Electoral Studies 77 (2022) 102471
Available online 26 April 2022
0261-3794/© 2022 The Author. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Does university make you more liberal? Estimating the within-individual
effects of higher education on political values
Ralph Scott
Department of Politics, School of Social Sciences, Arthur Lewis Building, University of Manchester, Manchester, M13 9PL, UK
A R T I C L E I N F O
Keywords:
Political values
Prejudice
Authoritarianism
Education
University
Selection effects
Longitudinal
Panel estimation
Two-way fixed effects
Linear mixed effects modelling
A B S T R A C T
An individual’s level of education is increasingly significant in explaining their political attitudes and behaviour,
with higher education proposed as a new political cleavage. However, there is limited evidence on the causal
effect of university on political attitudes, due to self-selection into educational pathways. Addressing this gap,
this article estimates the change in political values that occurs within individuals who graduate from university
by applying longitudinal modelling techniques to data from the 1970 British Cohort Study, overcoming the
selection problem by accounting for time-invariant confounding. It provides the first causal estimate of higher
education specifically, finding that achieving a degree reduces authoritarianism and racial prejudice and increases
economic right-wing attitudes. This has important implications for the study of politics: as populations
become more highly educated on average, we should expect continuing aggregate value change towards lower
levels of authoritarianism and racial prejudice, with significant consequences for political behaviour.
1. Introduction
Within established democracies, an individual’s level of education is
increasingly significant in explaining their political attitudes and
behaviour (Norris and Inglehart, 2019), with some suggesting that
higher education in particular could represent a new political cleavage
in Western Europe (Ford and Jennings, 2020). One potential explanation
for this observed effect of higher education is through changes to political
values. Studies have shown that these values are valid consistent
constructs that vary by education level (Evans et al., 1996) and are more
often influential on party identification than the other way around
(Evans and Neundorf, 2018). This analysis focuses on three such values:
racial prejudice, understood as hostility towards racial out-groups
(Kinder and Kam, 2010, p. 8); authoritarianism, or support for social
order over individual liberty (for example, the death penalty or harsh
sentences for criminals); and economic Left-Right values which captures
views on inequality and the role of the state in the economy (Evans et al.,
1996).
Competing explanations are given for the particular effect of higher
education on these different values, whether it is the graduate premium
leading to higher earnings, and so lower support for redistribution
(Surridge, 2016); the liberalising influence of faculty and university
culture (Dey, 1996); the impact of peer socialisation in what remain
relatively elite institutions (Mendelberg et al., 2017); or the effects of
increased cognitive sophistication (Gelepithis and Giani, 2022). However,
given self-selection into educational pathways (Persson, 2015), an
important first step is to determine whether this difference is causally
attributable to university. Previous studies have attempted to address
this selection problem by leveraging exogeneous changes in educational
participation attributable to variation in policy regimes (Bullock, 2021;
Cavaille and Marshall, 2019; Marshall, 2016) or randomised encouragement
designs (Sondheimer and Green, 2010). However, these
well-identified studies do not estimate the specific effect of higher education,
instead tending to focus on the effect of increased years of
secondary education.
It is this gap that this article addresses, adopting a longitudinal
approach to estimate the change in values that occurs within individuals
who graduate from university. Specifically, it applies two-way fixed
effects and linear mixed effects models to data from the 1970 British
Cohort Study to account for all time-invariant confounding, finding that
achieving a university degree reduces an individual’s authoritarianism
and racial prejudice and makes an individual more economically rightwing.
In so doing, it makes three main contributions. Firstly, it provides
the first causal estimate of the effect of university specifically on these
values. Second, it does so for a validated multi-item scale for authoritarianism
for the first time, which matters given the growing significance
of this ‘second’ dimension for contemporary politics. Third, it
extends the previous causal literature in this area by providing an
E-mail address: ralph.scott@manchester.ac.uk.
Contents lists available at ScienceDirect
Electoral Studies
journal homepage: www.elsevier.com/locate/electstud
https://doi.org/10.1016/j.electstud.2022.102471
Received 25 June 2021; Received in revised form 25 March 2022; Accepted 7 April 2022
Electoral Studies 77 (2022) 102471
2
estimate of the effect of higher education among a more nationally
representative sample, therefore allowing us to be more confident in
drawing general conclusions. The findings have important implications
for the study of public opinion and political attitudes, as it suggests that
as the population becomes more highly educated on average, we should
expect current trends of aggregate value change towards lower levels of
authoritarianism and racial prejudice to continue, with significant
consequences for political behaviour.
2. The significance of political values
From an initial position of scepticism about the existence of underlying,
consistent attitudes among much of the voting public (Converse,
2006), there is now broad acceptance within political science that many
individuals hold essentially consistent underlying positions in terms of
racial prejudice, left-right economic and liberal-authoritarian social
values throughout their adult lives (Evans et al., 1996; Sears and Levy,
2013). In part, this change in perspective can be attributed to the difference
between searching for stability in individual survey items, which
do demonstrate inconsistency partly attributable to measurement error,
and stability in the latent constructs which selections of these items are
designed to estimate (Ansolabehere et al., 2008).
Existing empirical evidence on political socialisation suggests that
attitudinal orientations are developed early on in life but are subject to
change based on early adult and maturation experiences, becoming
relatively stable from that point on. Longitudinal studies, such as the
Political Socialisation Project, find considerable consistency within individuals
over the life course across issues including prejudice against
other races and the role of trade unions, with values particularly stable
after early adulthood (Stoker and Jennings, 2008). Similarly, analysis of
the General Social Survey in the US finds that, after the age of 30, individuals
demonstrate a high degree of consistency over time even at the
item-level (Kiley and Vaisey, 2020), although there are exceptions due
to period effects (for example, on the issue of gay marriage). Importantly,
in the British case at least, research designed to disentangle the
relationship between these ‘core values’ and party identification
through cross-lagged modelling suggests that the causal direction is
from values to party identification, with values possessing twice the
amount of stability (Evans and Neundorf, 2018).
To discuss the specific values sets under study here, racial prejudice
describes a general predisposition to favour perceived racial in-groups
and oppose out-groups, measured through scales addressing out-group
stereotypes, for example, how close an individual feels to people of
different ethnicities (Kinder and Kam, 2010). So measured, the construct
demonstrates a high degree of within-individual validity and stability
over a lifespan. As described by Fleishman (1988), the other two attitude
sets of interest were derived after a single dimension of political values
proved insufficient in explaining variation among the general population.
While the survey items used varies, in the British case the economic
Left-Right scale measures an individual’s degree of support for free enterprise
or state involvement in the economy, based on statements such
as “big business benefits owners at the expense of workers”, while the
social liberalism-authoritarianism scale measures attitudes towards
liberty and individual rights, such as “people who break the law should
be given stiffer sentences” (Evans et al., 1996). Factor analysis models
have validated this structure of political values (Fleishman, 1988; Heath
et al., 1994), and position on each of these scales is also shown to be
stable in individuals over time. While some treat racial prejudice and
authoritarianism as part of the same ‘cultural values’ construct
(Fleishman, 1988), following Sobolewska and Ford (2020) I treat it
separately, and justify this through Confirmatory Factor Analysis, presented
in appendix 6 in the supplementary material.
This validity and within-person stability makes understanding the
role of higher education in the formation of these values important in its
own right: if an effect is present once accounting for selection, then
university attendance plays an important role in an individual’s political
socialisation, with implications for values formation in the aggregate
given increasing participation rates. However, additional motivation is
provided by the importance of these values for understanding contemporary
political behaviour. Analysis of the British Election Study finds
that party competition is increasingly taking place more along the social
liberal-authoritarian and racially prejudicial axes rather than the
traditionally dominant economic Left-Right axis (Fieldhouse et al.,
2020), with the electoral ‘shock’ of Brexit accelerating a pattern of
realignment across these values axes. Furthermore, decompositional
analysis suggests that the relationship between education and vote
choice is primarily mediated by its effect on cultural values (Simon,
2021).
This underlines the importance of understanding whether higher
education is truly driving the apparent differences in these values, or
whether the observed relationship is a result of selection. If education
itself plays a role, then the increasing rate of participation in higher
education will have significant implications for both political values and
behaviour in the future; if the effect is attributable to selection, then we
must look to other social divides to explain the changing basis of party
competition.
3. Theoretical perspective and hypotheses
3.1. Education effects
The important influence of early adulthood socialisation on these
values raises the question of the role that education might play in their
formation. This section discusses the evidence for the relationship between
education and each value set in turn, drawing on this to formulate
hypotheses to be tested in the analysis.
3.1.1. Racial prejudice
In general, those with higher levels of education demonstrate lower
levels of racial prejudice. This is attributed by Kinder and Kam to the
“knowledge, principles, and experiences that together act as a counterweight
to the ‘natural’ inclination toward prejudice” (Kinder and Kam,
2010, p. 35). Cross-cultural analysis of the World and European Values
Surveys by Drazanov´a (2017) indicates this appears to be the case across
contexts, finding a strong effect of education on social tolerance (defined
in opposition to prejudice), although this does depend on the strength of
the liberal-democratic tradition in the country.
Research designs aimed at estimating a causal effect also tend to
draw these conclusions, while investigating racial outgroup prejudice
and related constructs such as opposition to immigration. A series of
studies applying fuzzy regression discontinuity designs (RDD) to European
Social Survey data, where policy changes to the compulsory education
age are used to instrument for the effect of education, reinforce
these findings. Focusing on racial prejudice, Gelepithis and Giani (2021)
find significant reductions in prejudice among the ethnic majority as a
result of the extension to the secondary-school leaving age. Looking
instead at anti-immigration attitudes, Cavaille and Marshall (2019) find
that increased years of schooling reduce overall anti-immigrant sentiment,
and Nunziata and d’Hombres (2016) draw similar findings from
an equivalent research design across a broader range of European
countries. In line with the conclusions of Hainmueller and Hopkins
(2014) both analyses conclude that this effect is not purely attributable
to the economic benefits of higher levels of education (placing the more
highly educated at less risk of labour market competition from migrants),
but also due to changes in cultural outlook.
However, other studies contradict these findings. Using a similar
RDD research design, but with a more focused case study of Norway,
Finseraas et al. (2018) find no significant effect of increased years of
education on attitudes towards immigration. Lancee and Sarrasin (2015)
also dispute a causal role for education in reducing opposition to
immigration through a longitudinal panel design, finding little variation
within individuals over time, even during educational transitions (ie
R. Scott
Electoral Studies 77 (2022) 102471
3
progressing to university), suggesting instead that the observed differences
are attributable to selection effects. In addition, while
well-identified, it is worth noting that the RDD studies discussed here are
limited by variable and relatively weak first stage results, with the
extension of the compulsory school age by one year generally leading to
an increase in school participation of around four months on average.
This suggests the treatment effects identified by these studies may be
quite local to those who would otherwise have left education at a young
age.
With all of this in mind, we can formulate the following hypothesis to
be tested in the analysis:
H1. achieving a university degree leads to a reduction in racial
prejudice.
3.1.2. Authoritarianism
The observed relationship between education and authoritarianism
can be traced back at least as far as Lipset, who in seeking to understand
high levels of authoritarianism among members of the working-class,
argued that this sprung from “greater suggestibility, absence of a sense
of past and future, inability to take a complex view, difficulty in
abstracting from concrete experience, and lack of imagination” Lipset
(1959, p. 492) resulting both from lower levels of education and a more
precarious economic position.
Scholars since have reinforced this finding of a negative relationship
between education and authoritarianism. For example, Surridge (2016)
accounts for numerous possible selection effects in analysis of the 1970
British Cohort Study, and finds a large reduction in authoritarianism
among graduates. In addition, given that the liberalising effect of higher
education differs by the degree subject of study (with humanities and
social science students the most liberal), she concludes that the change
can be attributed to education. Through a similar design (albeit
adjusting for fewer confounders) but also analysing data from the prior
1958 birth cohort study, Paterson (2009) draws the same conclusions.
Despite this precedent, there are relatively few analyses designed to
produce a causal estimate of the effect of education specifically on
authoritarianism, with those that do exist focused more on related issues.
An early example is provided by Dee (2004), where the effect of
more years of schooling is instrumented through the introduction of
restrictions on child labour and the proximity of two-year colleges,
drawing the conclusion that more education leads to greater support for
freedom of speech. In their study, Nunziata and d’Hombres (2016) find
that those who received more years of secondary education were more
likely to express liberal attitudes on sexuality. Campbell and Horowitz
(2016) account for the effects of family background by comparing sibling
pairs where one is a university graduate and the other is not through
Study of American Families data, finding that graduates are significantly
more supportive of civil liberties and gender equality.
Other research has investigated the effect of education on related
values through panel designs. For example, through analysis of the
Citizenship Education Longitudinal Study, Janmaat (2018) finds that
those who pursue an academic route for post-16 education develop
higher levels of support for liberal democratic principles (including
freedom of speech and tolerance of others). In their study of changes in
moral dispositions using panel data from the National Study of Youth
and Religion, Bro´ci´c and Miles (2021) find that graduates in the humanities
and social sciences become more morally progressive during
the course of their studies, for example, being more likely to agree that
moral principles should be updated based on societal changes, while also
finding similar differences in terms of a greater regard for the individual
over social order (albeit with a weaker design). Yet the issue of there
being no study which provides a causal estimate of the effect on a
validated authoritarianism scale remains.
Still, based on this we can propose the following hypothesis:
H2. achieving a university degree leads to a reduction in
authoritarianism.
3.1.3. Economic left-right values
Finally, previous research also identifies a relationship between education
and economic values, with graduates tending to be more
economically right-wing. Weakliem (2002) demonstrates this association
from analysis of World Values Survey data, concluding that the
effect of education is to make a person more individualistic, rendering
them opposed to higher taxation and a bigger role for the state in the
economy, but more supportive of an individual’s right to choose in social
matters, as discussed above. Yet this analysis does not address the selection
problem, not least as in most jurisdictions we would expect the
wealthier to be both more likely to hold economically right-wing views
and to have the opportunity to continue in education.
Surridge (2016) finds that this relationship persists while adjusting
for confounders including an individual’s family background and their
pre-university values. She also suggests that this difference between
graduates and non-graduates is in large part attributable to the
socio-economic position enabled by higher education, as once the mediators
of income and occupation are accounted for the relationship is
substantially weaker. In contrast, drawing on the European Social Survey
to provide a more comparative analysis, Gelepithis and Giani (2022)
emphasise the importance of ideas in explaining the effect of education
on economic values, as they find the effect remains even when accounting
for income and labour market position. They identify this effect
as being specifically attributable to higher education, and echo Weakliem
in concluding that: “university education fosters norms of inclusion,
while eroding norms of solidarity” (2022, p. 2).
A series of studies using research designs appropriate for determining
causal effects reinforce this finding. Using legal changes in compulsory
schooling in the US as an instrument, Bullock (2021) finds that more
years of secondary schooling makes an individual more economically
right-wing. Marshall (2016) employs a fuzzy RDD design on British
Election Study data to estimate that each additional year of schooling
leads to an increase of 12 per cent in probability of voting Conservative,
which, as it appears to be mediated by increases in income, can be
considered a good proxy for right-wing economic values. Finally, a study
of US college students using a panel design with matching (Mendelberg
et al., 2017), finds that those attending colleges with a critical mass of
affluent students became more economically right-wing during their
time there (addressing numerous other potential explanatory factors),
an effect they attribute to norm socialisation.
We can therefore propose the following hypothesis:
H3. achieving a university degree makes an individual more
economically right-wing.
3.2. Selection effects
Despite this foregoing evidence for education effects, an alternative
argument is that the observed differences are more attributable to selection
(Lancee and Sarrasin, 2015). Drawing on the approach of Surridge
(2016) and other literature we can hypothesise selection effects
that may be confounding the relationship between education and values,
such as differences in demographics, family background or cognitive
skill. For example, the difference in values could be explained by
compositional differences by gender and ethnicity in who attends university;
parental occupation, income and education; or indeed the
cognitive selectivity of higher education, as some have argued that:
“when considering social, moral, and political situations, those with
greater cognitive skill are able to form more individualistic and
open-minded (i.e. anti-authoritarian) attitudes than those of lesser
cognitive ability” (McCourt et al., 1999, p. 987). It is therefore important
to ensure that the effect of education identified in this analysis is robust
to these influences, by accounting for them where appropriate in the
estimation.
R. Scott
Electoral Studies 77 (2022) 102471
4
4. Methodology
4.1. Estimand
Lundberg et al. (2021) argue that every quantitative study should
explicitly state its estimand, that is, the target quantity of the research.
Doing so provides a clear link from the theory through to the observed
data and estimation strategy, which enables identification of this
quantity under a set of assumptions. In this spirit, let it be clear that the
theoretical estimand of this study is the causal effect of graduating from
university on each of authoritarianism, racial prejudice and economic
Left-Right values, as set out in the hypotheses.
As the focus is on the within-individual change among those who
achieve a degree, this treatment effect can be considered the Average
Treatment Effect on the Treated (ATT), which within the potential
outcomes framework assumes that “the average level of the outcome
under the control for those in the treatment is equal, on average, to the
average level of the outcome under the control for those in the control
group” (Morgan and Winship, 2015, p. 159). When applied to causal
analysis of panel data, as in the Difference-in-Difference (DiD) approach
to estimation, this is known as the parallel or common trends assumption
(Angrist and Pischke, 2008, p. 171). This assumes the trajectory of
those who receive treatment would have otherwise been similar to those
in the control condition, with the difference in these trajectories being
the estimate of the causal effect. A relaxation of this assumption is the
conditional parallel trends assumption (Callaway and Sant’Anna, 2021),
which suggests that if the two groups appear to follow similar trends
once conditioned on pre-treatment confounding variables, there is no
threat to causal inference.
As will become clearer in the discussion of the data and methods, the
empirical estimand of this study is the difference in the change in these
values (the ‘within-individual’ effect) between those who do and don’t
graduate from university, conceptualised as the treatment. This estimation
will account for the effects of selection, in addition to adjusting
for the influence of pre-treatment confounders in some specifications.
This will be the total, unmediated effect of holding a university degree,
which could be attributable to any number of the explanations provided
by previous research on the topic, whether social and economic position
post-treatment, norm socialisation during study, or the effect of learning
on values. However, crucially, the contribution of this analysis is to
robustly separate out the effect of selection from that of the treatment
itself.
4.2. Estimation strategy
In a review of the evidence on the effect of education on political
participation, Persson (2015) makes plain that the difficulty of randomly
assigning individuals to different educational paths often entails the use
of sophisticated methods for causal inference with observational data to
produce a less biased estimate of the effect of education (although see
Sondheimer and Green (2010) for examples of an experimental
approach). In an attempt to do so through a longitudinal mixed model
using the Swiss Household Panel, Lancee and Sarrasin (2015) dispute a
causal role for education in reducing opposition to immigration, finding
little variation within individuals over time, even during educational
transitions (ie progressing to university), suggesting instead that the
observed differences are attributable to selection effects. This is a significant
finding, as if the effect of higher education is reducible to selection,
then we would not expect increased participation in university
to affect political attitudes in the aggregate. Yet the study explores a
range of educational transitions at once, with only 7 per cent of cases
attending university, meaning that these estimates may lack sufficient
power to determine a precise null effect.
To discuss the models in more detail, a common approach to DiD
estimation is to apply the two-way fixed effects model (TWFE), formalised
as follows:
yit = αi + γt + βxit + εit (1)
Here the subscripts i and t signify individuals and timepoints
respectively, while y is the dependent variable. As described by Imai and
Kim, the average within-individual change attributable to a time-variant
treatment (represented by βxit) is estimated by including fixed effects for
both the unit (αi) and time (γt) and so the model “accounts for both unitspecific
(but time-invariant) and time-specific (but unit-invariant) unobserved
confounders in a flexible manner” Imai and Kim (2020, p. 2).
The result is that the estimated effect of the time-varying treatment is
not subject to bias attributable to unobserved confounding, enabling
greater confidence that is what is observed is a causal effect.
However, a downside of this approach is that it provides no information
on the effect of time-invariant covariates, including the ‘between’
effect (ie the average differences between those who are treated
and those who are not) (Bell et al., 2019). It also assumes a homogeneous
treatment effect, which is liable to produce biased estimates if the
effect varies across units (de Chaisemartin and D’Haultfœuille, 2020).
An alternative is to adopt the hybrid approach of Lancee and Sarrasin
(2015), and combine fixed and random effects in a panel design. Such a
model, a version of which was first formalised by Mundlak (1978) and
further developed by Bell et al. (2019) as the Random Effects
Within-Between model (REWB), can be expressed as:
yit = μ + γt + β1W (xit − xi) + β2Bxi + β3zi + vi0 + vi1(xit − xi) + εit0 (2)
The key parameter here is β1W, the within-individual effect,
measured as the time-specific demeaned treatment status. This estimates
the change in the outcome within the individual attributable to a change
in treatment condition, accounting for the effects of time (γt). Wooldridge
(2021) has demonstrated the equivalence of this ‘within-effect’
and the TWFE estimator. β2B is the between effect, parameterised as the
individual-level grand mean of treatment, which estimates the overall
difference between individuals who are and are not treated at the point
of treatment, and therefore the effects of selection. The effects of additional
time-invariant covariates (which are themselves between-effects)
can be incorporated as fixed effects as represented by β3. Finally, the
model estimates two random effects: first, a random intercept for each
individual (vi0 , while μ expresses the mean intercept value) and second,
a random slope for the effect of treatment (vi1xit). This latter allows
heterogeneous treatment effects and incorporates this variation into the
estimation of the standard errors.
As shown, both models account for all unobserved time-invariant
confounding and therefore selection effects. The causal identification
is thereby strengthened by focusing on the change within individuals
across the two treatment groups over time (less the effects of time itself).
While the strength of the TWFE is its simplicity, the REWB enables two
further extensions to test the robustness of the findings. First it explicitly
accounts for pre-treatment differences in the outcome and observed
covariates between individuals at the point of treatment. This supports
the partial relaxation of the parallel trends assumption to a conditional
parallel trends assumption, conditioning on those pre-treatment covariates
expected to confound the relationship between higher education
and the outcomes. A further test of this is presented in appendix 4 of the
supplementary material, where the models are fitted on entropybalanced
data (Hainmueller, 2012), weighted so that the treated and
control groups have the same values on the pre-treatment outcome and
covariates, producing broadly similar results.
Second, the use of random effects parameters in the REWB does not
simply model out the variance attributable to the grouping variables as
in the TWFE approach, but instead models this assuming the variance
follows a normal distribution (Bell et al., 2019). In the approach taken
for this analysis, the mixed effects models include random intercepts for
the outcome (factoring in each individual’s unique starting point) and
random slopes for the treatment, allowing for variation in the effect of
treatment, thereby reducing bias (Long, 2011, chap. 5). In addition, the
R. Scott
Electoral Studies 77 (2022) 102471
5
estimation of these parameters is carried out with ‘partial pooling’,
whereby the estimates (for example, the intercept for a given individual)
recognise the dependence within the data and shrink the more extreme
estimates towards the mean (Gelman and Hill, 2007).
For this reason, the results section will present results estimated
through both TWFE and REWB modelling. The TWFE models are fitted
using the lfe package for R (Gaure, 2013), while the REWB models are
fitted using the lmer package (Bates et al., 2015b), with the parameters
estimated using restricted maximum likelihood estimation with the
nloptwrap optimiser using the BOBYQA algorithm. On fitting the models,
the estimated random slope for the within-effect parameter for racial
prejudice was singular, and so following various attempted modifications
to the model and optimiser, this was dropped from the final model
following the recommendations of Bates et al. (2015a). The standard
error for this parameter estimate should therefore be interpreted with
greater caution – although not overly so, given the minimal difference
with the standard error for the within-effect in the entropy-balanced
model which contains a random slope (in appendix 4). Further details
on the models and diagnostics are available in appendix 3.
5. Data
The 1970 British Cohort Study (BCS) is a birth cohort study following
approximately 18,377 people born in England, Scotland and Wales in
one week in April 1970 (TNS BMRB, 2013) – the target population is
adjusted at each wave due to changes attributable to migration and
mortality. Since then there have been eleven waves of data collection,
with data collected from the cohort members themselves, as well as their
parents, midwives and schools at appropriate stages of their development.
The study initially focused on childbirth and health but expanded
to include a broader range of topics including political values and voting
behaviour in later waves. This makes the dataset incredibly rich for
longitudinal analysis, as we can observe individual change over time,
with data prior to a life event not reliant on respondent recall.
The dataset also has limitations, primarily due to its nature as a
longitudinal study. The BCS suffers from attrition over time, partly from
natural attrition such as mortality and emigration, and partly due to
non-response (Plewis et al., 2004). For example, of an expected sample
of 18,377, 17,196 cohort members participated in the birth survey, 11,
615 participated at the age of 16 and 9841 at 42. As can be seen in Fig. 1,
due to additional effort made by the survey team on particular waves,
the attrition is not monotonic. Survey elements within each wave are
also prone to missingness, as not all respondents provide a valid
response to each question, an issue which particularly affects the attitudes
questions at 16 (as these were included on a self-completion survey).
More information on the item-level missingness is presented in
appendix 2.
Missing data is addressed in this study firstly by performing the
analysis on a complete case dataset across the sweeps included,
comprising the respondents who provided a response to all of the variables
included in the analysis from birth to the age of 42, which reduces
the total sample to 1520. In addition, to address potential bias attributable
to systematic missingness, analysis was also performed on 75
multiply imputed datasets (N = 15,874), with the imputation carried out
according to the procedure described for the British Cohort Studies by
Silverwood et al. (2020). The results using the imputed data are presented
alongside the main results, while more detail on the imputation
approach is included in appendix 2. This multiply imputed sample is
smaller than that of the original population and the sample who
responded to the birth survey due to natural attrition, with cohort
members who are known to have died or emigrated permanently not
imputed, following the recommendation of Silverwood et al. (2020).
There are further design issues in the composition of the sample
which affect interpretation: while efforts were made to incorporate new
migrants to Britain during their childhood (contacting them through
their school), it was not possible to do this since, meaning that the ethnic
diversity of the sample is not fully representative of the target population
(Elliott and Shepherd, 2006). Finally, a further restriction on generalising
from these findings are the historical and geographic specificity
of the sample, meaning that we can only be confident that the results
relate to the specific target population (ie this generation in Britain) in
the first instance.
5.1. Treatment variable
As this analysis is focused on the effect of graduating from university
on political values, a binary measure of holding a university degree or
higher qualification is coded as the treatment variable. In each adult
wave of the BCS, cohort members are asked their highest qualification,
whether academic or vocational. Those who reported having a degreelevel
or higher qualification (according to the ISCED-2011 classification
(UNESCO, n.d.)) at each age point are included in the treatment
condition, with all others placed in the control condition. As it refers to
graduate status, the treatment is considered irreversible, so that once
someone has reported holding a university degree they remain in the
treatment condition until the end of the panel. Table 1 presents how the
ISCED classification maps onto qualifications reported in the survey,
while Fig. 2 presents the proportions in the complete case sample at each
age point.
As we might expect, the proportion of those in the treatment group
grows over time, as cohort members gain qualifications over their lifetime.
This varying composition of the treatment and control groups has
implications for the empirical estimand. The treatment effect as estimated
is essentially a weighted average of the effects of these different
levels of education on the outcomes, with the effect of achieving no or
some school-leaver qualifications (and sub-degree vocational qualifications)
grouped together. The potential impact of this can be inferred
from Fig. 3, where a standalone OLS regression analysis enables comparison
of mean outcomes (with uncertainty) across ISCED levels at the
Fig. 1. Number of valid responses to the 1970 BCS at each wave.
Table 1
ISCED-2011 classification of British qualifications.
ISCED Example qualifications
1 No qualifications, left school at 12 or younger
2 Fewer than 5 good GCSEs, CSEs or O-levels; fewer than 5 good Ordinary or
Standard grades (in Scotland); NVQ level 1
3 5 or more good GCSEs, CSEs or O-levels, or any A/AS-levels; 5 or more good
Ordinary or Standard, or any Higher or Advanced Higher grades (in
Scotland); NVQ level 2 or 3
4 Higher National Certificate, NVQ level 4
5 Diploma of Higher Education, Higher National Diploma, NVQ level 5
6 Bachelor’s degree
7+ Master’s degree, postgraduate diploma or certificate (eg PGCE), PhD
R. Scott
Electoral Studies 77 (2022) 102471
6
age of 30, where those holding first degrees (level 6) are the reference
category. In addition, an alternative specification where the control
group excludes those without good school-leaver qualifications (ISCED
level 2) are presented as a robustness check in appendix 5.
Here and throughout, a higher value on the outcome indicates higher
levels of racial prejudice and authoritarianism, and more economically
right-wing values, and the error bars indicate the 95% confidence interval.
As can be seen, across the outcomes, the differences between
graduates and non-graduates are generally the most significant. This
difference tends to be strongest when compared with those achieving
levels 2 and 3, the most populous of the non-graduate groups, and less so
with those achieving level 5 qualifications. However, there are notable
differences between those achieving some good school-leaving qualifications
(level 3) and those who do not (level 2), which does mean the
effects of achieving a degree on both racial prejudice and economic
values are attenuated if those without these qualifications are excluded
from the control group (see appendix 5 for more).
In addition, postgraduates are significantly different in their values
from graduates, and although this is generally in the expected direction,
it is not for economic values, where postgraduates are more left-wing
and therefore closer to non-graduates. We should note that this may
mute the estimated effect of higher education on economic values – yet
as the intended estimand is the total effect of gaining a degree, including
pursuing postgraduate education, this group should remain in the
treatment condition. The decision to focus the analysis on this binary
difference in qualification status is not just informed by the underlying
data but also by the previously discussed research interest in the differences
between graduates and non-graduates, and with a view to
preserving statistical power by not disaggregating the treatment across
different levels of qualification. But it is important to bear in mind that it
may obscure patterns that exist underneath these larger groupings.
5.2. Pre-treatment covariates
To address the conditional parallel trends assumption, we must first
identify the pre-treatment covariates that might confound the relationship
between university attendance and adult political values. The
choice of variables is informed by the theoretical discussion of selection
effects in section 3.2. These, plus their age point of measurement, are
presented alongside the type of selection they represent in the model in
Table 2. Through their inclusion as fixed effects (ie time-invariant effects)
in the REWB modelling, the causally prior nature of these variables
will reduce any bias attributable to these. Note that ethnicity was
not reported prior to the age of 30, but is assumed to be pre-treatment as
a relatively static demographic variable.
5.3. Outcome variables
The three values scales of interest are measured at four age points
with varying items in the 1970 BCS. Standardised scales across time are
produced by calculating the arithmetic mean of the constituent survey
items, then scaling this value to have a mean of zero and a standard
deviation of one at each age point. As summarised in Table 3, the scales
over time demonstrate a good degree of internal reliability – with
Cronbach’s alphas above the conventional threshold of 0.7 (Cheng et al.,
2012; Cronbach, 1951). An alternative approach, producing scales
through inverse covariance weighting is presented as a robustness check
in appendix 5 of the supplementary material.
However, there are a few issues of note. First, there is no measure of
racial prejudice at the age of 26, so this timepoint is excluded from that
model. Second, the internal reliability of the authoritarianism measure
is weak at the age 42 – for this reason a separate analysis is performed on
the consistently included death penalty item as a robustness check,
demonstrating similar relationships and included in appendix 5. Finally,
while there is some consistency over time in the items, there are divergences,
particularly in the items measuring Left-Right economic
values in the age 16 and 26 surveys. The internal reliability of these
scales over time provides some reassurance in this regard. Nonetheless,
measurement error attributable to these differences is addressed primarily
by the scaling at each age point – meaning that any changes over
time reflect differences in relative position, rather than absolute value.
As the estimand of interest is the difference in the within-individual
effect between graduates and non-graduates and not the levels themselves,
and the analysis is performed on a balanced panel (making the
relative changes meaningful), these measurement differences are not
considered a threat to inference.
In addition, the fixed effects for age (and time, as the two are colinear
Fig. 2. Proportion achieving each ISCED-2011 qualification level by age.
Fig. 3. Political values by ISCED level of qualification aged 30.
N = 1520.
Table 2
Pre-treatment covariates included as fixed effects.
Category Covariates Age measured
Demographic selection Gender Birth
Ethnicity 30
Parental selection Parental social class Birth
Parental education 5
Mother’s authoritarianism 5
Ability selection Maths ability 10
Reading ability 10
R. Scott
Electoral Studies 77 (2022) 102471
7
in cohort data) will capture variance attributable to differing approaches
to measurement over time, which also means they cannot be simply
interpreted as age or period effects. Finally, the estimation of random
intercepts and slopes in the REWB estimation allows individual-level
variance in the starting point and effect of treatment to be modelled,
meaning that variance between individuals attributable to these differences
in measurement are accounted for in the standard errors and
therefore the significance testing. Further details on the constituent
items and analysis of the measurement properties of these scales
including Confirmatory Factor Analysis are available in appendix 6.
A further potential source of bias in measuring these outcomes is
social desirability: as respondents could consider the items sensitive, in
particular those addressing racial prejudice, and adjust their responses
accordingly. This could be a particular threat to validity if, in line with
previous research on some political measures such as voter turnout
(Bernstein et al., 2001; Karp and Brockington, 2005), those with higher
levels of education are more likely to give the perceived socially desirable
response, as this would make it appear as though the effect of
university is larger than it actually is. In contrast, if the more educated
feel less need to obscure their true feelings, as has been found with some
measures of racial prejudice in a political context (Heerwig and McCabe,
2009), then this could mute the apparent effect of achieving a degree.
To address this, the fact that these items are consistently measured
within the BCS data through self-completion surveys provides some
reassurance, as research into survey modes has found that these
methods, where the respondent is not prompted by a researcher, tend to
elicit a ‘truer’ response (Presser and Stinson, 1998). However, it is also
worth noting that meta-analytic research seeking to quantify the effect
of social desirability bias on reported racial prejudice is sceptical about
its influence (Blair et al., 2020): suggesting that those in the ethnic
majority who hold prejudiced views tend to be comfortable expressing
these in anonymous surveys. Indeed, some argue against considering
social desirability effects as measurement error at all, advocating instead
for “dual process models of prejudice [which] incorporate motivation to
conform to antiprejudice norms directly into our understanding of what
constitutes prejudice and what drives prejudiced behaviour” (Blinder
et al., 2019). Given the challenges in addressing possible social desirability
bias in secondary data, the outcomes, and racial prejudice in
particular, should therefore be considered as representing this dual
process model.
Having discussed the measurement properties of the outcome scales,
we can now observe the relationships between our variables prior to
accounting for selection. The mean of each political value by treatment
condition over time is presented in Fig. 4. As can be seen, based solely on
the mean differences, graduates are starkly less authoritarian and
racially prejudiced, although they appear to become more so and
therefore closer to non-graduates with age. The picture for economic
values is more complex, with graduates appearing more left-wing at 26
(although not significantly so compared with their starting position),
and then more right-wing at 30. It is possible this is a real effect,
reflecting a lag in the graduate premium to earnings, or an artefact of
measurement as discussed above. Yet in any case, these differences are
not estimating the change within individuals and therefore accounting
for time-invariant confounding, as we will now go on to do to provide a
more robust estimate of the effect of higher education.
6. Results
As set out in the methodology, the theoretical estimand of interest for
this study is the effect of graduating from university on racial prejudice,
authoritarianism and Left-Right economic values. This is estimated
empirically as the average difference in values before and after
achieving a degree within individuals who do so during the panel – and
so can be interpreted as the effect of university on values less any timeinvariant
confounding. Three estimators are employed to give confidence
in the results: the standard TWFE estimator; a REWB estimator
which accounts for observed pre-treatment confounding and includes
random effects parameters to allow for heterogeneous treatment effects;
and finally, a REWB estimator applied to 75 multiply imputed datasets
to account for non-random attrition in the cohort study data.
Fig. 5 presents these estimates of the effect of university for each of
the values under study. As the outcomes have been standardised and the
treatment is binary, the x axis can be interpreted as an effect size
(measured in standard deviations), while the standard errors used to
estimate the 95 per cent confidence intervals represented by the error
bars take account of the clustering within individuals. In addition, the
multiple imputation estimates have been pooled according to Rubin’s
rules (Rubin, 1987). The full results, including the random effects parameters
for the REWB estimates, are presented in appendix 1, and more
detail on the approach to the multiple imputation in appendix 2.
To take the hypotheses in turn and start with racial prejudice: the
results indicate that achieving a degree has a moderate and statistically
Table 3
Outcome scales example items and across-time reliability.
Outcome Example statement Std.
Alpha
Racial prejudice (higher
score ¼ more prejudiced)
- I would not mind if my child went to
a school where half the children were
of another race
0.81
Authoritarianism (higher
score ¼ more
authoritarian)
- For some crimes the death penalty is
the most appropriate sentence
0.82
Economic Left-Right (higher
score ¼ more right-wing)
- Government should redistribute
income from the better off to those
who are less well off
0.79
Fig. 4. Mean political values for graduates and non-graduates by age.
Fig. 5. Estimated treatment effects of university attendance on political values.
N (TWFE and REWB) = 1520, N (MI REWB) = 15874, M = 75.
R. Scott
Electoral Studies 77 (2022) 102471
8
significant negative relationship with this outcome, meaning that
graduates become less prejudiced on average, accounting for the effects
of selection. The results are remarkably consistent across the three estimators,
suggesting that graduating from university results in a small
reduction in prejudice, of around 0.15 of a SD. We can therefore accept
hypothesis 1. Moving on to review the results for authoritarianism, we
can see that higher education demonstrates its strongest effect here. The
results show a significant reduction of around 0.3 of a SD among those
who achieve a degree. This effect is corroborated with greater precision
by the multiple imputation estimates. This provides strong support for
hypothesis 2.
Finally, the results indicate that achieving a degree renders an individual
significantly more economically right-wing, although the estimators
differ in the exact magnitude of the effect. Those using the
complete case sample estimate a similarly-sized effect to that on racial
prejudice, of just over 0.15 SD – yet the REWB model applied to the
multiply imputed data suggests a much larger effect, of around 0.35 SD.
This larger effect may be attributable to differential attrition affecting
those in the control condition, where cohort members from more
disadvantaged backgrounds (and who therefore would be less likely to
achieve a degree and more likely to hold economically left-wing views)
were less likely to continue with the study. This result provides evidence
in support of hypothesis 3, and is in line with previous literature, where
authors have attributed this change both to peer socialisation and to
economic allocation effects.
Additional perspective can be provided on these findings by plotting
the estimated marginal means from the TWFE modelling over time to
provide a sense of the modelled linear trajectories, as in Fig. 6. This
shows the strength of the effect of higher education on authoritarianism,
and the weaker but still significant effect on the other two outcomes.
Given that these are predicted based solely on change in individuals over
time and therefore adjusting for time-invariant confounding, these results
provide strong evidence for the effect achieving a degree has on
political values, and authoritarianism in particular. However, it also
reveals something interesting about the nature of the effect, by suggesting
that the effect of university is quick to materialise but then
stable, with political values relatively unchanging over the rest of the life
course after early adulthood. This lends strength to the idea that higher
education plays an important role in political socialisation and values
formation.
7. Discussion
This study has presented new analysis of the causal effect of higher
education on political values, finding that individuals become less
authoritarian, less racially prejudiced and more economically rightwing
due to achieving a degree. It makes three main contributions. Its
primary contribution is to estimate the causal effect of university specifically.
Much of the previous causal literature has focused on additional
years of secondary schooling as the treatment, understandable
given the potential for natural experiments linked to policy changes. But
this new insight on the effect of university specifically is valuable in
understanding the increasingly divergent political behaviour of graduates,
not least given their growing numbers in much of the world.
Secondly, it provides the first causal estimate of the effect of education
on an established measure of authoritarianism. Despite the wellevidenced
cross-sectional relationship between education and authoritarianism,
few studies have employed a research design enabling causal
identification of these effects. Those that have employed such a design
(eg Bro´ci´c and Miles, 2021) have looked at adjacent values such as moral
relativism and freedom of speech rather than a validated multi-item
scale. This matters given the growing significance of this ‘second’
dimension for contemporary politics. This study addresses that gap,
providing good evidence that graduates become less authoritarian as a
result of university.
Third and finally, the data used for this analysis extends the previous
causal literature in this area by providing an estimate of the effect of
higher education among a more nationally representative sample,
therefore allowing us to be more confident in drawing general conclusions.
As noted above, previous research (Cavaille and Marshall, 2019;
d‫׳‬Hombres and Nunziata, 2016) making use of policy changes in the
compulsory participation age to address the selection problem through
instrumental variable and regression discontinuity designs are
well-identified, but tend to find variable and quite small effects of the
policy change itself, suggesting that the education effects they estimate
are specific to a less representative subpopulation who are prone to
educational disengagement. This study estimates the effect of graduating
from university across the whole cohort, irrespective of background
or the nature of the higher education experience itself.
There are, however, a few caveats to the interpretation. One is the
potential for time-variant confounding to be influencing these results,
particularly as the post-treatment outcomes are measured at the age of
26, which may be some years after graduation for some respondents. It’s
worth noting again that these estimates are the total effect of university
of political values, including any downstream effects such as differences
in socio-economic position, peer socialisation, geographic mobility and
family formation. This should be borne in mind in the interpretation,
particularly when considering the effects on economic values, which are
likely to be sensitive to changes in occupation and income.
Another is the specificity of the sample. This is British data, justifiably
chosen for the quality of the cohort study, but the findings cannot
be assumed to be true of other populations where education systems and
political values differ. Also, by its nature this analysis is of just one
cohort, and so the findings cannot be confidently generalised without
being replicated in other generations, as both the nature of education
and political values change over time. Therefore, similar analysis with
alternative high-quality longitudinal data will enable greater generalisation
of these effects, as well as understanding of how they vary across
settings and generations.
Yet having estimated the effect of higher education on political
values, a question that remains for future study is why education is
demonstrating this effect: to unpick the mechanism. Based on the values
trajectories of graduates, the effect appears to be more likely one of
socialisation: of rapid change during university and then stabilisation
during later life. But from the existing literature it is not clear the extent
to which this is attributable to a process of peer socialisation, identity
formation, changes in socio-economic position, greater cognitive sophistication
or some combination of all of these. Yet with the significance
of higher education in politics only growing, addressing this
question is vital in understanding how it will affect political values and
behaviour in the future.
Indeed, these findings indicate that the more polarised politics
observed in Britain since the EU referendum may be partly attributable
to the values differences from differential higher education participation
across the generations. Given the participation rate in higher education
continues to grow among under-30s in Britain (now being well in excess
of 50 per cent of that population (Department for Education, 2020)), this
Fig. 6. TWFE-estimated marginal means of political values for graduates and
non-graduates by age. N = 1520.
R. Scott
Electoral Studies 77 (2022) 102471
9
suggests that we should expect to see continued long-term aggregate
value change towards lower levels of authoritarianism and racial prejudice,
as the population becomes more highly educated on average.
However, during this transition, it is also likely that values divides will
continue to intensify in the short to medium-term, with significant
consequences for political behaviour.
Funding
Support for this research was provided by the Economic and Social
Research Council (grant no. ES/P000665/1).
Declaration of competing interest
The author declares none.
Data availability
Replication data and code is available at: https://doi.
org/10.17632/kpv6r63gmc.1
Acknowledgements
The author thanks Rob Ford, Ed Fieldhouse, and two anonymous
reviewers, for their very valuable comments on earlier versions of this
article. In addition, thanks are due to Jack Bailey, Stephanie Chen, Zac
Greene, Nicole Martin, Jon Mellon, Chris Prosser, Rosie Shorrocks and
Nick Turnbull, as well as colleagues at the 2021 Midwest Political Science
Association annual conference, the Essex PhD Colloquium and the
Manchester Democracy and Elections cluster for helpful feedback at
various stages of the research process.
Appendix A. Supplementary data
Supplementary data to this article can be found online at https://doi.
org/10.1016/j.electstud.2022.102471.
References
Angrist, J.D., Pischke, J.-S., 2008. Mostly Harmless Econometrics: an Empiricist’s
Companion. https://doi.org/10.1080/14697688.2015.1080490.
Ansolabehere, S., Rodden, J., Snyder, J.M., 2008. The strength of issues: using multiple
measures to gauge preference stability, ideological constraint, and issue voting. Am.
Polit. Sci. Rev. 102, 215–232. https://doi.org/10.1017/S0003055408080210.
Bates, D., Kliegl, R., Vasishth, S., Baayen, H., 2015a. Parsimonious Mixed Models.
https://doi.org/10.48550/arXiv.1506.04967 arXiv.
Bates, D., M¨achler, M., Bolker, B.M., Walker, S.C., 2015b. Fitting linear mixed-effects
models using lme4. J. Stat. Software 67. https://doi.org/10.18637/jss.v067.i01.
Bell, A., Fairbrother, M., Jones, K., 2019. Fixed and random effects models: making an
informed choice. Qual. Quantity 53, 1051–1074. https://doi.org/10.1007/s11135-
018-0802-x.
Bernstein, R., Chadha, A., Montjoy, R., 2001. Overreporting voting: why it happens and
why it matters. Publ. Opin. Q. 65, 22–44. https://doi.org/10.1086/320036.
Blair, G., Coppock, A., Moor, M., 2020. When to worry about sensitivity bias: a social
reference theory and evidence from 30 Years of list experiments. Am. Polit. Sci. Rev.
114, 1297–1315. https://doi.org/10.1017/S0003055420000374.
Blinder, S., Ford, R., Ivarsflaten, E., 2019. Discrimination, antiprejudice norms, and
public support for multicultural policies in Europe: the case of religious schools.
Comp. Polit. Stud. 52, 1232–1255. https://doi.org/10.1177/0010414019830728.
Bro´ci´c, M., Miles, A., 2021. College and the “ culture war ”: assessing higher education’s
influence on moral attitudes. Am. Soc. Rev. https://doi.org/10.1177/
00031224211041094.
Bullock, J.G., 2021. Education and attitudes toward redistribution in the United States.
Br. J. Polit. Sci. 51, 1230–1250. https://doi.org/10.1017/S0007123419000504.
Callaway, B., Sant’Anna, P.H.C., 2021. Difference-in-Differences with multiple time
periods. J. Econom. 225, 200–230. https://doi.org/10.1016/j.
jeconom.2020.12.001.
Campbell, C., Horowitz, J., 2016. Does college influence sociopolitical attitudes? Sociol.
Educ. 89, 40–58. https://doi.org/10.1177/0038040715617224.
Cavaille, C., Marshall, J., 2019. Education and anti-immigration attitudes: evidence from
compulsory schooling reforms across western Europe. Am. Polit. Sci. Rev. 113,
254–263. https://doi.org/10.1017/S0003055418000588.
Cheng, H., Bynner, J., Wiggins, R., Schoon, I., 2012. The measurement and evaluation of
social attitudes in two British cohort studies. Soc. Indicat. Res. 107, 351–371.
https://doi.org/10.1007/s11205-011-9852-3.
Converse, P.E., 2006. The nature of belief systems in mass publics (1964). Crit. Rev. 18,
1–74. https://doi.org/10.1080/08913810608443650.
Cronbach, L.J., 1951. Coefficient alpha and the internal structure of tests. Psychometrika
16, 297–334. https://doi.org/10.1007/BF02310555.
d‫׳‬Hombres, B., Nunziata, L., 2016. Wish You Were Here? Quasi-experimental evidence
on the effect of education on self-reported attitude toward immigrants. Eur. Econ.
Rev. 90, 201–224. https://doi.org/10.1016/j.euroecorev.2016.02.007.
de Chaisemartin, C., D’Haultfœuille, X., 2020. Two-way fixed effects estimators with
heterogeneous treatment effects. Am. Econ. Rev. 110, 2964–2996. https://doi.org/
10.1257/aer.20181169.
Dee, T.S., 2004. Are there civic returns to education? J. Publ. Econ. 88, 1697–1720.
https://doi.org/10.1016/j.jpubeco.2003.11.002.
Department for Education, 2020. In: Participation Measures in Higher Education:
Academic Year 2018/19 [WWW Document], 4.2.21. https://explore-education-stati
stics.service.gov.uk/find-statistics/participation-measures-in-higher-educatio
n/2018-19.
Dey, E.L., 1996. Undergraduate Political Attitudes: an examination of peer, faculty, and
social influences. Res. High. Educ. 37, 535–554. https://doi.org/10.1007/
BF01724937.
Drazanov´a, L., 2017. Education and Tolerance: A Comparative Quantitative Analysis of
the Educational Effect on Tolerance. Peter Lang, Bern, Switzerland.
Elliott, J., Shepherd, P., 2006. Cohort profile: 1970 British birth cohort (BCS70). Int. J.
Epidemiol. 35, 836–843. https://doi.org/10.1093/ije/dyl174.
Evans, G., Heath, A., Lalljee, M., 1996. Measuring left-right and libertarian-authoritarian
values in the British electorate. Br. J. Sociol. 47, 93–112.
Evans, G., Neundorf, A., 2018. Core political values and the long-term shaping of
partisanship. Br. J. Polit. Sci. 50, 1263–1281. https://doi.org/10.1017/
S0007123418000339.
Fieldhouse, E., Green, J., Evans, G., Mellon, J., Prosser, C., Schmitt, H., Van Der Eijk, C.,
2020. Electoral Shocks: the Volatile Voter in a Turbulent World. Oxford University
Press, Oxford.
Finseraas, H., Skorge, Ø.S., Strøm, M., 2018. Does education affect immigration
attitudes? Evidence from an education reform. Elect. Stud. 55, 131–135. https://doi.
org/10.1016/j.electstud.2018.06.009.
Fleishman, J.A., 1988. Attitude organization in the general public: evidence for a
bidimensional structure. Soc. Forces 67, 159–184. https://doi.org/10.1093/sf/
67.1.159.
Ford, R., Jennings, W., 2020. The changing cleavage politics of western Europe. Annu.
Rev. Polit. Sci. 23, 1–20.
Gaure, S., 2013. Lfe: linear group fixed effects. R J 5, 104–116. https://doi.org/
10.32614/rj-2013-031.
Gelepithis, M., Giani, M., 2022. Inclusion without solidarity: education, economic
security, and attitudes toward redistribution. Polit. Stud. 70, 45–61. https://doi.org/
10.1177/0032321720933082.
Gelepithis, M., Giani, M., 2021. In: Education and the Intergroup Discrimination
Paradox. SocArXiv. https://doi.org/10.31235/osf.io/2v96u.
Gelman, A., Hill, J., 2007. Data Analysis Using Regression and Multilevel/Hierarchical
Models, first ed. Cambridge University Press, Cambridge.
Hainmueller, J., 2012. Entropy balancing for causal effects: a multivariate reweighting
method to produce balanced samples in observational studies. Polit. Anal. 20, 25–46.
Hainmueller, J., Hopkins, D.J., 2014. Public attitudes toward immigration. Annu. Rev.
Polit. Sci. 17, 225–249. https://doi.org/10.1146/annurev-polisci-102512-194818.
Heath, A., Evans, G., Martin, J., 1994. The measurement of core beliefs and values: the
development of balanced socialist/laissez Faire and libertarian/authoritarian scales.
Br. J. Polit. Sci. 24, 115–132. https://doi.org/10.1017/S0007123400006815.
Heerwig, J.A., McCabe, B.J., 2009. Education and social desirability bias: the case of a
black presidential candidate. Soc. Sci. Q. 90, 674–686. https://doi.org/10.1111/
j.1540-6237.2009.00637.x.
Imai, K., Kim, I.S., 2020. On the use of two-way fixed effects regression models for causal
inference with panel data. Polit. Anal. 1–11 https://doi.org/10.1017/pan.2020.33.
Janmaat, J.G., 2018. Educational influences on young people’s support for fundamental
British values. Br. Educ. Res. J. 44, 251–273. https://doi.org/10.1002/berj.3327.
Karp, J.A., Brockington, D., 2005. Social desirability and response validity: a
comparative analysis of overreporting voter turnout in five countries. J. Polit. 67,
825–840. https://doi.org/10.1111/j.1468-2508.2005.00341.x.
Kiley, K., Vaisey, S., 2020. Measuring stability and change in personal culture using panel
data. Am. Soc. Rev. 85, 477–506. https://doi.org/10.1177/0003122420921538.
Kinder, D.R., Kam, C.D., 2010. Us against Them: Ethnocentric Foundations of American
Opinion. University of Chicago Press, Chicago.
Lancee, B., Sarrasin, O., 2015. Educated preferences or selection effects? A longitudinal
analysis of the impact of educational attainment on attitudes towards immigrants.
Eur. Socio Rev. 31, 490–501. https://doi.org/10.1093/esr/jcv008.
Lipset, S.M., 1959. Democracy and working-class Authoritarianism. Am. Soc. Rev. 24,
482–501. https://doi.org/10.2307/2089536?ref=search-gateway:
ad599b2b60cb317df7ccb86109578235.
Long, J.D., 2011. Longitudinal Data Analysis for the Behavioral Sciences Using R. SAGE
Publications, Thousand Oaks, California.
Lundberg, I., Johnson, R., Stewart, B.M., 2021. What is your estimand? Defining the
target quantity connects statistical evidence to theory. Am. Soc. Rev. 86, 532–565.
https://doi.org/10.1177/00031224211004187.
Marshall, J., 2016. Education and voting conservative: evidence from a major schooling
reform in Great Britain. J. Polit. 78, 382–395. https://doi.org/10.1086/683848.
R. Scott
Electoral Studies 77 (2022) 102471
10
McCourt, K., Bouchard, T.J., Lykken, D.T., Tellegen, A., Keyes, M., 1999.
Authoritarianism revisited: genetic and environmental influences examined in twins
reared apart and together. Pers. Indiv. Differ. 27, 985–1014. https://doi.org/
10.1016/S0191-8869(99)00048-3.
Mendelberg, T., McCabe, K.T., Thal, A., 2017. College socialization and the economic
views of affluent Americans. Am. J. Pol. Sci. 61, 606–623. https://doi.org/10.1111/
ajps.
Morgan, S.L., Winship, C., 2015. Counterfactuals and Causal Inference: Methods and
Principles for Social Science, second ed. Cambridge University Press, New York.
Mundlak, Y., 1978. On the pooling of time series and cross section data. Econometrica
46, 69–85.
Norris, P., Inglehart, R., 2019. In: Cultural Backlash: Trump, Brexit, and Authoritarian
Populism. Cambridge University Press, Cambridge. https://doi.org/10.1093/ia/
iiz097.
Paterson, L., 2009. Civic values and the subject matter of educational courses. Oxf. Rev.
Educ. 35, 81–98. https://doi.org/10.1080/03054980802351801.
Persson, M., 2015. Education and political participation. Br. J. Polit. Sci. 45, 689–703.
https://doi.org/10.1017/S0007123413000409.
Plewis, I., Calderwood, L., Hawkes, D., Nathan, G., 2004. National child development
study and 1970 British cohort study technical report: changes in the NCDS and
BCS70 populations and samples over time. Cent. Longitud. Stud. 1–39.
Presser, S., Stinson, L., 1998. Data collection mode and social desirability bias in selfreported
religious attendance. Am. Soc. Rev. 63, 137–145.
Rubin, D.B., 1987. Multiple Imputation for Non-response in Surveys. Wiley, New York.
Sears, D.O., Levy, S., 2013. Childhood and adult political development. In: Huddy, L.,
Sears, D.O., Levy, J.S. (Eds.), Oxford Handbook of Political Psychology. Oxford
University Press, New York. https://doi.org/10.1093/oxfordhb/
9780199760107.013.0003.
Silverwood, R., Narayanan, M., Dodgeon, B., Ploubidis, G., 2020. In: Handling Missing
Data in the National Child Development Study: User Guide [WWW Document],
7.5.21. https://cls.ucl.ac.uk/wp-content/uploads/2020/04/Handling-missing-data-
in-the-National-Child-Development-Study-User-Guide.pdf.
Simon, E., 2021. Explaining the educational divide in electoral behaviour: testing direct
and indirect effects from British elections and referendums 2016–2019. J. Elections,
Public Opin. Parties 1–21. https://doi.org/10.1080/17457289.2021.2013247.
Sobolewska, M., Ford, R., 2020. Brexitland: Identity, Diversity and the Reshaping of
British Politics. Cambridge University Press, Cambridge.
Sondheimer, R.M., Green, D.P., 2010. Using experiments to estimate the effects of
education on voter turnout. Am. J. Pol. Sci. 54, 174–189. https://doi.org/10.1111/
j.1540-5907.2009.00425.x.
Stoker, L., Jennings, M.K., 2008. Of time and the development of partisan polarization.
Am. J. Pol. Sci. 52, 619–635. https://doi.org/10.1111/j.1540-5907.2008.00333.x.
Surridge, P., 2016. Education and liberalism: pursuing the link. Oxf. Rev. Educ. 42,
146–164. https://doi.org/10.1080/03054985.2016.1151408.
TNS BMRB, 2013. Technical Report of the 1970 British Cohort Study: Age 42 Survey
(2012-2013) 1–42.
UNESCO, n.d. ISCED Mappings [WWW Document]. URL http://uis.unesco.org/en/isced
-mappings (accessed 3.5.22).
Weakliem, D.L., 2002. The effects of education on political opinions: an international
study. Int. J. Publ. Opin. Res. 14, 141–157. https://doi.org/10.1093/ijpor/14.2.141.
Wooldridge, J.M., 2021. Two-way fixed effects, the two-way Mundlak regression, and
difference-in-differences estimators. SSRN Electron. J. https://doi.org/10.2139/
ssrn.3906345.
R. Scott