Winners and losers of 'globalization':
is the effect of education on political
attitudes mediated by job situation?
A study in 27 European Union
countries
Bas Torenvliet 6041973 b.torenvliet@gmail.com Master Thesis University of AmsterdamResearch Master Social Sciences Empirical Analytical Track Supervisor: Dr. Theresa Kuhn
Second reader: Prof. Dr. Wouter van der Brug Date: 10 July 2015
Die Philosophen haben die Welt nur verschieden interpretiert; es kommt aber darauf an, sie zu verändern (Karl Marx, 1845)
Acknowledgements
I would like to thank Theresa Kuhn for her supervision during the writing of my thesis. Your critical reading of my earlier drafts has definitely improved this final version. I also want to thank Wouter van der Brug for being second reader. Annike, I want to thank you for our working together. In Amsterdam, but especially in Berlin, I had a wonderful time. Last, but not least, papa & mama, I want to thank you for all your help during these six year of studying. You and Carolien were there, especially in difficult times. Without you, I would not have even started the Research Master.
I hope that you, the reader, will enjoy reading this thesis and I hope we can discuss it afterwards.
Bas Torenvliet, July 2015
Contents
Title page 1 Motto 2 Acknowledgements 3 Abstract 5 Introduction 6Theory & hypotheses 8
Data & method 15
Results 22
Conclusion and outlook 40
Appendices 44
Abstract
Since the introduction by Kriesi et al. (2006) of the term winners and losers of globalization, many have accepted, challenged or changed the original thesis. However, the material base by which different political attitudes of those winners and losers could be formed is not studied often. In this thesis, it is argued that the lower educated are the losers of globalization, while the higher educated are the winners of globalization. Moreover, this study hypothesises that the effect of education is mediated by job situation, as the higher educated have a better job situation than the lower educated. The effect of education on attitudes towards different political issues, concept images and political concerns in 27 European Union countries is confirmed and is found to be very robust. Furthermore, the effect of education is significantly mediated by job situation for the issues regarding immigration and the European Union. Therefore, this thesis provides new evidence that the winners and losers of globalization are indeed formed by educational differences and economic situation, and that differences in attitudes are especially apparent for the new cleavage.
Keywords: education, political attitudes, job situation, globalization, winners and losers, European Union, ordinal logistic regression, mediation
Introduction
In 2006, Kriesi et al. observed a new cleavage between the winners and losers of
globalization in Western Europe.1 They argued that winners and losers have different
political attitudes and vote choices, because the consequences of globalization are rather different for those two groups (Kriesi et al., 2006). Furthermore, the losers feel threatened by other cultures and international and supranational institutions that undermine the national political culture. These latent potentials are articulated by (new) political parties, which changes the structure of the national political space.
Ever since, many have accepted, challenged or changed the original thesis (Van der Brug & Van Spanje, 2009; Bornschier, 2010a; Bovens & Wille, 2010; Oesch & Rennwald, 2010; Burgoon, 2013; Hakhverdian, et al., 2013; Teney et al., 2013; Kuhn, et al., 2014). All these authors focus on political participation, political attitudes and/or vote choice between the winners and losers of globalization, but the material base by which those different attitudes and vote choices are formed, are less well studied. Some researchers (Van der Brug, 2007; McLaren, 2002; Teney, et al., 2013) have even transformed the thesis into a more subjective gap between winners and losers of globalization, based on different identities.
As this thesis studies the latent political potentials only (and not the positions of political parties in the national political space), it is able to zoom in on the winners
and losers of globalization.2 More precisely, this thesis will focus on the different
economic risks and opportunities that winners and losers face and how this is related to political attitudes.
In the theoretical section, I will identify the winners and losers of globalization. It is argued that the higher educated could be regarded as the winners of globalization, while the lower educated could be regarded as the losers of globalization. The first hypothesis is that different education groups have different attitudes towards political issues and concept images. Moreover, they also have different political concerns. In particular, I will focus on issues, concept images and concerns regarding the two
1
However, the term was not completely new. Kapstein (2000) did already mention winners and losers of globalization. Betz speaks of the winners and losers in post-industrial society (1994: 29-32).
2
I have written a paper on this subject before for the course Politics of Globalization, lectured by Theresa Kuhn. This paper tried to model the effect of time, globalization and the world economic crisis on (national) unemployment levels for different educational groups in the European Union.
cleavages identified by Kriesi et al. (2008): the traditional economic cleavage and the new integration/demarcation or cultural cleavage.
Subsequently, I hypothesise that the lower educated have a worse job situation than the higher educated and that this will lead them to have different political attitudes and concerns. Lastly, the third hypothesis states that the effect of education on political attitudes is stronger for the new cleavage than for the traditional economic cleavage, as issues related to this new cleavage have gained much more salience in recent years.
In the data & methods section, I will introduce the Eurobarometer data and the methods used. In the results-section, I first test the effect of education on 11 political issues, 15 concepts images and 14 political concerns. If the effect of education is found, I test if the effect is significantly mediated by job situation using the KHB-method (Breen, et al., 2013). To test the third hypothesis, I use Mokken scale analysis to identify if we can regard the issues and concept images to belong to the traditional economic cleavage or the new cultural cleavage.
In the conclusion, I will discuss the results in relation to the hypotheses stated in the theory section. I also discuss the limitations of this study, and how this could be improved in future research. Additionally, I will reflect on how this study relates to other findings in political science research. It is very important to better understand the effects of education, material inequalities and their (causal) interrelation on political attitudes and behaviour, as they do not only reflect outcomes, but they are also related to social, economic and political opportunities. Therefore, this thesis should be seen as an intensification of Kriesi et al.'s original thesis. Moreover, it should be seen an update of it, as the book was published in 2008, just before the global economic crisis and the following European debt crisis set in.
Theory & hypotheses
In their article Globalization and the Transformation of the National Political Space, Kriesi et al. argue that there is a new cultural cleavage between the winners and losers of globalization (2006: 922). The winners are in favour of international integration and they support growing connections and interdependence between nations, cultures and economies. At the same time, the losers of globalization want to protect themselves through protectionist measures, because they feel threatened by these same global forces. The latent potentials among the electorate are effectively translated by old and new political parties, which leads to a complete transformation of the national political space (Kriesi et al., 2006). In their subsequent book, Kriesi et al. (2008) confirm their original thesis with additional empirical evidence from Austria, France, Germany, the Netherlands, Switzerland and the United Kingdom. In the book they give more attention to political attitudes and behaviour of voters. However, the aim of the book is to examine the relationship between voters’ attitudes and the positioning of political parties in the national political space and the structural
changes that are induced by this interaction (Kriesi, et al., 2008: 20).3
This thesis will only deal with the so-called demand side of politics, i.e. political attitudes and behaviour of voters. I fully agree with Kriesi et al. that issues only become a political struggle if parties will exploit a cleavage to become such a
struggle4: in this respect, the demand side of politics forms a necessary, but not
sufficient condition for the transformation of the political space (2008: 23-24). Until then, the attitudes of voters are latent political potentials. However, because I only deal with the demand side of politics, I do not need to fit the voters and parties on the same ideological dimensions (Van der Brug & Van Spanje, 2009: 310). In this way, I am able to zoom in on the winners and losers of globalization and their political attitudes to get a more complete understanding of the mechanisms at work.
Kriesi et al. argue that this new cleavage cuts across the two traditional cleavages that were in place after World War 2 (2008: 4): the traditional economic cleavage and the religious cleavage. According to Kriesi et al., the economic or class cleavage is quite different from the new cleavage between winners and losers of globalization (2008: 6). Logically, the class cleavage is a cleavage between classes,
3
See also Kitschelt (1994), Kitschelt & McGann (1995) and Bornschier (2010b) for the structural changes in the national political space.
4
while the new cleavage is not a struggle of interest between classes per se. Moreover, they argue in the book that the class divide is not that relevant anymore and that economic issues have lost salience (Kriesi et al., 2008: 264).
The other traditional cleavage was structured by conservatism on one side and liberal and open attitudes on the other. This cleavage was mainly structured by religious attitudes. They argue that the new winners/losers divide changed especially this cleavage. Religious issues have lost salience as a consequence of secularization, while new struggles about global cultures and immigrants have emerged (Kriesi et al., 2008: 258-261). The change of this traditional cleavage into the new integration/demarcation cleavage transformed the national political spaces of Western European countries.
Kriesi et al. argue that education is an important predictor if someone is a
winner of loser of globalization (2006: 922; 2008: 7).5 High educated people do
benefit from new opportunities of globalization, like the global job market, studying abroad and new destinations to spend holidays. However, low educated people are threatened by foreign labour and do not go abroad, because they lack foreign language skills and/or money. This leads them to have different attitudes towards
immigrants and international or supranational cooperation like the European Union.6
However, Kriesi et al. do not differentiate between those who actually are winners or losers of globalization (objectively) and those who feel a winner or loser (subjectively). They argue that material (objective) and cultural (subjective) factors are mutually reinforcing, and that these factors are not perceived differently by individuals (Kriesi et al., 2008: 8).
The effect of education on attitudes towards political issues is not new. Lipset argues that education leads to a better understanding of democratic values in general (such as tolerance towards the opposition). Moreover, it also leads to more open attitudes towards minority groups in particular (1960: 56). Almond & Verba (1963) identify education as the single most important predictor of having certain political attitudes. They argue that this effect is partly direct (through teaching as such), but
5
Another important predictor they identify is the economic sector that someone works in: the ones that benefit from an open economy and world trade are winners, while those that are threatened by it are considered as losers. I will come back to this.
6
See Kriesi et al., 2008: 260-264 for the empirical evidence of the effect of education on political attitudes.
also indirect: people develop different attitudes through different situations one finds him- or herself in (Almond & Verba, 1963: 379).
Bourdieu (1986) sees this social capital (acquainted through education) as a form of accumulated labour in a more Marxist sense. Just as wealth, social capital is unequally distributed and reproduced over different classes (Bourdieu, 1986: 252). In
Distinction, he argues that education is an important predictor to which class one
belongs (2010, org. 1984: 15).
Piketty argues in Capital that the richest 1% is indeed high educated, but not higher educated than many others. Therefore, he concludes that education is not the right factor to focus on (2013: 315). However, he tries to construct a new class divide between capital and labour. In Europe, the effect of education on income and wealth
is very clear (Website Eurostat 1).7 Moreover, Piketty argues that the (material)
interests of the dominant class (the richest 1% in the world) are better taken care of,
which is not my purpose here.8
In recent studies, we find that (in the Western world) education is a good predictor of political attitudes, as well in sociology (Pascarella & Terenzini, 1991; Nie, et al., 1996; Pallas, 2000; Weakliem, 2002; Klamijn & Kraaykamp 2007) as in political science (Kitschelt, 1994; Kriesi et al., 2008; Stubager, 2008, 2010; Kuhn, et
al., 2014; Lancee & Sarrasin, 2015).9 In this thesis, I will test the effect of education
on 11 different political issues, 15 so-called concept images and 14 political concerns. Therefore, the first hypothesis is:
Hypothesis 1: The lower educated have different political attitudes than the higher educated. They also have different political concerns.
However, how does education lead to these different attitudes and concerns? Many have addressed this causal puzzle before (Hyman & Wright, 1979; Niemi & Junn, 1998; Pallas, 2000; Ichilov, 2003; Kalmijn & Kraaykamp, 2007; Stubager, 2008). Kriesi et al. do not address this issue, as they only want to identify the winners and
7
Besides, higher educated Europeans are less at risk of poverty (European Commission, 2013).
8
This thesis is about the origins of political attitudes and not about political participation, political representation or political power of different education groups. Of course, these questions are no less relevant (see also the conclusion & outlook of this thesis).
9
Some even argue that this ‘education gap’ has increased over time (Van der Werfhorst & De Graaf, 2004; Bovens & Wille, 2010; Hakhverdian, et al., 2013).
losers of globalization (2008: 8). To be more precise, they want to argue which citizens perceive themselves as being winners or losers, by having different attitudes (especially on the ‘new’ cleavage).
It is very important to identify the precise mechanisms through which education leads to different attitudes, as a diploma as such will not cause you to have different political attitudes. I will differentiate between the more objective or material aspects of globalization (‘being threatened by globalization’) and the more subjective
or cultural aspects of globalization (‘feeling threatened by globalization’).10
Globalization is not a causal actor that threatens individuals; globalization should rather be seen as a process (Held, et al., 1999; Dreher, 2006). However, in this process of opening up borders and the expansion of global capitalism, it is predicted by international trade theory that skilled labour benefits from increased international
trade, while unskilled labour loses (Rogowski, 2000; Berger, 2000; Margalit, 2012).11
Rodrik (1997) argues that in the global economy, low skilled workers face higher material risks than high skilled workers. Additionally, high skilled workers are more mobile to benefit (materially) from globalization (Baumann, 1998). This means that they have resources to move across countries, between jobs and on the social ladder.
These skills are mainly obtained through formal education (Rosenau, 2003; Carbonaro, 2007). Thus, the level education does have severe consequences for the risks and opportunities that one faces on the job market (see also Blossfeld et al., 2011). Some will interpret this education gap as one of the mechanisms by which social classes are formed (Wright, 1997; Oesch & Rennwald, 2010) and/or reproduced (Bourdieu, 1986; Erikson & Goldthorpe, 2002). Buchholz et al. conclude therefore that "present-day societies hence still can be characterised as class societies” (2009: 67). Others explicitly distinguish between the effect of education and class (Van der Werfhorst & De Graaf, 2004; Kalmijn & Kraaykamp, 2007).
10
It remains difficult to argue that the first is fully objective while the second is fully subjective, because they are (of course) interrelated. I use the distinction similar to the economic competition mechanism and cultural competition mechanism identified by Kriesi et al. (2008: 5-7). I will not discuss the third mechanism they mention: political competition, as this is more about democratic government as such (e.g. political trust). An important difference between Kriesi et al and this thesis is that they do not test for these mechanisms.
11
These authors rearranged the Stolper Samuelson-theorem by replacing high skilled labour for capital and low skilled labour for land. The original theorem predicts that in a world of capital and land (which are not mobile between countries), if the relative (world) price of capital goes up, the capital owners win at the expense of the land-owners (Stolper & Samuelson, 1941).
It is beyond the scope of this thesis to solve the issue if education groups could be regarded as (new) classes or that education groups should be viewed as something
different.12 However, as noted before, Kriesi et al. argue that the new winner/loser
divide cuts through the traditional class cleavage, because the winners are those who work in market-oriented (open) sectors and the losers are those who work in sheltered sectors. They rely mainly on Schwartz (2001) here. Schwartz’ main argument is about firms and nations as a whole (and not about individuals), although he observes that the amount of unskilled labour in sheltered sectors is relatively large (2001: 42).
Therefore, we can conclude that they are the same individuals: low educated people that face high risks and have few opportunities on the job market (as a consequence of the expansion of global capitalism). So, we can consider them the losers. The higher educated are more secure, and could therefore be regarded as the winners. Furthermore, Kitschelt argues that higher educated have more control and autonomy over their work situation (1994: 17). They translate these experiences to the political realm, where they tend to have more liberal views (see also Kohn & Schooler, 1969; Stubager, 2008).
The mediation effect of education and economic situation on political attitudes is not tested explicitly for this wide range of attitudes before. Some test the mediation effect for attitudes towards welfare state policies and/or preferences for social democratic parties (Rueda, 2005; Rehm, 2009; Rehm, et al., 2012; Margalit, 2013;). Others do focus on the losers of globalization (Sniderman, et al., 2004) or radical right wing voting (Kitschelt & McGann, 1995; Swank & Betz, 2003; Rydgren, 2007; Bornschier & Kriesi, 2013). Hooghe & Marks (2004) tested the effect of both identity and economic situation on attitudes towards European integration and conclude that both are important.
Some do not find an effect of economic situation on political attitudes. However, Stubager (2008) used other variables, i.e. social class, income and employment sector to measure the effect of economic situation on attitudes towards immigrants. Ivarsflaten & Stubager (2013) used economic strain as a mediating variable. So, they were only able to measure if people face economic problems and do not leave open the possibility of giving a positive answer (i.e. being able to differentiate between winners and losers more accurately). Hainmueller & Hiscox
12
However, to control for occupation, I include it as an independent variable (see data & methods).
(2006) also claim that there is no mediation effect of education and economic situation on attitudes towards international trade, but they only test this mediation effect for employed versus not employed citizens.
This thesis will test the mediation effects of education and job situation for 11 political issues, 15 concept images and 14 political concerns. It is proposed that the lower educated face a worse job situation and have therefore other attitudes than the higher educated. I do not expect the whole education effect to be mediated, as there
are also other mechanisms in play, like e.g. differences in identity and socialization. 13
Hypothesis 2 states the following (see also figure 1):
Hypothesis 2: These different political attitudes are partly explained by different job situations that lower and higher educated face. So, the effect of education on political attitudes and concerns is mediated by job situation (figure 1).
Figure 1 – The effect of education mediates partly through job situation.
I will test the effects of education and job situation for a wide range of attitudes. Therefore, I am able to distinguish between attitudes that refer to the traditional economic cleavage (such as redistribution and attitudes towards the welfare state) and to more cultural issues, including attitudes towards the European Union.
The growth of support for anti-immigrant parties and recent debates about immigration, integration and the European Union have emphasized this new (or
13
Thus, I will not differ between the direct and indirect effect of education (Almond & Verba, 1963; Stubager, 2008), but I will try to establish a causal path of how education leads to different attitudes.
transformed) cleavage in Western Europe (Bornschier & Kriesi, 2013). Bornschier argued in 2007 that, although the new winners/losers divide finds its origin in the employment structure, it is especially relevant for the new cleavage. Van der Werfhorst & De Graaf also find a larger effect of education on cultural issues (2004: 228).
However, in the light of the recent economic (world) crisis and subsequent European debt crisis, we could expect that the traditional class cleavage has regained salience. As Hofstadter (2002) argues, during recessions, economic issues will prevail over more cultural issues. Moreover, Bornschier concludes in 2010(b) that the winners and losers also have different attitudes towards socio-economic issues.
If the first (and second) hypothesis is confirmed, I can identify for which issues the effect of education (and job situation) is stronger. Because this thesis has started with the argument of Kriesi et al. (2008), who highlighted the new cultural cleavage, I have stated the following hypothesis:
Hypothesis 3: The (mediated) effect of education (and job situation) will be stronger for cultural issues, because the new cultural cleavage (as outlined by Kriesi et al., 2008) is more prominent nowadays than the traditional economic cleavage.
Data & method
To test the hypotheses, I use Eurobarometer 82.3 of November 2014 (European
Commission, 2014).14 This is the most recent Eurobarometer data at the time of
analysis. Eurobarometer is the official survey questionnaire conducted by the European Commission (2014) twice a year since 1973. This study includes
respondents from 27 European Union countries.15 For every country, around 1000
respondents were interviewed.16 The total number of respondent used for the analyses
is 26,874. I did not apply weights, because it is uncommon to do so in this kind of research.
The dependent variables
The dependent variables are 11 political issues, 15 concepts images and 14 political concerns. In appendix 1, you find an overview of the political issues, for which people had to say if they agree with the statement or not or that they are positive about it or not. All issues are coded the same way (low codes refer to being negative/do not agree and higher numbers refer to being positive/do agree). The issues represent attitudes towards economic issues that refer to the traditional class cleavage and attitudes that refer to the new cultural cleavage. It is true that many of them are related to the EU, which are, according to the theory, related to the new cleavage (I will come
back to this in the results section).17
Appendix 2 shows the same for 15 concept images, for which respondents had to say if the term brings up something positive or negative. These are coded the same way (1 is very negative). Appendix 3 shows an overview of the concerns that Europeans have (1 indicates that a respondent did mark this subject as being an important concern that the country faces and 0 means that a respondent did not mark
14
All statistical analyses are done in Stata 13. Using the do-file, one should be able to replicate the results. The do-file is available upon request.
15
The countries included are: Belgium, Bulgaria, Czech Republic, Denmark, Germany, Estonia, Ireland, Greece, Spain, France, Hungary, Italy, Cyprus, Latvia, Lithuania, Luxembourg, Malta, the Netherlands, Austria, Poland, Portugal, Romania, Slovenia, Slovakia, Finland, Sweden and the United Kingdom. Former East and West Germany are coded as Germany, and Northern Ireland is coded as United Kingdom. Croatia is a Member State since 2013, but I decided not to include the Former Yugoslavia countries. Also, Iceland, Turkey and the Turkish Cypriot Community are excluded from analysis.
16
Highest number of respondents is Germany (1610) and the lowest number of respondents is Cyprus (500).
17
this concern as being important). Some issues and concept images have a relatively high number of missing values, of which I come back in the results section.
I did not use the left-right scale as a dependent variable, because it is very difficult to assess how people interpret the left-right scale, especially with respect to the two dimensions as outlined by Kriesi et al. (2008): do they use attitudes towards immigrants, attitudes towards the welfare state and redistribution or a combination of both to identify themselves or parties as being left or right (see De Vries et al., 2013 for the dynamics involved).
The independent variables
The most important independent variable is education level. Education level was measured by asking respondents how old they were when they stopped full-time
education. Eurobarometer already recoded the variable into five categories.18 I
recoded this variable into three categories: low education (respondents that do not have any formal education and those who stopped at the age of 15 or younger); middle education (respondents that left school at 16-19); and high education (those
who left school at 20 or later), as is common in this kind of research.19 4,618
respondents are low educated, 11,224 respondents are middle educated and 10,616
respondents are high educated.20 I will use dummy variables to estimate the effect of
having a middle education or a high education, whereas low education is the reference category.
Some researchers argue that the new cleavage cuts between high educated and the rest, i.e. low and middle educated citizens (Stubager, 2008; Bovens & Wille, 2010). Therefore, we will not be surprised if only the effect of having a high education is found.
To measure the respondent’s job situation, I use the survey question: “how
would you judge your current personal job situation”? (see appendix 4 for descriptive
statistics). One could argue that this is rather a subjective measure of economic situation. However, at least, it shows which people feel insecure and secure in their
18
In this way, outliers (e.g. someone who stops full-time education at age 70) cannot affect the estimation of the effect (this person will be in category high education, just as someone who left school at 22).
19
As all respondents younger than 18 were excluded from analysis, I assume that all students of 18 years will study for at least another two years (i.e. obtain a high education level).
20
current job (which is something different than feeling threatened by globalization culturally). Furthermore, using this question, I am able to distinguish between people that have a bad and a good job situation on a 4-point scale, so we do not only distinguish between people having problems or not (like economic strain or unemployment). Using unemployment as the mediating variable, we can only compare those that are already unemployed and those who are not. Using job situation,
we are able to distinguish between those at risk and those who feel secure.21
Therefore, job situation is in my view the most appropriate measure to use.
Control variables
However, to control for unemployment, I use a dummy-variable (1=employed, 0=unemployed). Additionally, I control for occupation using dummy variables for being self-employed, a manager, a white collar worker, a manual worker, a house person (without a paid job), being retired or unable to work and being a student, whereas being unemployed is the reference category. This variable is already a recode available in the Eurobarometer dataset (see appendix 5 for which occupations belong to which category). As said, occupation is not included to solve issues related to class, but rather to test the robustness of the education-effect and the effect of job situation.
Besides, I control for age and gender. Age is considered to have an effect on political attitudes and it is common to include age in these types of analyses (Down & Wilson, 2013). The same goes for gender: women have different attitudes than men (Harteveld, et al., 2015). However, some even argue that young people and women are considered to be economic vulnerable groups too, and that this leads to different attitudes towards the welfare state (Häusermann, et al., 2014). Therefore, it is important to control for these personal characteristics. Gender was coded as a dummy (1=men, 0=women). Age was measured on a continuous scale. All respondents younger than 18 are excluded from analysis.
Another possible control variable is ethnicity or having a migrant background, as immigrants have other political attitudes (and face higher economic risks) than
21
Using risk of being unemployed per education group cannot be used either, because it would result in perfect multicollinearity between education level and those unemployment rates.
native citizens.22 Unfortunately, I was not able to control for ethnicity, as it is not included in Eurobarometer. Moreover, the only variable that measures nationality or something related to that gives only 105 respondents who answered that they come from another country outside the EU. Moreover, in the literature referred to in this
thesis, ethnicity is not a control variable either.23
Method, assumptions and model fit
Because all issues and concept images are measured on an ordinal scale, I use ordinal logistic regression to estimate the effect of education, job situation and the control variables. For the political concerns, I use binary logistic regression, because respondents did only raise a concern or not.
The ordinal scales of the issues and concept images mean that we cannot assume that the distances between the categories are equal (Treiman, 2009: chapter 14). Moreover, the number of categories is limited for these scales (between 3 and 5), so an Ordinary Least Squares (OLS) regression would lead to incorrect standard errors and incorrect hypotheses testing (because of heteroscedasticity and non-normally distributed errors).
Therefore, the log-odds are estimated24, just as for a binary logistic regression.
However, the log-odds now refer to being in the same or a higher category versus being in a lower category (Treiman, 2009: chapter 14). Ordinal logistic regression makes use of cut-off points which indicate thresholds that need to be crossed to get to the next category (in the tables shown, I refer to them as T1, T2, and so on.). In fact, ordinary logistic regression is a set of (number of categories-1) binary logistic regressions.
Consequently, ordinal logistic regression has an important assumption: the proportional odds assumption, also called the parallel regression assumption (Brant, 1990). This assumption states that the proportional odds are the same, i.e. for every cut-off point (or binary logistic regression), the effect is exactly the same (so we can
22
For political attitudes see Maxwell (2010) and Reeskens & Van Oorschot (2015). For economic risk see Büchel & Frick (2005).
23
Which is remarkable, as attitudes towards immigrants and immigration are studied often (including this thesis). Mewes & Mau (2013) and Reeskens & Van Oorschot (2015) are an exception to this rule.
24
I choose to present the log-odds, because the direction and size of the effects can be interpreted very intuitively. However, I also present some odds ratios and predicted probabilities (see results).
interpret the effect for all categories simultaneously). Using the Brant-test (1990), we are able to test this assumption for the whole model and for the effects separately. If the Brant test gives a significant result, the assumption is not met. If this is the case, we should use a generalized ordered logistic model, which relaxes the assumption of the proportional odds. This model estimates the effect for each cut-off point separately.
I do not make use of a multilevel (random effects model) framework, because we cannot assume that people from different countries are randomly selected from one (European) population. Moreover, we cannot assume that unobserved country-level variables are distributed independently from the observed variables (i.e. zero conditional mean assumption of regression). Using country-dummy’s (measured as nationality) to control for a country’s specific context is therefore most appropriate
(Greene, 1990). Additionally, country robust standard errors are used.25
The absence of multicollinearity (i.e. a linear relationship between independent variables) is an important assumption of all types of regression analyses, because multicollinearity inflates standard errors and causes coefficients to be less efficient (Agresti & Finlay, 2009: 456). However, meditational effects imply a linear relationship between independent variables (in this case, between education and job situation, see below), so we can expect multicollinearity to exist. Agresti & Finlay argue that a Tolerance value of 0.1 or lower indicates multicollinearity can be a problem (2009: 456). I did estimate Tolerance values for all independent variables
used and no multicollinearity problems are present.26
For each issue, concept image and concern, I estimate 3 models. In the first model I estimate the effect of education, while controlling for age, gender and nationality. In the second, the effect of job situation (with controls age, gender and nationality) is estimated. In the last model, I estimate the effect of both, while controlling for age, gender and nationality. I re-estimated the models using employment and occupation as control variables, and also did some other alternative specifications, to examine if the effects of education and job situation are robust.
25
Unfortunately, Stata is not able to calculate the Wald 𝜒𝜒2 to test the significance of the model when using these country robust standard errors. Therefore, I use a likelihood-ratio test to test if the model is a significant improvement compared to a null-model (with an intercept only).
26
The lowest tolerance values are 0.255 (dummy for being retired) and 0.278 (dummy for having a rather good job situation). This means that 74.5% of the variance of the dummy for being retired is explained by the other independent variables.
Baron & Kenny (1986) argue that for a mediation effect to found, you first have to establish a significant effect of the independent variable on the mediating variable. Appendix 6 shows the effect of education on job situation. The effect is significant and quite big: when one has a higher education, one is expected to find
oneself in a better job situation.27
To test if job situation mediates the effect of education on political attitudes, I use the KHB-method. Because the variance structure of two logistic models is not the same when including new variables, we have to account for these rescaling effects
(Breen et al., 2013). The KHB method does exactly this.28 We now get a model with
the mediation variable (the full model) and a model without the mediating variable (reduced model). We can also estimate the difference between these two models and
assess if the difference (i.e. the mediation-effect) is significant. 29
To assess the model fit, I use (McFadden’s) Pseudo R2, which compares the
likelihood of the estimated model to a null-model without independent variables.30 I
also estimated the Adjusted Count R2, which indicates the percentage of reduction of
the error in the prediction. Kalmijn & Kraaykamp (1999) notified that attitudes research yields modest R-squares, so I do not expect them to be very large. To compare the models, Akaike’s Information Criterion (AIC) is used. If the AIC is lower, the model fit is better (Burnham & Anderson, 2002).
Mokken scale analysis
Beside the ordinal logistic regressions, I use Mokken scale analysis to examine if we can identify if attitudes towards these issues and concept images are formed along one or two latent dimension(s). A (latent) dimension could not be measured. However, it is the abstract configuration that structures people’s answers on different concrete
questions.31 Therefore, a causal relationship between dimensions and concrete
27
The effects of age and gender are also significant, but small. It is also clear that employed Europeans have a better job situation than unemployed Europeans.
28
See Lagana et al. (2014) for an application of the method.
29
I also tried to run a Sobel Test (Sobel, 1986), but this test does not work properly for ordinal (i.e. dummy) independent variables.
30
McFadden’s Pseudo R2 is the default Pseudo R2 estimated by Stata. The Pseudo R2 is not the explained variance, because we predict probabilities, so there is no residual as with OLS-regression. However, 1 indicates a perfect fit, and 0 indicates that your model has no predictive power.
31
For example, to identify if you are a winner or loser of globalization (the dimension), we can measure your education, job situation, and so on (the concrete questions).
answers is assumed (Jacoby, 1991). We can use these dimensions or scales to confirm or reject the third hypothesis, that the effect of education is stronger for cultural issues about immigration and the European Union.
The most easy procedure to create such a scale is Cronbach’s Alpha, which assesses if the variables correlate in such a way that they could be interpreted as being one scale. Cronbach’s Alpha above 0.7 indicates a good scale. However, there are some problems with Cronbach’s Alpha, as it does not provide a good measure for the internal structure of the test that the items are part of one dimension (Sijtsma, 2009; Van der Eijk & Rose, 2015).
Therefore, I will make use of Mokken scale analysis. Mokken scale analysis is helpful in assessing if different issues and concept images can be regarded as being formed by one latent dimension. The H-value of Mokken is better able than
Cronbach’s Alpha to differentiate if the scale is appropriate to use.32 Mokken scales
are less stringent than Rasch scales, and lead to the same results in almost all cases (Van Schuur, 2003: 40). For Mokken scale analysis, only binary items could be used (Sijtsma, et al., 1990). Therefore, I recoded all dependent variables into binary variables (0=negative, 1=positive). For the issues about working and living in every EU country and the general image of the EU, respondents could have given a neutral answer. This neutral answer is coded as being positive.
Once the scales are obtained, I will estimate the effects of education and job situation and controls on these scales.
32
An H-value above 0.3 indicates that the scale has a good fit (Sijtsma, et al., 1990). It also provides a significance-test.
Results
First, I will discuss the effects of education and job situation on 11 political issues. Then, I discuss the effects of education and job situation on 15 concept images. Subsequently, I use Mokken scale analysis to identify to which dimensions the issues and concept images are related. Finally, I discuss the effects of education and job situation on 14 political concerns.
The effects of education and job situation on 11 issues.
To be able to estimate the effect of education correctly and compare the models
properly, I estimated the effect of education using the KHB-method (see table 1).33
The effect of education in the reduced model (without job situation) is positively significant for attitudes towards immigration outside the EU (issue 1), inside the EU (issue 2), globalization is an opportunity for economic growth (issue 3), right to work and live in the EU is a good thing (issue 4 and 5), feeling an EU-citizen (issue 6), general image about the EU (issue 7) and the imposition of a tax on financial transactions (issue 11). For attitudes towards the issue if the EU creates more jobs (issue 8), only the effect of higher education is significant. The effect is also positive. For the statement that the private sector is a better place to create jobs (issue 9) and the statement that reducing public debt is no priority now (10), the effect of education is not significant. However, for the last (issue 10), there is a high number of missing values (54%), which indicates that many respondents did not know what to answer. This result is therefore not completely statistically valid, as missing values can lead to biased results.
The effect of education is largest for issue 6 (feeling an EU-citizen). For the middle educated, the log odds are 0.500 higher that one feels an EU-citizen, compared to low educated. For the higher educated, the log odds are 1.070 higher, compared to low educated. Alternatively, we could state that the odds to feel an EU-citizen increase by 1.644 for middle educated and by 2.867 for high educated with respect to
33
In appendix 7-17, you find the ordinal logistic regressions without using the KHB-method. The results do not differ much with the results in table 1. However, here you find the effect of age, gender and job situation. The AIC, Brant-test, Log-likelihood and Adjusted Count R2 are also given.
Tab le 1 - Th e ef fec t o f ed uc at io n o n 1 1 p ol iti cal is su es u sin g th e K H B-m et hod . Not es : J ob s itu at io n i s th e m ed iat in g v ar iab le . C on tr ol le d f or ag e, g en de r an d c ou nt ry o f liv in g. Es tim at es ar e l og o dd s, c ou nt ry rob us t s tan dar d e rr or s i n p ar en th ese s. P se ud o R 2 of fu ll m od el is s ho w n. * p < 0 .0 5, ** p < 0. 01, *** p < 0. 001 (t w o-tai le d te st ) Sou rc e: E ur ob ar om et er 82. 3 ( Eu ro pe an C om m iss io n, 2 014) Is su e 1 Is su e 2 Is su e 3 Is su e 4 Is su e 5 Is su e 6 Is su e 7 Is su e 8 Is su e 9 Is su e 10 Is su e 11 Im m ig ra tion ou ts id e E U Im m ig ra tion in si de E U G lo ba - liz at io n i s op por tu ni ty W or k i n th e E U Liv e in th e E U Feel in g EU -cit iz en Ge ne ral fe elin g EU EU cr eat es jo bs Pr iv ate se ct or be tt er De bt has n o pr io rity Tax o n fin an cia l tr an sa ct io ns M id dle ed uca tion Red uc ed 0. 165 * 0. 295 *** 0. 239 ** 0. 159 * 0. 194 ** 0. 500 *** 0. 307 *** 0. 148 0. 018 0. 155 0. 113 * m od el (0. 075) (0. 055 ) (0. 086 ) (0. 069 ) (0. 06 2) (0. 063 ) (0. 061 ) (0. 094 ) (0. 073 ) (0. 079 ) (0. 052 ) Fu ll 0. 124 0. 239 *** 0. 183 * 0. 110 0. 151 * 0. 417 *** 0. 220 *** 0. 081 -0. 009 0. 141 0. 106 * m od el (0. 075) (0. 054 ) (0. 085 ) (0. 067 ) (0. 061 ) (0. 063 ) (0. 059 ) (0. 093 ) (0. 075 ) (0. 080 ) (0. 052 ) Dif fe re nc e 0. 042 0. 056 0. 056 0. 049 0. 043 0. 083 0. 088 0. 067 0. 027 0. 014 0. 007 (0. 043) (0. 061 ) (0. 075 ) (0. 058 ) (0. 050 ) (0. 085 ) (0. 092 ) (0. 075 ) (0. 039 ) (0. 052 ) (0. 010 ) H igh educa tion Red uc ed 0. 593 *** 0. 853 *** 0. 450 *** 0. 633 *** 0. 668 *** 1. 070 *** 0. 751 *** 0. 405 *** 0. 125 0. 138 0. 172 *** m od el (0. 110) (0. 071 ) (0. 083 ) (0. 073 ) (0. 095 ) (0. 065 ) (0. 066 ) (0. 100) (0. 0778) (0. 090 ) (0. 050 ) Fu ll 0. 505 *** 0. 729 *** 0. 317 *** 0. 520 *** 0. 571 *** 0. 892 *** 0. 560 *** 0. 256 ** 0. 058 0. 100 0. 157 ** m od el (0. 110) (0. 071 ) (0. 081 ) (0. 073 ) (0. 097 ) (0. 065 ) (0. 063 ) (0. 098 ) (0. 079 ) (0. 090) (0. 050 ) Dif fe re nc e 0. 088 * 0. 123 * 0. 132 0. 113 0. 096 8 0. 178 * 0. 191 * 0. 149 * 0. 067 0. 038 0. 015 4 (0. 044) (0. 062 ) (0. 075 ) (0. 060 ) (0. 051 ) (0. 086 ) (0. 093 ) (0. 07 6) (0. 040 ) (0. 053 ) (0. 013 ) N Pse ud o R 2 19 ,87 2 0. 05 20 ,06 9 0. 05 18 ,36 1 0. 05 20 ,84 3 0. 05 20 ,80 7 0. 05 21 ,06 2 0. 05 21 ,05 2 0. 04 19 ,58 3 0. 06 19 ,19 0 0. 02 9, 832 0. 03 18 ,41 2 0. 04
Figure 2 – Predicted probabilties for feeling an EU citizen (with 95% confidence intervals) from model 6.1 (education).
Figure 3 – Predicted probabilties for feeling an EU citizen (with 95% confidence intervals) from model 6.3 (education & job situation).
.04 .06 .08 .1 .12 .14 P robabi lit y
low middle high
Education
1 = No, definitely not
.1 .15 .2 .25 .3 P robabi lit y
low middle high
Education
2 = No, not really
.4 .42 .44 .46 .48 P robabi lit y
low middle high
Education
3 = Yes, to some extent
.15 .2 .25 .3 .35 .4 P robabi lit y
low middle high
Education
4 = Yes, definitely
Source: Eurobarometer 82.3 (European Commission, 2014)
0 .05 .1 .15 .2 .25 P robabi li ty
low middle high
Education 1 = No, definitely not
.1 .2 .3 .4 P robabi li ty
low middle high
Education 2 = No, not really
.3 .35 .4 .45 .5 P robabi li ty
low middle high
Education 3 = Yes, to some extent
.1 .2 .3 .4 .5 P robabi li ty
low middle high
Education 4 = Yes, definitely
Source: Eurobarometer 82.3 (European Commission, 2014)
Very bad job situation Rather bad job situation Rather good job situation Very good job situation
Figure 4 – Predicted probabilties for attitude towards immigrants from outside the EU (with 95% CIs) from model 1.1 (education).
Figure 5 – Predicted probabilties for attitude towards immigrants from outside the EU (with 95% confidence intervals) from model 1.3 (education & job situation).
.15 .2 .25 .3 .35 P robabi lit y
low middle high
Education 1 = very negative .4 .42 .44 .46 .48 P robabi lit y
low middle high
Education 2 = fairly negative .2 .25 .3 .35 P robabi lit y
low middle high
Education 3 = fairly positive .03 .04 .05 .06 .07 .08 P robabi lit y
low middle high
Education
4 = very positive
Source: Eurobarometer 82.3 (European Commission, 2014)
.1 .2 .3 .4 .5 P robabi lit y
low middle high
Education 1 = Very negative .4 .42 .44 .46 .48 P robabi lit y
low middle high
Education 2 = Fairly negative .15 .2 .25 .3 .35 P robabi lit y
low middle high
Education 3 = Fairly positive .02 .04 .06 .08 .1 P robabi lit y
low middle high
Education
4 = Very positive
Source: Eurobarometer 82.3 (European Commission, 2014)
Very bad job situation Rather bad job situation Rather good job situation Very good job situation
Figure 6 – Predicted probabilties for EU creates jobs (with 95% confidence intervals) from model 8.1 (education).
Figure 7 – Predicted probabilties for EU creates jobs (with 95% confidence intervals) from model 8.3 (education & job situation).
.1 .12 .14 .16 .18 .2 P robabi lit y
low middle high
Education 1 = Totally disagree .32 .34 .36 .38 .4 .42 P robabi lit y
low middle high
Education 2 = Tend to disagree .34 .36 .38 .4 .42 .44 P robabi lit y
low middle high
Education 3 = Tend to agree .06 .08 .1 .12 P robabi lit y
low middle high
Education
4 = Totally agree
Source: Eurobarometer 82.3 (European Commission, 2014)
.1 .15 .2 .25 .3 .35 P robabi lit y
low middle high
Education 1 = Totally disagree .3 .35 .4 .45 P robabi lit y
low middle high
Education 2 = Tend to disagree .2 .3 .4 .5 P robabi lit y
low middle high
Education 3 = Tend to agree 0 .05 .1 .15 P robabi lit y
low middle high
Education
4 = Totally agree
Source: Eurobarometer 82.3 (European Commission, 2014)
Very bad job situation Rather bad job situation Rather good job situation Very good job situation
lower educated.34 This means that middle educated Europeans are more than 1.5 times more likely and high educated Europeans are almost 3 times more likely to feel an EU-citizen than low educated Europeans.
Figure 2 shows the predicted probabilities of being in each of the four categories of feeling an EU-citizen. For these predicted probabilities, we have to specify for which scores on the independent variables, we want to estimate the outcome on the dependent variable. I estimated the predicted probabilities for all
education groups, for men who are 50 years old in Germany.35 Figure 2 shows that
the low educated (12.04%) are more likely to definitely not feel an EU-citizen than the high educated (4.56%). The middle educated are in between (7.68%) Moreover, the lower educated are less likely than the high educated to feel an EU-citizen definitely (19.00% versus 40.20 %). Figure 4 and 6 show the same for attitudes towards immigrants from outside the EU and the issue if the EU creates more jobs.
The effect of having a better job situation is also significant for most issues (only for the issues reducing public debt is not a priority now and the imposition of a tax on financial transactions it is not). The effect is quite linear and large (see appendices 7-17). The effect of age and gender are significant for some issues, but the effects are small.
The KHB-estimations in table 1 also show if the mediation-effect is significant; if the difference between the effect of education in the reduced model (excluding job situation) and the full model (including job situation) is significant, we can conclude that the effect of education is significantly mediated by job situation. For the middle educated (as compared to the low educated), the mediation effect is not significant for
any of the issues.36 However, for the high educated (as compared to the lower
educated), the mediation effect is significant for the issues about immigration from outside and inside the EU (issue 1 and 2), feeling an EU-citizen (issue 6), general image about the EU (issue 7) and EU creates more jobs (issue 8).
34
Odds ratios are calculated by taking 𝑒𝑒 to the power of the effect. Odds ratio’s indicate a positive effect when they are larger than 1 (the odds increase) and a negative effect when they are smaller than 1 (the odds decrease). In tables, however, I think it is easier to interpret the log odds, because you directly interpret the negative sign for negative effects.
35
The mean age in the data-set is 51.41. Germany is the biggest country of Europe, with the most respondents in the data-set and it lies in the middle of Europe. Moreover, because one part of the country finds its origin in Western Europe and the other in Eastern Europe, it best reflects Europe as a whole.
36
Moreover, the effect of having a middle education does not decrease much when including job situation.
We can calculate how much of the effect of education is mediated by dividing the difference between the reduced and the full model by the reduced model*100%. So, for immigration from outside the EU, 0.088/0.593*100%= 14.85% of the effect of education is (significantly) mediated by job situation. For immigration from inside the EU, 14.47% of the effect of being high educated is mediated by job situation. For the others three issues, respectively 16.66% (feeling an EU-citizen), 25.47% (general image about the EU), and 36.85% (EU creates more jobs) of the effect of being high
educated is mediated by job situation.37
The effect of job situation is confirmed by estimating predicted probabilities. Figure 3 shows that low educated people with a very bad job situation have a probability of 21.97% to definitely not feel an EU-citizen. High educated who have a very good job situation, have a probability of 3.39% to say this. The probability to definitely feel an EU-citizen is 9.52% for low educated Europeans who have a very bad job situation and 45.75% for high educated who have a very good job situation. However, for high educated who have a very bad job situation, the probability to definitely feel an EU-citizen is still 20.41%, while for low educated who have a very good job situation, it is 25.69%.
Figure 5 and 7 show the same for attitudes towards immigrants from outside the EU and EU creates more jobs. The effect of education on the probability to have a negative or positive attitudes towards these issues is somewhat smaller when including job situation. The effect of job situation itself is evident in figure 3, 5 and 7. It is notable that the effect of job situation is quite linear, i.e. a very bad job situation leads to a higher chance of having a negative attitude towards these issues, while having a rather bad, rather good and very good job situation leads for every category to a more positive attitude towards these issues. Moreover, the differences between the effects of these dummy variables are comparable.
These findings imply that for some issues (the immigration issues and some of the issues concerning the European Union), part of the effect of (high) education is mediated through job situation. However, the larger part of the education-effect is not mediated: educational differences still exist when job situation is included. The
Pseudo R2 is around 4-5% in all models. The Adjusted Count R2 is on average
37
For these calculations, I used the unrounded estimations, so calculations from the table by hand could be slightly different.
somewhat smaller. The lower AIC for the models including education and job situation, confirm the effects of education and job situation on political attitudes.
The overall Brant test (appendices 7-17) signifies that for all 11 issues, we cannot assume that the effects are the same for every cut-off point. However, evaluating the Brant test for the effect of education only, it signifies that this is only the case for the immigration issues (issue 1 and 2) and the issue about EU creating jobs (issue 8). Moreover, re-estimating the effect of education using generalized ordered logistic models, emphasizes the robustness of the education-effect (not shown). Although the effect differs somewhat for different cut-off points the effect stays on average the same.
I re-estimated all three models using employment and occupation as control variables (not shown). Being employed does have a positive significant effect on all issues (except for reducing public debt is not a priority now and tax on financial
transactions should be imposed).38 The dummy variables for the effect of occupation
are sometimes significant, but there is no clear pattern here. Furthermore, I included interaction-terms to measure the interaction between education and job situation, between education and employment and between education and occupation. However,
this did not yield any interesting results.39 I also estimated the effect of job
expectations on these 11 issues. This yielded on average the same results. However, the effect of job expectations is somewhat smaller (not shown).
For all issues, almost all country-dummy’s are significant (not shown). Most of them are positive, so other Europeans are on average more positive about all
images than Belgians.40 Swedish citizens are most positive about immigrants from
inside and outside the EU (they are on average 4.5 times more likely to be more positive about immigrants). Italians are most negative about these issues. Both their attitude towards immigrants and their attitudes towards the EU are substantially more negative than in other European countries.
38
However, when including job situation, the effect of employment is insignificant.
39
The interaction-terms for education and employment are never significant. The interaction between having a (very) good job situation and education is sometimes significant (but for the other job situation-categories the interactions are not significant). The interaction between education and occupation is significant for some occupations, although many of these are highly unlikely groups (like high educated manual workers).
40
The exception is attitudes towards a tax on financial transactions, which is opposed more in other European countries than in Belgium.
The difference between Western Europe and Eastern Europe is most clear for issues about having the right to work and live in every Member State of the EU and the issue about if the EU creates more jobs: citizens of Czech Republic, Estonia, Hungary, Latvia, Lithuania, Poland, Slovakia, Slovenia, Bulgaria and Romania are much more positive about these features of the European Union, although their general image about the EU is not more positive. They do also not feel more an European citizen than Western Europeans.
To test if the effect of education differs between countries, I re-estimated the effect of education using a multilevel framework. The effect of education in the
multilevel model (not shown) is comparable to the analyses presented above.41
Including a random intercept and random slope in these multilevel models do not change the effect of education. Moreover, the standard deviations of the random slopes and the random intercepts are very small. This confirms that the effect of
education is stable across countries.42 Therefore, the effect of education on these 11
political attitudes is found to be very robust. None of the alternative specifications did change the effect of education substantively.
The effects of education and job situation on 15 concept images
Table 2 shows the effect of education on 15 concept images using the KHB-method.43
The effect of education is not significant for how European citizens think about large companies and trade unions (although not all Europeans are positive or negative, see appendix 2). For all other concept images, the effect of high education is positively significant. Regarding small and medium enterprises (SME-S), competitiveness, competition, flexibility and solidarity, the effect of having a middle education is also positively significant (but the effect of middle education is smaller than the effect of high education). In general, the effect of education is less strong than for the issues.
41
As outlined in the methods-section, using a multilevel framework is not completely statistically valid, because we cannot assume that differences between European countries are random.
42
Re-estimating the effects for Austria, France, Germany, the Netherlands and the UK (the countries studied by Kriesi et al. (2008) yielded no different results either. Switzerland is not a member of the European Union, so no data for Switzerland is available in Eurobarometer
43
le 2 – Th e e ffe ct o f e du cat io n o n 1 5 c on ce pt im ag es u si ng th e K H B-m et hod . ot es : J ob s itu at io n is th e m ed ia tin g v ar ia bl e. C on tr ol le d f or a ge , g en de r a nd c ou nt ry o f l iv in g. tim at es a re l og od ds , c ou nt ry ro bu st s ta nd ar d e rr or s i n pa re nt he se s. Ps eu do R 2 of fu ll m od el is s ho w n. < 0 .0 5, ** p < 0 .0 1, *** p < 0 .0 01 (t w o-ta ile d t es t) urce : E ur ob ar om et er 8 2. 3 (E ur ope an C om m issi on, 2014) 1 3 3 4 5 6 7 8 9 10 11 12 13 14 15 La rg e co mp . S-ME S We lfa re st at e Co m pe ti-tiv en es Free trad e Pr ot ec - tio nis m G lo ba - liz at io n Li bera - liz at io n Co mp e -t iti on Tr ad e un io ns Ref orm Pu bli c se rvi ce Fle xi - bilit y Se cu rit y So li-da rit y id dle ucatio n uc ed -0 .0 08 0. 17 3 ** 0. 11 4 0. 13 1 * 0. 08 4 -0 .0 70 0. 10 4 0. 12 0 0. 13 2 ** -0 .0 18 0. 12 5 * 0. 04 4 0. 14 7 * 0. 03 2 0. 11 1 * od el (0 .0 62 ) (0 .0 65 ) (0 .0 60 ) (0 .0 51 ) (0 .0 62 ) (0 .0 49 ) (0 .0 84 ) (0 .0 62 ) (0 .0 46 ) (0 .0 84 ) (0 .0 61 ) (0 .0 67 ) (0 .0 60 ) (0 .0 65 ) (0 .0 48 ) ll m od el -0 .0 63 0. 13 1 * 0. 09 1 0. 07 4 0. 03 6 -0 .0 97 0. 05 8 0. 06 9 0. 07 4 -0 .0 25 0. 07 0 -0 .0 02 0. 10 0 0. 00 1 0. 08 4 (0 .0 62 ) (0 .0 62 ) (0 .0 59 ) (0 .0 51 ) (0 .0 60 ) (0 .0 50 ) (0 .0 84 ) (0 .0 61 7) (0 .0 45 ) (0 .0 82 ) (0 .0 59 ) (0 .0 66 ) (0 .0 58 ) (0 .0 65 ) (0 .0 46 ) iff eren ce 0. 05 5 0. 04 2 0. 02 3 0. 05 8 0. 04 88 0. 02 7 0. 04 7 0. 05 1 0. 05 8 0. 02 4 0. 05 6 0. 04 6 0. 04 7 0. 03 1 0. 02 70 (0 .0 69 ) (0 .0 63 ) (0 .0 37 ) (0 .0 73 ) (0 .0 67 ) (0 .0 32 ) (0 .0 56 ) (0 .0 68 ) (0 .0 70 ) (0 .0 26 6) (0 .0 63 ) (0 .0 56 ) (0 .0 66 ) (0 .0 44 ) (0 .0 43 ) ig h ucat io n uc ed -0 .0 41 0. 54 9 *** 0. 32 9 *** 0. 41 2 *** 0. 34 9 *** -0 .2 35 ** 0. 22 4 * 0. 28 7 *** 0. 38 5 *** 0. 05 7 0. 47 2 *** 0. 23 7 ** 0. 44 9 *** 0. 28 0 ** * 0. 43 0 *** od el (0 .0 66 ) (0 .0 67 ) (0 .0 58 ) (0 .0 69 ) (0 .0 63 ) (0 .0 89 ) (0 .0 90 ) (0 .0 78 ) (0 .0 66 ) (0 .0 88 ) (0 .0 64 ) (0 .0 76 ) (0 .0 64 ) (0 .0 69 ) (0 .0 56 ) ll m od el -0 .1 72 ** 0. 44 1 *** 0. 26 9 *** 0. 27 7 *** 0. 23 0 *** -0 .2 95 ** * 0. 12 1 0. 16 7 * 0. 25 2 *** 0. 00 7 0. 34 9 *** 0. 13 0 0. 33 2 *** 0. 20 1 ** 0. 36 0 *** (0 .0 63 ) (0 .0 61 ) (0 .0 55 ) (0 .0 68 ) (0 .0 59 ) (0 .0 85 ) (0 .0 88 ) (0 .0 77 ) (0 .0 62 ) (0 .0 85 ) (0 .0 61 ) (0 .0 73 ) (0 .0 64 ) (0 .0 68 ) (0 .0 56 ) iff eren ce 0. 13 1 0. 10 8 0. 06 0 0. 13 6 0. 11 9 0. 06 0 0. 10 3 0. 12 1 0. 13 2 0. 05 1 0. 12 3 0. 10 7 0. 11 6 0. 08 0 0. 07 06 (0 .0 70 ) (0 .0 64 ) (0 .0 40 ) (0 .0 74 ) (0 .0 68 ) (0 .0 35 ) (0 .0 57 ) (0 .0 69 ) (0 .0 71 ) (0 .0 30 ) (0 .0 64 ) (0 .0 57 ) (0 .0 67 ) (0 .0 45 ) (0 .0 44 ) 19 ,8 50 19 ,0 40 19 ,5 84 19 ,7 77 19 ,6 10 16 ,6 60 18 ,2 26 17 ,6 08 19 ,8 83 19 ,3 24 19 ,4 38 20 ,1 22 19 ,4 95 20 ,5 39 20 ,0 55 ud o R 2 0. 03 0. 07 0. 05 0. 03 0. 03 0. 06 0. 04 0. 03 0. 03 0. 05 0. 03 0. 04 0. 05 0. 05 0. 04
The education-effect is not significantly mediated by job situation.44 The effect of job situation is, however, positively significant for all concept images. Age and gender are significant for some concept images, but the effects are again small.
The Pseudo R2 is around 3-5% in all models.
Using a Brant-test, we see that the effect of education is the same for all
cut-offs points.45 The effect of job situation differs significantly between cut-offs points,
but this difference is not very large (and the ordinal logistic regression estimations gives the average effect in all cases). When re-estimating the models including employment and occupation, we find that the effect of unemployment is significant for some issues, but never when including job situation (not shown). The effects of occupation differ, but almost all of them make sense (e.g. manual workers are significantly more negative about liberalization and free trade). More importantly, the effect of education does not change. In conclusion, the higher educated are more positive about all these concept images. Moreover, people who have a better job situation are also more positive about these concept images.
The positive effect of high education could be due to the fact that many refer to neo-liberal economic concepts, which could be a reason why higher educated are more positive about these images (they are the winners of globalization). For some concept images, we have to be careful to interpret the results. For small and medium enterprises, globalization, protectionism, liberalization and trade unions, the percentage of missing values is above 10%, which could lead to biased results.
The effects of education and job situation on two dimensions
To identify if the concept images and issues relate to the dimensions as outlined by Kriesi et al. (2008), I used Mokken scale analysis. Mokken scale analysis indicates that from these 11 issues and 15 concept images, 2 scales can be constructed. In the first scale, 4 cultural issues are included: immigration outside the EU, immigration inside the EU, work in every EU country and live in every EU country. The H-value
44
Although not significant, a large part of the effect of high education is mediated for globalization (45.98%), public services (45.15%), liberalization (42.16%), free trade (34.10%), competitiveness (33.01%) and competition (31.43%).
45
For some concept images, the effect of education differs significantly between groups, but the effect does not differ substantially for every cut-off point (or the effect is not significant at all).
is 0.64 and significant.46 As these issues are about immigration and the EU, we can conclude that these issues refer to the cultural cleavage as outlined by Kriesi et al. (2008).
In the second scale, 14 concept images and 1 issue (globalization is an opportunity for economic growth) can be included. Only the concept image of trade
unions cannot be included in this scale. The H-value is 0.34 and significant.47 We can
regard this dimension to refer to the economic cleavage, because both the issue and most concept images are rather about economic and traditional issues about redistribution and the economy. All remaining issues and the concept images of trade unions cannot be included in one of the two scales or a third scale, because H-values are not higher than 0.30.
Table 3 shows the effects of education and job situation on the cultural scale. The cultural scale consists of only 4 issues (and has therefore only 7 categories), so an ordinal logistic regression is estimated. The effect of middle education and high education are positively significant. Middle educated are 1.3 times more likely than low educated to be positive about immigrants and the EU. High educated are 2.1 times more likely than low educated to be more positive about these cultural issues. The effect of age is very small, but significant. The effect of gender is not significant.
Estimating the effects using the KHB-method (not shown) yields the same coefficients for education. Moreover, it reveals that the effect of high education is significantly mediated by job situation (13.90% of the effect of education is mediated through job situation). The effect of job situation is also positively significant (table
3). The effects are hence comparable to the effects on the issues separately. The R2 of
the models remain low (around 3%), but the AIC decreases when including both
education and job situation, indicating a better model fit of model 12.3 compared to model 12.1 and 12.2.
The second scale consists of 14 concept images and 1 issue. This results in a scale with more than 10 categories, for which OLS-regression is appropriate to use. Table 4 shows the results. The effect of middle education is significant in the first model. However, when including job situation, the effect of middle education is not significant. The effect of high education is significant in both models, but the effect is almost cut in half when including job situation.
46
Cronbach’s Alpha gives a scale-reliability of 0.73.
47
Table 3 – 3 ordinal logistic models estimating the effects of education and job situation on the cultural scale based on 4 cultural issues.
Notes: Estimates are log odds, country robust standard errors in parentheses. * p < 0.05, ** p < 0.01, *** p < 0.001 (two-tailed test)
Source: Eurobarometer 82.3 (European Commission, 2014
Model 12.1 Model 12.2 Model 12.3 Education Job situation Education & job
situation Education
(reference is low education)
Middle education 0.246*** 0.199*** (0.058) (0.060) High education 0.750*** 0.648*** (0.071) (0.078) Age -0.006** -0.01*** -0.007*** (0.002) (0.002) (0.002) Male 0.004 -0.004 -0.005 (reference is female) (0.033) (0.035) (0.033) Job situation
(reference is very bad situation)
Rather bad situation 0.194** 0.152*
(0.060) (0.060)
Rather good situation 0.614*** 0.521***
(0.077) (0.076)
Very good situation 0.780*** 0.651***
Country dummy’s not shown (reference is Belgium) (0.079) (0.078) T1 -1.059*** -1.275*** -0.753*** (0.172) (0.140) (0.184) T2 -0.596*** -0.812*** -0.286 (0.171) (0.134) (0.180) T3 -0.563*** -0.779*** -0.253 (0.171) (0.134) (0.180) T4 0.631*** 0.415*** 0.954*** (0.126) (0.096) (0.136) T5 0.680*** 0.464*** 1.004*** (0.123) (0.094) (0.134) T6 1.568*** 1.349*** 1.900*** (0.108) (0.101) (0.122) N 21,170 21,170 21,170 Log likelihood -32,139.0 -32,169.4 -32,005.7 Pseudo R2 0.031 0.030 0.035 Adjusted Count R2 0.030 0.029 0.034 AIC 64,296.0 64,358.7 64,035.4
