in the given
state (after adjusting for undecided voters—see below), the two marginal
densities for p1 are depicted in Figure 1. Under these choices for the bis,
Senator Kerry had a 0.374 probability of winning Texas, while President
Bush had a 0.208 probability of winning Massachusetts. These probabilities
seem too high given both the historical voting patterns of these states and
that these states are the home states of the candidates (Lewis-Beck & Rice,

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

708 American Politics Research

Figure 1
Marginal Probability Density Function for p1 in Massachusetts
and Texas When Sum of Prior Parameters Is 4

P
1983). Another choice would be to let 4k = 1 bk = 400, in which case Sena-
tor Kerry has a 0.002 probability of winning Texas, while President Bush
has a 10–15 probability of winning Massachusetts (see Figure 2). These
probabilities seem too low given historical trends. Perhaps a value in
between
P4 these two choices would be more ideal. Toward this effort, setting
k=1 b k = 40 gives Senator Kerry a 0.177 probability of winning Texas
and gives President Bush a 0.010 probability of winning Massachusetts (see
Figure 3).1 Of these values (4, 400, and 40), 40P appears to be the most sub-
stantively grounded, so we move forward with 4k = 1 bk being set to 40.
Our priors for the Republican and Democratic candidates involve the nor-
mal vote for all states except Alaska and Hawaii (and Washington, D.C.). For
the combined third-party candidates, we chose the mean of the prior to be
equal to the combined third-party vote in 2000. The prior mean for each of
the Republican and Democratic candidates for president was taken to be the
normal vote reduced by half of the third-party vote for that state.
To incorporate the effect from voters who declared that they were
undecided, let us assume that the prior mean for undecided voters is 3%.
This is an assumption in the purest sense, but it also seems reasonable as a
prior given the percentage of undecided voters in our various polls. Of
course, our method is not wedded to this value, and users are free to incor-
porate different values as they see fit. We could base this figure on older
polls or even an intuition of current trends since the last election.

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

 Rigdon et al. / Bayesian Model of Presidential Election 709

Figure 2
Marginal Probability Density Function for p1 in Massachusetts
and Texas When Sum of Prior Parameters Is 400

Figure 3
Marginal Probability Density Function for p1 in Massachusetts and
Texas When Sum of Prior Parameters Is 40

Thus, the prior parameters for a given state were chosen to be

 
1
b1 = 40ð1 − 0:03Þ NV1 − C3 ,
2
 
1
b2 = 40ð1 − 0:03Þ NV2 − C3 ,
2
b3 = 40ð1 − 0:03ÞC3 ,
b4 = 40ð0:03Þ,

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

710 American Politics Research

Figure 4
Marginal Density of p1 for Florida

where NVi is the normal vote for candidate i and C3 is the proportion of
the 2000 vote for third-party candidates. Using these parameters, the prior
distribution for a given state is
  
1
p = ðp1 , p2 , p3 , p4 Þ ∼ DIRICHLET 38:8 NV1 − C3 ,
2
  
1
38:8 NV2 − C3 , 38:8 C3 , 1:2 :
2
Illustrations of priors. The backdrop of the 2000 election may be used
to illustrate how this prior works with real data. In 2000, 49% of Flori-
dians voted for Governor Bush, 49% voted for Vice President Gore, and
2% voted for some other candidate. Therefore, the prior distribution for p
in Florida is
p = (p1 , p2 , p3 , p4 ) e DIRICHLET(19:012, 19:012, 0:776, 1:2):

From Figure 4, we can see that the marginal prior probability density of
p1 and p2 are identical in this case.
To obtain an expression for the likelihood function, let X = (X1 , X2 ,
X3 , X4 ) denote the random vector of sample proportions in a state poll for
p. Therefore, for a survey with n respondents,

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

 Rigdon et al. / Bayesian Model of Presidential Election 711

X = (X1 , X2 , X3 , X4 ) e MULTINOMIAL ðn, p1 , p2 , p3 , p4 Þ,

where the joint probability density of X is

n! x
X
4
gðX|pÞ = px1 px2 p 3 px4 , pk = 1, xi ≥ 0:
x1 !x2 !x3 !x4 ! 1 2 3 4 k=1

As another example, consider a SurveyUSA poll (conducted in October

2004) in Florida that had 742 respondents. In this poll, 364 expressed an
intention to vote for President Bush, 356 favored Senator Kerry, and the
final 22 respondents were undecided. Using these polling data, the likeli-
hood function for Florida voters is

742! X
4
gðX|pÞ = p364 p356 p0 p22 , pk = 1:
364! 356! 0! 22! 1 2 3 4 k=1

Bayes’s theorem can be used to obtain the posterior distribution for p. In

particular,

h(p | X) = CB f (p) · g(X | p),

where

h(p | X) is the posterior distribution density (for the probability of win-

ning a state);
CB is a proportionality constant, chosen so that the prior integrates to 1;
f(p) is the prior distribution density (in this case, Dirichlet);
g(X | p) is the likelihood function density (in this case, multinomial).

Therefore,

b − 1 b2 − 1 b3 − 1 b4 − 1 n! x x x x
hðp| XÞ = CB c p11 p2 p3 p4 p 1p 2p 3p 4
x1 ! x2 ! x3 ! x4 ! 1 2 3 4
x + b3 − 1 x4 + b4 − 1
X
4
= C px11 + b1 − 1 px22 + b2 − 1 p33 p4 , pi ≥ 0,8 i, pi = 1,
i=1

where
P 
4
 k = 1 ðbk + xk Þ
C = Q4
k = 1 ðbk + xk Þ

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

712 American Politics Research

Figure 5
Marginal Density for p1 (solid line) and p2 (dotted line)

is a constant of proportionality. This result means that the posterior distribution

for p given X (polling data results) is distributed as DIRICHLET(x1 + b1 ,
x2 + b2 , x3 + b3 , x4 + b4 ), which translates to DIRICHLET(383.012, 377.012,
0.776, 23.2) for the Florida data. The marginal densities of p1 and p2 are dis-
played in Figure 5.
Incorporating the role of undecided voters. Our method also includes a
means for incorporating the uncertainty presented by voters who remain
undecided up until election time. These variable voters are accounted for
by different swing scenarios, which we term the ‘‘no swing scenario,’’
the ‘‘Republican swing scenario,’’ and the ‘‘Democratic swing sce-
nario.’’ In a no swing scenario, 48% of declared undecided voters vote
for the Republican candidate and 48% vote for the Democratic candi-
date. Alternatively, there may be a Republican swing in which 52%
vote for the Republican candidate and 44% vote for the Democratic can-
didate, or a Democratic swing in which 44% vote for the Republican
candidate and 52% vote for the Democratic candidate. Note that in each
of the three scenarios, 4% of the undecided voters are assumed to vote
for a third-party candidate, which may or may not capture the reality,
particularly if undecided voters ultimately abstain from the election.
This is an assumption of the model that can be modified by the user if

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

 Rigdon et al. / Bayesian Model of Presidential Election 713

the user feels that it is far flung from any conceivable reality. After per-
using exit poll results for our election of interest, we felt comfortable
with this choice.
The posterior probability that a particular candidate wins a state can be
computed from the joint posterior distribution of the ps. For the no swing
scenario, we need only the posterior of (p1 , p2 ). For example, the prob-
ability that the Republican candidate wins a state is

Pðp1 + 0:48p4 > p2 + 0:48p4 |xÞ = Pðp1 > p2 |xÞ

Z 1/2 Z 1 − p2
= f12 ðp1 , p2 |xÞ dp1 dp2 :
0 p2

For the Republican swing scenario, the posterior probability must be com-
puted as a triple integral over the appropriate region. Figure 6 shows the
region in the p1 p2 plane for a fixed value of p4; p4 then goes from 0 to 1.
As this figure suggests, the triple integral must be broken into two parts,
giving

Pðp1 + 0:54p4 > p2 + 0:46p4 |xÞ = Pðp1 + 0:08 p4 > p2 |xÞ

Z 1 Z 0:5 − 0:54 p4 Z p1 + 0:08 p4

= f124 ðp1 , p2 , p4 |xÞ dp2 dp1 dp4
0 0 0

Z 1 Z 1 − p4 Z 1 − p4 − p1
+ f124 ðp1 , p2 , p4 |xÞ dp2 dp1 dp4 :
0 0:5 − 0:54 p4 0

The posterior probability under the Democratic swing scenario is similarly

computed.
Note that this analysis assumes that there is no third-party candidate
who has a nonnegligible chance of winning a state. If a viable third-party
candidate were present, we would need to make assumptions about the
swing of undecided votes among the other candidates. The posterior prob-
ability that
p1 + ðRepublican swingÞ · p4 > p2 + ðDemocratic swingÞ · p4

and
p1 + ðthird-party swingÞ · p4 > p3 + ðthird-party swingÞ · p4

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

714 American Politics Research

Figure 6
Region of Integration in the p1 p2 Plane
for the Republican Swing Scenario

would then need to be computed. Although our model formulation can

incorporate the impact of a third-party candidate, this feature was not
needed in 2004 because no third-party candidate mounted a viable candi-
dacy in this particular race.
Implementation. To obtain the probability that a candidate will win the
electoral votes in any given state, we must compute a series of fairly
involved formulas. For example, to implement the no swing scenario, we
must evaluate a double integral. The Republican swing and the Demo-
cratic swing scenarios involve triple integrals. These particular integrals
cannot be solved analytically. Instead, they must be approximated numeri-
cally. Accordingly, implementation of our model requires a software
package that is able to compute numerical approximations for integrals. In
addition, because these computations involve both very large numbers

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

 Rigdon et al. / Bayesian Model of Presidential Election 715

(from the Gamma function) and very small numbers (from the fractions
raised to large powers), the numerical evaluations are prone to inaccura-
cies unless a high degree of precision is maintained throughout the compu-
tation. We used WinBUGS to obtain the probabilities that each candidate
would win a state. Kaplan and Barnett (2003) cleverly designed a dynamic
programming algorithm to compute the probability distribution for the
number of Electoral College votes that a candidate will receive. This
dynamic programming algorithm, which uses the state-by-state probabil-
ities, was implemented in MATLAB. Implementation of our model, how-
ever, can be completed in other software packages as well. For example,
Mathematica could be used to evaluate the integrals.

Forecasting Presidential Elections

We now illustrate the use of our estimators for forecasting the 2004
U.S. presidential election. Our state-level surveys for this election were
reported by Real Clear Politics, an independent company that gathers
information from numerous publishing companies and organizes it for
public consumption on its Web site. The polls were gathered from a
variety of companies including Zogby, Rasmussen Reports, NBC, The
Wall Street Journal, the American Research Group, Fox News, and
Survey USA, among others.
For each state, the posterior distribution for the proportion of voters
who will vote for a candidate, the prior distribution for p, and the likeli-
hood distribution for X are all explicitly given in the previous section. The
posterior probabilities, shown in Table 2, can be used to compute the
probability that a candidate wins a state. A candidate who wins a state is
awarded the number of Electoral College votes associated with that state
(with the exception of Maine and Nebraska). To estimate the probability
that President Bush wins a state, the posterior probability that p1 > p2 was
computed, which is equivalent to assuming that all third-party candidates
have a zero probability of winning a state. In 2004, no third-party candi-
date was in a position to win a state, though third-party candidates were
influential in close states.
Once the probabilities of winning each individual state have been com-
puted, we used the Kaplan and Barnett (2003) dynamic programming
algorithm to compute the probability distribution for the Electoral College
votes. More formally, in this stage, we number the states (including

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

716 American Politics Research

Table 2
Posterior Probabilities That President Bush Wins Each State in 2004
Dem. No Rep. Dem. No Rep.
State Swing Swing Swing State Swing Swing Swing

Alabama 1.00 1.00 1.00 Nebraska 1.00 1.00 1.00

Alaska 1.00 1.00 1.00 Nevada .98 .98 .99
Arizona 1.00 1.00 1.00 New Hampshire .64 .66 .67
Arkansas .87 .88 .89 New Jersey .35 .50 .64
California .00 .00 .00 New Mexico .22 .23 .24
Colorado .81 .83 .84 New York .00 .00 .00
Connecticut .00 .01 .01 North Carolina .98 .98 .99
Delaware .01 .03 .06 North Dakota 1.00 1.00 1.00
Florida .36 .37 .38 Ohio .92 .94 .96
Georgia .99 .99 .99 Oklahoma 1.00 1.00 1.00
Hawaii .42 .49 .57 Oregon .05 .06 .07
Idaho 1.00 1.00 1.00 Pennsylvania .35 .37 .40
Illinois .00 .00 .00 Rhode Island .00 .00 .00
Indiana 1.00 1.00 1.00 South Carolina 1.00 1.00 1.00
Iowa .21 .22 .24 South Dakota 1.00 1.00 1.00
Kansas 1.00 1.00 1.00 Tennessee 1.00 1.00 1.00
Kentucky 1.00 1.00 1.00 Texas 1.00 1.00 1.00
Louisiana 1.00 1.00 1.00 Utah 1.00 1.00 1.00
Maine .00 .01 .01 Vermont .00 .00 .00
Maryland .00 .00 .00 Virginia .87 .87 .88
Massachusetts .00 .00 .00 Washington .14 .15 .16
Michigan .06 .07 .07 Washington, D.C. .00 .00 .00
Minnesota .04 .05 .06 West Virginia .97 .97 .98
Mississippi .99 .99 1.00 Wisconsin .05 .06 .07
Missouri .90 .91 .91 Wyoming 1.00 1.00 1.00
Montana 1.00 1.00 1.00

Note: Normal vote data were unavailable for Alaska, Hawaii, and Washington, D.C. We used
the 2000 presidential vote outcome as the prior for these states and Washington, D.C.

Washington, D.C.) from 1 to 51, and let p(k) be the probability that the
candidate wins state k and vk be the number of Electoral College votes for
state k. Let P(i, k) be the probability that the candidate wins exactly i Elec-
toral College votes in States 1, 2, . . . , k. Then, P(i, k) can be computed
for k ≥ 2 via the following recurrence relation:

ð1 − pðkÞ ÞPði, k − 1Þ + pðkÞ Pði − vk , k − 1Þ if i ≥ vk
Pði, k Þ =
ð1 − pðkÞ ÞPði, k − 1Þ if i < vk ,

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

 Rigdon et al. / Bayesian Model of Presidential Election 717

with boundary conditions (i.e., k = 1),

(
1 − p(1) if i = 0
Pði, 1Þ = p(1) if i = v1
0 otherwise.
The probability that a candidate wins a total of i electoral votes is given
by P(i, 51);P hence, the probability that the candidate wins the presidency
is given by 538i = 270 P(i,51). The posterior estimates of the probability of a
candidate winning state k (described in the Bayesian Estimators section) is
used in place of p(k) in this algorithm, based on available polling data.
Figures 7, 8, and 9 report empirical probability density functions for
the number of Electoral College votes won by the Republican candidate
under the no swing scenario, Republican swing scenario, and Democratic
swing scenario, respectively. For the no swing scenario, the probability
that President Bush wins the 2004 election is 0.63, with the expected num-
ber of Electoral College votes a little more than 277. In the final tally, Pre-
sident Bush won 286 Electoral College votes. The probability that he
would have received at least 300 Electoral College votes is 0.16. For the
Republican swing scenario, the probability that Bush would win the 2004
election is 0.71, with the expected number of Electoral College votes a lit-
tle more than 282. The probability that he would have received at least
300 Electoral College votes is 0.22. For the Democratic swing scenario,
the probability that Bush would win the 2004 election is 0.54, with the
expected number of Electoral College votes a little more than 272. In
addition, the probability that he would have received at least 300 Electoral
College votes is 0.11. These results suggest that the poll data accurately
indicated a very tight election.
The output from our model includes three different potential out-
comes to capture last-minute campaign changes under the assumption
that a small percentage of votes are truly swing votes and remain up for
grabs even in the last hour of the campaign with the release of critical
information. Surprising last-minute news releases are possible and cer-
tainly on the strategy radar of the campaigns. If the campaigns cruise
into Election Day with the same messages they have touted throughout
the election, then the no swing scenario is the most likely. If the Republi-
cans are somehow able to mount a concerted and effective effort in the
last hours to swing some voters in their direction, then the Republican
swing distribution would be the most relevant. Similarly, the Democratic
swing distribution would be the distribution of interest if the Democrats
were able to mount an effective final-hour campaign. These last-minute

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

718 American Politics Research

Figure 7
Estimate of Distribution of Electoral Votes
in 2004 for President Bush Under the No Swing Scenario

efforts are commonplace and occur after the polls on which our estimates are
based. Accordingly, supplying the three different outcomes gives us a way to
account for relevant information that transpires outside of our data window.
Unlike non-Bayesian forecasting models, our model supplies a poster-
ior distribution for the resulting Electoral College tally. The distribution
allows us to make a point estimate for the Electoral College vote, if such
precision is requested, while also allowing one to examine the uncertainty
attached to our estimates. Note as well that the final Electoral College tally
need not match the results obtained by predicting each state separately
and then simply summing up the tally from each state separately. If there
are several states where the outcome is close, but leaning to Bush, we
would not necessarily expect Bush to win each of these states even though
a single point estimate for each state would fall in his favor. In our polls,

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

 Rigdon et al. / Bayesian Model of Presidential Election 719

Figure 8
Estimate of Distribution of Electoral Votes in 2004 for
President Bush Under the Republican Swing Scenario

in 12 of the states Bush and Kerry were within 5 percentage points of each
other. In addition, there were 5 more states where the candidates were 6
points apart. In this election, every state predicted to be in Bush’s camp,
except New Hampshire, was computed to have at least an 80% probability
of voting for Bush. In contrast, there were 208 electoral votes in total from
states that had at least an 80% probability of voting for Senator Kerry.
That is, accordingly to our analysis, Senator Kerry had to win more of the
battleground states than President Bush in order to win the election.

Discussion

In the end, what matters is the Electoral College vote. We argue, then,
that election forecasting efforts should be directed at the Electoral College

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

720 American Politics Research

Figure 9
Estimate of Distribution of Electoral Votes in 2004 for President
Bush Under the Democratic Swing Scenario

vote rather than the popular vote. Attention to the Electoral College
changes forecasting in significant ways. Any prediction involves many
predictions, one for every state (and Washington, D.C.) rather than a sin-
gle prediction for the national popular vote. More data are involved as
well as more analysis. Our method is focused not on single state predic-
tions or even a single national prediction, but on finding an accurate esti-
mate of the distribution of electoral votes across the nation. From this
distribution, one may obtain a point estimate and measures of uncertainty
for the final election outcome.
We have presented a methodology for predicting the outcome of the
U.S. presidential election that uses a Bayesian estimation approach that
incorporates polling data. Our model includes the effect of third-party can-
didates and declared undecided voters as input to the Kaplan and Barnett

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

 Rigdon et al. / Bayesian Model of Presidential Election 721

(2003) dynamic programming algorithm to build the Electoral College

vote probability distribution for each candidate. There are other ways to
proceed as well. For instance, we could have used the individual state pre-
dictions to generate the Electoral College distribution via Monte Carlo
simulations. This is an alternative method that would also yield an Elec-
toral College distribution. However, our approach using the dynamic pro-
gramming algorithm is advantageous because it computes the probability
mass function for the Electoral College votes exactly. An infinite number
of Monte Carlo simulations would yield the same distribution, whereas a
smaller number of simulations would include more variability. Plainly,
there are different ways to proceed. Our point is not that we have identi-
fied the only method of obtaining an Electoral College distribution or that
all methods need to be Bayesian in nature, but that the Electoral College
distribution is the entity of interest in forecasting models. We have
demonstrated a simple means of identifying this distribution that is sub-
stantively consistent and allows us to incorporate finer details that may
affect the election outcome, especially in the types of close elections that
have typified presidential elections recently. We illustrated the predictive
capability of our particular estimation procedure on actual elections. Our
methods are flexible and also applicable to different scenarios such as
House and Senate races as long as state-by-state polling data are available
for these races.
Given the recent trends in presidential elections, the methodology
presented provides a rigorous approach for transforming state-level
polling data into presidential election forecasts. By incorporating third-
party candidates and undecided voters (using swing voter effects), the
emotionally heated peaks and valleys of tightly contested elections are
transformed into a rational probabilistic representation of the likely
winner of the race. Surely, as polling data change, so do the predic-
tions, and our methodology can be used at any point in the campaign
cycle to compare and capture these changes and allocate appropriate
weight to them so as to provide reasonable predictions even in the most
volatile elections.

Note
P
1. We did examine other values for 4k = 1 bk , ranging from 10 to 70. Values around 40
yield similar results, whereas values significantly larger or smaller than 40 may not be
appropriate.

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

722 American Politics Research

References
Abramowitz, A. I. (2004). When good forecasts go bad: The time-for-change model and the
2004 presidential election. PS: Political Science and Politics, 37, 745-746.
Althaus, S. L., Nardulli, P. F., & Shaw, D. R. (2002). Candidate appearances in presidential
elections. Political Communication, 19, 49-72.
Belenky, A. S. (2008). The good, the bad, and the ugly: Three proposals to introduce the
nationwide popular vote in U.S. presidential elections. Michigan Law Review First
Impressions, 106, 110-116.
Bennett, R. W. (2006). Taming the Electoral College. Stanford, CA: Stanford University
Press.
Campbell, J. E. (2004). Forecasting the presidential vote in 2004: Placing preference polls in
context. PS: Political Science and Politics, 37, 763-767.
Campbell, J. E. (2005). Introduction—Assessments of the 2004 presidential vote forecasts.
PS: Political Science and Politics, 38, 23.
Converse, P. (1966). The concept of a normal vote. In A. Campbell, P. Converse, W. Miller,
& D. Stokes (Eds.), Elections and the political order (pp. 9-39). New York: John Wiley.
Doherty, B. J. (2007). Elections: The politics of the permanent campaign: Presidential travel
and the Electoral College, 1977-2004. Presidential Studies Quarterly, 37, 749-773.
Edwards, G. C. (2004). Why the Electoral College is bad for America. New Haven, CT: Yale
University Press.
Gaines, B. J. (2001). Popular myths about popular vote—Electoral College splits. PS: Politi-
cal Science and Politics, 34, 70-75.
Gelman, A., Katz, J. N., & Tuerlinckx, F. (2002). The mathematics and statistics of voting
power. Statistical Science, 17, 420-435.
Goux, D. J., & Hopkins, D. A. (2008). Empirical implications of Electoral College reform.
American Politics Research, 36, 824-843.
Grofman, B., Brunell, T., & Campagna, J. (1997). Distinguishing the difference between
swing ratio and bias: The U.S. Electoral College. Electoral Studies, 16, 471-487.
Grofman, B., & Feld, S. L. (2005). Thinking about the political impacts of the Electoral Col-
lege. Public Choice, 123, 1-18.
Hansen, J. M. (2008). Equal voice by half measures. Michigan Law Review First Impres-
sions, 106, 100-105.
Hiltachk, T. W. (2008). Reforming the Electoral College one state at a time. Michigan Law
Review First Impressions, 106, 90-95.
Hirsch, S. (2008). Awarding presidential electors by congressional district: Wrong for Cali-
fornia, wrong for the nation. Michigan Law Review First Impressions, 106, 95-100.
Holbrook, T. M. (2004). Good news for Bush? Economic news, personal finances, and the
2004 presidential election. PS: Political Science and Politics, 37, 759-761.
Jackman, S., & Rivers, D. (2001). State-level election forecasting during election 2000 via
dynamic Bayesian hierarchical modeling (Working Paper). Stanford, CA: Stanford Uni-
versity, Department of Political Science.
Kaplan, E., & Barnett, A. (2003). A new approach to estimating the probability of winning
the presidency. Operations Research, 51(1), 32-40.
Leib, E. J., & Mark, E. J. (2008). Democratic principle and Electoral College reform. Michi-
gan Law Review First Impressions, 106, 105-110.

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

 Rigdon et al. / Bayesian Model of Presidential Election 723

Lewis-Beck, M. S., & Rice, T. W. (1983). Localism in presidential elections: The home state
advantage. American Journal of Political Science, 27, 548-556.
Lewis-Beck, M. S., & Tien, C. (2004). Jobs and the job of president: A forecast for 2004. PS:
Political Science and Politics, 37, 753-758.
Lockerbie, B. (2004). A look to the future: Forecasting the 2004 presidential election. PS:
Political Science and Politics, 37, 741-743.
Longley, L., & Dana, J. (1992). The biases of the Electoral College in the 1990s. Polity, 25,
123-145.
Nardulli, P. F. (2005). Popular efficacy in the democratic era: A re-examination of electoral
accountability in the United States, 1828-2000. Princeton, NJ: Princeton University
Press.
Norpoth, H. (2004). From primary to general election: A forecast of the presidential vote. PS:
Political Science and Politics, 37, 737-740.
Rae, D. W. (1972). Political consequences of electoral laws. New Haven, CT: Yale Univer-
sity Press.
Rathbun, D. P. (2008). Ideological endowment: The staying power of the Electoral College
and the weaknesses of the national popular vote interstate compact. Michigan Law
Review First Impressions, 106, 117-122.
Shaw, D. R. (2006). The race to 270. Chicago: University of Chicago Press.
Stokes, D. E. (1962). Party loyalty and the likelihood of deviating elections. Journal of Poli-
tics, 24, 689-702.
Tokaji, D. P. (2008). An unsafe harbor: Recounts, contests, and the Electoral College. Michi-
gan Law Review First Impressions, 106, 84-88.
Whitaker, L. P., & Neale, T. H. (2004). The Electoral College: An overview and analysis of
reform proposals (Tech. Rep. RL30804). Washington, DC: Congressional Research
Service.
Wlezien, C., & Erikson, R. S. (2004). The fundamentals, the polls, and the presidential vote.
PS: Political Science and Politics, 37, 747-751.

Steven E. Rigdon is Professor of Statistics and Graduate Program Director at Southern Illi-
nois University, Edwardsville. His research interests include reliability, quality control, statis-
tical computing, and election modeling.

Sheldon H. Jacobson is a professor in the Department of Computer Science at the University

of Illinois at Urbana-Champaign. His research interests are in the area of operations research,
with applications to societal problems of national interest. His research has been supported
by the National Science Foundation.

Wendy K. Tam Cho is associate professor in the Departments of Political Science and Sta-
tistics and senior research scientist at the National Center for Supercomputing Applications
at the University of Illinois at Urbana-Champaign.

Edward C. Sewell is a professor and assistant chair of the Department of Mathematics at

Southern Illinois University, Edwardsville. His current research interests are combinatorial
optimization and health applications.

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009

724 American Politics Research

Christopher J. Rigdon is a doctoral student in the Department of Industrial, Systems, and

Operations Engineering at Arizona State University. His research interests include design of
experiments, quality control, and statistical applications in sports.

Downloaded from http://apr.sagepub.com at UNIV OF ILLINOIS URBANA on June 15, 2009