Professional Documents
Culture Documents
SSRN 3327181
SSRN 3327181
Abstract
A large number of studies have analysed the role of individual and geographical determinants
of the Leave vote in the recent EU referendum, but several questions remain. One key
unresolved issue is the extent to which there is interaction between individual characteristics
and the geographical context. For instance, much of what has been attributed to the role of
individual characteristics in determining the Brexit vote may have been mediated by the
geographical context, and vice versa. Space, and the people who live in it, are in continuous
interaction, which requires a careful conceptual and empirical treatment of the issues of
composition and sorting. We therefore ask: to what extent did individuals with very similar
individual characteristics vote differently in different places? We use data from the
British Election Study, and a non-parametric matching approach, to analyse whether
comparable individuals voted differently in areas with particular economic, social, and cultural
characteristics. We find that composition effects account for less than half of the observed
constituency-level variation in the vote, while the remaining contextual effects are driven almost
entirely by cultural factors.
i Corresponding Author. Department of Land Economy and Pembroke College, University of Cambridge
(ma405@cam.ac.uk).
ii Department of Land Economy, University of Cambridge (oo263@cam.ac.uk); Research Institute of Industrial
Economics, Stockholm.
In “The revenge of the places that don’t matter”, Rodriguez-Pose (2018) argues that the rise of
populism in certain regions is due to persistent poverty, economic decay, and a lack of economic
opportunities. Individuals who choose or are obliged to live in lagging and declining regions
manifest their discontent through voting for anti-establishment candidates. A similar argument
is made by Dorling and Tomlinson (2019), who highlight the additional burden of austerity
policies since the financial crisis in 2009, which has disproportionately affected the poorest
areas. This point is also noted in many recent academic and media reports, highlighting the
recent and stark urban-rural disparities in voting behaviour (Johnston et al., 2018; Rossi, 2017;
Lichter and Ziliak, 2017), as well as the divide across neighbourhoods within cities and towns
(Barber and Imai, 2014). An analysis of the striking regional disparities observed in the EU
referendum vote faces two potential methodological problems. First, if individuals with certain
views and political preferences are selected (either through birth or subsequent migration) into
economically-lagging areas, it is difficult to identify the extent to which their political behaviour
can be attributable to their individual characteristics vis-à-vis their regional context. Simply
controlling for both individual and regional characteristics would not allow us to isolate the
impact of the regional context, since the two sets of characteristics are not orthogonal to each
other. Second, the clustering of individuals with similar characteristics is likely to lead to further
feedback mechanisms that may amplify their original political views, as individuals are
influenced by their peers and through their social networks. As noted by Goodwin and Heath
The remainder of the paper is organised as follows. Section 2 discusses the theoretical and
empirical literature and presents our conceptual framework; Section 3 discusses the data
sources, variables included in the analysis, and our empirical design; Section 4 presents our
results; Section 5 concludes.
Immediately following the shock result of the referendum vote, there was intensive coverage in
both the traditional and online media. This includes reports providing statistics and maps,
showing correlations between the EU referendum vote and regional determinants, including
demographic characteristics and area-level measures of economic prosperity (Ashcroft, 2016;
Burn-Murdoch, 2016; Harris, 2016). More recently, a significant academic literature has
emerged in a number of academic fields, aiming to analyse and understand the determinants
of the vote across demographic groups, and across regions.
In an early and comprehensive study, Becker et al. (2017) use a machine-learning method to
analyse which sets of variables best predict the referendum result across local authority districts,
and across the 107 English wards for which there is available data. The factors considered
include exposure to the EU through immigration, trade, and access to structural funds, the
local impact of austerity policies, demographic variables and education, economic structure
and local labour markets conditions, and events on the referendum day (e.g., rainfall). The
results indicate that the most key regional determinants are demographic characteristics (age
At the individual level, a comprehensive Joseph Rowntree Foundation report finds that the
most significant individual driver of the Leave vote is low education, followed by poverty,
unemployment, financial insecurity, and being in a low-skilled and low-income job. The report
also finds that the gap in the Leave vote between high- and low-skilled individuals is much
larger in low-skilled areas than in high-skilled areas, indicating that there is an interaction
between individual- and area-level characteristics (Goodwin and Heath, 2016b). In a related
paper, Alabrese et al. (2018) use both individual and regional data to show that the Leave vote
is associated with “older age, white ethnicity, low educational attainment, infrequent use of
smartphones and the internet, receiving benefits, adverse health and low life satisfaction”
(Alabrese et al., 2018). Interestingly, the authors find that using both individual and regional
variables adds predictive power to the model, and that the model predicts especially well in
areas with extreme referendum vote shares (that is, those that are more homogeneous in terms
of voter views). They also find that culture or identity factors are likely to be important, given
that “closeness to a political party” explains a significant proportion of the classification errors
in their model.
The cultural context of each location appears to have had a significant effect on the propensity
to vote Leave, and this is manifested in anti-immigration and anti-establishment views. Using
data from the European Values Study and the British Election Study, Arnorsson and Zoega (2018)
show that higher net migration regions are more likely to be suspicious of immigrants, more
reluctant to have them as neighbours, and more likely vote for Leave. Similarly, a number of
studies have shown that although the Leave vote was lower in more diverse locations with a
higher proportion of non-UK born residents, it was higher in areas of high immigration growth
(Goodwin and Milazzo, 2017; Kaufmann, 2017). Other studies show that perceptions of higher
immigration are associated with greater support for Leave, and that voters do not generally
distinguish between EU and non-EU migrants (Goodwin and Milazzo, 2017; Blinder and
Markaki, 2018). Moreover, there is a clear spatial pattern to anti-immigration views, which
tend to be strongest in neighbourhoods which are themselves relatively homogenous, but which
The impact of these individual characteristics is also likely to be qualified by regional level
variables. For instance, Lee et al. (2018) show that geographically-immobile individuals living
in areas experiencing rapid economic or demographic change were significantly more likely to
vote Leave (relative to mobile individuals), but the same is not true for areas experiencing little
change. Similarly, Gordon (2018) finds that individuals in graduate occupations were less likely
to support populist parties in areas with large immigrant populations, as were individuals who
are members of organisations with a national focus, such as trade unions or religious
institutions.
One important source of economic grievance that may have affected the EU referendum vote
is the differential impact of globalisation on regions and groups, and this is the focus of a
number of recent studies. For instance, Colantone and Stanig (2018) find that greater exposure
to trade shocks (in the form of imports from China) result in a higher share of the Leave vote,
after controlling for other regional factors. Moreover, once the trade shock is accounted for,
there is no further statistically significant role for immigration in predicting the vote.
Interestingly, they find that individuals tend to react to the local economic conditions,
Another second important factor is the extent to which areas and individuals have been affected
by the government spending cuts which followed the 2009 crisis (known as the “austerity”
policies). As argued by Dorling (2016) and Dorling and Tomlinson (2019), the impact of
austerity, and existing health and income inequalities, are likely to have been significant
contributors to the decision to vote Leave, particularly among Northern England lower-income
households. To address the deterministic impact of austerity on the vote, Fetzer (2018)
investigates the extent to which they explain regional variations in political preferences, such
as support for the populist UKIP party, anti-establishment views, and popular support for
Leave. The results suggest that, were it not for these welfare reforms, the outcome of the
referendum might have been a victory for Remain. His analysis also shows that exposure to
austerity policies is a key activating factor that increases individual propensities to vote Leave,
given existing economic grievances (Fetzer, 2018). There is also evidence that the Leave result
was driven by a significantly higher turnout in deprived and segregated areas (Bartle et al.,
2017).
While outside the scope of this paper, we also note that the grievances resulting from economic
decline and austerity may have been further amplified by the role of both the traditional and
the digital media, with micro- and geo-targeted adverts forming a large proportion of campaign
spending during the EU referendum campaign (Levy et al., 2016; Liberini et al., 2018).
Conceptually, the question of the relative importance of composition and contextual effects in
the EU referendum can be decomposed into three parts:
2. Are all individuals with similar personal characteristics equally affected, or does the impact
vary by location?
3. To what extent do these personal and neighbourhood experiences translate into a Leave
vote?
The first question relates to the effects of globalisation (in the form of EU integration) on living
standards and wellbeing. There is a large literature on this topic, starting with the Stolper-
Samuelson Theorem (developed in 1941), which argues that trade openness can have
significant distributional effects, with workers in the importable-goods sector being negatively
affected. An important and recent finding in the literature is that these negative effects increase
as the process of liberalisation advances, so that the removal of the final trade barriers have the
greatest distributional impacts (Rodrik, 2017). In principle, the gains from trade could be used
to compensate workers who have lost out. However, in practice, providing adequate
compensation is often politically difficult. In the UK context, the negative impact of trade
liberalisation on low-skilled workers came at a time of significant government spending, which
included cuts to public services, welfare benefits, and public sector jobs.
Two other aspects of globalisation may also have contributed to the negative distributional
effects. An increase in international capital mobility has heightened the likelihood of financial
crises, while an increase in immigration may have depressed wages among low-skilled workers
in certain occupations during the recent economic downturn. Whether these negative impacts
can be attributed directly to EU membership is debatable. Unlike other free trade agreements
such as NAFTA, EU membership led to the development of complex cross-border supply
chains, which protected manufacturing jobs that might otherwise have been lost. In addition,
EU membership includes a component of redistribution (the Cohesion policy), which mitigated
some of the effects of UK government spending cuts. As discussed in the previous section, it is
of course possible that individuals attributed the negative effects of globalisation and other
changes (such as automation) to EU membership, and these perceptions may have been further
amplified by the media, the growing influence of UKIP, and the EU referendum campaign.
An important issue is the extent to which geographical patterns in the Brexit vote are a result
of contextual effects resulting from observation, social interaction, emulation, or political
mobilisation (Weaver, 2014), or whether the observed effects are due to self-selection of
individuals into specific neighbourhoods. Gallego et al. (2016), for instance, find that contextual
effects tend to be relatively weak, and that self-selection on demographic characteristics such as
age, marital status, and caring responsibilities, is more important in explaining neighbourhood
effects in preferences, although this appears to have been less of a factor in explaining the EU
referendum vote (Kaufmann, 2017; Lee et al., 2018). There may also have been an element of
solidarity in voting Leave, by voters who are not otherwise significantly affected by job-market
prospects (for instance, retired voters).
The third question is concerned with whether these tendencies to vote Leave translate into a
Leave vote in the EU referendum. In particular, did some voters feel more motivated to vote
than others, and did this affect relative turnout across different groups and locations? The
political science literature has identified a number of factors that influence the likelihood to
vote in a referendum. Individual voting intentions are affected by party affiliation, popularity
of the party leaders, support for the governing party, and the level of education of the voters,
with more educated and older voters generally supporting the government position (de Vreese
and Semetko, 2004). On the other hand, unique nature of British democracy (based on a First
Past the Post system) may have led to a higher turnout among voters whose choices are typically
unlikely to reach the threshold needed to elect an MP (Goodwin and Heath, 2016b; Johnston
et al., 2016).
We use individual-level data from the British Election Study (BES), a nationally representative
survey of electoral intentions and behaviour, which also includes a range of economic, political,
demographic, and cultural variables. The BES has been run before and after every general
election since 1964 (and at other salient times) and consists of two main components: (i) a
residence-based random sample of individuals, contacted using face-to-face interviews, with
around 3,000-4,000 individuals per wave, and (ii) a panel sample of individuals surveyed 10
times since 2014, conducted online, with around 30,000 individuals per wave. The attrition
rate between panel waves is small, over 80% remain in the panel between waves. There are
three particularly relevant EU referendum waves in the panel: Wave 8 (pre-referendum), Wave
9 (immediately post-referendum), and Wave 10 (autumn of 2016). The panel survey is
conducted by YouGov based on their internet panel, and is representative of the UK by region,
economic, and demographic characteristics, and also adjusted to account for various response
biases.
We focus on Wave 9, the immediate post-referendum wave of the survey, in order to obtain
the most accurate (in terms of recollection and other biases) measure of how individuals voted
in the EU referendum. The total number of individual observations available is 30,036,
although the final sample size for the analysis is 17,382 individuals, due to item non-response
for some of the key variables of interest, and after excluding individuals aged below 18 years of
age.1
An added complication is the fact that unlike in general elections, the EU referendum votes
were tallied mostly at the local authority district level, although in some areas there are also
ward level results (a smaller spatial unit which can be aggregated up into constituencies). In
order to analyse the results at the constituency level, we are therefore forced to use estimates
rather than actual votes. We use the best available constituency-level estimates of the overall
Leave vote developed by Hanretty (2017), and note that these estimates are not central to our
analysis, but are used for reference purposes only.
All of the individual level variables used in the analysis are taken from Wave 9 of the BES (these
variables are discussed in detail in Section 3.3). The constituency-level aggregates are taken
from the following data sources: (1) non-UK born residents in 2001 and 2011, residents from
EU accession states in 2011, and residents educated to Level 4 (university level) in 2011, from
the Census; (2) median weekly gross pay for full-time residents, 2005 and 2015, and hourly
gross pay percentiles, 2015, from the Annual Survey of Hours and Earnings; 3 (3) working age
benefits claimants, 2015, from the Department of Work and Pension’s Work and Pensions
Longitudinal Study; (4) economically inactive working-age residents, 2015, from the Annual
Population Survey; (5) UKIP vote share and turnout in the 2015 UK Parliamentary General
Election, from the Electoral Commission; (6) average views regarding the cultural impact of
immigration (on a Likert scale, ranging from 0 “bad” to 7 “good”), taken from estimates
constructed by Hanretty and Vivyan (2015), based on Waves 1 and 2 of the 2015 BES
Combined Panel; (7) constituency-level Leave vote estimates, taken from Hanretty (2017), as
discussed above.
10
There are two important challenges in estimating the influence of contextual or neighbourhood
effects on individual outcomes, in this case, on the likelihood of voting Leave in the EU
referendum. First, there are likely to be observed and unobserved individual characteristics that
affect both the individual’s electoral choice, and the decision to live in a particular location.
Second, the choice of location is likely to be affected by individual outcome variables (such as
income, access to credit, education), which in turn affect the electoral choice. We therefore
need a method that is able to disentangle the composition from the contextual effects, and to
deal with the resulting self-selection bias.
Our empirical strategy consists of using a non-parametric matching method, Coarsened Exact
Matching (CEM), that allows us to control for pre-treatment confounding influences in the
selection of individuals into different types of constituencies (where location in a type of
constituency is the “treatment”). Intuitively, we temporarily coarsen any continuous variables
and assign individuals to a small number of categories. For instance, the continuous variable
age is assigned to the categories <25 years, 26-35 years, 36-45 years, 46-55 years, 56-65 years,
66-75 years, and 76+ years. We then sort individuals into strata, so that within each strata all
individuals have the same value for the coarsened matching variable, and obtain weights that
can be used to generate a balanced sample in further analysis (with strata containing only
treated or non-treated individuals having a zero weight).4 CEM matching offers a number of
advantages over other matching methods. It is a form of exact matching through its use of the
coarsening process, which allows the degree of imbalance to be set in advance of the analysis.
It also restricts the matched data to areas of common support and provides a simple way (via
the use of weights) to use the resulting matched sample in regression analysis (Iacus et al, 2008;
Blackwell et al., 2009).5
Our analysis proceeds in two steps. In a first step, we analyse whether there are differences in
the propensity to vote Leave across Leave and Remain voting areas, for individuals with similar
characteristics. Here the “treatment” is whether an individual lives in a constituency that
predominantly voted to Leave (i.e., with a 50% or higher Leave vote), or not. This part of the
analysis allows us to test directly for composition effects, that is, whether the referendum result
is due to differences in the individual characteristics across Leave and Remain voting areas
11
𝑃𝑟 𝑌 = 1
𝑙𝑛 = 𝛼 + 𝛽𝑇 + 𝛾 1 𝑋 + 𝜀 (1)
1 − 𝑃𝑟 (𝑌 = 1)
where Y is an outcome variable that equals one if the individual voted Leave in the referendum,
and zero if they voted Remain, T is a dummy variable that equals one in the individual is
“treated”, and zero otherwise, X is a matrix of pre-treatment characteristics that are coarsened
in the CEM process, included in un-coarsened form to control for any remaining imbalance,
α is a constant term, β is the SATT estimate, γ is a vector of coefficients, and ε is an error term.
The results are shown in Table 2, which also shows the non-weighted estimates (uncorrected
for composition effects) for comparison purposes. The estimates reported show the effect of a
one unit increase in the explanatory variable on the odds ratio. A coefficient greater than one
therefore indicates a positive effect, and a coefficient lower than one indicates a negative effect.
In a second step, we analyse whether similar individuals voted differently in districts with
high/low values of variables identified in the literature as being particular important in
explaining the EU referendum result. In particular, we calculate the SATT for the following
treatments. First, we analyse whether individuals in UKIP-voting areas have a higher
propensity to vote Leave, in order to test whether the rise of Euroscepticism at the constituency
level explains the Leave vote (Becker et al., 2017; Fetzer, 2018). Second, we test the relevance
of immigration-related variables, both in terms of actual immigration measures (immigration
growth and EU accession country immigrants), and views about immigration (anti-immigration
values). This is to test whether immigration is a strong motivator for the Leave vote, and
whether it is the number of immigrants, or perceptions of immigration, that matter in
12
In a third step, we extend the analysis to consider “multichotonous” treatments, that is,
treatments that involve more than one variable. We are thus able to analyse whether, for
instance, individuals in the areas left behind by globalisation (the “places that don’t matter”)
where more likely to vote Leave if they also had high levels of immigration. More specifically,
in this section of the analysis, we estimate the difference in the propensity to vote leave across
areas with (i) high/low values of wage growth and immigration, and (ii) high/low values of skill
levels and immigration.7 The first combination (wage growth / immigration) tests the “revenge
of the places that don’t matter” hypothesis, while the second combination tests the view that
high skilled areas have cultural norms in favour of diversity. The model to be estimated now
becomes:
𝑃𝑟 𝑌 = 1
𝑙𝑛 = 𝛼 + 𝛽4 𝑇4 + 𝛽5 𝑇5 + 𝛽6 𝑇6 + +𝛾 1 𝑋 + 𝜀 (2)
1 − 𝑃𝑟 (𝑌 = 1)
where the treatment now takes the form of four binary variables T1–T4, capturing
combinations of high/low values for the two variables of interest, with T1 (Low/Low) the
13
Our aim is to match very similar individuals living in different locations in order to improve
the balance of our sample. Since there is a trade-off between ensuring a good match and
minimising the loss of observations, we match on personal characteristics that have been
identified in previous work as being strongly associated with political views and voting
behaviour.8. Our matching variables are:
Age: the individual’s age in years. This is our only continuous variable, and we include the un-
coarsened variable in the logistic model to control for any remaining imbalance.
Education: a series of dummy variables indicating the individual’s highest qualification, out of (i)
no qualifications, (ii) GCSE D-G grades, a low school-leaving qualification, (iii) GCSE A*C
grades, a high school-leaving qualification, (iv) A-level or equivalent, a post-16 college
qualification, (v) undergraduate degree or equivalent, and (vi) postgraduate degree or
equivalent.
Ethnicity: we include a dummy variable to indicate whether the individual is white (British), vs.
all other ethnicities.
Employment circumstances: we include the three dummy variables for the employment
circumstances that are most highly correlated with the propensity to vote Leave: employed,
student, and retired.
Household income: we include a dummy variable to capture whether the individual’s household
has a gross annual income equal to or above £35,000.
Personality: we include two variables to capture the individual’s personality traits, based on the
Big Five psychometrics framework. The two traits that are most highly correlated with the
propensity to vote Leave are Openness and Conscientiousness, and the BES survey provides a
14
Political efficacy: a dummy variable that is equal to one if the individual strongly agrees or agrees
with the statement “politicians don't care what people like me think”, and zero otherwise.
Daily newspapers: a categorical variable indicating which daily newspaper the individual reads
regularly, out of Express, Daily Mail, Mirror, Sun, Telegraph, Guardian, Independent, Times,
local papers, other (including smaller circulation national newspapers), and none. This variable
allows us to control for the effects of the material in the news media on individual voting
intentions, and also to match individuals to others with similar cultural views which are not
otherwise controlled for by the other matching variables. There is extensive evidence to show
that the editorial line of British newspapers was starkly political during the referendum
campaign, and that they may have influenced the propensity to vote Leave (Levy et al., 2016).
4. Results
We first use descriptive statistics to analyse whether there are differences in the Leave vote for
similar individuals, across Leave and Remain voting constituencies. The results are shown in
Table 1.
The table illustrates the nature of composition issue. For instance, only 29% of graduates voted
Leave in Remain-voting areas, but 46% of graduates voted Leave in Leave-voting areas. This
could be either because graduates in Leave-voting areas have other personal characteristics
(such as age, income, personality, etc.) which are associated with a higher propensity to vote
Leave, or it could be that they are influenced by contextual factors. It is interesting to note that
there is less variation in the Leave vote for individuals with other levels of qualification, for
instance for individuals without a degree (for which there is only a 11% difference in the Leave
vote across locations). Note also that the difference in the Leave vote between employed and
unemployed individuals is relatively small in Leave-voting areas (3.5 percentage points), but
significantly greater in Remain voting areas (9 percentage points), while students have fairly
homogenous voting intentions across locations.
15
Table 2 shows our basic result (in the form of the SATT), where the “treatment” is being
resident in a Leave or Remain constituency. Our initial, unweighted, estimate of the SATT is
2.13, indicating that the odds that an individual living in a Leave-voting constituency will vote
Leave are 2.13 times greater than the equivalent odds for an individual living in a Remain-
voting constituency. After matching on a wide range of personal characteristics, we re-run the
regression model using CEM weights to control for non-random selection into treatment; the
results show that the SATT is still positive and significant, although smaller at 1.42. For very
similar individuals, we therefore find that the odds that an individual will vote Leave are 1.42
times higher in Leave-voting constituencies than in Remain-voting constituencies. This
suggests that although a large proportion (around a third) of the Leave vote can be explained
by composition effects alone, there is still a considerable proportion (around two thirds) that is
unaccounted for.
We next investigate several of possible explanations in turn, and the results are shown in the
remainder of Table 2. We find that there is a positive and statistically significant SATT resulting
from living in a high-UKIP-vote constituency, and the increase in the odds of voting Leave
(1.39) due to this treatment is similar in magnitude to the effect of living in a Leave-voting
16
The next set of estimates compare high- and low- immigration constituencies. We find that the
odds of voting Leave are substantially lower in high-EU immigration constituencies and in high
immigration growth areas (although slightly less significantly in the latter case), contrary to the
popular view, and the findings in some of the studies discussed in Section 3.2, that the Leave
vote was driven primarily by an increase in immigration. Moreover, after controlling for
composition effects (using CEM weights), this result disappears entirely in both cases. We find
that there is no statistically significant effect on the odds of voting Leave for individuals located
in high immigration growth, or high EU-migration, areas, relative to those living in low
immigration areas. Interesting, however, we do find that living in a constituency that has strong
anti-immigration views significantly increases the odds of voting Leave, even after controlling
for composition effects. Individuals in anti-immigration constituencies have 1.34 higher odds
of voting Leave than those in more pro-immigration areas. This finding suggests that cultural
(as opposed to economic) factors may explain a significant proportion of the Leave vote.
Our next set of results considers the “revenge of the places that don’t matter” and the
“metropolitan elites” theories, whereby places that are ageing and in industrial decline, and
places that are young, highly-skilled, and innovative, have very different collective cultures and
values, in a way that also influences individual voting behaviour. We find that the odds of voting
Leave are significantly higher in low-turnout and low-wage-growth areas, although the
magnitudes are small (an increase in the odds of 1.14 and 1.08, respectively). We also find that
the impact of cultural factors, such as low turnout (reflecting low confidence in the political
system), is more significant than the impact of economic factors, such as wage growth and the
working age inactivity rate. Moreover, we find that individuals living in high-skill constituencies
have significantly lower odds of voting Leave, regardless of their own skill levels and employment
circumstances.
We test two additional hypotheses: that the propensity to vote Leave was driven by the austerity
cuts of recent years, and separately, that it was driven by growing inequality. We should note,
as discussed in Section 3.2, that our measures for these two variables are imperfect, due to a
lack of data on inequality and government spending at this level of disaggregation.
Nevertheless, our results show that, after controlling for composition effects, individuals living
17
We have so far considered estimates that compare two sets of locations, but contextual effects
may work along more than one dimension. We next consider two-way interactions between
the variables identified as having the greatest effects on voting behaviour, with a focus on the
“revenge of the places that don’t matter” and the “metropolitan elites” theories. These results
are shown in Table 3, where all the estimates are based on models that have been adjusted for
composition effects (using CEM weights). The effects shown therefore reflect contextual effects
only.
First, we consider the increase in the odds to vote Leave arising from an interaction between
wage growth and immigration, based on the “revenge of the places that don’t matter” and the
“immigration activates populist vote in economically vulnerable communities” arguments for
the Leave vote. Our reference category are low wage growth and low immigration
constituencies, and we consider three measures of immigration: immigration from EU
accession countries, immigration growth rates, and anti-immigration views. Our results show
that relative to the reference category, individuals living in high immigration growth areas had
similar odds of voting Leave, as did individuals in high EU immigration areas, after controlling
for composition effects. However, we also find higher odds of voting Remain for individuals in
high wage and high immigration areas (mainly diverse metropolitan areas). We also find strong
contextual effects when considering areas with strong anti-immigration, with individuals living
in these areas having between 1.6-1.7 times higher odds of voting leave (relative to the reference
category of low wage growth and low anti-immigration views). Interestingly, this effect appears
to be strongest for high-wage growth areas (relative to low-wage growth areas). This suggests
that immigration is an activating factor in wealthier constituencies, with high-immigration
growth rates and levels decreasing the odds of voting Leave in high-wage growth constituencies,
while anti-immigration views increase the odds of voting Leave in similarly high-wage growth
areas.
18
5. Conclusions
Following the surprising result of the EU referendum in the UK, a now significant literature
has investigated the individual and regional determinants of the vote. The factors identified
include (at the individual level) age, education, occupation, psychological traits, and political
views, and (at the regional level), demographics, immigration, austerity policies, exposure to
trade shocks, economic deprivation, and cultural values. Moreover, it is clear that there are
interaction effects between these individual and regional determinants, as shown in a number
of studies (Alabrese et al., 2018; Goodwin and Heath, 2016b; Gordon, 2018; Lee, 2018). An
under-researched question in this context is whether the observed regional variations are due
to differences in the composition of the electorate across regions, or whether they are due to
the contextual effects of regional characteristics on individual voting intentions, or a
combination of the two.
We aim to fill this gap in the literature by using a non-parametric matching approach to
compare the voting preferences of very similar individuals living in different types of
constituencies. We use individual-level data form the post-referendum wave of the British
Election Study (Wave 9) and analyse the vote across Leave and Remain-voting constituencies,
as well as across constituencies with high and low values of: UKIP support, electoral turnout,
wage growth, immigration growth, EU immigration levels, inactivity rates, skills, inequality,
and anti-immigration views. Our approach allows us to quantify the relative size of the
composition and contextual effects, and the increase in the odds of voting Leave for very similar
individuals living in different types of constituencies.
19
Our analysis also suggest that the most important contextual determinants of the Leave vote
are cultural rather than economic in nature, possibly reflecting the presence of strong social
interaction, solidarity, and emulation effects in voting behaviour. For instance, we find that
individuals in constituencies with strong anti-immigration views had significantly higher odds
of voting Leave, and that this effect is stronger in high-wage-growth areas (1.7 times higher vs.
1.6 times for those in low-wage-growth areas). Similarly, living in a high-EU-immigration
constituency is associated with an increase in the odds of voting Leave if the constituency is also
a low-skills one (1.3 higher odds), but is associated with a decrease in the odds of voting Leave if
it is also a high-skills one (1.1 lower odds). The latter type of constituency is typically located in
a diverse metropolitan area or in a university town.
Finally, we note that the observed contextual effects could also result from targeted advertising
during the referendum campaign, with resources likely to have been directed to areas with
potential Leave voters, given their economic and social contexts (for instance, economically-
declining areas, with low skill levels, high levels of deprivation, and ageing populations). This
might explain the significant increase in turnout in areas typically associated with low political
engagement. While the data necessary to test this hypothesis are not currently available, the
role of turnout may turn out to be very significant in explaining the contextual effects in the
EU referendum.
20
citizens (excluding Ireland, Cyprus, and Malta, although EU citizens can vote in local elections), and some other
categories (members of the House of Lords, prisoners, those convicted of electoral offences).
3 The median weekly gross pay data are deflated using the Consumer Price Index, compiled by the Office for
National Statistics.
4 The degree of balancing achieved can be measured using a multivariate imbalance measure (denoted L ), which
1
is based on the difference between a multi-dimensional histogram of all pre-treatment matching variables in the
treated group vs. in the non-treated group. An improvement in the balance is indicated by a fall in the size of this
measure.
5 Our CEM models were estimated using the imb and cem routines in Stata (Blackwell et al., 2009).
6 We would ideally model the impact of austerity using estimates for the spending cuts using area-based estimates
of the reduction in local government spending (Fetzer, 2018; Gray and Barford, 2018). However, these estimates
are only available at the Local Authority District (and equivalent) level, as this is the level at which local
government spending decisions are made. Previous research has shown that the spending cuts fell
disproportionately on the most deprived areas, which have the lowest revenues, are more reliant on the
government grant, while also facing significant welfare spending commitments (Innes and Tetlow, 2015). Areas
with a high proportion of welfare recipients were most heavily affected by the spending cuts, and we use the
percentage of the working-age population in receipt of welfare payments in 2010, before the start of the austerity
programme, as a proxy for the local impact of the cuts.
7
We use four categories: Low/Low, Low/High, High/Low, and High/High constituencies, according to whether
the constituency for each individual has below- or above-average values for the two variables of interest. Low/Low
the reference (omitted) category.
8
Table A1 in the appendix provides descriptive statistics for the matching variables included in the analysis.
21
Alabrese, E., Becker, S. O., Fetzer, T., & Novy, D. (2018). Who voted for Brexit? Individual
and regional data combined. European Journal of Political Economy, 56, 132-150.
Alesina, A., Carloni, D. and Lecce, G. (2011). The electoral consequences of large fiscal
Adjustments. NBER Working Paper Series, No. w17655.
Arnorsson, A. & Zoega, G. (2018). On the causes of Brexit. European Journal of Political Economy,
55, 301-323.
Ashcroft, M. (2016). “How the United Kingdom voted on Thursday... and why.” Available
from http://lordashcroftpolls.com/2016/06/how-the-united-kingdom-voted-and-why.
Barber, M., & Imai, K. (2014). Estimating neighborhood effects on turnout from geocoded
voter registration records. Mimeo, Princeton, NJ.
Bartle, J., Birch, S., & Skirmuntt, M. (2017). The local roots of the participation gap: Inequality
and voter turnout. Electoral Studies, 48, 30-44.
Becker, S. O., Fetzer, T., & Novy, D. (2017). Who voted for Brexit? A comprehensive district-
level analysis. Economic Policy, 32(92), 601-650.
Blackwell, M., Iacus, S. M., King, G., & Porro, G. (2009). cem: Coarsened e xact matching in
Stata. The Stata Journal, 9(4), 524-546.
Blinder, S., & Markaki, Y. (2018). Public Attitudes Toward EU Mobility and Non-EU Immigration: A
Distinction with Little Difference (Vol. 10). Working Paper, Reminder Project, August 2018.
https://www. reminderproject. eu/wp-content/uploads/2018/08.
Burn-Murdoch, J. (2016). “Brexit: voter turnout by age.” FT Data Blogs, Financial Times, 24
June 2016. Available from http://blogs.ft.com/ftdata/2016/06/24/brexit-demographic
-divide-eu-referendum-results/.
Carter (2018). Fear, Hope and Loss. Understanding the Drivers of Hope and Hate. Hope Not Hate,
London.
Colantone, I., & Stanig, P. (2018). Global competition and Brexit. American Political Science
Review, 112(2), 201-218.
Crescenzi, R., Di Cataldo, M., & Faggian, A. (2018). Internationalized at work and localistic
at home: The ‘split’ Europeanization behind Brexit. Papers in Regional Science, 97(1), 117-
132.
Dorling, D. (2016). Brexit: the decision of a divided country. British Medical Journal, 354, i3697.
Dorling, D., & Tomlinson, S. (2019). Rule Britannia: BREXIT and the End of empire. London:
Biteback Publishing.
22
23
24
25
Table 1. Leave vote (% of total) across Leave and Remain voting constituencies, by group based
on personal characteristics.
26
SATT SATT
(no weights) (CEM weights)
27
28
29
Univariate imbalance:
Matching summary
0 1
All 6606 9911
Matched 3422 4874
Unmatched 3184 5037
Univariate imbalance:
30
Pre-matching L1 Post-matching L1
Leave area 0.837 0.622
High-UKIP vote area 0.815 0.618
High immigration growth area 0.830 0.632
High EU migration area 0.845 0.630
Anti-immigration area 0.835 0.622
Low turnout area 0.822 0.622
Low wage growth area 0.812 0.622
High working-age inactivity area 0.819 0.605
High skills area 0.827 0.636
High welfare claimants area 0.813 0.619
High inequality area 0.814 0.610
31