Master Thesis University of Groningen Faculty of Economics and Business MSc International Economics & Business 18-06-2019

(1)

Master Thesis

University of Groningen Faculty of Economics and Business MSc International Economics & Business

18-06-2019

WHAT DOES INCREASED GLOBALIZATION AND INEQUALITY DO WITH OUR (S)ELECTION? - A STUDY OF POPULISM IN EUROPE

Abstract

There is a medial discussion about whether globalization and populism are related. This study sheds light on the possibility that inequality mediates the effect of certain parts related to trade to a vote for populistic parties. The sample is 26 European countries over the period 1980-2016 with the dependent variable populist vote shares. Multiple regression analyses lead up to a mediation analysis performed with a supplementary Sobel-Goodman test. There is no evidence found for inequality as a mediating factor, but a finding is that increased inequality positively influences voting on populism.

Keywords: Trade, Inequality, Populism.

(2)

2

(i)

Abbreviations

AIC Akaike Information Criterion

BIC Bayesian Information Criterion

FE Fixed Effects

GDP Gross Domestic Product

NI National Income

OLS Ordinary Least Squares

PWT Penn World Tables

RE Random Effects

R&D Research and Development

R2 R-squared

SS Stolper-Samuelsson

WID World Inequality Database

(4)

2

1. Introduction

There is a change in the world, as we knew it. Former advanced countries advantage gets smaller and globalization increase with both praise, criticism (Stiglitz, 2017) and consequences as more international fragmentation of production and more flexible labor laws (Rodrik, 2018). The middle class in Europe identify themselves with the well-educated upper-middle class because of equivalent consumption possibilities. When the former loses their jobs, dissatisfaction grows because of raised economic insecurity and decreased income position relative to the rest of the country. This creates a want for change, to get politicians to make it better for the ‘common’ people, which might take the form of a vote for populism. It can be about temporary resentment and a protest against single issues, but it can also represent a reply to structural problems (Betz, 1993). With a more divided society rising simultaneously along a more globalized world, with consequences as more fragmentation of production, increased import competition and less labor protection, an overarching research question arises:

Is globalization-induced inequality a reason for increased voting for populism?

The topics are of interest both for an economist such as for the society aggregated since it affects political outcomes and economic policies. Trade1 with the correct policies will raise GDP in all trading countries giving everyone a bigger part (Autor, Dorn and Hanson, 2016) and decrease inequalities between countries but with increasing within-country inequality (Bourguignon, 2015). One factor for this is the vertical fragmentation of production (offshoring2) (Baldwin, 2016) which displace workers in rich countries (Helpman, Itskhoki and Redding, 2010). The gain in unemployment, decrease wages and create inequality, which elevates tensions between groups and make it harder for communities to reach consensus in political questions and education will suffer (Keeley, 2015). This can be seen in the diverging middle class when social mobility decrease (OECD, 2015) and it becomes an even bigger issue when citizens feel that there are different rules for different people (Rosanvallon, 2016). Hence, with economic shocks from globalization, support for populism and anti-globalization increase (Autor et al., 2016; Colantone and Stanig, 2018b) with a common proposal to impose barriers for trade. It will hurt parts of the world more than it creates gains for the native population and in general, Rodrik (2018) argues, populism is not something economists are fond of because of unsustainable and thoughtless policies.

There are many studies on globalization and particularly the economic aspects. However, only a minor part focus on the relation between trade and voting behavior directly (Colantone and Stanig, 2018b). The examination of the consequence of economic globalization, in this case, theorized by offshoring and import competition3 that are assumed to follow the theory of job polarization. Hence, increased economic globalization will lead to unemployment for workers in the industrial sector. The findings in this study will add further evidence for a possible relation between these phenomena to this small group of evidence.

1_{Further, when the word trade is used it refers to international trade of goods and services.}

(5)

3

The method to investigate the research question is multiple regression analyses to find an answer to four hypotheses. The regressions lead up to a mediation analysis, the causal four-step mediation proposed by Baron and Kenny (1986) with a supplementary Sobel-Goodman test. Mediation analysis is widely used within social sciences, but not yet so much within economics (see for example Heckman and Pinto, 2015; Hacibedel, Mandon, Muthoora and Pouokam, 2019) and it is an intuitive instrument when examining channels or influences between two variables. The data consist of measures of populism, inequality and the consequence of economic globalization, measured as the number of workers within industrial and manufacturing sectors. Conclusions drawn are that inequality is not found to be a possible mediating factory and there is no evidence on a relation between economic globalization and inequality nor populism. However, there is evidence found on a relation between inequality and voting on populism.

The rest of the thesis is structured in the following way. First, there will be a discussion of former literature on the three areas trade, inequality, and populism. Second, the methodology and data will be discussed. Third, the results are on display and analyzed and finally, there will be a conclusion.

2. Literature review

2.1. Globalization and the backlash

The phenomenon of globalization is on the agenda and have been for some time, with both positive and negative consequences for both high- and low-income countries. IMF (2002) describe globalization as a process through which an ever-increasing flow of goods, services, ideas, people and capital lead to economic interdependence of countries worldwide. For high-income countries, from around the year 1990, their share of high-income has reversed for the benefit of countries with lower income and it is now back where it was in 1914. This decline is a plausible explanation as to the origin for anti-globalization attitudes rising in advanced economies (Baldwin, 2016) such as Europe and mainly the Western part.

(6)

4

regulation. Trade agreements have moved the power from the employees to the employers with the consequence that organized labor lost its political influence (Rodrik, 2018). In advanced stages of globalization, trade agreements tend to be more about redistribution than to increase the overall economic pie, and this might be a potential reason why it is becoming politically contagious (Rodrik, 2018).

With globalization and liberalization of trade if the correct policies are executed, increased trade will raise GDP for all countries involved and create a larger economic pie giving everyone a bigger part (Autor et al., 2016) and thus there should be net welfare gains. Despite this, it also creates losers through economic shocks by shaping the labor markets in advanced economies. There are jobs offshored and there are jobs replaced by technology, both share the property that they are easy to codify (Blinder, 2009; Autor, Dorn and Hanson, 2015) which makes it hard to disentangle the origin (Rodrik, 2018). Both can, however, be used as reasons for job polarization which occurs when there is growth in both high-skilled (professional, managerial) and low-skilled (personal services) jobs with declining employment for the middle-skilled workers, which in the manufacturing sector is doing routine production jobs (Goos and Manning, 2007).

One hypothesis explains job polarization as a technological progress that replaces routine, codifiable jobs in the middle of the wage distribution (Autor, Levy and Murnane, 2003). Autor and Dorn (2013) found evidence for this in the US with employment losses within routine task-intensive occupations but at the same time employment growth within abstract (problem-solving and organizational tasks) and manual task-intensive occupations. However, in the manufacturing sector, even the manual and abstract jobs suffered employment losses rather than gains. Together these effects had a significant negative overall employment effect (Autor et al., 2015). A second hypothesis is that offshoring has created the change in the composition of jobs in advanced countries (Blinder, 2009) for which Goos, Manning and Salomons (2009) finds weaker evidence. Job polarization has been occurring in both the United States and in Europe since the early 1990s with an increase in high and low-paid jobs (Goos, Manning and Salomons, 2009) and a shrinking middle class in Europe with the strongest cases in Austria, France and Sweden (Peugny, 2019).

Globalization itself is intrinsically neither good nor bad, but with the power of it, both are possible (Stiglitz, 2017). Nonetheless, there has been a backlash on globalization, as can be argued to be predominantly a backlash on economic globalization and trade. This backlash can take many forms, but the one further discussed here is populism or populistic parties. Mudde (2007) defines radical right populist parties with three particular characteristics separating them from ‘mainstream’ parties. Nativism, stating that members of the native group should be the only population in a country. Authoritarianism, that there is a general custom to be uncritical toward authoritative figures in a person’s own group but at the same time punish figures outside of the group. Lastly, populism, which should be understood as a ‘thin-centered ideology’4_who see the society as divided into two homogenous and hostile groups, ‘the common people’ versus an enemy who can be in the form of bureaucrats, elites or immigrants (Rodrik, 2018) and which

4_{‘The particular ideas under its command are of limited scope, complexity and ambition when measured}

(7)

5

argues that the policies proposed should be according to the general will of the people. The difference for left-wing populist parties is that they lack the characteristic of nativism, but instead have (in the case of Europe) an EU-skeptic orientation (Armingeon et al., 2018). Finally, in a populistic democracy, the most important aspect to take into consideration is the peoples ‘general will’, and not, for example, human rights (Mudde, 2007).

2.2. Economic globalization, inequality, and populism

Although liberalization of trade creates gains, there is always parts of the population that is going to be damaged. When low-wage countries have entered into world trade, domestic wages, industrial employment (Artuç, Chaudhuri, and McLaren, 2010) and occupations (Ebenstein, Harrison, McMillan, and Phillips, 2014) have been depressed in advanced industrial countries due to import competition and offshoring, which can both be connected to job polarization. Further, import competition is now seen as the essential force responsible for the decline in labor shares since the end of the 1980s (Elsby, Hobijn and Sahin, 2013), and the increased trade flows after China joined WTO 2001 magnified this. It has been a considerable factor for the decline in manufacturing employees and weekly earnings in the US (Autor, Dorn and Hanson, 2013; Autor et al., 2015). On the other hand, Feenstra, Ma and Xu (2017) found the opposite, that areas with higher exposure to Chinese trade tend to benefit more from increased exports to China and that the effects roughly cancel out. It might however not only be manufacturing workers affected by import competition. Colantone and Stanig (2018b) found that service workers and employees in the public sector also was affected by Chinese imports, which seem odd since they could be assumed to be protected from foreign competition.

In general, (economic) globalization tend to negatively spur support for right-wing populistic parties and an anti-globalization agenda. There is evidence found in both the case of trade shocks from China as for rising import competition and immigration (Swank and Betz 2003; Autor et al., 2016; Colantone and Stanig 2018b). In the European case, the rise of populism started in the late 1980s associated with globalization (Rodrik, 2018) and in the examples of Germany and France; globalization has essentially fed its own opponents. The workers that are exposed to rising imports have shown their unhappiness by voting for nationalist alternatives to globalization on the right-edge (Dippel, Gold and Heblich, 2016; Malgouyres, 2017). A generalization that can be made to Western Europe where imports from China have boosted the support for populism and in British regions, where the most affected by Chinese import competition also were the regions voting the most for Brexit, which can be seen as more support for populistic politics (Colantone and Stanig, 2018a, 2018b).

(8)

6

mainstream established parties that have former experience of times with economic hardship. Lubbers, Gijsberts and Scheepers (2002) find that radical right support was stronger among unemployed, blue-collar workers with less education but on the other hand, Inglehart and Norris (2016) found the opposite, that voting for populism was stronger for the upper class. However, the effects seem to be particularly strong in the Eurozone area. The reasoning is that the institutions are too rigid and thus have a greater inability to react to globalization shocks. This makes voters frustrated with the institutions and vote for populism (Guiso, Herrera, Morelli and Sonno, 2019).

Concluded, although there are some mixed results on why populism occurs, the direction summarized is that economic globalization and trade is a part of it. With job polarization happening, unemployment increases and hence resentment and dissatisfaction that lead to a vote for populistic parties. Generally, trade is a suitable black sheep and a reason for why this vote materialize. It is easy to use to point towards other nationalities and economies as the problem. To lose a job to peers that compete under the same regulations is one thing, but it is different to lose it to other nationalities with lower safety standards and labor costs at their advantage (Rodrik, 2018).This reasoning gives background to the first hypothesis that will be tested:

H1: Economic globalization leads to increased populism.

Globalization has influenced wages and contributed to the relative wage increase for skilled relative unskilled workers whereas trade and offshoring have had a negative impact on certain workers’ salaries (Helpman, 2017). The relation between trade and wage inequality is, however, a much-debated topic within economics (Helpman, Itskhoki and Redding, 2010) for which OECD (2012) stated in a report that globalization could widen income inequality within their member countries due to offshoring. Recent studies have started to give a bigger role to trade in the diversion of wages and growth of skill premium (and not technology as the main source anymore) (Rodrik, 2018) even though Helpman (2017) argues that the aggregated effect has been moderate and that globalization only explains a small part of the wage inequality increase within countries.

(9)

De-7

regulations have only reinforced consequences from globalization with increased performance-related pay and share-options creating exceptional incomes at the top (Nolan, 2017).

When the effect from trade to inequality is analyzed through neoclassical trade theory and the Stolper-Samuelsson theorem, there are two goods and two factors where the owner of one of the factors will necessarily be worse off. The factor that is used more intensively in the production of the good that is imported must see declining real earnings. If applied to the United States and Europe which are advanced industrial powers but labor scarce, the result is predicting that the low-skilled workers should be worse off when there is trade liberalization (Rodrik, 2018) and hence inequality should rise (and vice versa for labor-intensive countries). In general, the theorem states that trade always produces losers although it will be better for the world. Finally, the results found on the relation between trade and inequality are widely mixed. Li, Squire, and Zou (1998) do a cross-country study and find no systematic and significant relationship between income distribution and trade liberalization. Förster and Toth (2015) do a comprehensive survey of cross-country studies on income inequality in the OECD area where the same finding is made, the evidence tends to be mixed between insignificance and inconclusive with contradicting results between studies (can be due to differing methods). However, a consequence of globalization has been deregulation of labor markets, which had a positive effect on employment, but it has also been associated with a rise of wage inequality in many countries (Förster and Toth, 2015). Maybe that is a reason why some scholars have found that globalization and especially trade flows increase income inequality (Dreher and Gaston, 2008; Bergh, Kolev and Tassot, 2017). It is essentially as common to find that globalization has influenced increasing income inequality such as it did not have any effect (Marsh, 2016). There is no clear direction on evidence from former studies, and the relation between globalization and inequality is debated. In the form of the SS-theorem, there should be increased inequality in advanced industrial economies, but it might not be the case empirically. When trade increase or production get offshored, it can be expected that the relative wage for skilled workers goes up. Together with import competition, it also creates job polarization and therefore can make middle-skilled workers lose their income. These people are to a high extent in the middle class, which decline when the skill premium grow. Following the reasoning of Wolfson (1994), the decline of the middle-class is just another way of stating polarization in the income distribution, and thus hypothesis two is:

H2: Economic globalization leads to increased inequality.

(10)

8

decrease in the medium run. The opposite is true if the bottom 20% in the income distribution increase their share for which a channel might be more accumulation of human capital (Ostry, Berg and Tsangarides, 2014; Dabla-Norris, Kochhar, Suphaphiphat et al., 2015). Cingano (2014) states that if inequality had not grown since the 1980s in many OECD countries, the GDP growth would have been much higher and there exists substantial evidence that growth is stronger in countries with a stable middle-class (Brueckner and Lederman, 2018). If inequality decrease opportunities for the coming generation, it can become a channel for populism. Some evidence supports this, but they are often based on cross-country comparisons and since countries are heterogeneous with multiple potential factors affecting mobility, it is hard to draw strong conclusions from a comparison like this (Nolan, 2017)

Pástor and Veronesi (2019) develop a model where populism grows when the economy grows. The voters dislike inequality and particularly extreme consumption of the elites. Growth aggravates inequality and as a response, voters elect populists promising to end globalization. Empirical evidence from Brexit and the American election 2016 supports the model where the backlash is inevitable because the agents are assumed to dislike inequality. However, evidence on inequality as a determinant of populism voting have been mixed and with diverging explanations. Han (2016) finds mixed results; one of them is that there are different effects on rising income inequalities depending on what socio-economic status an individual inhibits. However, when just examining income inequality, it is not found being a determinant of right-populism. Coffé, Heyndels and Vermeir (2007) found that in Belgium, increased inequality reduced the support for right-wing populistic parties because the lower socio-economic classes voted for left-wing parties instead because their policies were aimed towards the poor, and hence represented their interest the best. No evidence at all was found for income inequality as a determinant of populism when Jesuit, Paradowski, and Mahler (2009) did a study on eight Western European countries. This concludes with that populism has grown in countries where inequality has been more or less stable over time (for example France and Austria) as well as in countries where inequality has increased (Nolan, 2017).

The indicated can give a hint that it might not be about inequality per definition, but rather ‘positional deprivation’. Burgoon, van Noort, Rooduijn and Underhill (2019) define it as how much a certain voter’s decile has experienced growth in their real income relative to the development of other deciles in the income distribution of the country. People that have experienced less growth than other deciles have a higher propensity to support radical-right parties, which can be because these parties paint an easy picture on who to blame for the unjustness the relatively deprived voters feel. Hence, positional deprivation might be a critical source for right-wing populism in Europe (Burgoon et al., 2019). This theory can also be connected to the way populism see the society, as divided into two groups, the ‘pure people’ versus an enemy in the form of a ‘corrupt elite’ in this case. Thus, if the top quintiles have increased their share of total income over time, which seems to be the case (Lawrence, 2007; Pressman, 2015), and the lower parts have decreased, then there could be a possibility of positional deprivation.

(11)

9

inequalities and decreased possibilities in life. In Europe particularly with the history of a strong welfare state, it can give a feeling of that this is dissolving and in search for a change, they vote for populists. The evidence on a possible relation is mixed, but the evidence is not conclusively against it either. If assumed that a population does not like inequality, then voting for anti-globalization parties should increase following the model of Pástor and Veronesi (2019). Two other possible channels that inequality can give rise to populism via is decreasing economic growth and a decline of opportunities. Nevertheless, it can also be the same reasons as to why populism grows from economic globalization and trade, with workers getting unemployed and losing their wages. Derived from this discussion follows hypothesis three:

H3: With increased inequality, populism increases.

To follow up on the discussion so far, there are a few things to add to form the final hypothesis. The argument can be made that job polarization together with an increasing inequality has diverged the middle class. The “old middle class” as Iversen and Soskice (2019) call it has experienced declining wages and mainly consists of less-educated manual workers with low income. They argue that this group has low confidence in the educational system and that it will make it possible for their children to be a part of the new economy at all or at the same conditions as others, they are essentially concerned about the chances that their children will find a job. The new middle class, on the other hand, is well educated, has a cosmopolitan attitude with no need of any scapegoats and see themselves as the drivers of the economy (Iversen and Soskice, 2019). These differences might have made the old middle class more prone to vote populistic. This is at least what Iversen and Soskice (2019) find when comparing the old and new middle class. The old is supposedly eight times more prone to vote for a populistic party, with values and economic position considered.

The status of the old middle class can be argued to be a consequence of economic globalization, with offshoring and import competition that decrease employment and create job polarization. With more unemployment and fewer hours worked on average, wages decrease and there might rise a feeling that they are left behind in the new economy. When this is added to the feeling of decreased opportunities for the next generation this make the old middle-class vote for resentment politics of populist parties. Thus, it is argued here that inequality has mediated the effect of trade and created a vote for populism, for which hypothesis four is presented:

(12)

10

3. Data and Methodology

3.1. Data

As can be argued to be one of the bigger discussions in the contemporary political discourse, populism will be the dependent variable5. The variable is an aggregate of the share of votes for populistic parties on both the political left and right side. It is collected from the Comparative Political Data Set (Armingeon et al., 2018) which consist of political data on 36 OECD and/or EU members between the years 1960 and 2016. The data set divide parties into different party groups depending on their characteristics and a party is in the data if it has got at least 2% of the votes in the corresponding election. The parties defined as ‘right-populist’ in the data set is after a definition by Mudde (2007) who says there are three characteristics dividing these parties from ‘mainstream’ parties; nativism, authoritarianism, and populism (examples Sweden Democrats or National Front from France). What can be argued to be left-wing populism is the party group ‘protest’ in the data set as defined by Lane, McKay, and Newton (1997), which in recent movements have had characteristics such as EU-skeptic orientation but they lack nativism (Armingeon et al., 2018).6 To use both sides as an aggregate has increased relevance since politicians on both sides of the political spectrum have started to join forces to oppose the mainstream center, which De Vries (2018) use partly to argue for why the conventional left-right dimension is not relevant anymore.

The independent variables represent the consequences of import exposure and offshoring. It is manufacturing and industrial jobs in thousands of workers7, to depict job polarization that stems from increased economic globalization. Manufacturing is the main economic activity subject to job polarization, but the wider definition industrial jobs are also chosen as a second independent variable because of the richer data available, and hence they will both be used in the specifications. The data are gathered from Eurostat (2019) and exist from 1983-2016 (industrial) respectively 1992-2018 (manufacturing). The classification of industries used is NACE revision 1.1 1983-2008 with the totals for industry and manufacturing aggregated. However, for 2008-2016 and revision 2.0 there is no aggregate of the industrial sector. It is aggregated manually by using classifications B, C, D, E (manufacturing, mining and quarrying, and other industry) and F (construction) to follow the broad classification of industry in NACE revision 1.1 (Eurostat, 2008). For 2008, data exist according to both classifications, for harmonization; an average was created between the two values of the different revisions. Both are weighted for the total population (The World Bank, 2019) in thousands to decrease outliers and give an equitable picture of the circumstances, hence they end up as per capita. The choice to not weigh by labor force comes from that the income effect that emanates when losing a job affects the whole population’s tendencies to vote populist. A small bias might arise when the levels over time change because of workers that leave the labor force, but it should not affect the result substantially. Finally, when the exposure of economic globalization increases for a country, the number of jobs should go down following theory, which is the common trend over time.

5_{For individual country-trends over time, see appendix B.1.} 6_{For a list of parties, see appendix A.2.}

(13)

11

The mediating variable used as a proxy to measure inequality is the income share for the middle 20% of the population before taxation and will further be denoted as income share (mid-20),

8_{which is where the biggest drop in share of total income for the middle class over time has}

happened (Pressman, 2015). It is gathered from World Inequality Database (2019) because of the harmonizing data over time for many countries and that it does not rely for the most part on household surveys, which can create biases (World Inequality Database, 2019). Instead of using GDP, WID uses NI, which is GDP minus consumption of fixed capital plus the net foreign income. NI is more meaningful because, for instance, a country with a large GDP but extensive capital depreciation and foreign outflows do not have much income to distribute to its residents and citizens, which the NI reflect (World Inequality Database, 2019). The choice to use just inequality in income as a factor that might create populism is because it considers the forces arising from the difference in ownership of capital and capital income, but also the forces stemming from inequality in labor income (Piketty and Saez, 2014). The measure of the middle class by income fraction is one way to do it but there are a few imperfections in measuring it this way. It does not show what happens to the magnitude of the group in the middle of the income distribution, it cannot grow or shrink, and it is always the same size of the population (Pressman, 2015). Alternatives would be to define it normatively (Iversen and Soskice, 2019) or by upper- and lower-income limits. Another choice that can be discussed is the choice of pre-taxation. Post taxation can be argued to be a better choice because consumption possibilities might determine populism. However, after taxes, redistributions have already been done, and in the sample of Europe, this can affect the income share quite substantially and hence pre-taxation is the chosen variable. Finally, when the income share goes down, it can be assumed that inequality increase within a country.

There are several control variables used that all have been found to affect populism and inequality. Since technology might be a reason for job polarization, it will be controlled for, as a creator of populism and inequality. A well-used proxy for technological change is expenditure on R&D as percentage of GDP; with a limitation that innovations not always generate innovate output (Barbieri, Piva and Vivarelli, 2019) and it is collected from OECD (2019). From former studies on populism, there are clues telling that GDP per capita should matter, especially when there are changes and due to this, it is used as a control. It is measured as GDP per capita (constant 2010 US$) with the last year subtracted to create GDP per capita change and collected from The World Bank (2019). Further, as Rodrik (2018) argue, when trade barriers are low it tends to be more about redistribution when liberalizing more trade that makes globalization politically contagious, and thus aggregated public social expenditure as percentage of GDP is included as a control, collected from Eurostat (2019). Trade openness is used as a control as well and it is measured as the sum of exports and imports of goods and services as a share of GDP, collected from The World Bank (2019). Finally, to control for education, a human capital index is used. It is based on years of schooling and returns to education; hence, it is also displaying if education gives returns to the investment spent in time (and money). This variable is collected from PWT 9.0 (Feenstra, Inklaar and Timmer, 2015).

(14)

12

As previously discussed, many aspects are to consider when studying the determinants of populism. Therefore, there are two dummies used throughout the analysis. The entry of China into WTO from 2001 and financial crises (banking and systemic) which is one aggregate dummy compiled from a data set on global crises by Reinhart, Rogoff, Trebesch, and Reinhart (2019) based on that especially far-right populism have a tendency to increase after financial crises (Funke, Schularick and Trebesch, 2016). An example of the regressions performed together with corresponding control variables will be presented before the result section.

3.2. Sample

This study chose countries based on the criteria that they are inside Europe’s single market and that they are OECD members from the notion that all can be assumed to be advanced industrial countries. The choice of countries follows former studies on the phenomena’s approximately to be able to compare the results (see for example Swank and Betz, 2003; Goos, Manning and Salomons, 2009; Colantone and Stanig, 2018). These studies also find that the effect of job polarization and globalization shocks to vote populistic is mainly happening in (early) EU members; hence, a subsample of only EU members (from the year they entered) will be used to compare with the main results. This to see if the findings differ when excluding observations outside of the union. Even if a lot of literature is about the consequences for The United States, it is particularly not included because of their electoral system. It is a pronounced two-party system with minimal chance for a populistic party to arise; there is no option to vote for a populistic party since they are not on the ballot. However, the choice not to omit countries without successful populist parties (for example Portugal) decrease the risk of selection bias, but at the same time, it poses a problem of inference (Arzheimer and Carter, 2006). In the end, the amount of countries is 26 at most and less when using the subsample of EU. 9

The time aspect examined is 1980-2016 because of the rise of populism in Europe in the late 1980s (Rodrik, 2018), and that globalization as measured by trade flows started to increase. Another aspect is that consistent income inequality data for the countries in question was hard to find before 1980 but with a time aspect this large, the study should become more robust. However, since it is the elections that are the dependent variable, most of the time the observations per country will be every 3-5 years depending on their electoral system. Within the time aspect, it is important to be aware of exogenous factors that might have affected the dependent variable, where two is financial crises and China’s WTO membership 2001 for which dummy variables are introduced (as discussed).

3.3. Methodology

The four hypotheses will be tested empirically through four panel regressions in which country-specific and time-country-specific variables can be accounted for and conclude in a mediation analysis. The regressions are estimated to identify possible relations between economic globalization (measured as jobs), inequality and populism, conditioned on multiple control variables identified in the literature. A mediation analysis is essentially to make a hypothesis about a network of causality (Judd and Kenny, 2010). There are three major assumptions for mediation

(15)

13

except for the standard assumptions of linear models and these are reverse causal effects, measurement error in the mediator and omitted variables. It is assumed that there is an order for when the variables are taking place (MacKinnon, 2008) which the theory of that decreasing income share and jobs are leading to the human choice to vote for populism should be enough to withdraw concerns. The temporal precedence is taken care of by creating averages, one year before elections to make sure they take place in order (more on this later). Measurement errors should not exist because the data is gathered from trustworthy sources and controlled for miss measurements. At last, it might be that an omitted variable causes both changes in income share and populism, but controls are included to decrease this risk and the absence of omitted variables is neither a necessary condition (Pearl, 2014). It is also important to include the same control variables for both paths of the independent variable in the models; otherwise, they will not be controlled for entirely (Hayes, 2018)

The type of method for mediation is the causal four-step mediation analysis by Baron and Kenny (1986) with the goal to try to identify causal mechanisms suggested by theoretical claims (Imai, Keele and Tingley, 2010). It will estimate the role of causal mechanisms when populism is affected by jobs via income-share and reports results for both direct and indirect effects (Hicks and Tingley, 2011). The four-steps by which the causal mediation is performed is:

Step 1: Estimate c (Jobs -> populism) Step 2: Estimate a (Jobs -> income share)

Step 3: Estimate b (Income share -> populism, controlled for jobs) Step 4: Estimate c’ (Jobs -> populism, controlled for income share)

Figure 1 – The mediation paths

Jobs c Populism Income share a b Jobs Populism c’

(16)

14

is total mediation) (for example Kenny, Kashy, and Bolger, 1998; MacKinnon, Krull and Lockwood, 2000). After the four steps are carried out and the criterions are fulfilled, the different effects are calculated:

𝑇𝑜𝑡𝑎𝑙 𝑒𝑓𝑓𝑒𝑐𝑡 (𝑐) = 𝑑𝑖𝑟𝑒𝑐𝑡 𝑒𝑓𝑓𝑒𝑐𝑡 (𝑐′) + 𝑖𝑛𝑑𝑖𝑟𝑒𝑐𝑡 𝑒𝑓𝑓𝑒𝑐𝑡 (𝑎 ∗ 𝑏) (1)

What is likely to occur is partial mediation, which will appear when income share explains parts, but not all of the relation between jobs and populism. However, even if an effect is statistically significant it can simply be a type I error, where the effect is significant by chance in its own (MacKinnon, 2008), but it can also be that income share is, in fact, a proxy for the real variable that is the mediating force (Kraemer, Wilson, Fairburn and Agras, 2002). Nevertheless, there are flaws with the causal four-step analysis; it has low statistical power, do not test the significance of a specific indirect effect nor quantify the magnitude of the mediation effect (Hayes, 2018). Hence, as a supplement to the four-step analysis, the Sobel-Goodman test is used. It focuses on the product a*b to assess the significance of the indirect effect and not the individual paths because a*b will be equal to the discrepancy between the total and direct effect (Preacher and Hayes, 2008). It is however also with low power and a flaw because of the assumption that the sampling distribution of the indirect effect (a*b) is normal which tends to not be the case with a small sample like here (Hayes, 2009). The Sobel test uses standard errors for its calculations that assume that estimates of a and b are independent which is neither the case here because of using a panel structure (related to multilevel modeling). The outcome will therefore at best be evidence for approximate independence. Neither Sobel-Goodman or the causal four-step is used much anymore because of bootstrapping that has replaced it which is a better method to use for testing the indirect effect (MacKinnon, Lockwood and Williams, 2004). Among other factors, it does not impose a distribution of normality and has higher power for smaller samples (Preacher and Hayes, 2008).

It is recommended when using longitudinal (or panel) data in mediation models to acknowledge methodological limitations (Cole and Maxwell, 2003) which there are several here. First, traditional mediation analysis assumes the data to not be clustered (by countries). It can also be a problem doing mediation over time; the temporal properties can give issues such as correlated errors and lagged effects (Bolger and Laurenceau 2013). For the causal four-steps method, optimally the sample should be large to give the total indirect effect a normal distribution (Preacher and Hayes, 2008), bootstrapping is not used because of limitations in Stata10 and time constraints, but the more conservative Sobel-Goodman test. Hence, the outcomes from the mediation will not be optimal. For the empirical analysis, the criterions by Baron and Kenny (1986) are checked to see if mediation can be established, if not, only step two and three will be used to try to establish mediation as contemporary analysts believe is enough. If mediation is found, a Sobel-Goodman test is conducted to check the validity of the conclusions reached without it (Hayes, 2009). In general, the different methods to examine mediation agree in more than 90% of the cases (Hayes and Scharkow, 2013) and thus, this should give a clear view if there is a mediating relationship between economic globalization and populism via inequality.

(17)

15

Only election years are used because the interest is in how people change their votes for populism depending on several determinants. To be able to perform the type of mediation analysis explained, OLS regression must be used, although it is more common in similar studies to use Tobit to take care of all zero percent vote shares, and not violate the assumption of having a zero mean for the errors. To try to solve the problem of this assumption all variables except populism are averaged between elections, except the year of the next election but including the year of the former election. Therefore, when there is an election 2000 and the last was 1996, the variable will be averaged 1996-1999 and after this is done, non-election years are deleted from the data.11 This way the values included in the average will certainly happen before an election, which would not have been the case otherwise. The only variable that is averaged two years before an election is R&D because it is expected to take a longer time before showing an effect. When using averages, the highs and lows in the trends of variables are being flattened and hence smoother. This can both increase and decrease the chance of getting significant results because there is less noise in the observations. The data set is unbalanced, and with gaps due to elections in irregular intervals between countries. Hence, there is not the same amount of observations for every country and the number of observations will differ depending on which variables included in the specification.

Endogeneity might be a problem if there is a correlation between populism, income share, and jobs, it does not have to be causality (Longhi and Nandi, 2017). Between jobs and populism, there should not be endogeneity and neither between income share and populism because the focus lies on voting and not economic policy proposed by populists. That the election year predicted is excluded from the averages should further decrease the problem of endogeneity because it is about the future, the variables cannot be decided simultaneously. This is different to jobs and income share, for which there can exist endogeneity; a story could be that the income share goes down and therefore the demand for manufacturing products, which offsets industrial workers. However, the averages should decrease the risk of endogeneity also in this case. All specifications are checked for heteroscedasticity with the modified Wald test for groupwise heteroscedasticity and the finding is that there is heteroscedasticity throughout all specifications.12_{Thus, robust clustered standard errors (by country) will be used for all}

specifications to account for prevailing heteroscedasticity and repeated measurements of the same country. These standard errors also account for serial correlation and hence it can be assumed that there is no linear association between errors (Hill, Griffiths and Lim, 2012). The Breusch-Pagan test is performed to test if pooled OLS or RE should be used. The outcomes unanimously reject the null hypothesis meaning that RE is preferred.13_{To see if RE or FE is}

preferred, the Hausman test is executed as a director to provide an indication for if the relationship between is different from the relationship within countries. If the test shows that RE should be used, it can only be interpreted as that the between effect is not significantly biasing an estimate of the within effect (Bell and Jones, 2015). All 17 specifications are tested and the outcome is that in nine cases FE is favored and in eight RE.14_{However, to use a}

(18)

16

consistent causal four-step analysis the models and assumptions have to be as consistent as possible and hence FE will be used even if it is only the within effects that can be estimated. The arguments for the choice are twofold. One is intuitive, with countries it seems like a strong predictor that FE should be used because of the heterogeneity in characteristics (for example culture, geography, and legal systems). There will not be any heterogeneity bias since only within effects are estimated and all country-level variance and between effects are controlled out (Allison, 2009), captured in the error term and making the coefficient more precise (Kreuter and Kohler, 2012), but a lot of information is also lost this way (Bell and Jones, 2015). The other is from an econometric view if any of the explanatory variables in a RE model is correlated with the random errors, then the estimators are biased and inconsistent where for the FE estimator, on the other hand, it is still consistent even if this correlation applies. Finally, the rejection of the Hausman test can depend on the model specification. It can be misspecified and lack relevant explanatory variables or nonlinearities that are not captured (Hill, Griffiths and Lim, 2012). Year FE will not be used because of the limited degrees of freedom, and instead, two dummies will be used for the economic shocks that are expected to affect the dependent variable the most, financial crises and China’s entry to WTO.

One of the assumptions for OLS is a normal distribution for dependent variables (particularly for smaller samples because of the central limit theorem) which is controlled for in populism and income share.15 It is controlled for by using a histogram and a Skewness-Kurtosis test and populism is found to be non-normal. To take care of this it is transformed into logarithms, but because of the many occasions where there are 0% votes for populistic parties it is still skewed and not normally distributed. Hence, a non-parametric method should be used instead of OLS for optimal results, an example is quantile regression were the assumption of normality is not as strict for smaller samples. However, to understand how to execute mediation as optimal as possible took too much time and because of time constraints, quantile regression is not performed. This creates consequences for the regression such as that the calculation of p-values for significance tests might be distorted; thus, there will also be a shorter discussion of the results with the maximum probability level set at 1% for significant results to be on the safe side (Hubbard, 1978). The transformation into logarithms make the variance smaller and hence the confidence intervals, which makes it harder to find significance in the first place. Nevertheless, even when normality is violated, the OLS estimates preserve most of the desirable properties (Hubbard, 1978), except perfect inference (Gelman and Hill, 2006) and can still be the best linear unbiased estimator (Hill, Griffiths and Lim, 2012).

Finally, when doing mediation with panel data it is important to test the assumption of stationarity (Cole and Maxwell, 2003). Out of the available tests in Stata, the best one for handling unbalanced panel data with a finite amount of observations is the Fisher type unit root tests but it cannot be used when the data has gaps (Statacorp, 2013) which is the case after creating the averages. Hence, the test is performed before creating averages for all variables (populism in logs) and all of them except GDP per capita change and social expenditure seem to have stationarity.16 Thus, the assumption should hold at the most (even after creating 15_{See appendix C.4.}

(19)

17

averages), but not perfectly. To be aware of possible outliers in the data set it is checked for where the value is greater than two standard deviations from the mean. The outliers found is probably not because of miss measurements but they rather reflect reality. The exception is social expenditure for Latvia 1995, which is the first data point Latvia has for this variable and it is zero. When creating the average for the election 1995, it is only based on this value and the value seem unrealistic because two years after (the next averaging period) it is 15% and continues to increase afterward, hence it is dropped. No other observation is dropped causing no losses of observations. There are some interesting trends to highlight. The income share for the middle 20% in the Czech Republic at the beginning of the 1980s is very high, with almost 19% of the income going to this part of the population. For Estonia 1993-2001 the trend was the opposite with an extremely low part of the income going to the middle 20%, around 13%. Luxembourg is an extreme in trade openness, which is high throughout the whole sample. Last, Slovak Republic 1992-1997 had high levels of populism, something no other country had at this time.

3.4. Specification

The four specifications that will be used for the causal four-step analysis are:

𝑙𝑛 (𝑝𝑜𝑝𝑢𝑙𝑖𝑠𝑚)𝑖𝑡= 𝛽1𝐽𝑜𝑏𝑠𝑖𝑡+ 𝜃𝑖𝑡+ ∑ (𝐵𝑁𝑘 𝑘𝐶𝑖𝑡𝑘_𝑖𝑡)+ 𝛼𝑖+ 𝜖𝑖𝑡 (2)

𝐼𝑛𝑐𝑜𝑚𝑒 𝑠ℎ𝑎𝑟𝑒 𝑚𝑖𝑑20𝑖𝑡 = 𝛽1𝐽𝑜𝑏𝑠𝑖𝑡+ 𝜃𝑖𝑡+ ∑ (𝐵𝑁𝑘 𝑘𝐶𝑖𝑡𝑘_𝑖𝑡)+ 𝛼𝑖+ 𝜖𝑖𝑡 (3) 𝑙𝑛 (𝑝𝑜𝑝𝑢𝑙𝑖𝑠𝑚)𝑖𝑡= 𝛽1𝐼𝑛𝑐𝑜𝑚𝑒 𝑠ℎ𝑎𝑟𝑒 𝑚𝑖𝑑20𝑖𝑡+ 𝛽2𝐽𝑜𝑏𝑠𝑖𝑡+ 𝜃𝑖𝑡+ ∑ (𝐵𝑁𝑘 𝑘𝐶𝑖𝑡𝑘_𝑖𝑡)+ 𝛼𝑖+ 𝜖𝑖𝑡 (4) 𝑙𝑛 (𝑝𝑜𝑝𝑢𝑙𝑖𝑠𝑚)𝑖𝑡= 𝛽1𝐽𝑜𝑏𝑠𝑖𝑡+ 𝛽2𝐼𝑛𝑐𝑜𝑚𝑒 𝑠ℎ𝑎𝑟𝑒 𝑚𝑖𝑑20𝑖𝑡+ 𝜃𝑖𝑡+ ∑ (𝐵𝑁𝑘 𝑘𝐶𝑖𝑡𝑘_𝑖𝑡)+ 𝛼𝑖+ 𝜖𝑖𝑡 (5)

C is a vector of control variables (=0 for the baselines), θ is a set of two dummies, t is time, i is the country, α are country FE, 𝜖 is the error term and parameters are 𝛽. The baseline will be one per specification except for when jobs are the independent variable and then there will be two baselines, one for manufacturing jobs, and one for industrial jobs, in total seven baselines. When there are controls added to the baseline, this will be reflected in the tables of results and in total there are 17 specifications tested. To include the controls into the specifications is helping to ‘clean’ the effect jobs or income share has on the populism and hence make sure that the effect of other determinants is not driven through the variation. Irrelevant variables might have been included, which would reduce the precision of estimated coefficients (Hill, Griffiths and Lim, 2012) but the risk for this has been minimized through a process in four parts by which the specifications have been decided. First, by finding variables according to former theory. Second, to make the set of controls constant for at least both paths of the independent variable jobs to perform mediation analysis. Third, to test for as high-adjusted R2 as possible together with as low AIC/BIC as possible. Last, to reduce the specifications to consist of maximum one variable per approximately 15 observations because of the limited degrees of freedom. The expectations for the signs of the hypotheses are based on theoretical grounds and by plotting the relations17. The plots are created to be able to see if there exist a linear relationship between the main variables. By looking at them, mixed messages are shown. Because of all the zeros in

(20)

18

the case of populism (as discussed), it is hard to interpret if there is a linear relation when predicting populism, even though the fitted line gives a reason to think so. If following the fitted line, there should be a negative relation between jobs and populism. Between income share and jobs there is a positive relation, which is more pronounced in the case of industrial jobs. Finally, between income share and populism, the scatters are widely spread, but there is the possibility of a slightly positive relation. The expectations can be seen in figure 2 together with examples of how the averages are calculated jointly with controls for the hypotheses with the assumption of elections 1996 and 2000.

Figure 2 – Example of the mediation analysis with expectations

t: 1996-1999 _{H1: ( - )} t: 2000

Jobs Populism

Controls: - China’s entry WTO - Financial crises - Social expenditure - GDP per capita change

- Trade openness - Human capital - R&D (1995-1998)

t: 1996-1999

Income share mid-20

H2: ( + ) Controls: - Jobs

- China’s entry WTO - Financial crises - GDP per capita change - R&D (1995-1998)

H3: ( - )

t: 1996-1999 t: 2000

Jobs H4: ( + ) Populism

(21)

19

4. Results

The layout of this chapter is as follows; Descriptive statistics and correlations between variables will be first up for discussion followed by the results for the specifications assigned to the different hypothesis. The outcomes of these will be interpreted and connected to former findings before the mediation analysis are concluded together with the hypotheses and a general discussion.

Table 1 – Descriptive statistics

(1) (2) (3) (4) (5)

VARIABLES mean SD minimum maximum N

Populism 7.163 8.698 0 45.20 237

Ln(populism) 1.348 1.230 0 3.811 237

Manufacturing jobs per capita 0.077 0.023 0.024 0.148 153

Industrial jobs per capita 0.118 0.026 0.049 0.195 182

Income share mid-20 0.163 0.012 0.132 0.194 229

China’s entry WTO 0.464 0.500 0 1 237

Financial crises 0.405 0.492 0 1 237

Social expenditure 20.38 4.707 9.514 33.26 222

GDP per capita change 543.2 735.7 -1,895 3,999 220

Trade openness 88.41 44.56 34.53 332.7 229

Human capital 3.001 0.389 1.645 3.723 236

R&D 1.519 0.797 0.148 3.665 219

Number of countries 26 26 26 26 26

(22)

20

Table 2 – Correlation matrix

Pop Manufac Indust Income Social GDP Trade Human Financ China R&D

Ln(populism) 1

Manufacturing

jobs per capita -0.134 1 Industrial jobs per capita -0.147 0.954 1 Income share mid-20 0.356 0.145 0.052 1 Social expenditure 0.144 -0.123 -0.223 0.218 1 GDP per capita change -0.167 0.034 -0.002 0.032 -0.170 1 Trade openness 0.106 -0.104 -0.069 -0.026 -0.161 0.156 1 Human capital 0.256 0.217 0.235 0.192 0.044 -0.032 0.256 1 Financial crises 0.028 0.054 0.123 -0.063 -0.257 -0.228 0.406 0.083 1 China's entry WTO 0.150 -0.215 -0.114 -0.160 0.078 -0.143 0.294 0.317 0.228 1 R&D 0.207 -0.010 -0.074 0.341 0.499 0.075 -0.022 0.424 -0.280 0.129 1

Names of variables are shortened on the horizontal axis and values rounded to make the table more compact.

In the correlation matrix, there is one strong correlation, which is intuitively clear beforehand, and it is between manufacturing jobs per capita and industrial jobs per capita with a correlation of 0.954. The explanation is that manufacturing is a big part of the industrial sector, but this should not be a problem since the two variables never are used in the same specification. For the rest of the variables, there are no problematic correlations between the variables.18 There is a correlation of 0.356 between income share and populism, thus it can be expected that they increase together over time to a certain limit. Overall, it is hard to predict how the relations will be in the empirical analysis from this. There should however not be any inflated variance nor standard errors in the estimator because of correlations.

The four hypotheses are going to be discussed one by one, and the empirical results will be shown together with this. To prepare for the mediation that is executed after all four regression, every step is systematically discussed after corresponding regression. The low explanation power (R2) pervading the results when measuring populism can be explained by that it is essentially human behavior that is measured, what people will vote for and the determinants of their choice. Humans are harder to predict than other phenomena and hence, the R2 is reasoned to be lower.

(23)

21

4.1 Economic globalization and populism

Table 3 – Hypothesis 1

(1) (2) (3) (4) (5) (6) (7) (8)

VARIABLES Populism Populism Populism Populism Populism Populism Populism Populism

Manufacturing jobs per capita

-15.87 0.849 -18.61 -3.884 (10.26) (18.78) (10.84) (17.85) Industrial jobs per capita -9.400* -3.206 -10.87* -5.230 (4.449) (6.751) (4.151) (5.557) China’s entry WTO 0.165 -0.0432 0.391* -0.0849 0.192 -0.218 0.446* -0.199 (0.158) (0.226) (0.153) (0.220) (0.206) (0.274) (0.180) (0.251) Financial crises 0.0891 -0.0579 0.195 0.0299 0.0859 -0.0758 0.200 0.0336 (0.182) (0.172) (0.149) (0.135) (0.180) (0.170) (0.146) (0.125) Social expenditure -0.0386 -0.0807 -0.0758 -0.0991* (0.0590) (0.0477) (0.0592) (0.0463) GDP per capita change -0.000178 -0.000143 -0.000128 -0.000121 (0.000127) (0.000107) (0.000134) (0.000112) Trade openness -0.0109 -0.0138* -0.0128 -0.0126 (0.00888) (0.00644) (0.0108) (0.00678) Human capital 2.812 3.501** 4.810 4.344*** (1.864) (0.967) (2.403) (1.026) R&D 0.539 0.507 0.200 0.319 (0.362) (0.264) (0.354) (0.258) Constant 2.558** -6.151 2.154*** -6.664* 2.634** -10.27 2.182*** -8.282** (0.868) (6.269) (0.539) (2.881) (0.915) (7.448) (0.508) (2.766) Observations 145 140 173 164 118 115 146 139 Number of countries 26 26 26 26 23 23 23 23 Adjusted R-squared 0.077 0.138 0.128 0.258 0.098 0.176 0.157 0.309

Country FE YES YES YES YES YES YES YES YES

Subsample ALL ALL ALL ALL EU EU EU EU

Robust standard errors in parentheses. Populism in logarithms. Variables averaged as discussed. *** p<0.001, ** p<0.01, * p<0.05

(24)

22

countries studied. When the amount of industrial jobs per person has increased since the last elections, the share of votes on populism should become lower since there is a negative relation. When control variables are added (4), the significance of both industrial jobs and China’s entry into WTO disappears. Trade openness and human capital become significant (5% and 1%), however, with different signs. When trade openness increase, there should be a small decrease in populism and respectively when human capital increase so should populism. In the case of EU members, the specifications with manufacturing jobs, (5) and (6) keep being non-significant, while the baseline with industrial jobs (7) have the same significant variables (5%) as the full sample. An interesting difference arises in the specification (8) where social expenditure becomes significant (5%) and with a negative relation to populism, hence when social expenditure increase, populism goes down. Human capital becomes more significant (0.1%) compared to the full sample, but with the same sign, a positive relation. If taking care of the violation of the normality assumption and decreasing the significance level to 1% the only significant variables are in (4) and (8) and in both cases, it is human capital with a positive relation to populism, stronger in the case of only EU members. In general, the variable of jobs is not significant, and neither with controls added which indicates that there is no link to populism.

(25)

23

is less of a driver for populist left-wing than the right-wing and the relatively small effect show here might be because of the inclusion of the left-wing. It is, however, hard to find a complete and plausible explanation for this reversed sign, it can be simply because of the possible violation of normality, and then it is not significant at 1%. Finally, social expenditure seems to be another predictor of populism only within EU countries with a negative relationship to populism. It is a major part of the welfare system and with a dissolving or decreased welfare state; populist voting tendencies seem to increase in order to save ‘their own’. As Rodrik (2018) argue, the European backlash has part of the roots in the concern of an eroded welfare state, which has connections to increased immigration and not trade or economic globalization particularly (Hatton, 2016) which might be an explanation in this case.

(26)

24

4.2 Economic globalization and inequality

Table 4 - Hypothesis 2 (1) (2) (3) (4) (5) (6) (7) (8) VARIABLES Income share mid-20 Income share mid-20 Income share mid-20 Income share mid-20 Income share mid-20 Income share mid-20 Income share mid-20 Income share mid-20 Manufacturin g jobs per capita 0.0738 0.102 0.0963* 0.108 (0.0447) (0.0772) (0.0455) (0.0758) Industrial jobs per capita 0.0495 0.0319 0.0536 0.0290 (0.0308) (0.0466) (0.0291) (0.0441) China’s entry WTO -0.00251 -0.00144 -0.00450** -0.00129 -0.00260 -0.00111 -0.00510** -0.000713 (0.00123) (0.00143) (0.00141) (0.00138) (0.00147) (0.00144) (0.00161) (0.00141) Financial crises 0.000182 0.000644 -0.000222 0.000813 0.000412 0.000568 -8.06e-05 0.000792 (0.00145) (0.00125) (0.00135) (0.00101) (0.00131) (0.00126) (0.00125) (0.000951) Social expenditure 0.000170 7.27e-05 0.000206 9.62e-05 (0.000208) (0.000155) (0.000196) (0.000142) GDP per capita change

-1.18e-06* -1.37e-06** -1.19e-06* -1.36e-06**

(4.67e-07) (4.12e-07) (5.26e-07) (4.51e-07)

Trade openness

-7.13e-05 -5.74e-05 -7.39e-05 -7.14e-05

(3.69e-05) (3.90e-05) (4.69e-05) (4.28e-05)

Human capital -0.00401 -0.0134* -0.00698 -0.0161* (0.00968) (0.00564) (0.0114) (0.00583) R&D 0.00171 0.00119 0.00194 0.00146 (0.00199) (0.00179) (0.00182) (0.00183) Constant 0.156*** 0.167*** 0.158*** 0.202*** 0.154*** 0.173*** 0.158*** 0.209*** (0.00396) (0.0328) (0.00420) (0.0204) (0.00401) (0.0358) (0.00406) (0.0205) Observations 145 140 170 162 118 115 143 137 Number of countries 26 26 26 26 23 23 23 23 Adjusted R-squared 0.132 0.229 0.247 0.362 0.184 0.254 0.309 0.407

Country FE YES YES YES YES YES YES YES YES

Subsample ALL ALL ALL ALL EU EU EU EU

Robust standard errors in parentheses. Variables averaged as discussed. *** p<0.001, ** p<0.01, * p<0.05

(27)

25

(5%), but manufacturing jobs are still not significant. When GDP per capita change has been positive since the last elections, the middle 20% will see a negative change of their income share. For the baseline with industrial jobs (3), the only significant variable (1%) is China’s entry into WTO, which has affected the income share negatively. When adding control variables (4), jobs are still insignificant, and China’s entry becomes insignificant. Among the newly added controls, GDP per capita change and human capital are significant (1% and 5%). When the GDP per capita change or human capital increase, then the income share goes down. In the case of EU, specification (5) shows a difference in that manufacturing jobs per capita has become significant (5%) with a positive relation to income share when there are more jobs per capita, the income share for the mid-20 increases. Specification (6) shows the same variable significant as the full sample, only GDP per capita change, with a negative relation. The same for specification (7) with only China’s entry significant and with a negative relation. Finally, for specification (8) GDP per capita change and human capital are both significant and with a negative relation to income share as for the full sample. If decreasing the significance level to 1% because of the possible violation of the normality assumption, the significance is only found in (3) and (7) where China’s entry to WTO is significant and in (4) and (8) where GDP per capita change is still significant. Generally, the variables for jobs are not significant and it does not become either with control variables added.

Only in the baseline case for EU members (5) jobs seems to have a relation with the income share of the middle 20%. However, when controls are added, this is not the case but GDP per capita change is significant in every specification negatively to income share, which can be interpreted as economic shocks seem to be the only important factor out of the factors included together with human capital and there are no special differences between the two samples. When the economic shock has been positive since the last election, then the share of income for the middle 20% should decrease. The explanation for this is probably that what increases GDP, for example, more production is not to the benefit of the middle 20%. In more advanced countries the positive changes in GDP do not come from the sectors where the middle 20% would benefit such as the manufacturing sector. The positive shocks arise where high-skilled labor works such as in technology or in more manual tasks and service jobs such as the theory of job polarization suggest and what is found in Europe (Goos, Manning and Salomons, 2009). Hence, there is more of the income going to other shares of the population and it decreases for the middle 20%. Another point to make is that economic growth also tends to correlate over time with increased inequality (Piketty and Saez, 2006) which is according to the finding. Human capital also has a negative relation to income share, which is interesting, but following expectations. An explanation can be that when education increase and mainly returns to education, more of the generated profits of companies will go to the already well paid in the top of the income distribution, which will have the consequence to concentrate wealth over time (Saez and Zucman, 2016) with a smaller share left for the middle.

(28)

26

rejected, the second step for the mediation must be evaluated. Only the baseline for EU (5) has a significant relationship between manufacturing jobs and income-share (but not significant at 1%). The relation is positive with a value of approximately 0.0963 for path a, and hence a is different from zero. This means that there is some type of relation between the variables, but it is seemingly restricted to only EU countries and not the full sample.

4.3 Inequality and populism

Table 5 - Hypothesis 3

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

VARIABLES Populism Populism Populism Populism Populism Populism Populism Populism Populism Populism

Manufacturing jobs per capita

-12.94 1.059 -14.80 -5.278 (10.22) (15.57) (11.23) (16.65) Industrial jobs per capita -4.925 -0.262 -6.351 -3.326 (5.192) (6.700) (5.272) (6.032) Income share mid-20 -6.879 -39.66** -51.83*** -35.06** -46.45*** -33.23* -39.50* -49.64* -32.79* -43.27** (12.52) (11.88) (13.22) (11.09) (11.30) (14.94) (16.34) (17.65) (14.10) (13.53) China’s entry WTO 0.241 0.0658 0.0275 0.194 0.110 0.368 0.0894 0.139 0.230 0.225 (0.171) (0.150) (0.169) (0.158) (0.202) (0.215) (0.189) (0.207) (0.195) (0.234) Financial crises 0.385* 0.0963 0.00261 0.177 0.0659 0.182 0.102 -0.00642 0.187 0.0696 (0.160) (0.184) (0.184) (0.164) (0.154) (0.182) (0.197) (0.200) (0.172) (0.163) GDP per capita change -0.000245 -0.00025* -0.000224 -0.00024* (0.00012) (0.000105) (0.00013) (0.00011) R&D 0.504 0.335 0.218 0.157 (0.374) (0.266) (0.437) (0.273) Constant 2.232 8.751*** 9.046*** 7.424*** 8.463*** 6.301* 8.708** 9.381*** 7.084** 8.364*** (2.092) (1.954) (1.559) (1.738) (1.473) (2.469) (2.501) (1.900) (2.170) (1.755) Observations 227 145 142 170 164 155 118 116 143 138 Number of countries 26 26 26 26 26 23 23 23 23 23 Adjusted R-squared 0.068 0.116 0.181 0.136 0.213 0.172 0.125 0.167 0.148 0.213

Country FE YES YES YES YES YES YES YES YES YES YES

Subsample ALL ALL ALL ALL ALL EU EU EU EU EU

Master Thesis University of Groningen Faculty of Economics and Business MSc International Economics & Business 18-06-2019

Table of Contents

(i)

Abbreviations

1. Introduction

2. Literature review

3. Data and Methodology

4. Results