Making sense of pollsters’ errors. An analysis of the 2014 second-order
European election predictions
aDepartment of Social and Political Sciences, Università degli studi di Milano, Milano,
Via Conservatorio 7, 20122 Milano, Italy; firstname.lastname@example.org
PhD in Political Science (1990) from the University of Florence (Italy), full professor at the
University of Milano where he teaches Comparative politics. His main research topics are the
comparative analysis of the performance of democracies, electoral behavior and EU politics and
Pollsters have been recently accused of delivering poor electoral predictions. We
argue that one of the reasons for their failures lies in the difficulty of including an
updated deep understanding of electoral behaviour. Even if pollsters’ predictions
are not forecasts produced by models, the set of choices needed to produce their
estimates is not indifferent to a theoretical comprehension of electoral dynamics.
We exemplify this lack of theory by using an original dataset consisting of 1057
party*poll observations in the case of the last European election. Pollsters failed
to account for what we know about second-order elections, thus overestimating
government and big parties, which normally obtain poor results in European
elections, and underestimating new and Eurosceptic ones, which usually perform
Keywords: Election polls; Accuracy; Second-order elections; European politics
Pollsters have had some hard times lately. It seems that the more their work becomes
influential, even endogenously affecting what they are trying to predict, the less it is
reliable. They did not anticipate very diverse outcomes, such as Trump’s US
presidency, Brexit, Fillon winning the primaries for the Republican candidate in France,
the Conservative victory in the UK 2015 general election, the success of Likud in Israel
in that same year, or the partial stalemates produced in the Italian 2013 election and in
the German 2017 one. And the list could easily continue. Their failures are often
attributed to a set of contingencies, like a “perfect storm” conspiring against their
prediction capabilities.1 But there is probably a different set of factors involved.
Scholars have identified the technical elements that affect pollsters’ predictions
and make their work complex (Voss, Gelman and King 1995). Pooling the polls
(Jackman 2005) is often seen as a remedy against potential biases due to house effects,
and a strategy for tracking actual changes in voter support. However, it would not have
been enough in the cases cited above. Deriving predictions from pre-election surveys is
not a straightforward task. A series of corrections and adjustments are needed in order
to produce a plausible guess from a sample which is probably non-representative of
relevant characteristics of the population, not least the willingness to respond to
surveys. Post-stratification weights are applied to correct for that well-known problem.
“Pollsters also employ weights to adjust for the likelihood that respondents will cast a
ballot, the differential response to polls between supporters of different parties, and their
differential willingness to admit preferences to the pollsters. (…) Besides weighting,
there are other practices that can alter error variance, (…and) lead to a net reduction or
net increase in error variance compared to a purely random sample” (Fisher et al. 2011,
A set of implicit assumptions and consolidated knowledge is crucial in order to
apply those corrections and adjustments. This wisdom pertains to traditional electoral
behaviour, local dynamics, general trends, etc. It is mostly derived from previous
experiences, and should be constantly updated. Yet, if society moves through troubled
and unparalleled circumstances – a deep economic crisis, a series of external shocks,
new world dynamics – the past may be only marginally useful for predicting the future.
In that event, a deeper theoretical understanding is needed in order to realign that
wisdom with the new shaken practices. We argue that this is precisely what the above-
cited pollsters’ failures have in common. After a prolonged period of social unrest – due
to the Great Recession, international tensions, migration and globalization dynamics –
social scientists need to go back to some deeper theoretical understanding of the
dynamics investigated in order to regain the capacity to adjust their predictions
positively. It is probably not by chance that political scientists’ forecasts were closer to
the actual results of the US Presidential competition than pollsters’ predictions
In this article, we test this intuition by using ex-post the 2014 European elections
as a case study. Its advantage is that of having parallel competitions in 28 member
states, with very different polling firms making projections on a wide range of diverse
parties. Because of the complexity of multiple multi-party ballots, it is a crucial case
study, with the further benefit of having a well-established interpretation of voters’
behaviours in the theory of second-order elections. We argue that failing to account for
that theoretical understanding was a systematic part of the misfit between predictions
and actual results, which, though not as dense with consequences as in the cases cited,
was not trivial, with an average discrepancy of more than 2%. The systematic bias in the
predictions that we ascertain in what follows is thus the outcome of mechanically
applying long-established polling practices developed in several rounds of national
general elections; practices that are not (or no longer) appropriate to this specific
electoral context. If voting behaviour is at least partially context-dependent, with
citizens making up their minds in different moments and following different frames, the
same should be said of the polling practices and adjustments mentioned before by
Fisher and colleagues.
The article is organized as follows. In the following section, we present the
theory of second-order elections applied to European ballots. From that theory we
derive five hypotheses, which represent the symmetrical expectations of pollsters’ errors
regarding the electoral performances of different types of party. Next, we introduce our
original dataset collecting the predictions of several different polling firms in the 28 EU
member states in the event of the May 2014 common European election. We further
specify the independent and control variables that will be used in our econometric
models, and, for robustness, propose several different dependent variables measuring
the gap between predictions and actual results. Finally, we present the results of those
models, which will be further summarized and discussed in the concluding section.
Theory and Hypotheses
Scholars of European politics generally agree that elections for the European parliament
should be considered second-order elections (Marsh 1998, Reif and Schmitt 1980).
Citizens generally perceive them as less important than national general elections
because there is less at stake. Most importantly, notwithstanding the institutional
innovations introduced by the Lisbon Treaty, their results are not directly connected to
the investiture of a common European executive. In spite of the evolution of the EU,
and of the politicization of that governance level, most scholars agree that even the most
recent elections adhere to the second-order framework (e.g. Hix and March 2011;
Schmitt and Toygür 2016). As a consequence, voters may consider behaving in the
European polls differently from their habits during their national general elections.
First of all, they may simply decide not to vote. For this reason, the level of
turnout is generally lower in European elections. In the last 2014 European election, the
turnout in all member states except Belgium – which simultaneously ran its national and
local election – was lower than in their previous general elections. On average, the
difference was almost 25% of the voting age population. This probably does not
significantly affect the legitimization of the European parliament (van der Eijk and
Schmitt 2009), which in many countries is trusted more than the national counterpart,
but disturbs the work of pollsters for three different interconnected reasons. Firstly,
because it is more difficult to guess the behaviour of a smaller non-random sample of
the population, which is different from the one more usually surveyed for national
elections. Secondly, because turnout may affect the electorates of the diverse parties in
different ways, not necessarily coincident with what happens in domestic elections.
Lastly, because we do not know much about the uneven disposition of those electorates
to declare sincerely their lack of interest in participating in second-order elections.
This strand of literature suggests a second behavioural option which is a direct
consequence of the less-at-stake interpretation of second-order elections: that of voting
more sincerely. This conduct is normally considered favoured also by institutional
elements, such as the specific design of the electoral system which features a
proportional formula with comparatively large districts, as is confirmed by comparing
the so-called ‘effective threshold’ in the two types of elections. This set-up relatively
favours, or does not discourage, the performance of small, and even new parties.
However, indices synthesising the disproportionality of the system, normally associated
with strategic behaviours, do not entirely confirm that interpretation. The average
Gallagher index for European and national elections in our sample is not systematically
different, with some countries having a higher and others a lower apparent level of
permissiveness. Yet, the actual value for the European parliament could have been
endogenously produced by the increased competition amongst a larger number of
political groups, as demonstrated by the fact that the effective number of electoral
parties is substantially higher, almost 1 point, for the European election compared to the
national one, something that indirectly confirms the favourable institutional
environment of European elections for small and new parties.
Moreover, there are further, political more than institutional, reasons why we
should observe a greater success of new or usually marginalized parties during
European elections. The literature normally connects them with the specular suffering
of government parties: on the one hand, because the latter cannot appeal to any sense of
responsibility, or to the risk of wasting votes in those electoral appointments, as they
can profitably do in the national arena; on the other, because voters may profit from the
opportunity to experiment with a different vote. This means that citizens choose parties
closer to their preferences, yet even without any domestic coalitional potential, and they
empower leaders, political groups and MEPs whose actions and choices they will
evaluate in the less sensitive arena of Strasbourg and Brussels.
There is even a third argument that has been put forward to expect a reduced
consensus for governing parties during European elections. Several empirical studies, in
very different institutional contexts, have found evidence of a recurrent electoral cycle
for incumbents. After a period of honeymoon, they usually lose consent, at least
partially recovering it on the eve of a new election (Fisher 2014). This may depend on
the citizens’ strategic expression of discontent for the government performance, as the
traditional second-order theory goes, or be the side-product of a lack of information and
mobilization from behalf of incumbents that are mainly interested in the first-order
arena (Weber 2007).
Whatever the underlying mechanism, and unless a member state holds
simultaneous elections or has had its preceding national ballot very close to the
European one – something that happened in 2014 only for Belgium and Hungary – we
should expect government parties to lose in what approximates a mid-term election
European elections, at least the most recent ones, have one further feature that
distinguishes them from the several other national second-order consultations.
Potentially, part of the competition takes place along a supranational dimension that
confronts Europhiles, Eurosceptic and neutral positions. As a consequence, parties with
a clearer profile and position on EU integration should perform better in those elections.
We must accompany this statement with a word of caution, however, because one of the
major failures of the EU has been precisely its incapacity to foster any sincere debate on
European issues. Yet some recent studies have demonstrated that information makes a
difference between those who vote only according to domestic preferences, and those
who also consider issues of European integration (Hobolt and Wittrock 2011).
Furthermore, at least for 2014, the austerity policies of the EU against the Great
Recession undoubtedly affected the electoral debate in many countries (Schmitt and
Teperoglou 2015). This may have produced a stronger polarization of the electorate on
this, now more salient, dimension. It is a political dynamic that, due to the economic
circumstances of the Great Recession, should have favoured mostly Eurosceptic parties
(Giuliani and Massari 2017), with the UKIP being the clearest example of this tendency.
To synthesise, the literature on second-order elections suggests that in the
European polls, compared to national appointments: a) government parties lose,
especially if elections fall around mid-term; b) there is relatively more space for small,
and even new parties; c) parties with an unambiguous position on the EU, which in
2014 mostly meant Eurosceptic parties, comparatively win. Admittedly, these features
are not independent. Incumbent parties, for example, cannot be new (unless they formed
by fission or aggregation during the ongoing legislature), are mostly large in size, and
with a comparatively moderate if not positive attitude towards Europe. If our aim here
were to explain the level of parties’ support in European elections, as Hix and Marsh
(2011) did for the first seven appointments, and Schmitt and Toygür (2016) for the last
one, we would certainly need a multivariate model including all those variables at the
same time. Instead, we are here only exploring potential overlapping sources of
pollsters’ bias, and will thus mostly check them separately, following what also the
original literature on second-order elections did.
Therefore, if our intuition is right, and pollsters, accustomed to predicting
electoral results within a domestic frame, failed to take proper consideration of the
theoretical and empirical suggestions specific to second-order European elections, we
should observe an overestimation of parties that the literature expects to lose, and an
underestimation of those that it expects to win. Thus:
Hp. 1 Pollsters overestimated government parties,
Hp. 2 … mostly if the European election fell in the middle of the electoral cycle;
Hp. 3 Pollsters overestimated big parties,
Hp. 4 … and underestimated new ones;
Hp. 5 Pollsters underestimated Eurosceptic parties.
Data, Variables and Measures
In order to test our hypotheses, we collected all the final electoral predictions
formulated by different polling firms in each EU member state before the European
election held from 22 to 25 May 2014.3 In some cases, polls were published just before
the election, with 45% of them made public in the last 10 days before the electoral
appointment, and almost all of them during the last month. Eventually, we arrived at
1057 party*poll observations for which we had both a prediction and an actual result.
Mainly from those two pieces of information, we computed several different
measures of our dependent variable, that is, the under- or over-estimation of a party’s
performance. The simplest of these indices is the Error of the prediction, i.e. the
difference between the estimate and the actual result, which assumes positive values in
the case of an overrated party, and negative ones in the case of an underrated one. Since
our hypotheses deal with the direction of the blunder, more than with its magnitude, we
also computed a dummy variable, Positive error, when the sign of the error reflected
overestimation of the party.4
The statistical soundness of each prediction was not our major concern, since we
were mostly interested in detecting potential cross-cutting inaccuracies. Therefore, even
systematic small differences between estimate and actual result, within the usual
sampling margin of error, contributed to, and were included in, our tests. However, we
decided to replicate our analyses by also adopting a more conservative approach. Using
the sample size to estimate the 95% confidence intervals of the estimated result of each
party, we verified when our error exceeded that random component, thus denoting a
biased prediction. We thus built two non-symmetric dummy variables, Positive bias
and Negative bias, assuming the value of 1, respectively, in the case of non-random
positive and negative prediction mistakes, and zero in all the other circumstances.
Finally, we took up a recent suggestion by Arzheimer and Evans (2014) in
regard to measuring the polling bias, which generalizes to a multi-party competition an
index originally proposed by Martin, Traugott and Kennedy (2005) for two-party races.
Their Accuracy index is the following:
where 𝑝 and 𝑣 represent respectively the proportion of the votes predicted by
the polls, and those actually obtained in the election by party i. “Positive values indicate
bias in favour of party i, whereas negative values imply bias against i” (33). Adopting
several measures for the same concept representing our dependent variable served as a
robustness check of our results.
Our independent variables are easily defined. Government and New parties were
measured as dummy variables relative to the incumbent executive, and to the preceding
national ballot. In most cases, the correspondence between the domestic and European
arenas was straightforward. In case of doubt, we mainly preferred to resolve it in favour
of novelty, thus producing a more conservative test for our hypotheses.5 Euroscepticism
is simply the reversed scale of the party positions on EU integration proposed by Chapel
Hills experts for 2014 (Bakker et al. 2015), here ranging from 1 (Strongly in favour) to
7 (Strongly opposed), only marginally complemented with qualitative information in
less than 3% of our observations, for minor parties excluded by their analysis. The Size
of the party was computed as the percentage of votes obtained in the preceding general
election, and the electoral Cycle as the ratio between the number of days since that
domestic appointment and the overall legal duration of the parliamentary term.
We further introduced a small set of standard control variables capturing typical
problems of electoral polling and predictions; and because of some missing information
on these controls the actual number of observations used in our models was reduced to
approximately 930 cases. First of all, we included in the right-hand side of the equation
the size of the Sample, and the Time between the date of the poll and the European
election computed in number of days (Jennings and Wlezien 2016). Next, we controlled
for the level of Turnout, which, as we have seen, is one of the challenging elements of
European elections, strongly affecting our prediction capacity. Finally, we checked the
Contemporaneous holding of other elections or referenda, and the presence of
Compulsory voting, two factors that may reduce the context-specific characteristic of
the European appointment.
Models and Empirical Results
Before testing our hypotheses thoroughly, we simply verified the plausibility of the
effect of being in government and being a new party – the only two dummy independent
variables involved in our conjectures – on pollsters’ predictions. If they overlooked
what we know about second-order elections, we should have observed the former
overestimated, and the latter underestimated in their predictions. As can be seen in
Table 1, the average difference between polls and results is in fact higher for
government parties compared to opposition ones, and the accuracy index is positive in
the former case and negative in the latter.6 The Anova tests confirmed that the
differences between the two values is always significant at a p<0.05 value. The same
applies to the comparison between the mean values for new and old parties, although
this time only the difference between values of the error was statistically significant,
whereas the accuracy index failed to reach the usual standard. Given the promising
results, and the tiny differences, we proceeded with more robust confirmations
including our control variables, and the whole range of indices presented in the previous
Table 1. Comparison of Error and Accuracy of the prediction for different party
N Mean Std. err. Mean Std. err.
Government 291 0.81 0.27 0.08 0.02
Opposition 766 0.10 0.10 -0.03 0.01
New 182 -0.05 0.16 -0.03 0.04
Old 866 0.37 0.12 0.01 0.01
We started by modelling the three dichotomic outcomes of having overestimated
predictions (looking for the causes of (a) positive errors and (b) positive biases) or
underestimated ones (explaining the origins of (c) negative biases). For this purpose, we
used separate logistic regressions, clustering the standard errors for each poll in order to
account for the non-independence of the observations within the same wave.7 Given our
effects-of-causes research design (Mahoney and Goertz 2006), only marginally
interested in the comprehensive explanation of the dependent variable, we only report
the coefficients for the covariates of interest of the three models, referring to the online
appendix for the complete results.
Table 2. Impact of different party characteristics on its over and under-estimation in the
Government Size New Euroscepticism
Positive error 0.317** 0.042*** -0.417** -0.031
(0.143) (0.007) (0.186) (0.032)
Positive bias 0.679*** 0.048*** -0.265 -0.171***
(0.163) (0.008) (0.242) (0.042)
Negative bias 0.019 -0.031*** 0.175 -0.065
(0.162) (0.009) (0.227) (0.044)
Constant and control variables not reported; Clustered standard errors in parentheses,
*** p<0.01, ** p<0.05, * p<0.1
Confirming our expectations, Table 2 shows a systematic positive difference between
predictions and results, with p<0.05 for government parties, and with p<0.01 for big
ones. More specifically, when the odds ratio is computed from the coefficients reported
in the table, being in government increases the probability of being overestimated by
pollsters by more than 37%, while each increase of 1% in size boosts that same
probability by more than 4%. At the same time, being a new party systematically
reduces, by 32%, the chances of being overrated, while the degree of Euroscepticism
does not seem to affect that result.
If we limit the positive outcomes to overestimations beyond the sampling error,
thus looking at positive biases, the results are mostly the same. Once again, the
coefficients of the models for being in government and for being a large party are both
positive and highly statistically significant. The probability of a positive bias is almost
double for incumbents compared to non-incumbents, and grows by 5% for each increase
of 1% in size. This time, the coefficient for being a new party is non-significant, though
with the correct sign8, whereas the position on the EU integration process is, as
expected, inversely associated with overestimation. For each point increase on the
seven-point scale of Euroscepticism there is an almost 16% decrease in the probability
of a positive bias in pollsters’ estimates. In the models explaining negative biases,
which are not simply the opposite of positive ones because of the grey area of errors
within the sampling margins, only the size of the party is statistically significant. The
coefficient is negative, as expected, since the higher the support in the preceding general
election, the lower the probability of being underestimated by pollsters.
Our hypotheses are largely confirmed by these first results, though it should be
noted that, in the last two models, the non-symmetric nature of bias reveals that a
systematic reduction of the probability of being overrated does not translate into a
specular increase in the chances of being underestimated, and vice versa. For
symmetric, and even more fine-grained hypotheses testing, we should turn to our last
two dependent variables, which consider not only the direction of the misjudgement but
also its extent. For this reason, using the same control variables and clustering the
standard errors as before, we ran a series of OLS regressions aimed at explaining the
gap between predictions and actual results, and the proposed index of accuracy. The
results are presented in Table 3.
Table 3. Impact of different party characteristics on the accuracy and on the error of the
Government Size New Euroscepticism
Error 0.551* 0.075*** -0.292 -0.032
(0.286) (0.012) (0.242) (0.052)
Accuracy 0.093*** 0.005*** -0.031 -0.014*
(0.027) (0.001) (0.055) (0.007)
Constant and control variables not reported; Clustered standard errors in parentheses,
*** p<0.01, ** p<0.05, * p<0.1
These analyses confirm the picture that emerged from the logit models. All the
coefficients have the expected signs, and in five cases out of eight they are statistically
significant. Euroscepticism is the only exception if we take the error between polls and
actual results as dependent variable, whereas being a new party systematically affects
neither the error nor the accuracy of the prediction. Turning to the positive evidence,
being incumbent yields on average a surplus estimate of half a point, while for every
increase of 10% in the votes obtained in the previous national election there is a ¾ point
of overestimation in the European one. Turning to the index of accuracy, which has a 4-
point empirical range, almost one tenth of a point of favouritism is generated by
incumbency, or by a party holding 25% of the votes at the national level, whereas
moving from the lowest extreme of the scale of Euroscepticism to the highest one
produces a discrimination of a similar magnitude.
Generally speaking, combining different econometric models with four ways of
measuring our dependent variable – the misjudgement of pollsters – yields not a perfect
but a sufficiently robust confirmation of our hypotheses. Big and incumbent parties
obtain favourable estimates, while new and Eurosceptic ones receive mostly adverse
predictions. There is only one of the proposed propositions that has not yet been tested
and that cannot be investigated with our direct models: the one suggesting that
government parties should be mostly overestimated the farther away they are from the
preceding or the successive national election (because the second-order theory expect
them to lose the most when the European appointment falls exactly in the middle of the
mandate; e.g. Reif 1984; Weber 2011). In order to test this hypothesis, we ran two
conditional models in which the electoral cycle interacted with our dummy variable
capturing the characteristic of being incumbent parties. Following the best practices,
more than reporting the coefficients, it was essential to plot the marginal conditional
effects for the whole observed range of the electoral cycle (Kam and Franzese 2007,
Brambor, Clark and Golder 2006). Since our hypothesis assumed a non-linear
relationship with the time from the preceding election, we modelled it using the square
of the cycle variable as interaction term. The results are illustrated in Figure 1.
Figure 1. Marginal effects of incumbency on the error (left panel) and on the accuracy
of the prediction (right panel) at different moments of the electoral cycle
Both graphs, representing the marginal impacts of incumbency on the delta between
polls and results, and on the accuracy index, confirm our intuition and the expectations
derived from the theory of second-order elections. In the first part of the electoral cycle,
the honeymoon effect prevents government parties from suffering the usual impact of
European appointments. Symmetrically, as shown by the confidence intervals
overlapping the no-effect horizontal line, they are not overestimated by pollsters, whose
predictions are made easier also by the closeness of the preceding electoral ballot. The
same happens in the last phase of the electoral cycle, when government parties usually
recover, and when, once again, pollsters’ predictions become more accurate, cancelling
any systematic bias in their favour. However, when European elections take place
during the central phases of the cycle, approximately between one and three years after
the preceding national appointment, government parties are more exposed to second-
-4 -2 0 2
10 30 50 70 90
Pct of electoral cycle
Marginal effects on the error of prediction
-.1 0 .1 .2 .3 .4
10 30 50 70 90
Pct of electoral cycle
Marginal effects on the accuracy of the prediction
order effects, and pollsters naively overestimate their actual support. This can be seen
from the positive marginal effects shown in both graphs, whose confidence intervals are
consistently above the horizontal zero line in the central portion of the chart.
The two graphs are far from being perfectly symmetrical, and especially the
point estimates of the one regarding accuracy resemble a linearly declining over-
prediction. This is probably due to the fact that two factors combine in producing those
marginal effects: on the one side, second-order elections modify political behaviours,
and, on the other, the closeness of other ballots provides additional information which is
probably biased by the contingent character of that appointment. In any case, what
matters more for our propositions, which reverse the predictions of the theory of
second-order elections, is that confidence intervals overlap with the null hypothesis at
the two extremes of the cycle, thus confirming also our final (second) hypothesis.
In this article, we have empirically checked one simple hypothesis: that pollsters
systematically underutilize the theoretical knowledge produced by the social sciences,
and by political scientists more specifically. By using the 2014 European election as a
test bed for that insight, we have suggested that pollsters erred in the direction opposite
to the expectations of the theory of second-order elections, which found confirmation
even in that event. For robustness, we have tested several measures, mostly confirming
When the theory predicts European success, as in the case of new and
Eurosceptic parties, pollsters generally underestimated their electoral performances,
thus confirming our hypotheses 4 and 5. When the theory suggests a relative European
failure, as for big and government parties, pollsters mostly overestimated their actual
electoral backing, as we suggested with our hypothesis 1 and 3. Including a conditional
effect of the electoral cycle better specifies the misfit between predictions and actual
results due to a failure to consider the incumbency effect in second-order elections, as
we actually implied with our hypothesis 2.
The modern world, and not only our political worlds, is becoming more and
more complex and less and less predictable (Waldrop 1992; Bertuglia and Vaio 2005).
That does not mean that, even in such complexity, equilibria do not form and regular
patterns cannot be identified. Bringing this very general idea to the daily work of
pollsters as social scientists, it is evident that the more complex and changing a social
and political environment is, the less we can simply rely on a correct sampling design.
Predicting electoral results “doesn’t just involve asking people whether they support
candidate A or candidate B. It also involves trying to determine whether respondents
will act on their preferences by casting a ballot at all. […] It is this extra step, […] that
is quite distinct from the principle of random sampling and good question design that
make survey research valid and reliable” (Gramlich 2017).
The problem is thus no longer, or not simply, statistical, and even house effects
are irrelevant, as we indirectly demonstrated with our cross-country and cross-polling
analysis. Several correcting strategies are needed, including all the possibilities offered
by modern data analytics. Yet, corrections due to firmly-established theories should be
the first to be taken into consideration. For the European appointment polls, we used the
theory of second-order election, which could be further complemented with more
general hypotheses regarding the support cycle of government parties (Fisher 2014;
Weber 2011) and the specific profile of late voters (Box-Steffensmeier et al 2015).
Going back to the examples mentioned in the introduction of this article, what
could be other theories in social and political science that are sufficiently established to
provide useful guidance to pollsters? The first obvious candidate is the theory of
economic voting, whose robustness has been proved in different contexts and time-
periods (Lewis-Beck and Stegmaier 2007; Stegmaier, Lewis-Beck and Park 2017;
Giuliani and Massari 2018). Its conjectures give an additional leverage to non-
incumbent parties, clarify the electoral costs for coalition partners, and may complement
other information regarding the expected turnout levels: three suggestions that certainly
apply to many of the prediction failures mentioned above.9
The theory of valence political competition (Curini 2018) is a second contender
for that role. The less ideological the competition, and the more issues such as
corruption, leadership capacities, and similar non-positional topics appear at the
forefront of the campaign, the more its hypotheses may help. At the intersection
between political communication and political psychology, the same could be said
about works on the personalization of electoral competition (Garzia 2014), on the self-
reinforcing effects of echo-chambers (Barberà 2015) – whose effects are exploited by
the analysis of social networks (Leiter et al. 2018) – and on the risks of spirals of silence
It should be remembered that we are thinking neither of substituting the mining
of public opinion, nor moving directly from predictions to forecasts. These theories may
help improve pollsters’ work much in the same vein as more traditional sociological
knowledge helped them frame the procedures necessary for post-stratification or for
non-respondents substitution in the past. Even “local” theories accounting for regular
behavioural patterns and voting traditions in certain countries may be of use, until our
increasingly fluid and unpredictable world endangers even those solid theoretical
Arzheimer, K., and J. Evans. 2014. “A New Multinomial Accuracy Measure for Polling
Bias.” Political Analysis 22 (1): 31-44.
Bakker, R., E. Edwards, L. Hooghe, S. Jolly, G. Marks, J. Polk, J. Rovny, M.
Steenbergen and M. Vachudova. 2015. 2014 Chapel Hill Expert Survey. Version
2015.1. Available on chesdata.eu. Chapel Hill, NC: University of North
Barberà, P. 2015. “Birds of the Same Feather Tweet Together: Bayesian Ideal Point
Estimation Using Twitter Data.” Political Analysis 23: 76–91.
Barnea, S. and G. Rahat. 2011. “ ‘Out with the old, in with the “new” ’: What
constitutes a new party?” Party Politics 17 (3): 303-320.
Bertuglia, C.S. and F. Vaio. 2005. Nonlinearity, Chaos, and Complexity. The Dynamics
of Natural and Social Systems. Oxford: Oxford U.P.
Bolleyer, N. and E. Bytzek. 2013. “Origins of party formation and new party success in
advanced democracies.” European Journal of Political Research 52 (6): 773–
Box-Steffensmeier, J., M. Dillard, D. Kimball and W. Massengill. 2015. “The long and
short of it: The unpredictability of late deciding voters.” Electoral studies 39:
Brambor, T., W. Clark and M. Golder. 2006. “Understanding Interaction Models:
Improving Empirical Analyses.” Political Analysis 14 (1): 63-82.
Campbell, J. E. 2016. “How Accurate Were the Political Science Forecasts of the 2016
Presidential Election?.” Cristal Ball, University of Virginias Center for Politics,
Curini, L. 2018. Corruption, Ideology, and Populism. The Rise of Valence Political
Campaigning. New York: Palgrave.
Döring, H., and P. Manow. 2015. Parliaments and Governments Database (ParlGov):
Information on Parties, Elections and Cabinets in Modern Democracies.
Available at http://www.parlgov.org/
Emanuele, V. and A. Chiaramonte. 2016. “A growing impact of new parties: Myth or
reality? Party system innovation in Western Europe after 1945.” Party Politics
Fisher, S. D., R. Ford, W. Jennings, M. Pickup, and C. Wlezien. 2011. “From Polls to
Votes to Seats: Forecasting the 2010 British General Election.” Electoral Studies
30 (2): 250–257.
Fisher, S. D. 2014. “Predictable and Unpredictable Changes in Party Support: A
Method for Long-Range Daily Election Forecasting from Opinion Polls.”
Journal of Elections Public Opinion and Parties 25 (2): 1-22.
Garzia, D. 2014. Personalization of Politics and Electoral Change. New York: Palgrave
Giuliani, M. and S. A. Massari. 2017. “The economic vote at the party level: Electoral
behaviour during the Great Recession.” Party Politics
Giuliani, M. and S.A. Massari. 2018. It’s the Economy, Stupid. Votare in tempo di crisi.
Bologna: il Mulino.
Gramlich, J. 2017. “Q&A: Pew Research Center’s president on key issues in U.S.
polling”. Pew Research Center. Fact Tank. News in Numbers: June, 16.
Hix, S., and M. Marsh. 2011. “Second-order effects plus pan-European political swings:
An analysis of European Parliament elections across time.” Electoral Studies 30
Hobolt, S. B., and J. Wittrock. 2011. “The Second-order Election Model Revisited: An
Experimental Test of Vote Choices in European Parliament Elections.” Electoral
Studies 30 (1): 29-40.
Jackman, S. 2005. “Pooling the Polls Over an Election Campaign.” Australian Journal
of Political Science 40 (4): 499–517.
Jennings, W., and C. Wlezien. 2016. “The Timeline of Elections: A Comparative
Perspective.” American Journal of Political Science 60 (1): 219–233.
Jennings, W., and C. Wlezien. 2018. “Election polling errors across time and space.”
Nature Human Behaviour 2 (4): 276–283.
Kam, C. D., and R. J. Franzese. 2007. Modeling and Interpreting Interactive
Hypotheses in Regression Analysis. Ann Arbor: The University of Michigan
Leiter, D, A. Murr, E. Rascón Ramírez and M. Stegmaier. 2018. “Social networks and
citizen election forecasting: The more friends the better.” International Journal
of Forecasting 34 (2): 235–248.
Lewis-Beck, M. S. 2005. “Election Forecasting: Principles and Practice.” The British
Journal of Politics and International Relations 7 (2): 145-164.
Lewis-Beck, M. S. and C. Tien. 2000. “The Future in Forecasting: Prospective
presidential models.” In Before the Vote: Forecasting American National
Elections, edited by J. Campbell and J. Garand, 133-144. Thousand Oaks: Sage.
Lewis-Beck, M. S. and C. Tien. 2012. “Election forecasting for turbulent times.” PS:
Political Science and Politics 45 (4): 625-629.
Lewis-Beck, M. S. and C. Tien. 2016. “The Political Economy Model: 2016 US
Election Forecasts.” PS: Political Science and Politics 49 (4): 661-663.
Lewis-Beck, M.S. and R. Dassonneville. 2015. “Forecasting elections in Europe:
synthetic models.” Research & Politics 2 (1): 1-11.
Lewis-Beck, M.S. and M. Stegmaier. 2007. “Economic models of voting.” In The
Oxford handbook of political behaviour, edited by R.J. Dalton and H.D.
Klingemann, 518-537. Oxford: Oxford U.P.
Lewis-Beck, M.S. and M. Stegmaier. 2014. “Election forecasting, scientific
approaches.” In Encyclopedia of social network analysis and mining, edited by
J. Rokne and R. Alhajj, 462-469, Heidelberg: Springer.
Litton K. 2012. “Party novelty and economic voting: the evidence from the EU
parliamentary elections.” Paper MPSA Conference, 12–15 April, Chicago (Il).
Litton K. 2015. “Party novelty: conceptualization and measurement of party change.”
Party Politics 21 (5): 712–725.
Mahoney, J., and G. Goertz. 2006. “A Tale of Two Cultures: Contrasting Quantitative
and Qualitative Research.” Political Analysis 14 (3): 227-249.
Marsh, M. 1998. “Testing the Second-Order Election Model after Four European
Elections.” British Journal of Political Science 28 (4): 591–607.
Martin, E., M. Traugott, and C. Kennedy. 2005. “A Review and Proposal for a New
Measure of Poll Accuracy.” The Public Opinion Quarterly 69 (3): 342-369.
Noelle-Neumann, E. 1993. The Spiral of Silence: Public Opinion - Our Social Skin.
Chicago: University of Chicago Press.
Reif, K. 1984. “National Electoral Cycles and European Elections, 1979 and 1984.”
Electoral Studies 3 (3): 244-255.
Reif, K., and H. Schmitt. 1980. “Nine Second-Order National Elections. A Conceptual
Framework for the Analysis of European Election Results.” European Journal
of Political Research 8 (1): 3-44.
Schmitt, H., and E. Teperoglou. 2015. “The 2014 European Parliament Elections in
Southern Europe: Second-Order or Critical Elections?.” South European Society
and Politics 20 (3): 287-309.
Schmitt, H., and I. Toygür. 2016. “European Parliament Elections of May 2014: Driven
by National Politics or EU Policy Making?.” Politics and Governance 4 (1):
Stegmaier, M., M.S. Lewis-Beck, and B. Park. 2017. “The VP-function: A review.” In
The SAGE Handbook of Electoral Behaviour, edited by K. Arzheimer, J. Evans
and M.S. Lewis-Beck, 584-605, London: Sage.
van der Eijk, C., and H. Schmitt. 2009. “Legitimacy and Electoral Abstentions in
European Parliament Elections.” In The Legitimacy of the European Union after
enlargement, edited by J. Thomassen, 208-224. Oxford: Oxford University
Voss, S. D., A. Gelman, and G. King. 1995. “The Polls. A Review. Preelection Survey
Methodology: Details from Eight Polling Organizations, 1988 and 1992.” Public
Opinion Quarterly 59 (1): 98-132.
Waldrop, M.M. 1993. Complexity: the emerging science at the edge of order and chaos.
New York: Touchstone
Weber, T. 2007. “Campaign Effects and Second-Order Cycles: A Top-Down Approach
to European Parliament Elections.” European Union Politics 8 (4): 509–36.
Weber, T. 2011. “Exit, Voice, and Cyclicality: A Micrologic of Midterm Effects in
European Parliament Elections.” American Journal of Political Science 55 (4):
1 In truth, some of these results have been attributed to the inability sufficiently to detail the
geographical distribution of vote intentions, so that the disproportionality of the electoral
system artificially inflated the gap between overall prediction and actual political result.
This was especially the case of the discrepancy between popular vote and Electoral
College in the US presidential election. A very recent article by Jennings and Wlezien
(2018), surveying polling errors in 45 countries for 75 years, demonstrate that there is not
a historical downward trend in the accuracy of predictions, thus suggesting a diffused and
media-inflated misperception of recent misses. The results of their analysis, checking for
the origin of absolute polling errors, does not invalidate our own directional and election-
2 In the article we have consistently followed the distinction between (pollsters’) predictions and
(models’) forecasts. Yet, if “forecasting requires more than curve-fitting” (Lewis-Beck and
Tien 2000, 98), since it is a theory-driven process (Lewis-Beck 2005), this does not mean
that polls simply require the computation of frequencies. The way in which the sample is
constructed and balanced, non-respondents are treated, and, eventually, estimates are
produced, require a set of conscious choices (Fisher et al. 2011) that can certainly profit
from a theory-laden interpretation of the process, and a consolidated knowledge of the
context-specific features of that election.
3 The list of polls, mostly run as CATI and typically with four times more contacts than actual
respondents, was completed by drawing on three different sources: a dataset published
shortly after the election by @electionista, the Pollwatch 2014 website run by Votewatch
Europe http://www.votewatch.eu/ , and the 28 pages of Wikipedia in their original
language devoted to the electoral appointment. We complemented that information by
using several sources, ranging among direct polls reports, newspapers or weekly journals,
EU official results http://www.europarl.europa.eu/elections2014-results/en/election-
results-2014.html, national election data and indices from the Parlgov project (Döring and
Manow 2015). The surveys asked for vote intentions in the imminent European election,
and thus respondents’ answers should have had a built in second-order effect, as we could
verify for some polls that asked in parallel the preferences for the European and a
hypothetical national ballot. The dataset is available in the author’s personal webpage.
4 There are just five occurrences for which, at the level of precision of our data, pollsters
perfectly predicted the result of a party. Given the negligible amount of these cases, the
incidence of a negative error can be simply considered as the mirror image of positive
errors, without needing to test them separately.
5 We acknowledge that this is a rather drastic simplification of the complex issue of what really
constitutes a “new” party, how to treat cases of split/merger of preceding groups (Bolleyer
and Bytzek 2013; Emanuele and Chiaramonte 2016), and if novelty is a dichotomous
property or is better captured by a continuous measure (Barnea and Rahat 2011; Litton
2012; 2015). We cannot further develop our conceptual analysis in this direction, but in
the empirical part we will check the robustness of alternative operationalizations in order
to shed some light on the actual meaning of “new” for the present hypotheses.
6 Polls often estimate the support of only a subset of cases, sometimes reporting collectively as
“others” the votes for the remaining parties, not considered in our analysis. This justifies
the positive error for both government and opposition parties.
7 For the analysis of bias, we also experimented with a single multinomial logistic model. Yet
our hypothesis does not actually translate into contrasting the opposite type of biases
against a null baseline, but each type against the remaining options together (i.e. no bias
plus wrong ones). For this reason, we preferred separate logistic models. Yet, the results
are mostly similar, with the partial exception of Euroscepticism, which seems to depress
both type of biases against the null hypothesis.
8 As anticipated in note 5, we also operationalized novelty in alternative ways instead of its
simply being absent in the previous, and usually nearest, general election. We checked for
parties taking part in their first European ballot, or for having just one previous national
electoral experience. We also tried a four-point scale, from zero to three, in which
“newness” was the reverse of a count variable measuring the number of previous national
and European ballots in which the party took part. The results of these alternative
operationalizations are reported in the online appendix (Table A.8). Interestingly, none of
these substitutes proves to have a significant coefficient. It is not being new just for the
European election that counts for the errors, and, for that matter, the “virginity” seems to
be immediately lost after the first experience both in the ballot and in the polls. This
indirectly confirms the importance of a contextual knowledge of the specific election, and
of the actors that take part in it.
9 In fact, the same theory is a usual component of any forecasting effort during and after the
Great Recession, including the mentioned US presidential election (Lewis-Beck and Tien
2012; 2016; Lewis-Beck and Stegmaier 2014; Lewis-Beck and Dassonneville 2015).