ArticlePDF Available

Social networks and citizen election forecasting: The more friends the better

Authors:

Abstract

Most citizens correctly forecast which party will win a given election, and such forecasts usually have a higher level of accuracy than voter intention polls. How do citizens do it? We argue that social networks are a big part of the answer: much of what we know as citizens comes from our interactions with others. Previous research has considered only indirect characteristics of social networks when analyzing why citizens are good forecasters. We use a unique German survey and consider direct measures of social networks in order to explore their role in election forecasting. We find that three network characteristics – size, political composition, and frequency of political discussion – are among the most important variables when predicting the accuracy of citizens’ election forecasts.
International Journal of Forecasting 34 (2018) 235–248
Contents lists available at ScienceDirect
International Journal of Forecasting
journal homepage: www.elsevier.com/locate/ijforecast
Social networks and citizen election forecasting: The more
friends the better
Debra Leiter a,Andreas Murr b,Ericka Rascón Ramírez c,Mary Stegmaier d,*
aUniversity of Missouri – Kansas City, United States
bUniversity of Warwick, United Kingdom
cMiddlesex University, United Kingdom
dUniversity of Missouri, United States
article info
Keywords:
Social networks
Election forecasting
Citizen forecasting
Public opinion
Political interest
Expectations
Germany
abstract
Most citizens correctly forecast which party will win a given election, and such forecasts
usually have a higher level of accuracy than voter intention polls. How do citizens do it? We
argue that social networks are a big part of the answer: much of what we know as citizens
comes from our interactions with others. Previous research has considered only indirect
characteristics of social networks when analyzing why citizens are good forecasters. We
use a unique German survey and consider direct measures of social networks in order to
explore their role in election forecasting. We find that three network characteristics –
size, political composition, and frequency of political discussion – are among the most
important variables when predicting the accuracy of citizens’ election forecasts.
©2017 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.
1. Introduction
In most elections, the majority of citizens are able to
predict the election winner correctly, regardless of who
they plan to vote for (Lewis-Beck & Skalaban, 1989;Lewis-
Beck & Tien, 1999;Miller, Wang, Kulkarni, Poor, & Osh-
erson, 2012;Murr, 2011, 2015, 2016). Most US citizens
typically predict correctly not only which presidential can-
didate will win their state, but also who will win the pres-
idency (e.g., Graefe, 2014); and most British citizens are
usually correct about both which party will win their con-
stituency and which will garner a parliamentary majority
(e.g., Lewis-Beck & Stegmaier, 2011;Murr, 2016). How do
they do it?
A small body of work suggests that social networks are
a big part of the answer. Since much of what we know as
citizens comes from our social networks (e.g., Huckfeldt &
Sprague, 1995), we base our election predictions – like so
*Corresponding author.
E-mail address: StegmaierM@missouri.edu (M. Stegmaier).
many of our beliefs – on information from other people
in our network (Lewis-Beck & Tien, 1999;Meffert, Hu-
ber, Gschwend, & Pappi, 2011;Uhlaner & Grofman, 1986).
However, previous studies on social networks and citizen
forecasting accuracy have been hampered by the lack of
direct measures of social network characteristics, relying
instead on indirect or proxy measures. For example, Lewis-
Beck and Tien (1999) find that people with higher levels
of education are better able to predict who will win. This
is probably because people with higher levels of education
are more likely to develop skills in acquiring and processing
information. These authors also intimate that a person’s
level of education tells us something about the size of their
network, with more educated individuals possessing larger
networks. Meffert et al. (2011) and Uhlaner and Grofman
(1986) use electoral differences between the citizen’s elec-
toral district and the national level to capture the network’s
partisan composition indirectly, because the surveys that
they use do not collect measures of social network party
leanings. However, these indirect measures may miss
important aspects of the effect of social networks on citizen
forecasting.
https://doi.org/10.1016/j.ijforecast.2017.11.006
0169-2070/©2017 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.
236 D. Leiter et al. / International Journal of Forecasting 34 (2018) 235–248
This study uses direct measures of citizens’ network
sizes and compositions, along with other network charac-
teristics, in order to build a more complete model of citizen
forecasting. Using a unique cross-sectional survey that col-
lected both citizen election forecasts and direct measures
of several social network characteristics in Germany in the
autumn of 1990, we demonstrate that social networks have
as much predictive power of citizen forecasting accuracy
as the predictors identified as most important by previous
research, namely vote intention and political interest. In
addition, we show which social network characteristics
have predictive power for influencing election forecasts
(size, political composition, and frequency of discussion)
and which do not (heterogeneity and level of expertise).
In addition, we also provide guidance for future surveys as
to what network measure to include in order to improve
the accuracy of citizen election forecasts. Using a cross-
validation exercise, we demonstrate that a single, abbrevi-
ated measure of the network size improves out-of-sample
predictions.
2. Why citizen forecasts?
As the field of election forecasting has grown, schol-
ars have experimented with many different measures and
methods in an attempt to find the most accurate pre-
dictors (for reviews, see Lewis-Beck & Stegmaier, 2014;
Stegmaier & Norpoth, 2017). Such models often include
voter intentions or government approval ratings a few
months prior to the election as a gauge of the electorate’s
preferences.1Such variables can be found in models of elec-
tions in the US (Campbell, 2016;Erikson & Wlezien, 2016),
Britain (Ford, Jennings, Pickup, & Wlezien, 2016;Stegmaier
& Williams, 2016) and Germany (Jérôme, Jérôme-Speziari,
& Lewis-Beck, 2017;Norpoth & Gschwend, 2017), among
others. Both the approval and vote intention items re-
flect the respondent’s personal assessment of the incum-
bent government or the candidates. However, a developing
branch of the election forecasting literature has begun
to utilize electoral expectations, measured by the ques-
tion, ‘‘who do you think will win the election?’’ This ap-
proach is referred to as ‘‘citizen forecasting’’, and has been
used for election prediction in both the US (Graefe, 2014;
Lewis-Beck & Skalaban, 1989;Lewis-Beck & Tien, 1999;
Murr, 2015) and Britain (Lewis-Beck & Stegmaier, 2011;
Murr, 2011, 2016).
In such citizen forecasting models, the survey responses
are aggregated to the level of prediction, whether the
national level or the constituency level, and most often,
citizens get it right. For instance, in their pioneering study,
Lewis-Beck and Skalaban (1989) looked at citizen fore-
casts of eight US presidential elections between 1956 and
1984. They found that, on average, 69% of citizens fore-
cast the election winner correctly, but that the majority
of citizens forecasted 75% (six out of eight) of the elec-
tions correctly. In other words, moving from individual to
aggregate forecasts improved the accuracy from 69% to
1In addition to voter intention polls or approval ratings, such models
often include economic performance measures, the number of terms the
party has held office, and previous election results.
75% – an increase of six percentage points. Their two main
findings – that most citizens forecast correctly most of the
time, and that groups forecast better than individuals –
have subsequently been replicated at two different levels
(subnational and national) and in two countries (Britain
and United States); see for example Graefe (2014), Lewis-
Beck and Stegmaier (2011) and Murr (2011,2015,2016).
In addition to demonstrating that citizen forecasts are
accurate, several studies have also shown that citizen
forecasts are more accurate than any other forecasting
approach, including voter intention polls. Using national-
level data from the last 100 days before each of the seven
US presidential elections between 1988 and 2012, Graefe
(2014) compared the relative accuracies of citizen fore-
casts, voter intentions, prediction markets, expert surveys,
and quantitative models. He found that citizen forecasts are
better than any other approach at forecasting both election
winners and vote shares. Similarly, Murr, Stegmaier, and
Lewis-Beck (2016) used national-level data from the 48
months before each of the 18 British general elections
between 1950 and 2015 to compare the relative accuracies
of citizen forecasts and voter intentions, and found that
citizen forecasts are better than voter intentions at fore-
casting both the winning party and its seat share.
As Murr (2015) has shown, the accuracy of citizen fore-
casts can even be increased by weighting and delegating
the individual forecasts optimally based on the citizens’
competence (e.g., Grofman, 1975;Kazmann, 1973;Shapley
& Grofman, 1984). The method involves two steps: first,
predict the probability that each citizen will forecast cor-
rectly; then, delegate the forecasting to the most compe-
tent citizen and weight their forecasts according to their
level of competence. Using data from eleven US presiden-
tial elections between 1952 and 2012, Murr (2015) showed
that this increases the forecasting accuracy of both the
candidates’ vote shares in a state and which candidate will
carry the state. Thus, being able to predict the chance of
a citizen forecasting the election correctly is crucial for
improving the forecasting accuracy.
3. Why can citizens forecast correctly?
The explanation as to why citizen forecasts are accu-
rate has two parts (Murr, 2017). The first part explains
why groups forecast better than individuals. This part rests
on the assumption that individuals forecast better than
chance on average, and the second part explains why in-
dividuals are able to do so.
Murr (2011) explains the fact that groups predict better
than individuals based on Condorcet’s jury theorem and
its generalizations (Condorcet, 1785). Condorcet demon-
strates the conditions under which the group decisions
reached by a plurality rule are better than, equal to, or
worse than individual decisions. His proof assumes that
(i) the group faces two alternatives, one correct and one in-
correct, (ii) the kgroup members vote independently of one
another, and (iii) each member has one vote and the same
probability pof choosing the correct alternative. Then, the
probability of a correct group decision by a majority vote
is
P=
k
m=k/2+1(k
m)pm(1p)km.
D. Leiter et al. / International Journal of Forecasting 34 (2018) 235–248 237
Table 1
The probability of a correct majority vote from kmembers with an indi-
vidual probability of getting it right of p.
k=3k=5k=7k=9
p=0.6 0.6480 0.6826 0.7102 0.7334
p=0.7 0.7840 0.8369 0.8740 0.9012
p=0.8 0.8960 0.9421 0.9667 0.9804
p=0.9 0.9720 0.9914 0.9973 0.9991
He shows that if each member chooses the correct alter-
native with more than a 50% probability, the probability
of a correct group decision approaches unity as the group
size increases to infinity (‘‘wisdom of crowds’’). He also
shows that if each member chooses the correct alternative
with less than a 50% probability, the probability of a correct
group decision approaches zero as the group size increases
to infinity (‘‘folly of crowds’’).
Although Condorcet’s jury theorem refers to group sizes
approaching infinity, even small groups show the effect of
aggregating individual choices. Consider a group of three
independent members, each with an 0.6 probability of
choosing the correct alternative. This group chooses the
correct alternative using a majority vote if at least two of
the three members vote correctly. Using the above formula,
the probability of a correct group decision is P=3×0.62×
0.4+0.63=0.648, an increase in accuracy of about five
percentage points. This probability increases as the group
size increases: it is 0.6824 with five independent members,
0.7102 with seven members, 0.7334 with nine members,
and so on. In other words, even though individually mem-
bers may be only slightly better than chance in getting it
right, collectively they may choose the correct alternative
with almost certainty, if the group has enough members.
Table 1 displays the probabilities of a correct group deci-
sion for different individual probabilities of getting it right
(p=0.6, 0.7, 0.8, and 0.9) as well as different group sizes
(k=3, 5, 7, and 9).
In deriving his theorem, Condorcet made three assump-
tions: each member chooses between only two alterna-
tives, votes independently of the others, and has the same
probability of voting correctly. Since the publication of
his theorem, several other authors have relaxed each of
these assumptions in turn and generalized the theorem
accordingly. The theorem still holds even with more than
two alternatives (List & Goodin, 2001), which is important
because many elections involve voters choosing between
more than two parties. Further, Ladha (1992) generalizes
the theorem to correlated votes. This is relevant because
citizens might share the same information, talk to each
other, or tend to ‘‘groupthink’’ (e.g., Janis, 1982). Finally,
Grofman, Owen, and Feld (1983) prove that the theorem
still holds if members differ in their probability of getting
it right as long as they are all better than chance on average.
This is important because Lewis-Beck and Skalaban (1989)
show that citizens vary in their probability of making a
correct forecast. In summary, these generalizations make
the theorem useful for explaining why groups of citizens
predict better than individuals.
Since the explanation of why groups predict better than
individuals rests on the fact that individuals predict better
than chance on average, the next step is to explain why this
is the case. Murr (2017) explains the fact that individuals
predict better than chance based on Uhlaner and Grof-
man’s Contact Model (Uhlaner & Grofman, 1986). Echoing
Condorcet’s jury theorem, the Contact Model proves the
conditions under which a citizen’s forecast, reached by
choosing the party that is supported by the plurality of
information available to the citizen, will be better than,
equal to or worse than chance. The proof assumes that
the citizen is forecasting a two-party election, receives
and accepts pieces of information from the environment
independently of one another, and counts each piece of
information equally.
The Contact Model implies that if a citizen receives
and accepts only information that is consistent with her
vote intention (‘‘selective sampling’’), citizen forecasts will
always be better than chance on average, though always
as informative as voter intentions. However, if a citizen
receives and accepts information that is representative of
the public’s voter intentions (‘‘random sampling’’), citizens
will always be better than both chance and voter intentions
on average. As the number of randomly sampled bits of in-
formation increases to infinity, the probability of a correct
forecast approaches unity. In other words, citizens will do
better than chance and voter intentions, as is indeed the
case, as soon as they receive and accept at least some infor-
mation that is representative of the public’s vote intention
(e.g., Graefe, 2014;Lewis-Beck & Skalaban, 1989).
Because much of what we know as citizens comes from
interpersonal communication, we argue both that citizens’
social networks predict their election forecasts, and that
these networks offer the representative information that
is necessary to enable them to forecast better than chance.
4. Social networks and citizen forecasts
The study of social networks—the social context
through which individuals are tied to others—has shed
light on both the way and the extent to which friends,
family, neighbors, and peers influence electoral belief for-
mation and voting behavior. In addition to learning from
previous cohorts and personal experience (Blais & Bodet,
2006;Manski, 2004) and the media (Entman, 1989), net-
works provide contextual information that both allows
voters to form expectations about elections and influences
their choices. Meffert et al. (2011), for example, analyze
various factors that influence electoral expectations, such
as political motivations (knowledge and interest), rational
and strategic considerations (the perceived distance be-
tween parties), and social context (regional differences, as a
proxy for personal networks), and how these expectations
influence voting behavior. The authors find that voters
can form reasonable expectations about the winning party
and that these beliefs are used to cast ‘‘fairly sophisticated
votes’’, such as strategic coalition voting.
Complementarily, Pattie and Johnston (1999) showed
that conversations with partisan discussants influence vote
decisions, and can even lead citizens to switch their vote
to another party. Similarly, Huckfeldt and Sprague (1991)
showed that vote preferences are not determined only by
voter characteristics, but also by their discussant partners’
characteristics and political preferences; while Nickerson
238 D. Leiter et al. / International Journal of Forecasting 34 (2018) 235–248
(2008) provided evidence regarding the influence of cou-
ples on voting behaviors. Other studies have shown that
variations in the composition and size of an individual’s
network affect their political attitudes and the amount of
political information they have, which in turn affect their
behavior and their beliefs (Huckfeldt, 2007;Huckfeldt &
Mendez, 2008;Mutz, 1998;Partheymüller & Schmitt-Beck,
2012;Pietryka, 2015).
But how do people form electoral expectations? Citi-
zens may gather information and update their beliefs about
electoral victories based on: (1) their network members’
characteristics, by observing how other members act and
think about political, social and economic matters; (2) di-
rect information from their network, by discussing who
they think will win the election and which party they
support; (3) previous electoral experiences; and (4) the
news and opinion polls.
The very nature of social networks makes this source
of information more likely to influence citizen electoral
expectations and behavior than other sources such as the
news media or polls. For instance, Schmitt-Beck and Mack-
enrodt (2010) show that personal communication appears
to be more influential regarding turnout in a German local
election than mass communication. Despite the fact that
the media and polls may provide more reliable and bal-
anced information about the electoral environment than
social networks, information from social networks may
provide more personalized information by using language
and terms that are closer to the local context and more
familiar.
While the news media and polls are passive sources of
information, social networks give citizens the chance to
actively disagree with dissonant information and to learn
from it by debating with network members. Hence, though
all sources of information may be complementary, social
networks provide citizens with the opportunity to engage
in back-and-forth debate and to learn from disagreements.
As was suggested by McClurg (2006), social networks can
encourage higher levels of political involvement, as well
as an increased openness to differing viewpoints. In other
words, people can learn from their networks.
The magnitude of a network’s influence on citizens’
beliefs about who will win the election may depend on the
network’s size, frequency of political discussion, political
expertise and composition (heterogeneity), along with ad-
ditional sources of political information.2Citizens who are
embedded in larger social networks may have an advan-
tage in forecasting elections, as they frequently have higher
levels of political knowledge (Kwak, Williams, Wang, & Lee,
2005). In addition, the larger the social network, the more
likely it is to reflect the vote intentions of the population,
making the aforementioned indirect inference more accu-
rate (Banerjee & Fudenberg, 2004).3
Citizens without a network (isolated citizens) are likely
to form their beliefs about who will win the election based
2Similarly, Millner and Ollivier (2016) discuss three main factors that
determine the public’s beliefs in the context of environmental policies:
individual inference (how the updating of beliefs takes place), social
learning and media.
3This is true only if the most important agent’s influence diminishes
as the number of network members increases (Golub & Jackson, 2010).
on media or poll information, as well as on their own
electoral preferences. However, if these citizens are in-
correct in their belief of who will win the election, they
lack the social contextual pressure or ability to update
their expectations. In contrast, citizens with initially wrong
or uncertain beliefs who are embedded in networks may
retrieve information from their network in order to revise
their expectations using information about their network’s
voting preferences (Chandra, 2009).
Having large networks may influence beliefs and be-
havior, but the information that citizens obtain from them
should be updated frequently. The more political discus-
sions that citizens have with their network, the more in-
formation they collect from its members and the more
they will be able to remember it. Additional information
may also render the network’s information more salient
than the citizen’s own information when it provides the
citizen with new information. Moreover, the increased fre-
quency of discussion encourages citizens to become more
informed, thus improving their ability to forecast (Eveland,
2004;Eveland & Hively, 2009).
Both informed and uninformed citizens use networks to
gather information about the political system and elections
(Pietryka, 2015). They seek out political experts to help
them evaluate an election, even if they do not share the
same partisan affiliation. Citizens are more likely to be
influenced by those who they perceive as having exper-
tise (Ahn, Huckfeldt, & Ryan, 2014;Huckfeldt & Mendez,
2008;Huckfeldt, Pietryka, & Reilly, 2014;Ryan, 2011) than
by non-experts. Thus, these experts within the network
should help improve citizens’ forecasting accuracy by pro-
viding accurate, if still biased, information. Political exper-
tise can also help in recognizing dissonant information and
rejecting it (McClurg, 2006).
In general, social networks play a role in both the dis-
semination of information and the acquisition of informa-
tion that reduces ambiguity (Ahn et al., 2014;Eveland &
Hively, 2009;Finkel & Smith, 2011;Manski, 2004). How-
ever, in some cases, the information acquired from social
networks may actually decrease the likelihood of a cor-
rect election prediction. When a political network leans
toward the losing parties, or a citizen is unsure of how
other network members will vote, this will undermine the
citizen’s ability to offer an accurate election prediction.
Those embedded in homogeneous networks may assume
that there is a greater support for a political party than in
fact exists, and such networks may also reinforce ‘‘wishful
thinking’’; thus, citizens belonging to these networks may
overestimate the chances of victory of a party that actually
has little chance of success.
While political disagreement in networks persists even
in multiparty electorates (Huckfeldt, Ikeda, & Pappi, 2005;
Huckfeldt & Johnson, 2004), individuals frequently find
themselves in social networks with other like-minded indi-
viduals. The homogeneity (homophily) of the network may
either increase or decrease the likelihood of a successful
forecast. Individuals in heterogeneous networks tend to
show higher levels of political knowledge, as they fre-
quently seek out additional information when they interact
with those who do not share their views, which should
improve their electoral forecasts (Eveland & Hively, 2009).
D. Leiter et al. / International Journal of Forecasting 34 (2018) 235–248 239
Fig. 1. Voting intentions, 1987–1990.
Source: Forschungsgruppe Wahlen
(2017).
Additionally, to the extent that individuals rely on their
networks to act as representative samples, more homoge-
nous networks, particularly those which are allied with an
unlikely winner, will decrease the likelihood of a correct
forecast. Thus, in such cases, the inclusion of more people
in a person’s network will not add new information. As
such, social networks may improve citizens’ ability to make
accurate electoral forecasts, but this depends on the size
and composition of the networks.
5. Data and measures
The 1990 German federal election offers a unique elec-
toral context in which to examine how social networks
predict citizens’ ability to forecast the election, as it pro-
vides a direct comparison between citizens with long-term
democratic experience (West Germans) and citizens who
were new to democratic elections (East Germans), without
varying the institutional or electoral context. West Ger-
many held its first democratic election on 14 August 1949,
whereas East Germany did not hold its first democratic
election until 18 March 1990. The 2 December 1990 Bun-
destag election was the first Federal Republic of Germany
election for East Germans, who had voted only four months
earlier to unify with West Germany.
The governing Christian Democratic Union (CDU) had
been losing support ever since its electoral victory in Jan-
uary 1987 (Fig. 1). This loss benefited the main opposition
party, the Social Democratic Party (SPD), which then led the
polls from October 1987 to September 1989. However, the
CDU started to recover midway through the electoral cycle,
and led again for the first time in October 1989, beginning
a period of uncertainty about whether the CDU or the
SPD would win the subsequent election. From March 1990
onward, it looked increasingly likely that the CDU would be
victorious in December. They won the East German general
election in March, leading the SPD by 19 percentage points.
In April, Oskar Lafontaine, the SPD candidate, fell victim
to an assassination attempt and was unable to campaign
for three months. From August onwards, opinion polls
showed the CDU to be in the lead, due largely to the public
perception that the CDU was the party best able to handle
the economic consequences of unification (Pulzer, 1991).
However, even though the outcome was fairly certain, as
we discuss in the next section, not everyone correctly fore-
cast a CDU win.
We examine how social networks predict the ability
of citizens to forecast by using the 1990 German section
of the Comparative National Elections Project, a cross-
national survey that collects both traditional individual-
level data and information on the respondents’ media,
organizational, and (most importantly for this project) so-
cial network characteristics (Gunther, Beck, Magalhães, &
Moreno, 2015;Gunther, Puhle, & Montero, 2007). The Ger-
man section of this survey relies on face-to-face interviews
in the pre-election period (October and November 1990),
and includes a network battery that asked respondents to
name up to five people with whom they discuss important
matters. Our sample includes a total of 1547 respondents,
of whom 487 are from East Germany. This survey uniquely
(to the best of our knowledge) provides both information
on the character and extensiveness of a respondent’s social
networks and the respondent’s electoral forecast.
To measure the ability of citizens to forecast the winner
of the election correctly, we rely on a survey item that
asks respondents whether they believe that a CDU-led or
an SPD-led government is likely to win the election, or
they do not know.4Based on the previous literature, we
code all respondents who predict a CDU victory as correct
forecasters, and all other respondents as incorrect. The ma-
jority of respondents forecast the winner correctly; how-
ever, approximately 25% of West Germans and 18% of East
Germans made incorrect forecasts about the election. It is
notable here that the East Germans were better forecasters
than the West Germans, despite their limited experience
with democratic elections.
We differentiate between uncertain and inaccurate an-
swers by creating a categorical variable, where those who
4The question wordings can be found in the Appendix A.
240 D. Leiter et al. / International Journal of Forecasting 34 (2018) 235–248
Table 2
Summary statistics of network variables.
West Germans East Germans
Average SD Min Max Average SD Min Max
Network size 2.46 1.21 1 5 2.67 1.24 1 5
Network discussion 1.64 0.80 0 3 2.39 0.63 0 3
Network expertise 1.08 0.54 0 2 1.27 0.50 0 2
Network left (proportion) 0.31 0.40 0 1 0.27 0.37 0 1
Network unknown (proportion) 0.29 0.41 0 1 0.28 0.40 0 1
Network heterogeneity 0.42 0.43 0 1 0.43 0.43 0 1
answer SPD are treated as inaccurate, those who respond
with ‘don’t know’ are uncertain, and correct CDU forecasts
are treated as the reference category. While the propor-
tions of inaccurate forecasts are similar between East and
West Germans, with 9.9% and 9.5% respectively forecasting
an SPD victory, more than 15.7% of West Germans were
uncertain about the election outcome, compared to only
8.9% of East Germans.
We test how social networks predict the accuracy of
election forecasts by examining four network characteris-
tics: network size, frequency of political discussion in the
network, political expertise in the network, and network
ideology (heterogeneity). The network size ranges from 1
to 5,5and is based on the number of discussants that the
respondent named in the network battery.6The frequency
of political discussion measures how often, on average, the
respondent discusses political matters with members of
their network, based on the respondent evaluation, ranging
from always to never (network discussion). Network ex-
pertise is based on the average evaluation of each network
member’s level of political knowledge. Network ideology
is measured as both the proportion of the network that
the respondent believes will vote for a left-leaning party
(network left), and the proportion of the network for whom
the respondent does not know the political party prefer-
ence (network unknown).7Finally, network heterogeneity
is operationalized as one minus the absolute difference
between the proportions of left- and right-leaning mem-
bers in the respondent’s network.8While increases in the
network size, frequency of discussion, network expertise,
5We exclude respondents without a discussant because the other
network characteristics cannot be calculated for them.
6Since the creation of this survey in 1990, there has been a growing
scholarly discussion about network size generators. Although Marsden
(2003) demonstrates that less than 10% of respondents generate more
than five names, and Merluzzi and Burt (2013) provide evidence suggest-
ing that five is a cost-effective number of network responses, Eveland,
Hutchens, and Morey (2013) argue that the type of name generator used
in this survey consistently underestimates the network size. However,
given our theoretical expectation, we argue that this underestimation
provides a conservative test for our hypotheses. In addition, summary
network measures cannot measure network characteristics other than the
size (Eveland et al., 2013).
7While there could potentially be concerns regarding projection ef-
fects when using respondents’ evaluations of their discussion partners’
party preferences, previous research has demonstrated that voters are
surprisingly accurate at identifying their discussion partners’ political
preferences (Huckfeldt & Sprague, 1995).
8The proportion of right-leaning members is one minus the propor-
tions of left-leaning members and members for whom the respondent
does not know the political party preference. Respondents with equal
proportions of left- and right-leaning members in the network reach the
and network heterogeneity may be expected to improve
the ability of the respondent to forecast the outcome of
the election correctly, the network ideology, particularly
for left-leaning networks, may decrease the likelihood of a
correct election forecast, as was suggested in the previous
section. Table 2 displays summary statistics of the network
variables.
In addition to these network variables, we also con-
sider other factors that previous studies have suggested
might predict the accuracy of the forecasts (e.g., Lewis-
Beck & Tien, 1999;Meffert et al., 2011): political, me-
dia, and demographic factors, as well as the number of
days before the election that the survey interview took
place. We capture individual partisanship and the effects
of ‘wishful thinking’ by creating three dummy variables
based on the respondent’s reported vote intention on the
second ballot, namely SPD voters, CDU voters, and voters
who are uncertain about how they will vote, with minor
party supporters being treated as the reference category.9
We also control for self-reported levels of political inter-
est, attention to television news, and attention to news
in newspapers. The sociodemographic measures that we
include are gender, age (transformed into four quartiles),
and education (transformed into three categories). Finally,
since the survey was conducted over a number of weeks,
we account for the number of days before the election that
the respondent was surveyed.
Because we argue that social networks provide citizens
with information that helps them to forecast correctly, it
is instructive to examine how our network measures differ
from other measures related to information, such as formal
education, political interest, and media attention (TV and
print). We measure how network characteristics relate to
these other informational measures by calculating Pear-
son’s correlation coefficient r(Table 3). While there is an
association between network characteristics and political
education, interest, and media attention, most of the time
it is either very weak (|r|<0.20) or weak (0.20 |r|<
0.40). This means that while many people have personal
highest value of one on the measure, indicating complete heterogene-
ity, while respondents with network members of only one ideological
direction reach the lowest value of zero on this measure, indicating
complete homogeneity. Respondents with ideologically mixed networks
reach values between these two extremes.
9Germany uses a mixed member proportional electoral system, which
provides voters with the opportunity to cast both a candidate vote (first
ballot) and a party vote (second ballot) for the Bundestag, with the party
vote determining the overall share of seats in the legislature. This latter
measure of vote intention forms the most comparable measure between
East and West Germany, as partisanship was not asked of East German
respondents.
D. Leiter et al. / International Journal of Forecasting 34 (2018) 235–248 241
Table 3
Correlation between network characteristics and education, political interest, and media attention.
Education Political interest TV news attention Print news attention
Network size 0.20 0.18 0.12 0.12
Network frequency 0.33 0.46 0.43 0.31
Network expertise 0.22 0.34 0.28 0.24
Network left 0.01 0.09 0.03 0.04
Network unknown 0.05 0.13 0.11 0.10
Network heterogeneity 0.01 0.06 0.06 0.06
characteristics (e.g., a low political interest) that might
make accurate forecasts less likely, they nevertheless have
social network characteristics (e.g., many discussants) that
might make accurate forecasts more likely. In other words,
for many citizens, their social network may potentially
compensate for their lack of information from the media,
while for others it may correct or complement the media
information they receive. The weak correlation between
network characteristics and political interest, education
and media attention, together with our theoretical argu-
ments, justify us in considering network characteristics as
additional predictors of a citizen’s forecasting accuracy.
The regression analyses reported below weight the re-
spondents by inverse sampling probability in East and
West, because East Germans were oversampled relative to
their proportion of the population, and cluster the standard
errors by sampling point.
6. Results
6.1. Correct and incorrect forecasts
First, we examine the variables that predict the accu-
racy of Germans’ election forecasts. The outcome in the
logit model shown in Table 4 is whether the respondent
forecasted the CDU victory correctly or not. In this anal-
ysis, incorrect forecasts include both responses that the
SPD would win and ‘‘don’t know’’ answers. While we are
most interested in the differences in forecasting accuracy
between respondents with different social network char-
acteristics, looking at other variables that could predict the
forecast accuracy enables us to compare these results to
those of the handful of other studies that have looked at
the characteristics of accurate forecasters.
The results of the binary logit model in Table 4 indi-
cate that social networks predict the forecast accuracy in
ways that are consistent with our expectations, even when
controlling for a host of other political, media, and demo-
graphic characteristics.10 We observe that both the number
of people in the respondent’s network and the frequency of
political discussion have positive and statistically signifi-
cant coefficients. This means that both having more people
in the network and having more frequent discussions in
the network make a positive difference to the probability of
a correct forecast. Conversely, we observe that the shares
of the network with left or unknown political leanings
have negative and statistically significant coefficients. This
10 These computations and those that follow were performed on a Mac
OS X 10.11.6 with Stata/SE 12 using the logit, mlogit, margins, and lincom
commands.
Table 4
Correct forecast of CDU victory pooled binary logit estimates.
Log-odds
Estimate (Std. error)
Constant 0.89 (0.65)
East 0.09 (0.19)
Age 0.09 (0.06)
Female 0.01 (0.13)
Education 0.06 (0.13)
Political interest 0.29** (0.09)
TV news attention 0.03 (0.09)
Print news attention 0.03 (0.07)
SPD voter 0.03 (0.18)
CDU voter 2.10** (0.27)
Undecided voter 0.53** (0.25)
Days until election 0.01 (0.01)
Network size 0.22** (0.07)
Network discussion 0.19*(0.11)
Network expertise 0.24 (0.16)
Network left 0.77** (0.24)
Network unknown 0.59*(0.35)
Network heterogeneity 0.15 (0.33)
N1547
Note: Standard errors are clustered by sampling points. The data are
weighted by inverse sampling probabilities in East and West.
*p<0.10.
** p<0.05.
means that the larger the share of the network with left or
unknown political leanings, the less likely the respondent’s
forecast is to be correct. The coefficients of both network
expertise and network heterogeneity are in the expected
positive direction, but miss conventional levels of statisti-
cal significance.
Of the other variables, only a few of the coefficients
attain statistical significance. We corroborate the findings
of earlier studies that respondents with higher levels of
political interest are more likely to make accurate forecasts,
and find evidence that CDU voters are more likely to fore-
cast a CDU victory correctly than the excluded ‘‘minor party
vote’’ category. We also observe that respondents who say
that they don’t know for whom they will vote (undecided
voters) are also more likely to forecast correctly than minor
party voters, though the coefficient is smaller than for CDU
voters. In contrast, SPD voters are just as likely to get it right
or wrong as minor party voters.
Notably, the coefficient of the ‘‘East’’ variable, which is
designed to capture systemic differences between East and
West Germans in this pooled analysis, is not statistically
significant. Furthermore, the demographic variables, me-
dia exposure, and number of days before the election are
not predictive of the forecasting accuracy.
242 D. Leiter et al. / International Journal of Forecasting 34 (2018) 235–248
Fig. 2. Difference in expected probabilities for pooled binary logit model. Note: Difference in expected probabilities between two respondents with
maximum and minimum values of the indicated predictor while holding the remaining predictors constant at their median value. The predictors are sorted
by increasing difference in expected probability, for network characteristics and controls separately. Bold segments indicate 90% confidence intervals and
thin segments indicate 95% confidence intervals.
We investigate the results of the full binary model and
the subsequent multinomial logit models further by com-
puting first differences (King, 1989, pp. 107f). First differ-
ences estimate how much the fitted values would differ
on average when comparing two respondents who have
different levels of a given predictor while being identical
on all other variables. We compute first differences by
subtracting the expected probability of an outcome given
the maximum value of a predictor from the expected prob-
ability given its minimum value, holding all other variables
at their median.
Fig. 2 provides a visual assessment of the differences
between the expected probabilities of a CDU forecast when
comparing two respondents who have the minimum and
maximum levels of a given predictor, while holding all of
the other variables at their medians. The bold lines depict
the 90% confidence intervals around the point estimates of
the differences in expected probabilities, while the thinner
and slightly longer lines show the 95% confidence range.
The predictive power of the social network variables is
apparent here, reinforcing the importance of the network
characteristics. The network size and ideological leanings
show large differences in the expected probability of fore-
cast accuracy, differences that are rivaled only by political
interest and respondent vote intention for the CDU or not
known. For instance, if we compare a respondent who has
five network members with someone who has only one
network member (the maximum and minimum values for
network size), we expect the respondent with the larger
network to have a 15 percentage point higher chance of
making a correct forecast on average. As another example,
if we compare a respondent whose network consists of
only left-leaning members with someone whose network
consists of no left-leaning members, we expect the one
with the more left-leaning network to have a 16 percent-
age point smaller chance of making a correct forecast on
average. (Table A1 in the online appendix provides the
D. Leiter et al. / International Journal of Forecasting 34 (2018) 235–248 243
differences in expected probabilities and their confidence
intervals that correspond to this figure.)
6.2. Correct and incorrect forecasts and the ‘‘don’t knows’’
Next, we recognize that not all ‘‘wrong’’ forecasts are
the same. A respondent could either provide an incorrect
forecast of an SPD victory, or report not knowing who will
win, and the covariates that predict these results are likely
to be different. We assess this by estimating multinomial
logit models where those who offer incorrect (SPD) or
uncertain (don’t know) responses are assessed separately
relative to those who forecasted correctly. We estimate this
both for the pooled survey and in the form of an interactive
model where we assess whether differences exist between
East and West Germans when it comes to the coefficients
of the various predictors.
Again, we explore the results of the multinomial logit
model further by computing first differences.11 Fig. 3
presents the differences in expected probabilities and the
corresponding confidence intervals for each predictor and
each forecast (CDU, SPD, don’t know), based on the esti-
mates of the pooled multinomial logit model (full results
are reported in Table A2 in the online appendix). We ob-
serve that the social network variables differ in their pre-
dictive power across the three distinct forecasts. In general,
we observe that respondents whose networks displayed a
higher share of left or unknown leanings or lower levels
of expertise were more likely to provide an incorrect SPD
forecast. In contrast, respondents who had less frequent
discussions with those in their network were more likely
to give ‘‘don’t know’’ responses. Specifically, if we compare
a respondent whose network has five members to some-
one whose network has one member, we expect that the
respondent with the larger network will be 14 percentage
points more likely to make a correct CDU forecast, 11 per-
centage points less likely to give a ‘‘don’t know’’ response,
and 2 percentage points less likely to make an incorrect
SPD forecast. In other words, we expect that respondents
with varying network sizes will differ in their chances of
giving a CDU forecast or a ‘‘don’t know’’ response, but that
they will be similar in their chances of giving a SPD forecast
on average. In summary, the larger the network, the more
accurate and certain citizen forecasts are. (Table A3 in the
online appendix reports the differences in probabilities and
the values of the 95% confidence intervals.)
Fig. 3 also shows large differences in expected proba-
bilities for respondents who differ in their vote intentions
and levels of political interest. Comparing two respondents
with high and low levels of political interest, we expect
that the one who is more interested in politics will have a
27 percentage points higher chance of correctly forecasting
the CDU to win and a 27 percentage points lower chance of
a ‘‘don’t know’’ response, but will not differ in the proba-
bility of an incorrect SPD forecast on average.12
11 Online Appendix Table A2 reports the full results of our pooled and
interactive multinomial logit models.
12 While we cannot reject the null hypothesis that the first difference
for political interest is the same for network size related to a CDU response
(b=0.13; Std. Err. =0.09; z=1.54) or a SPD response (b=0.04, Std. Err.
=0.03; z=1.34), we can reject the null hypothesis for a ‘‘don’t know’’
response (b= −0.17; Std. Err. =0.08; z= −1.98).
So far, the binary and multinomial logit models and
the difference in expected probability figures have demon-
strated that social network characteristics are highly pre-
dictive of the accuracy of an election forecast, and can help
us to distinguish between incorrect forecasts and respon-
dent uncertainty. These network measures, in addition to
political interest and vote intentions, by far outperform
demographics and media variables. The number of days
before the election that the interview took place is not pre-
dictive of the type of prediction given by the respondent.
6.3. Allowing the coefficients to vary between East and West
Germans
German reunification ended 40 years of political di-
vision between East and West Germany. It has been of
general interest to describe the similarities and differences
in public opinion and behavior between East and West
Germans in order to understand the extent to which the
country has developed a unified political culture (e.g., Fal-
ter, Gabriel, Rattinger, & Schoen, 2006;Fuchs, Roller, &
Wessels, 2002;Gabriel, 1997;Gabriel, Falter, & Rattinger,
2005;van Deth, Rattinger, & Roller, 2000). In our context,
we expect East Germans to rely more on social network
information than West Germans, given the challenges that
new democracies are likely to face, such as weak partisan
cues, low levels of partisan identification, and volatile vot-
ers (Baker, Ames, & Renno, 2006). Hence, we now examine
whether the coefficients of our predictors differ between
the East and West.
We examine possible heterogeneous coefficients be-
tween East and West by following the recommendations
of Tsai and Gill (2013) on interactions in generalized lin-
ear models. We first add product terms between each
of the predictors and the East dummy variable to the
pooled multinomial logit regression equation (the last two
columns of Table A2 display the estimates of this interacted
multinomial logit model), then calculate first differences
of the predictors, separately for East and West. Finally,
we compare the first differences of a predictor between
East and West to assess the statistical significance and
magnitude of the interaction (Figure A1 and Table A4 in the
online appendix show all of these first differences).
By following this procedure, we found statistically sig-
nificant interactions for only two network variables (the
size of the network and the share of the network with left
political leanings) on just one outcome (‘‘don’t know’’). In
other words, of the 18 possible interactions – six network
variables multiplied by three outcomes – 16 are statisti-
cally insignificant. Since we would expect one to be statisti-
cally significant by chance out of 20 such comparisons, we
do not want to emphasize the differences that we found.
Thus, the results of the interacted model suggest that there
are no major differences in how network characteristics
predict forecast accuracy between East and West Germans:
social networks predict the forecast accuracy in the same
way for both groups.13
13 We also considered possible interactions between the most impor-
tant predictors (Gelman & Hill, 2007, p. 69): network size, network dis-
cussion, and network left, as well as political interest and vote intention.
244 D. Leiter et al. / International Journal of Forecasting 34 (2018) 235–248
Fig. 3. Difference in expected probabilities for the pooled multinomial logit model. Note: Difference in expected probabilities of a CDU, don’t know, or
SPD forecast between two respondents with maximum and minimum values of the indicated predictor, while holding the other variables constant at their
median value. Predictors are sorted by increasing difference in expected probability on giving a CDU response, separately for network characteristics and
controls. Bold segments indicate 90% confidence intervals and thin segments indicate 95% confidence intervals.
7. A simple network measure for improving accuracy of
out-of-sample predictions
The analysis above described which citizens were most
likely to forecast the election correctly. Next, we would
like to provide guidance for people who want to use citi-
zen forecasts to forecast future election outcomes. As has
been mentioned, aggregated citizen forecasts are most ac-
curate when individual forecasters are weighted by their
forecasting competence. The analysis above improves the
researcher’s ability to identify which individuals to weight
We tested whether the network variables interact with each other or
with the other predictors, again following the procedure recommended
by Tsai and Gill (2013). (In the online appendix, Tables A5, A6 and A8 show
the estimated regression models, while Tables A7 and A9 and Figure A2
to A6 show the first differences.) We found one statistically significant
interaction: the importance of the frequency of discussion decreases with
higher levels of political interest for the outcomes CDU and don’t know.
more heavily: because social network characteristics pre-
dict forecasting competence, future aggregated citizen
forecasts can be made more accurate by using these net-
work characteristics to calculate the individual weights.
However, network batteries take a great deal of space
on a questionnaire. The survey that we used in our analysis
included five questions identifying network members, as
well as follow-up items for each member thus identified,
measuring their political preference, expertise, frequency
of discussion, etc. Is including network batteries in new
surveys worthwhile in terms of improving the election
forecasting accuracy? Below, we show that even a single,
abbreviated measure of the network size – asking citizens
how many people they discussed an important personal
matter with – improves out-of-sample predictions.
We compared the out-of-sample predictive accuracies
of all possible subsets of the predictors considered above,
with three modifications. First, as the response variable, we
chose whether the citizen correctly forecasted the winner
D. Leiter et al. / International Journal of Forecasting 34 (2018) 235–248 245
(0 =‘‘no’’; 1 =‘‘yes’’), excluding the response ‘‘don’t know’’
because only actual forecasts can be weighted. Second, we
considered the network size (0 =‘‘no discussants’’ to 5 =
‘‘five discussants’’) as the only network characteristic. We
do this because the above descriptive analysis found the
size to be correlated strongly with the forecasting accuracy,
and because this predictor also applies to citizens without
a discussant, while the other network characteristics apply
only to citizens with at least one discussant. (Excluding
‘‘don’t knows’’ and including citizens without networks
changes the number of observations to 1592.) Finally, we
replaced the three vote intention predictors with a single
dummy variable indicating whether a citizen forecasted
the same party to win as the one they intended to vote for
(0 =‘‘no’’; 1 =‘‘yes’’). We do this because this predictor
can be used without the researcher knowing in advance
which party will win (Murr, 2015). This leaves us with ten
predictors: east, age, female, education, political interest,
TV news attention, print news attention, forecast intention,
days until election, and network size.
We used k-fold cross-validation (e.g., Murr, 2015;Ward,
Greenhill, & Bakke, 2010) to compare the out-of-sample
predictive accuracies of all 210 =1024 possible subsets of
predictors. Cross-validation splits the data randomly into
kfolds. It first fits the models to the k1 folds and then
tests them on the kth one, iterating these two steps from
1 to kto get a distribution of the predictive accuracy. We
set k=10, which is the typical value in the literature, and
repeated k-fold cross-validation with ten different splits.
We measured the predictive accuracy based on the area
under the receiver operator characteristic curve (AUROC),
which is a common measure of accuracy in the forecasting
literature for binary classification tasks (e.g., Murr, 2015;
Ward et al., 2010). An AUROC value of 50% indicates a
random classifier and a value of 100% indicates an optimal
classifier. The AUROC can be interpreted as the probability
that a randomly chosen correct citizen forecaster will be
ranked as more likely to be correct than a randomly chosen
incorrect citizen forecaster (Fawcett, 2006).
Including the network size as a predictor improved the
predictive accuracy (Table 5). Overall, the model with the
largest AUROC of 62.57% included only five of the nine
predictors: age, TV news attention, forecast intention, days
until election, and network size. In contrast, the best model
excluding the network size achieved an AUROC of 61.40%
– 1.17 percentage points lower than the best model includ-
ing the network size. Averaging across all 1024 models,
the AUROC of models including the network size was 1.4
percentage points higher than that of models excluding
the network size. In comparison, only forecast intention
and age had larger increases, of 3.98 and 3.22 percentage
points, respectively. Including some predictors even de-
creased the predictive accuracy on average. For instance,
the AUROC of models including print news attention was
an average of 0.17 percentage points lower than that of
models excluding print news attention. This all demon-
strates that it is worthwhile to include the network size as
a measure on new surveys because it does a better job of
predicting forecasting competence than many commonly
available measures (e.g., print news attention). As elections
grow increasingly competitive and election results grow
tighter, even minor improvements in forecasting measure-
ments may be critical in increasing the forecast accuracy.
8. Conclusion
This study has examined how social networks predict
the ability of citizens to forecast the election winner cor-
rectly when controlling for other variables such as politi-
cal interest, gender, education, media attention, and vote
intention. Specifically, we have found that citizens who
have larger social networks and engage in more frequent
political discussions are better at forecasting the winner
than people who do not share these network characteris-
tics. Our analysis also shows that the political leanings of
the network matter too. Those whose networks contained
a higher proportion of left-wing party supporters were
less likely to forecast (correctly) that the right-wing CDU
would win. Furthermore, respondents who were unsure of
their friends’ party preferences were less likely to provide
correct forecasts. Essentially, voters with extensive, com-
municative, and varied groups of friends – and, of course,
neighbors, colleagues, family members, and peers – are
best able to forecast the election winner accurately.
Finding such robust results for social network char-
acteristics might be a surprise in this particular election,
given that public opinion polls at the time of the survey
in autumn 1990 pointed to a decisive CDU victory. We
view this particular election as a conservative test of our
social networks theory. With the election all but a fore-
gone conclusion, one might expect the predictive power
of social networks for the respondents’ forecasts to be
limited. However, even in this context, networks demon-
strably predicted citizens’ forecasts. In more competitive
elections, where there is a greater degree of uncertainty
about the likely winner, social networks and their charac-
teristics would probably play an even more important role
in predicting voters’ election forecasts.
In addition to examining the predictive power of so-
cial network characteristics for election forecasts, we have
also considered how experience with democratic elections
might predict citizens’ abilities to give an accurate forecast,
based on whether a respondent resided in East or West
Germany. Perhaps surprisingly, East Germans were more
likely to forecast the victor correctly than West Germans.
Also, while we might have expected that having had less
democratic experience would mean that networks were
more important for East Germans than for West Germans
in predicting their expectations – given the challenges
faced by new democracies (e.g., weak partisan cues, low
levels of partisan identification and volatile voters, as was
discussed by Baker et al., 2006), our analysis indicates that
no such differences exist.
The robustness of our findings in both East and West
Germany suggests that the predictive power of social
networks should be present in both new and estab-
lished democracies. However, since the institutional and
political contexts of the 1990 German election are the
same for both regions, future research should examine
whether social networks predict citizen forecasts similarly
in countries with different party systems and electoral
rules.
Future research could study how the internet and the
emergence of online social networks have influenced citi-
246 D. Leiter et al. / International Journal of Forecasting 34 (2018) 235–248
Table 5
Out-of-sample accuracy of all 1,024 possible subsets of variables for predicting correct forecasts (0 =‘‘no’’; 1 =‘‘yes’’) by 1,592 citizens before the German
Bundestag election in 1990 using binary logistic regression.
Rank Predictors (0 = ‘‘excluded’’; 1 = ‘‘included’’)
East Age Fem. Educ. Pol. int. TV Print Forec. intent. Days Net. size AUROC
1 0 1 0 0 0 1 0 1 1 1 62.57
2 0 1 0 0 1 0 0 1 1 1 62.48
3 0 1 0 0 0 0 0 1 1 1 62.36
4 0 1 0 0 1 0 0 1 0 1 62.32
5 0 1 0 0 0 1 0 1 0 1 62.28
6 0 1 0 0 1 1 0 1 1 1 62.27
7 0 1 0 0 0 1 1 1 1 1 62.21
8 0 1 0 0 0 0 1 1 1 1 62.19
9 0 1 0 1 0 1 0 1 1 1 62.18
10 0 1 0 1 0 0 0 1 1 1 62.18
...
91 0 1 0 0 0 1 0 1 1 0 61.40
...
Note: Entries are sorted by decreasing average area under the receiver operator characteristic curve (AUROC) in 10-fold cross-validation across ten
repetitions and by decreasing the number of predictors. Due to space constraints, only the ten best models are presented, along with the best model
without the network size as a predictor.
zens’ forecasting abilities. Some studies have shown that
the internet has neither increased nor decreased social
capital, but instead supplemented it (e.g., Wellman, Haase,
Witte, & Hampton, 2001). Hence, citizens still seem to
bond (form closer connections with others) and bridge
(form ties across social groups) to the same extent as be-
fore. Other studies have shown a large overlap between
offline and online social networks (e.g., Subrahmanyam,
Reich, Waechter, & Espinoza, 2008), meaning that the net-
works elicited using electoral surveys are likely to be a
subset of those captured in online social networks. Online
platforms are likely to increase citizens’ abilities to fore-
cast because they provide a wider access to information
without additional cost. They enable citizens to be updated
about their networks’ electoral preferences without face-
to-face discussions, and allow citizens to be informed about
all of their network members, even those who are distant
from the most influential people in their network.
A final lesson of our analysis is that social network
characteristics, and questions on citizen forecasting, are
important elements in electoral surveys, and that their ex-
clusion may inhibit our understanding of political learning
and decision making. The size and composition of social
networks are associated with citizens’ ability to forecast
elections correctly, and understanding how and why cit-
izens estimate the winners of elections correctly will be
critical as the demand for political forecasting continues.
In the absence of measures of social network character-
istics, we cannot predict or utilize these forecasts fully.
In addition, understanding citizen forecasting also reveals
something important about how social networks predict
political learning. The size and ideological make-up of net-
works compete with other factors in predicting whether
citizens can make correct inferences about not just local,
but also national, political trends. In summary, just as social
networks help us to understand citizen forecasting, citizen
forecasting informs us about how social networks predict
contextual learning and political knowledge.
Appendix A. Question wording appendix
Outcome variable
Forecasting
From the present point of view: who would you say will
win the next general election: The CDU/CSU or a coalition
government led by CDU/CSU, or the SPD or a coalition
government led by the SPD?
Network variables
Network size
From time to time, most people discuss important per-
sonal matters with other people. Looking back over the last
six months, who are the people with whom you discussed
an important personal matter?
Network frequency
When you talk with these persons, how often do you
discuss political questions? Would you say almost always,
sometimes, seldom, or never?
Network expertise
How much do these persons, in your opinion, know
about politics: much or very much, average, less much?
Network ideology
Which party do you think would these persons vote for
in the general election of 2 December this year?
Individual level
Vote choice
Second Vote: Which party will you vote for with your
second vote?
Political interest
Generally Speaking: How much are you interested in
politics? Would you say: very much, much, so-so, some-
what, or not at all?
TV news attention
How attentively do you follow [television] news reports
on political events in Germany and other countries? Would
D. Leiter et al. / International Journal of Forecasting 34 (2018) 235–248 247
you say: very attentively, attentively, less attentively, or
not attentively at all?
Print news attention
Regardless of how often you read your daily newspaper:
How attentively do you read the reports on the political
events in Germany and other countries? Would you say:
very attentively, attentively, less attentively, or not atten-
tively at all?
Education
What education level do you have?
Age
Please tell me what month and year you were born
Gender
Sex of Respondent: Man or Woman.
Appendix B. Supplementary data
Supplementary material related to this article can be
found online at https://doi.org/10.1016/j.ijforecast.2017.
11.006.
References
Ahn, T. K., Huckfeldt, R., & Ryan, J. B. (2014). Experts, activists, and demo-
cratic politics: Are electorates self-educating? Cambridge University
Press.
Baker, A., Ames, B., & Renno, L. R. (2006). Social context and cam-
paign volatility in new democracies: networks and neighborhoods
in Brazil’s 2002 elections. American Journal of Political Science,50(2),
382–399.
Banerjee, A., & Fudenberg, D. (2004). Word-of-mouth learning. Games and
Economic Behavior,46(1), 1–22.
Blais, A., & Bodet, M. A. (2006). How do voters form expectations about
the parties’ chances of winning the election? Social Science Quarterly,
87(3), 477-493.
Campbell, J. E. (2016). The trial-heat and the seats-in-trouble forecasts of
the 2016 presidential and congressional elections. PS: Political Science
and Politics,49, 664–668.
Chandra, K. (2009). Why voters in patronage democracies split their
tickets: strategic voting for ethnic parties. Electoral Studies,28(1),
21–32.
Condorcet, Marquis de M. J. A. N. d. C. (1785). Essai sur l’application de
l’analyse a la probabilité des descisions redues a la pluralité de voix. Paris:
De l’Imprimerie Royale.
Entman, R. M. (1989). How the media affect what people think: an infor-
mation processing approach. The Journal of Politics,51, 347–370.
Erikson, R. S., & Wlezien, C. (2016). Forecasting the presidential vote with
leading economic indicators and the polls. PS: Political Science and
Politics,46, 669–672.
Eveland, W. P. (2004). The effect of political discussion in producing in-
formed citizens: the roles of information, motivation, and elaboration.
Political Communication,21(2), 177–193.
Eveland, . W. P., & Hively, M. H. (2009). Political discussion frequency,
network size, and ‘heterogeneity’ of discussion as predictors of po-
litical knowledge and participation. Journal of Communication,59(2),
205–224.
Eveland, W. P., Jr., , Hutchens, M. J., & Morey, A. C. (2013). Political network
size and its antecedents and consequences. Political Communication,
30(3), 371–394.
Falter, J. W., Gabriel, O., Rattinger, H., & Schoen, H. (Eds.). (2006). Sind wir
ein Volk? Ost- und Westdeutschland im Vergleich. Munich: Beck.
Fawcett, T. (2006). An introduction to ROC analysis. Pattern Recognition
Letters,27(8), 861–874.
Finkel, S. E., & Smith, A. E. (2011). Civic education, political discussion, and
the social transmission of democratic knowledge and values in a new
democracy: Kenya 2002. American Journal of Political Science,55(2),
417–435.
Ford, R., Jennings, W., Pickup, M., & Wlezien, C. (2016). From polls to votes
to seats: forecasting the 2015 British general election. Electoral Studies,
41, 244–249.
Forschungsgruppe Wahlen, Mannheim (2017). Partial cumulation of
politbarometers 1977-2015 GESIS Data Archive, Cologne. ZA2391
Data file Version 700. http://dx.doi.org/10.4232/1.12733.
Fuchs, D., Roller, E., & Wessels, W. (Eds.). (2002). Bürger und Demokratie in
Ost und West: Studien zur politischen Xultur und zum politischen Prozess.
Wiesbaden: Westdeutscher Verlag.
Gabriel, O. W. (Ed.). (1997). Politische Orientierungen und Verhaltensweisen
im vereinigten Deutschland. Opladen: Leske + Budrich.
Gabriel, O. W., Falter, J. W., & Rattinger, H. (Eds.). (2005). Wächst zusam-
men, was zusammengehört? Stabilität und Wandel politischer Einstel-
lungen im wiedervereinigten Deutschland. Baden-Baden: Nomos.
Gelman, A., & Hill, J. (2007). Data analysis using regression and multi-
level/hierarchical models. Cambridge University Press.
Golub, B., & Jackson, M. O. (2010). Naive learning in social networks and
the wisdom of crowds. American Economic Journal: Microeconomics,
2(1), 112–149.
Graefe, A. (2014). Accuracy of vote expectation surveys in forecasting
elections. Public Opinion Quarterly,78(S1), 204–232.
Grofman, B. (1975). A comment on ‘Democratic theory: a preliminary
mathematical model’. Public Choice,21, 99–103.
Grofman, B., Owen, G., & Feld, S. L. (1983). Thirteen theorems in search of
the truth, theory and decision. Theory and Decision,15(3), 261–278.
Gunther, R., Beck, P. A., Magalhães, P. C., & Moreno, A. (Eds.). (2015). Voting
in old and new democracies. Routledge.
Gunther, R., Puhle, H.-J., & Montero, J. R. (2007). Democracy, intermediation,
and voting on four continents. Oxford University Press.
Huckfeldt, R. (2007). Unanimity, discord, and the communication of public
opinion. American Journal of Political Science,51(4), 978–995.
Huckfeldt, R., Ikeda, K., & Pappi, F. U. (2005). Patterns of disagreement
in democratic politics: comparing Germany, Japan, and the United
States. American Journal of Political Science,49(3), 497–514.
Huckfeldt, R., & Johnson, P. E. (2004). Political disagreement: The survival of
diverse opinions within communication networks. Cambridge Univer-
sity Press.
Huckfeldt, R., & Mendez, J. M. (2008). Moths, flames, and political engage-
ment: managing disagreement within communication networks. The
Journal of Politics,70(1), 83–96.
Huckfeldt, R., Pietryka, M. T., & Reilly, J. (2014). Noise, bias, and expertise
in political communication networks. Social Networks,36, 110–121.
Huckfeldt, R., & Sprague, J. (1991). Discussant effects on vote choice:
intimacy, structure, and interdependence. The Journal of Politics,53(1),
122–158.
Huckfeldt, R., & Sprague, J. (1995). Citizens, politics and social communi-
cation: Information and influence in an election campaign. Cambridge
University Press.
Janis, I. L. (1982). Groupthink: Psychological studies of policy decisions and
fiascoes. Boston: Houghton Mifflin.
Jérôme, B., Jérôme-Speziari, V., & Lewis-Beck, M. S. (2017). The grand
coalition reappointed by Angela Merkel on borrowed time. PS: Political
Science and Politics,50, 683–685.
Kazmann, R. G. (1973). Democratic organization: a preliminary mathe-
matical model. Public Choice,16(1), 17–26.
King, G. (1989). Unifying political methodology: The likelihood theory of
statistical inference. Cambridge: Cambridge University Press.
Kwak, N., Williams, A. E., Wang, X., & Lee, H. (2005). Talking politics and
engaging politics: an examination of the interactive relationships be-
tween structural features of political talk and discussion engagement.
Communication Research,32(1), 87–111.
Ladha, K. K. (1992). The Condorcet Jury Theorem, free speech, and corre-
lated votes. American Journal of Political Science,36(3), 617–634.
Lewis-Beck, M. S., & Skalaban, A. (1989). Citizen forecasting: can voters
see into the future? British Journal of Political Science,19(1), 419–427.
Lewis-Beck, M. S., & Stegmaier, M. (2011). Citizen forecasting: can UK
voters see the future? Electoral Studies,30(2), 264–268.
248 D. Leiter et al. / International Journal of Forecasting 34 (2018) 235–248
Lewis-Beck, M. S., & Stegmaier, M. (2014). US presidential election fore-
casting. PS: Political Science and Politics,47, 284–288.
Lewis-Beck, M. S., & Tien, C. (1999). Voters as forecasters: a micro-
model of election prediction. International Journal of Forecasting,15(2),
175–184.
List, C., & Goodin, R. E. (2001). Epistemic democracy: generalizing the
Condorcet Jury Theorem. Journal of Political Philosophy,9(3), 277–306.
Manski, C. F. (2004). Social learning from private experiences: the dynam-
ics of the selection problem. The Review of Economic Studies,71(2),
443–458.
Marsden, P. V. (2003). Interviewer effects in measuring network size using
a single name generator. Social Networks,25(1), 1–16.
McClurg, S. D. (2006). The electoral relevance of political talk: examining
disagreement and expertise effects in social networks on political
participation. American Journal of Political Science,50(3), 737–754.
Meffert, M. F., Huber, S., Gschwend, T., & Pappi, F. U. (2011). More than
wishful thinking: causes and consequences of voters’ electoral expec-
tations about parties and coalitions. Electoral Studies,30, 804–815.
Merluzzi, J., & Burt, R. S. (2013). How many names are enough? Identifying
network effects with the least set of listed contacts. Social Networks,
35(3), 331–337.
Miller, M. K., Wang, G., Kulkarni, S. R., Poor, H. V., & Osherson, D. N. (2012).
Citizen forecasts of the 2008 U.S. presidential election. Politics and
Policy,40, 1019–1052.
Millner, A., & Ollivier, H. (2016). Beliefs, politics, and environmental policy.
Review of Environmental Economics and Policy,10(2), 226–244.
Murr, A. E. (2011). ‘Wisdom of crowds’? A decentralized election fore-
casting model that uses citizens’ local expectations. Electoral Studies,
30(4), 771–783.
Murr, A. E. (2015). The wisdom of crowds: applying Condorcet’s Jury
Theorem to forecasting US presidential elections. International Journal
of Forecasting,31(3), 916–929.
Murr, A. E. (2016). The wisdom of crowds: what do citizens forecast for
the 2015 British general election? Electoral Studies,41, 283–288.
Murr, A. E. (2017). ‘Wisdom of crowds’. In K. Arzheimer, J. Evans, & M. S.
Lewis-Beck (Eds.), Handbook of political behavior (pp. 835–860). Sage.
Murr, A. E., Stegmaier, M., & Lewis-Beck, M. S. (2016). Citizen forecasting
v. pocketbook forecasting: are two heads better than one? Paper
presented at the Midwest Political Science Association Conference.
Mutz, D. C. (1998). Impersonal influence: How perceptions of mass collectives
affect political attitudes. Cambridge University Press.
Nickerson, D. W. (2008). Is voting contagious? Evidence from two field
experiments. American Political Science Review,102(1), 49–57.
Norpoth, H., & Gschwend, T. (2017). Chancellor model predicts a change
of the guards. PS: Political Science and Politics,50, 686–688.
Partheymüller, J., & Schmitt-Beck, R. (2012). A ‘social logic’ of demobiliza-
tion: the influence of political discussants on electoral participation at
the 2009 German federal election. Journal of Elections, Public Opinion
and Parties,22(4), 457–478.
Pattie, C., & Johnston, R. (1999). Context, conversation and conviction: so-
cial networks and voting at the 1992 British general election. Political
Studies,47(5), 877–889.
Pietryka, M. T. (2015). Accuracy motivations, predispositions, and social
information in political discussion networks: accuracy motivations in
political discussion. Political Psychology,37(3), 367–386.
Pulzer, P. (1991). The German federal election of 1990. Electoral Studies,
10(2), 145–154.
Ryan, J. B. (2011). Social networks as a shortcut to correct voting. American
Journal of Political Science,55(4), 753–766.
Schmitt-Beck, R., & Mackenrodt, C. (2010). Social networks and mass
media as mobilizers and demobilizers: a study of turnout at a German
local election. Electoral Studies,29(3), 392–404.
Shapley, L., & Grofman, B. (1984). Optimizing group judgmental accuracy
in the presence of interdependencies. Public Choice,43(3), 329–343.
Stegmaier, M., & Norpoth, H. (2017). Election forecasting. In R. Valelly
(Ed.), Oxford bibliographies in political science. Oxford University Press.
Stegmaier, M., & Williams, L. (2016). Forecasting the 2015 British election
through party popularity functions. Electoral Studies,41, 260–263.
Subrahmanyam, K., Reich, S. M., Waechter, N., & Espinoza, G. (2008).
Online and offline social networks: use of social networking sites by
emerging adults. Journal of Applied Developmental Psychology,29(6),
420–433.
Tsai, T., & Gill, J. (2013). Interactions in generalized linear models: theo-
retical issues and an application to personal vote-earning attributes.
Social Sciences,2, 91–113.
Uhlaner, C. J., & Grofman, B. (1986). The race may be close but my horse
is going to win: wish fulfillment in the 1980 presidential election.
Political Behavior,8(2), 101–129.
van Deth, J., Rattinger, H., & Roller, E. (Eds.). (2000). Die Republik auf dem
Weg zur Normalität? Wahlverhalten und politische Einstellungen nach
acht Jahren Einheit. Opladen: Leske + Budrich.
Ward, M. D., Greenhill, B. D., & Bakke, K. M. (2010). The perils of policy
by p-value: predicting civil conflicts. Journal of Peace Research,47,
363–375.
Wellman, B., Haase, A. Q., Witte, J., & Hampton, K. (2001). Does the Internet
increase, decrease, or supplement social capital? Social networks,
participation, and community commitment. American Behavioral Sci-
entist,45(3), 436–455.
Debra Leiter is assistant professor in the Department of Political Science
at the University of Missouri-Kansas City. Her research focuses on the
intersection of elections, voters, and political parties, with an emphasis
on Western European electorates.
Andreas Murr is assistant professor of quantitative political science in
the Department of Politics and International Studies at the University of
Warwick. His research focuses on election forecasting, the voting behavior
of immigrants and the selection of party leaders.
Ericka Rascón Ramírez is lecturer in development economics at the
Department of Economics at Middlesex University London. Her research
focuses on how behavioral insights influence child development and
young people’s behavior, as well as predict electoral and educational
outcomes.
Mary Stegmaier is assistant professor in the Truman School of Public
Affairs at the University of Missouri. Her research focuses on voting
behavior, elections, forecasting, and political representation in the U.S.
and abroad.