Author: Bart van den Berg 10420274 July 2018
TRADE, INCOME INEQUALITY AND THE GOVERNMENT
An empirical analysis on the effect of trade globalization on income inequality in OECDcountries in 1970-2016
Master thesis Economics University of Amsterdam
Amsterdam School of Economics, MSc Economics Track: International Economics & Globalization
Word count: 14,691 Thesis Supervisor: N.J. Leefmans Second Reader: Dr. D.J.M. Veestraeten
Statement of Originality
This document is written by Student Bart van den Berg who
declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Abstract
This thesis uses a self-constructed dataset on all 35 OECD countries, spanning from 1970-2016. The effect of trade globalization on the income inequality and the effect of trade globalization on income inequality dependent on the government for OECD countries for 1970-2016 is researched. A LSDV regression with country fixed and time fixed effects is employed. Trade globalization is found to have a narrowing effect on the income distribution of low taxation OECD countries, most notably from 1988-2016. Trade globalization does decrease income inequality a year later, but not 5 years later. The decreasing effect trade globalization has on income inequality one year later is smaller, or completely diminishes for high taxation countries.
Table of Contents
1. Introduction ... 5 1.1 Introduction ... 5 1.2 Stylized facts on income inequality and trade ... 6 1.3 Research Questions ... 8 2. Literature Review ... 10 2.1 Theoretical Literature Review ... 10 2.1.1 The effect of trade on inequality ... 10 2.1.2 Influence of the effect of trade on inequality by taxes ... 12 2.2 Empirical Literature Review ... 14 3. Methodology ... 20 3.1 Econometric Model ... 20 3.2 Choice of countries ... 21 3.3 Timespan ... 21 3.4 Econometric assumptions ... 22 4. Variable construction and data sources ... 25 4.1 Income inequality ... 25 4.2 Trade openness ... 28 4.3 Power and will of the government to redistribute ... 28 4.3.1 Tax revenue as a percentage of GDP ... 28 4.3.2 Progressiveness of Tax ... 29 4.4 Technological progress ... 29 4.5 GDP per capita ... 31 4.6 Education ... 31 5. Data description ... 33 6. Results ... 37 6.1 Testing procedure applied ... 37 6.2 Introduction to all regressions ... 38 6.3 Trade on income inequality ... 41 6.4 The effect of trade on income inequality, dependent on the government ... 42 6.5 The effect of the control variables ... 44 7. Conclusion ... 45 8. Bibliography ... 47 9. Appendices ... 49 Appendix A: descriptive statistics ... 49 Appendix B: development of all variables on the country level ... 50 Appendix C: grouping of high and low taxation countries ... 531. Introduction
1.1 Introduction
Trade and its domestic effects are a hot topic in both economics and politics. In populist circles, the call for trade protectionism is omnipresent, as trade can hurt inefficient domestic industries. In the Netherlands, views like this can be heard in the PVV and FvD platform. In America, protectionist views are present even at the very top of the administration. President of the United States Donald Trump raised the import tariffs on steel and aluminium to 25% and 10% respectively just two months ago. This raise aims to support the tormented domestic steel industry. Since 2000, employment in the steel industry has fallen from 135,000 jobs to 83,600 (BBC, 2018). Critics argue that the tariffs would fail to protect American jobs and would ultimately raise prices for consumers (ibid). This raise of tariffs by Trump is merely one example of the real world consequences of (mis)conceptions about the effects of trade.
There is little question about the effects of trade on economic growth. Most economists accept that open economies on aggregate and in the long run fare better than closed economies and that relatively open policies will boost development (Winters, McCulloch, & McKay, 2004). Furthermore, research shows an increase in trade to have an overwhelmingly positive effect on economic growth (Dollar & Kraay, 2004). The effects of trade on the domestic income distribution on the other hand, are more ambiguous and no consensus on these effects has yet been reached by scholars.
Trade openness has real-life effects on citizens. As in the American steel industry, people in import competing sectors may lose their jobs due to increased competition. A person working in the American tech industry on the other hand, may see her wage rise due to higher demand for American tech from abroad. Theoretically, competitive industries cash in on trade globalization due to higher demand from abroad. The other winners are consumers, who can choose from a larger selection of goods and services, at a lower price. The domestic losers are the non-competitive industries, as these industries will be outcompeted internationally.
1.2 Stylized facts on income inequality and trade
Let us examine the stylized facts on world trade and income inequality. World trade has traditionally been fluctuating from year to year, but overall it has shown an increasing trend since the end of World War II (UNCTAD, 2013). Especially during The Great Moderation, world trade increased tremendously.
In figure 1, one can see how the sum of exports and imports as a percentage of GDP in the OECD countries has changed between 1970 and 2016 (OECD, 2018b). Trade clearly increased on average, as we would expect from trade globalization and the increase in trade as a percentage of GDP has even accelerated after 1990. However, some dips are visible, for example during the oil crises, the internet bubble and the financial crisis of ’08. Even though there are some differences between countries individually, in almost every country an increasing trend is visible. The development of trade over time for all OECD countries individually is presented in appendix B.2.
Figure 1: the average trade over GDP of all OECD countries. Compiled by the author with data from OECD (2018b)
Autor (2014, p. 18) states, that the top few percentiles of households in most developed countries have seen their share of income rise since the 1980s. Martin (2013) publishes similar results. According to him, income inequality in most OECD countries has been rising since 1975. High income countries, on average, became more unequal and this trend holds true also for traditional egalitarian countries like the Scandinavian countries. Only some very inegalitarian countries saw a drop in income inequality over the past 20 years. To see if their
40 50 60 70 80 90 100 110 1970 1980 1990 2000 2010 2020 Tr ad e as a % o f G D P Year
Average trade in OECD countries
statements match the data in this thesis, let us look at the evolution of the Gini-coefficient in this sample for 1970-20141, graphed in figure 2.
Figure 2: the mean Gini-coefficient over time for all OECD countries. Compiled by the author using data from Solt (2016)
As mentioned before, the graph takes the average income inequality of all OECD countries where data is available for. Data is not available for all countries in all years, however. For a some countries – most notably Eastern European countries, the first data are only available in 19882. For Turkey and Iceland, the first observations are around 1990. The mean first observations of these countries do not significantly deviate from the average without these countries, so adding these countries does not really alter the average of all countries.
The Gini-coefficient of all OECD countries on average fell between 1970 and 1987, but it has risen continuously on average after. Figure 2 gives us a picture of the sample as a whole, but does not necessarily provide a lot of information about separate countries. Indeed, the differences between all countries could not be bigger. For example, the income inequality in the United Kingdom and Australia has constantly been rising since 1970, but the Gini-coefficients for Greece and Turkey show a decreasing trend. 3
Concluding, it can be stated that in the first part of the time window, roughly from 1970-1987, the income inequality on average decreased while the trade as a percentage of GDP
1 Although the regressions run to 2016, the graph of the average income inequality ends at 2014
because this is the last year for which data is available for all countries.
2 These countries are Czech Republic, Estonia, Latvia, Slovenia and Slovakia.
3 See appendix B.1 for the development of the Gini-coefficient of all countries separately.
28.5 29 29.5 30 30.5 31 31.5 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 Gi ni -co ei ff ici en t Year Evolution of the Gini-coefficient over time (mean of all countries)
showed a small increase. In the second part of the sample, from 1988-2016 on the other hand, we see a clear increase in the income inequality and a faster increase in the trade to GDP ratio too. At first sight, it seems puzzling that an increase in trade over GDP is associated with a decrease in income inequality from 1970 to 1987 while an increase in trade over GDP is associated with an increase in income inequality from 1988 to 2016. However, this is without controlling for other factors that may well explain this ‘puzzling’ observation. This is why control variables are incorporated in the econometric analysis of this study. It is one of the objectives of this thesis, to provide an answer to how trade globalization influences the income inequality. The full set of objectives is elaborated on in the next section.
1.3 Research Questions
This thesis aims to contribute an answer to the question if and when trade globalization leads to an increase or decrease in income inequality within developed countries. It does so by using empirical data in an LSDV regression. It builds on earlier research in two ways. Firstly, it extends the time window of earlier research. The last observation in all of the studies mentioned in the literature review is 2008, pre-financial crisis4. My data runs until 2016. Secondly, this study focuses solely on OECD countries, instead of the developing countries earlier studies predominantly focus on. More recent data and a clear focus make this thesis’s results more relevant compared to earlier studies and more applicable to developed nations.
The first research question is as follows: What is the effect of trade globalization on income inequality in OECD countries in 1970-2016? Since we have seen in the previous section that this effect may change over time, apart from investigating the effect for the full period, the sample is also divided into two subsamples. One group runs from 1970 to 1987 and the other from 1988-2016.
A second contribution to literature is about how governments may alter this effect. If trade globalization leads to an undesired widening of the domestic income gap, governments may act to mitigate this development. The power to do so, of course depends on the size of the government in terms of money. This thesis will be one of the first studies, after Milanovic (2005) and Lim and McNelis (2014), in which the government is not regarded as a sitting duck, but can actively take action to diminish the effects trade globalization has on income inequality. It will be the first study to include government interaction in such research on developed
countries. Obviously, the more financial resources a government has, the greater the opportunities are to revert the effect of trade openness on income inequality. The accompanying research question is: How does the tax base as a percentage of GDP influence the effect of trade globalization to income inequality in OECD countries in 1970-2016?
These two research questions will be answered in the following structure. Chapter two covers the literature review. I start with the theoretical literature review in which the leading theories on the relationship between trade and income inequality is described. Subsequently, I highlight earlier research and its results on the topic. In chapter three, I set out the methodology used: the choice of countries, the time window of the research, and the econometric model and its assumptions. Thereafter, chapter four introduces all used variables: how the variables are constructed and what the data sources are. Chapter five elaborates on the dataset with the first descriptive statistics. I present the procedure of testing as well as the results in chapter six. Chapter 7 offers a conclusion.
2. Literature Review
The literature review is divided into two sections. Section 2.1 contains the theoretical
literature review, where the leading economic theories on trade and inequality, and the way in which governments can change this effect are discussed. Section 2.2 consists of the empirical literature review.
2.1 Theoretical Literature Review 2.1.1 The effect of trade on inequality
The Heckscher-Ohlin model is the classical model to reason the effects of trade globalization on the domestic income distribution. The model assumes a two factor world – capital and labour – and constant returns to scale. Every country is abundant in either capital or labour. Furthermore, production factors are mobile between sectors but immobile across borders. Herein, the Heckscher-Ohlin theorem states that in a multiple country case “countries export goods whose production is intensive in factors with which the countries are abundantly endowed’’ (Krugman, Obstfeld, & Melitz, 2012, p. 121).
Within this framework, the Stolper-Samuelson theorem makes a prediction about the effects of trade on the domestic income distribution. According to this theorem, trade benefits the reward of the country’s abundant production factor and hurts the reward of the country’s scarce production factor. The mechanism at work here is that due to international trade the relative price5 of goods converges. In the capital abundant countries, before opening up to trade, capital intensive goods are relatively cheap compared to labour intensive goods because of the high availability of capital and the low availability of labour. In labour abundant countries this relative price is obviously the other way around6.
With the possibility of international trade, it is now feasible to export capital intensive goods to the labour abundant country, where the price of capital intensive goods is higher. The increased supply of capital in the labour abundant country decreases the price for the capital intensive good, but the decreased supply in the capital abundant country – part is now exported – increases the price of capital. The relative price of capital intensive goods will thus converge: It will increase in the capital abundant country and decrease in the labour abundant country.
5 The relative price of capital intensive goods is defined as !"#$% '( $)*#+), #-+%-.#/% 0''1.
!"#$% '( ,)2'3" #-+%-. /% 0''1..
For the price of labour intensive goods, again, the same happens but the other way around (Krugman, Obstfeld, & Melitz, p. 121).
Since the OECD countries are predominantly highly developed, the abundant factor in these countries is capital. This means that the capital owners will gain from trade liberalization, while the labourers will lose. Given that the capital owners had a higher income to begin with, trade globalization is expected to increase the income inequality in OECD countries.
The Heckscher-Ohlin framework makes a distinction between labour and capital, but it assumes labour to be homogeneous. In reality, there are vast differences in labour in terms of skill, wage and complementarity to capital. Haskel et al. (2012) sled the HO framework to incorporate varying amounts of talent to allow for capital-talent complementarity. This means that talented workers are more productive working with capital, but no more productive in unskilled tasks than the untalented (idem, p. 93) and wage inequality stems from this talent-capital complementarity.
In the capital abundant country case, an increased openness to trade will lead to a rise in the relative price of capital. According to Haskel et al. (idem, pp. 94-95), the wage of the untalented workers in the labour intensive sector falls, just like in the HO framework. The wage of the medium talented workers falls as well: the price of the capital they have to work with increases and they are not talented enough to command higher wages. The highly talented workers are the ones who gain from the increased price of capital. The increased price of capital is completely offset by the favourable productivity effect (ibid).
Thus, contrary to the classical HO framework, it is not the case that the wages of all workers fall due to an increase in the relative price of capital. The most talented workers will see their wages increase, but the wages of the average and untalented workers fall. Compared to the classical HO framework, not only capital owners will reap the benefits of trade liberalization, but skilled labourers do too. Given that both capital owners and highly talented individuals initially earn a higher income than unskilled labourers, according to the extension of the framework by Haskel et al. (2012), trade globalization will worsen the income gap, just like in the classical HO framework, in capital abundant countries like the OECD countries.
The Hecksher-Ohlin theorem and the extension by Haskel et al (2012) lack government intervention. Even though a change in the income distribution as a consequence of trade is predicted, the average income is expected to rise. Ergo, with the right policy measures, every individual can be made better off. It depends on the political power and will to redistribute income to determine whether or not everyone will actually be better off (Krugman, Obstfeld,
& Melitz, 2012, p. 122). For this reason, in this thesis a variable that measures the power and willingness of the government to redistribute will be included.
2.1.2 Influence of the effect of trade on inequality by taxes
Governments differ in size and ideology. Some governments heavily prefer income redistribution, while other governments would rather leave the income distribution practically unchanged. This different belief can of course influence how trade globalization affects income inequality. For this reason, it is important to incorporate the wishes of a government with respect to income redistribution. How these abstract wishes can be included will be discussed in the next part.
In the 1960s and 70s, governments employed widely different economic growth strategies. From the 1980s onwards however, a remarkable convergence of views developed around a set of policy principles aimed at boosting economic growth. This set of policies, called the Washington Consensus (WC) promoted free trade, fiscal discipline, competitive currencies, financial liberalization, privatization and deregulation (Rodrik, 2004, pp. 4-5). This trade-oriented thinking has lifted many people out of poverty, but it also left the vulnerable who are not lifted out weak, as it was accompanied by a contracting government.
The WC principles are still at the heart of our economic thinking, but over the years the desired set of policies for economic growth has expanded by a number of new desired policy measures. These measures include labour market regulations, social policies and anti-poverty programs to protect the poor from unbridled capitalism. The ‘augmented’ WC for promoting economic growth, thus is aimed at more pro poor growth.7
Figure 3: a graphical presentation of the one- dimensional original-augmented Washington Consensus Scale
Figure 3 shows the one-dimensional scale of choosing between the original and the augmented WC. This straight line does not depict any full combination of policy measures possible. Every set of policies is unique and some countries – for example South Korea and Taiwan in the 1980s – have successfully deviated from the consensus. But the Washington Consensus is widely accepted in mainstream thinking as the strategy to maximize economic growth, so that every country can roughly be located somewhere on the scale.
On what point on the one-dimensional original to augmented WC line a country is located depends on the preferences and the will of the population to protect the vulnerable. The further on the augmented side of the scale a country is located, the more active its government needs to be and the higher the government’s budget needs to be, as it is actively trying to protect the poor from vulnerability. Illustrating the effect of tax income on income inequality, Autor (2014, p. 3) names the multiple reductions in the top federal marginal tax rates in the U.S. as one of the reasons for a growing income inequality. This gives the government less power to level the income distribution and, on top of that, the wealthy part of the population pays less tax than it did before. Both developments widen the income gap.
The preference of the policymakers with respect to the possible spectrum depicted in figure 3 and the corresponding institutions in a country ultimately determine to what degree the poor will be protected against market forces like trade globalization. A country more on the original side will tax their citizens less, as it has less desire to redistribute. More on the augmented side, we will see countries with higher taxation and a higher degree of redistribution. The preferences of the government and institutions in the country determine to what degree the government policies affect changes in income inequality. Consequently, it determines what the net effect is of trade globalization on the income inequality.
2.2 Empirical Literature Review
In the following section, the methodology and results of 9 different papers on trade globalization and income inequality will be summarized. All papers but Lim & McNelis (2014) use unbalanced data sets, due to constraints to data availability. The vast majority of papers takes 1960 as the first year of observation, only Milanovic (2005) starts in 1988, using merely three years of observation (1988, 1993 and 1998), and Spilimbergo et al. (1999) start their observations in 1965. Lim & McNelis (2014) start in 1992. In this thesis, due to data constraints on trade and income inequality, the first year of observation will be 1970. The measured time period will run until 2016. This is a substantial extension of the mentioned papers in this chapter, as most of those papers do not extend their research any further than 1990.
In terms of measurement of income inequality, Lundberg & Squire (2003), Ravallion (2001), Barro (2000), Spilimbergo et al. (1999), Milanovic (2005) and Li et al. (1998) all use the Gini-coefficient database as constructed by Deininger & Squire (1996). This database is a collection of all available Gini-coefficients. The measurements differ in methodology and reliability. All aforementioned papers make their own distinction in what data to use from this database. Lim & McNelis (2014) use the Gini-coefficient but do not name a source. Dollar & Kraay (2002) use the income share of the bottom quintile, employing the WIDER database – the world income inequality database of the UN. Like the Deininger & Squire (1996) database, this database combines income inequality indicators from multiple sources, differing in quality and methodology. Lall et al. (2007) use a self-constructed database on income inequality. This database is not publically available.
Trade globalization is also operationalized in a number of different ways. The most widely used measures are export as a share of GDP (Ravallion, 2001; Li et al., 1998; Lall et al., 2007) and import plus export as a share of GDP (Milanovic, 2005; Dollar & Kraay, 2002; Barro, 2000; Lim & McNelis, 2014). Lundberg & Squire (2003) and Spilimbergo et al. (1998) use their own measure of trade openness, which adjusts for natural resource endowments. Even within these groups there are differences: for example, Dollar & Kraay (2002) use PPP terms while Ravallion (2001) and Li et al. (1998) convert everything to current dollars. Furthermore, Barro (2000) adjusts for physical country size. In most papers, different measures for liberalization – like the degree in which tariffs and capital controls are present – are added to increase robustness.
Most authors use similar control variables. The most used control variables are technological development, education, economic growth and current GDP per capita, controlled for PPP. The exact control variable used, and what indicator they are proxying, will be discussed later in this section where every paper is discussed in more detail.
Technological development is, according to Lall et al. (2007) the main driver of the increase in income inequality and therefore it has to be controlled for. The mechanism by which technological progress increases income inequality is that technology is inherently skill biased and technological change thus leads to an increase in the skill premium, increasing income inequality (idem, p. 45). Education is reported to be important because a well-educated population is better capable of switching from an outcompeted sector to a competitive sector. It is therefore better able to absorb shocks on the labour market. For a constant level of technology, better access to education would be expected to reduce income inequality as it allows a higher share of the population to engage in high-skilled labour and therefore reduces the amount of people with the lowest income because of their incapability to work with advanced technology (idem, p. 47).
GDP per capita is often used as a control variable, because it reflects the initial level of development of a country and the income inequality is thought to depend on the initial wealth in a country. Richer countries in terms of GDP per capita on average experience lower income inequality (The Guardian, 2017).
Milanovic (2005) and Barro (2000) use an indicator for democracy on the assumption that an increase in inequality is more easily reversed by the government in democracies. When inequality is high, this means there is a large number of people earning only a limited share of the nationwide income. As the median voter hypothesis predicts that politicians will choose the policy with which they can win the most votes, the larger the share of people earning less than the mean income (this is associated with a high Gini coefficient), the more votes a politician will win with redistributing policies. Milanovic (2005, p. 33) finds that democracy positively affects the income shares of the middle deciles.. Since there is little fluctuation in the level of democracy in OECD countries after 1970, it does not make sense to include this variable in this thesis. Barro (2000) furthermore uses an indicator for the rule of law because a better rule of law leads to more economic growth and more investments and both economic growth and investment ratios could affect the income inequality.
Most papers use more or less the same research strategy. Barro (2000) does different OLS regressions on panel data. His regressions run to 1990 and are classified on a 10-year interval. In some of the regressions he uses country fixed effects, in others he does not. For his
results – he has both developed and developing countries in his dataset and he also uses region dummies – it does not make much difference. The controls are government consumption as a percentage of GDP, a rule-of-law index, a democracy index, and primary, secondary and higher schooling. He finds a pro inequality effect of trade openness, which is most pronounced in poor countries. Furthermore, a more widespread access to education leads to lower inequality, as does a better rule of law and more democracy.
Lundberg and Squire (2003) use numerous different model specifications. They start their structural models with the OLS Seemingly Unrelated Regression Equations (SURE) model. Subsequently, they do an IV estimation with 3 Stage Least Squares (3SLS). They then proceed to their expanded models, the Quasi-reduced-form models. They use a Keane and Runkle 3 Stages Least Squares (KR-3SLS) IV estimation for this. This kind of model is used for panel data model with serial correlation when instruments are not strictly exogenous (Keane & Runkle, 1992). The reason they use such complicated models is that they want to jointly explain the effect of growth on inequality and vice versa. They find a general mild pro-inequality effect of trade of income pro-inequality, they furthermore note that a 10% improvement in their trade openness ratio increases the growth rate by 10% and worsens income inequality by only 1% in the short run. What will happen in the long run depends on which groups bear the adjustment costs (Lundberg & Squire, 2003, pp. 339-340). This ultimately depends on the government policy pursued, they do however no further testing t for the impact of different kinds of government policies.
Dollar and Kraay (2002) use a more straightforward approach of first differencing the natural logarithm of the income share of the first quintile and first differencing the mean income. subsequently, they add the sum of import and export over GDP to see if this ensures the same growth rates for the poor and the average citizen. They do this using OLS, 2SLS and a GMM estimator. Their results are insignificant, so they conclude they cannot reject the null hypothesis that an increase in trade leads to as much growth in income to the poor as it does to the rest of society.
For Ravallion (2001, p. 1181), the effect of trade openness on inequality is only of oblique importance to his research. For this reason, he only conducts a simple OLS regression of trade openness on income inequality, controlling for schooling, financial sector development, urbanization and the black market premium. He finds no direct effect, but he does find a strong negative interaction effect with initial GDP per capita. This implies that openness leads to higher inequality, but only in poor countries. This finding is at right angles to the
Stolper-Samuelson thesis that predicts a lower income inequality due to an increase of trade in labour abundant (poor) countries.
Spilimbergo et al. (1999) perform an OLS regression on panel data for both developed and developing countries. I will take the same approach in this thesis, but I will also control for time and country fixed effects. A second difference is that they include the richness of skilled and unskilled labour and the degree of capital available in a country. They subsequently interact these with their measure for trade openness. Their only additional control variables are GDP per capita and GDP per capita squared to test for the existence of the Kuznets curve. This is the third difference in methodology to this thesis, as I will incorporate a number of extra control variables. They find trade to have a pro-equality effect in capital rich countries. For robustness tests, Spilimbergo et al. (1999) repeat their procedure with the Gini coefficient as the dependent variable substituted for quintiles. This does not influence their results significantly. Their findings do not support the Stolper-Samuelson thesis, as this predicts trade to make the income distribution more unequal in capital abundant countries.
Li et al. (1998) use a standard panel data OLS regression with an AR (1) error specification to correct for serial correlation. Their independent variables include the initial mean years of schooling, the civic liberty index, the Gini coefficient for the distribution of land and a measure of financial development. Because the measure of financial development may suffer from endogeneity, they reestimate the base regression with IV using the lagged variables as instruments. The IV estimates are similar to the OLS and AR (1) ones. For their sensitivity analysis, they stepwise add a number of extra control variables: real GDP per capita, gross domestic investment ratio, urbanisation ratio, black market premium, terms of trade shocks, openness and per capita arable area. Their findings for trade openness are insignificant.
Lall et al. (2007) use an OLS panel data model, much like the model I will use. Contrary to this thesis, they are not just interested in the effect of trade globalization on income inequality, but also in the effect of financial globalization on inequality. Their independent variable is the natural logarithm of the Gini coefficient. They operationalize trade openness in a number of ways, of which the export to GDP ratio is the most prominent. The others are the sum of export and import divided by GDP and 100 minus the tariff rate. They make use of relatively many control variables. Although most other papers on the topic just use one variable to measure the level of education of the population, Lall et al. (2007) use two: population share with at least secondary education and average years of education. They find a better level of education to lead to lower income inequality. Their other two main control variables are technological progress, operationalized as the share of information and communications technology (ICT) in
the total capital stock and the level of development of the financial sector, measured by the credit to the private sector as a percentage of GDP. They find that most of the increase in income inequality can be explained by technological progress while the effect of the deepening of financial markets depends on the quality of institutions in a given country (idem, p. 48). Like Spilimbergo et al. (1999), the empirical model was also estimated using the income shares of the five quintiles instead of the Gini coefficient. They conclude that in advanced economies, trade globalization has led to lower income inequality. In developing countries, the rise in agricultural exports and tariff liberalization have contributed to an improvement in the income distribution.
Milanovic (2005) uses the share of income of the different deciles as his dependent variable.. He has two main control variables: the level of financial depth of M2 over GDP and an indicator for democracy. Further controls are added to increase robustness. These controls are government spending as a percentage of GDP and the real interest rate. Furthermore, controls for regional effects are added. In total, there are 129 countries and three years (1988, 1993 and 1998) Milanovic (2005) has data for. The model run in his research is the instrumental variable GMM estimation, where the lagged values of the explanatory variables are used as instruments (Milanovic, 2005, p. 30). He finds that increased openness reduces the income share of the bottom six deciles. The negative effect of openness is smaller in richer countries (Milanovic, 2005, p. 31). He furthermore finds that a 10% increase in the ratio of the government expenditure to GDP raises the bottom decile’s income share by 0.24%, which is about one-tenth of what the bottom decile receives on average (Milanovic, 2005, p. 33).
Lim and McNelis (2014) empirically test the relationship between income inequality, economic growth and openness based on an annual data set of 42 (developing) countries for the period 1992-2007. They use OLS panel estimates and they control for both country-specific and time-specific effects, the same strategy I will use in this thesis. The only overlapping countries with this thesis are Mexico and Turkey, however. Their main question is if economic growth is more likely to reduce income inequality in countries which are relatively open. The set of countries they perform their research on is divided into low income, lower-middle income and upper-middle income (emerging markets). High income countries are not incorporated in their research. They include GDP growth rates, relative per-capita GDP to world GPD per capita and the government spending to GDP ratio as control variables.
They find that trade openness has an inappreciable effect on the income inequality in the low income countries, but it has a significant income levelling effect for middle income
countries (Lim & McNelis, 2014, pp. 9-10). As noted before, Lim & McNelis do not perform any research on the more developed countries of which the OECD predominantly consists.
Furthermore, they report that higher government spending increases income inequality. This may seem puzzling, but they explain that in lower income countries public spending programs have much greater political lobbying and clientism, thus reducing the redistributive effect. Since the research conducted in this thesis is mostly on developed countries, it is unlikely that the same mechanism of government spending on income inequality is in play.
In conclusion, it can be stated that different authors publish widely different results due to differences in methodology, the time window of observations and the choice of countries. The standard economic theory of the Stolper-Samuelson thesis is not confirmed by earlier studies. Furthermore, the interaction between the effect of trade on inequality and the size and willingness of the government to interfere with this remains unclear.
3. Methodology
3.1 Econometric Model
For this thesis, a panel data Least Squares Dummies Variables (LSDV) regression will be conducted on an unbalanced data set of all current OECD countries for 1970-2016. The Gini-coefficient, representing the income inequality in a given country in a given year, will be regressed on trade as a percentage of GDP, taxation as a percentage of GDP, an interaction term between trade and taxation, the total factor productivity, the access to tertiary education and the GDP per capita. The following base model will be used:
4565#+= 89+ 8;∗ 5=>?@A + BC>?@A 4DE #++ 8F∗ GHC 4DE#++ 8I ∗ 5=>?@A + BC>?@A
4DE ∗ JK?LB HLB@HMB AHCHA5?6 NO==P #++ 8Q∗ GRE#++ 8S ∗ HTTBUU A? AB@A5H@P BNOTHA5?6#++ 8V∗ ln (4DE>T)#++[#+ \++ ]#+
The research will control for country fixed effects and time fixed effects, like in Lim & McNelis (2014). Time fixed effects control for variables that are constant across countries, but evolve over time (Stock & Watson, 2015, pp. 407-408). They will be incorporated to control, inter alia, for a change in the belief in the way society should be organised. In the 1980s, the general belief changed to having a more market driven society, with higher corresponding inequality. Cross-section wide changes in ideology like this will be captured in the time fixed effects term. Other major worldwide events, like the oil crises in the 70s, and the financial crisis of 2008, will also be captured by using time fixed effects. In the model, the time fixed effects are represented by \+.
Country specific effects are represented by [# in the regression equation. It allows us to control for variables that differ from one country to another, but are constant over time (Stock & Watson, 2015, p. 403). One can think of differences in culture and tradition that do not – or hardly – change over time but are different from one country to another and lead to a different preference of the desired income inequality.
With the inclusion of both country and time fixed effects, the estimated relationship between trade openness and income inequality is immune to omitted variable bias from variables that are constant either over time or across countries. However, many important determinants of income inequality do not fall into this category, which is why all the other control variables in the model are added.
3.2 Choice of countries
As noted before, research in this thesis will take place on the OECD countries. The Organisation for Economic Co-operation and Development (OECD) was founded on September 30th, 1961 and aims at promoting policies that will improve the economic and social well-being of people around the world. As of right now it has 35 members. Most of the member countries are advanced economies, but it consists of emerging markets like Mexico, Chile and Turkey too (OECD, 2017).
One of the main objectives of the OECD is to gather and analyse data of its member countries. This results in better data availability of indicators on OECD countries than on other countries. For this reason, the group of countries in the OECD is an attractive group of countries to do research on. Because not all countries are founding members of the OECD and more than 10 countries joined the OECD after 1970, which is the start of the time period researched in this thesis, data on all the variables used, is not always complete for all countries.
There is a number of countries interested in OECD membership and the OECD is likely to expand further in the coming years. For example, Colombia and Lithuania have already signed the accession agreement May 30th 2018 to become full members on. After ratification, these countries will become full members (Gurria, 2018). For this thesis however, only the current members will be incorporated in the analysis. Ergo, research will be conducted on the current 35 member states in this thesis.
3.3 Timespan
In this thesis, the time period 1970-2016 is used. In order to yield the most accurate result, it is important to have as many data points as possible for OECD countries. Due to constraints to data availability to one of the most pivotal variables employed – the trade openness variable runs back only to 1970 –1970 is taken as the first year of observation. Since the last available year of the Gini-coefficient using the Luxembourg standard is 2016, the last year of observation in this study is 2016. There is no ready to use database. Instead, a database is created combining several data sources.
Even though the time window 1970-2016 will be used, this does not mean there is data on all countries for all years. For all founding OECD members, data is well recorded back to 1970. For South-American and Eastern European countries however, a complete time series for 1970-2016 is harder to find. The dataset consists of 1,160 observations in which there is
data for all variables. If the dataset was perfectly balanced, there would have been 1,6458 observations.
Data availability is predominantly limited for the eastern European countries who joined the OECD from the 1990s. For Czech Republic, Estonia, Hungary, Latvia, Poland, Slovakia and Slovenia, the earliest available data is typically after 19909. Hence, these countries are only incorporated in the regression from 1990 onwards, even if the regression starts in 1970 for other countries.
3.4 Econometric assumptions
For the models used in this thesis to have efficient and consistent coefficients, a number of assumptions need to be fulfilled. Clustered standard errors are used in all regressions. These standard errors are robust to both heteroscedasticity and to correlation over time within an entity and are called heteroscedasticity- and autocorrelation-consistent (HAC) standard errors (Stock & Watson, 2015, p. 413).
The assumptions of the model are extended with two assumptions for fixed effects regression, compared to standard the OLS assumptions. The standard assumptions and the extensions are listed in the table below.
Table 1: the OLS and extended OLS assumptions for panel data regressions
The standard OLS assumptions (Stock & Watson, 2015, p. 247)
1. The error term has conditional mean zero given all regressors
2. All regressors are independently and identically distributed draws from their joint distribution
3. Large outliers for the regressors are unlikely 4. There is no perfect multicollinearity
The added assumptions for fixed effects regression (Stock & Watson, 2015, p. 412)
5. The error term has conditional mean zero, given all T values of X for that entity 6. All regressors are i.i.d. draws from their joint distribution
8 The time span consists of 47 years and there are 35 countries. 35*47=1,645.
Assumption 5 plays the same role as assumption 1, and implies there is no omitted variable bias. It is important to further note that the error term may not depend on any of the values of X for that entity, that is past, present or future values. Omitted variables stem from theory and it is something that cannot really be tested for. Variables predicted by theory to be of importance are incorporated. The control variables used in this thesis do not deviate considerably from the controls used in earlier research for developed countries. Furthermore, using time and country fixed effects prevents for omitted variable bias for variables constant over time or across countries. This assumption is therefore adhered to in the best way possible. Assumption 6 is the fixed effects equivalent of assumption 2. It holds if entities are selected by simple random sampling for the population. Under assumption 6, regressors are allowed to be correlated over time within an entity. Ergo, autocorrelation is allowed for. However, the error needs to be uncorrelated across entities. In this thesis, the entire population – all countries the OECD consists of – is taken instead of a random sample. This of course implies the same, as this ‘sample’ is highly representative for the population and that this assumption can be considered as fulfilled.
Assumption 3 and 4 remain the exact same for regressions with fixed effects. Under the least squares assumptions for panel data, the fixed effects estimator is consistent and is normally distributed when n is large. An n of at least and at most 1,160 should to suffice this condition at first glance. Using the Shapiro-Wilk test for normality, all regressors but the dummy variable proved to be normally distributed over the whole time window, as well as over the two separate time windows of 1970-1987 and 1988-2016. Furthermore, the error term is normally distributed.
The data does not suffer from multicollinearity, as variables containing multicollinearity are automatically omitted by Stata. The data can however suffer from near multicollinearity, which inflates the standard errors. An often used test to check for near multicollinearity is the Variance Inflation Factor (VIF) test. For fixed effects models however, the reported VIF scores are often extremely large (Jacobs, 2005). Since multicollinearity means that one predictor can be linearly predicted by another predictor – the predictors measure the same – we can use the correlation coefficient as an alternative to check for multicollinearity. Correlating the explanatory variables used on one another, only the first and fifth lagged value of trade over GDP have high correlation (>0.95). This makes sense, as trade does not change very suddenly. Because theory suggest there might be different effects for each, both were left in the regression. Of the other variables, the highest correlation is 0.38 between the natural
logarithm of GDP per capita and trade as a percentage of GDP. Hence, there is no near multicollinearity among those variables and assumption 4 is fulfilled as well as possible.
4. Variable construction and data sources
In this chapter, all the indicators incorporated in the research will be discussed. The dependent indicator is on income inequality. The explanatory indicators are trade openness and the power and willingness of the government to redistribute. The control indicators are technological progress, GDP per capita and education. For every indicator I will explain why it is important to include this in the research, how the variable is created from the indicator that proxies it and what the data sources are.
4.1 Income inequality
The dependent indicator in this research is income inequality. The first way in which income inequality is operationalized in this study is through the Gini-coefficient for disposable income. This is post-tax and post-transfer income, and this is what best reflects the real life consequences of changes in income.
The Gini coefficient is a measure for inequality developed by the Italian statistician and sociologist Corrado Gini in 1912 (Investopedia, 2016). It can be presented graphically through a Lorenz curve, as depicted in graph 4. In a Lorenz Curve the cumulative share of people from lowest to higher incomes is on the x-axis, and the actual share of income earned is displayed on the y-axis. The 45-degree line is called the line of equality since this would mean that the poorest 20% of the people earn 20% of the income and the 20% richest people also earn 20% of the income, so that the income is distributed totally evenly. The farther the line that represents the actual income distribution is away from the line of equality, the higher the income inequality is. The Gini-coefficient is calculated as the surface between the line of equality and the actual distribution (A), divided by the total surface under the line of equality (A+B). If there is total equality, the actual distribution equals the line of equality so that A=0 and the Gini-coefficient is 0. If one person earns all the available income in the country, A=A+B, so that the Gini-coefficient is 1. Logically, the value of the Gini-coefficient will always be between 0 and 1 and the higher the coefficient, the larger the inequality of income.
In this research, the value is multiplied by 100 so that the Gini-coefficient can be seen as the percentage of inequality.
When comparing Gini-coefficients internationally, a number of problems arise. Some data can be very unreliable due to vast measurement errors. Even if the data does not suffer from measurement error, it might still be hard to compare Gini-coefficients between different countries as there is no consensus on the way in which the Gini-coefficient needs to be calculated. Worldwide there are differences in the usage of household level or the personal level data, the usage of before tax or after tax income data and the usage of market income or consumption data.
To circumvent these difficulties, the Standardized World Income Inequality Database (SWIDD) by Solt (2016) will be used in this study. This database seeks to maximize comparability between countries, while still providing the broadest possible coverage of countries and years. Solt (2016) uses the Limburg Income Study (LIS)10 way of measuring as his standard. The LIS uses disposable income on the household level (Solt, 2016, p. 7). Data from all available other sources, like the OECD, World Bank, Eurostat, Deininger & Squire (1996) and Milanovic (2005), are subsequently converted to the LIS standard using model generated estimates11 to maximize comparability of the data. According to Solt (2016), in terms of coverage and comparability, the SWIDD is better suited for cross-national research on income inequality than any other available source.
10 The LIS is a non-profit organization whose mission it is to enable, facilitate, promote, and conduct
cross-national comparative research on socio-economic outcomes and on the institutional factors that shape those outcomes. One of the variables they collect data on is the Gini-coefficient (LIS, 2018).
Since the OECD themselves gather data on the Gini-coefficient, one might ask why this data is not primarily used in a thesis that is exclusively about OECD countries, as this data is likely to be highly comparable among countries. The answer is simple: the disadvantage of using just the OECD as source, is that for most countries there are only recent observations; that there are only 307 data points available. For comparison, the SWIDD reports over 1500 data points for OECD countries. The correlation between both variables will be taken to check for comparability.
Many of the studies mentioned in the literature review use Deininger & Squire (1996) for their cross country research of trade on income inequality. There are a number of reasons why this database is inferior to SWIDD. Firstly, Deininger & Squire (1996) gather all data they can find, but not all are of the same quality and some parts of the data can be very unreliable. Secondly, the data is quite outdated as it only spans up to 1994. The SWIDD’s first edition was launched in 2008 and it has been updated since every year. Thirdly, unlike Solt (2016), Deininger & Squire (1996) do not standardize their data, so not all data is measured on the same level and hence, coefficients might be inconsistent.
Even though the Gini-coefficient is the most widely used measure for income inequality, there are some shortcomings to this approach. Firstly, the measure depends on reliable GDP and income data. Shadow economies are not incorporated in the Gini-coefficient. If the size of the shadow economy differs per country, the Gini-coefficient does not give a reliable image. For OECD countries however, this is not expected to pose a problem since their shadow economies are relatively small. A second shortcoming is that two very different distributions of income (or Lorenz curves) can result for the same Gini-coefficient. Two different actual distributions may give the same surface. When the Gini-coefficient is used, there is no way of setting two countries with a different income distribution but with the same Gini-coefficient apart. Nevertheless, the Gini-coefficient remains to be the most important measure of domestic income inequality. In this regard, it is common practice in the literature to use the Gini-coefficient.
4.2 Trade openness
The main independent variable in this thesis is trade openness. All trade openness data is retrieved from OECD Data (OECD, 2018b). In line with earlier literature, this variable will be operationalized in a number of ways for robustness checks. The first way is by taking the sum of import and export in goods and services as percentage of GDP: trade over GDP. A country that trades more with other countries is more open and shows a higher percentage of the sum of import and export over GDP.
A number of the papers mentioned in the literature review, like Lall et al. (2007) reported this first way of operationalizing to be insignificant. Instead, they used only export in goods and services as a percentage of GDP. This is the second way in which the variable is operationalized. A third way is to look exclusively at import as a percentage of GDP. The differences between using import, export and total trade will be discussed in the results section.
4.3 Power and will of the government to redistribute
Of the papers discussed in the empirical literature review, only Milanovic (2005) and Lim & McNelis (2014) use a government variable. There are two reasons other papers do not incorporate government variables. The first reason these variables are excluded, is that most papers focus on either developing countries or on both developing and developed nations and comprehensive data on government variables is not available for developing countries (Lall et al., 2007, 49). A second reason government variables are not included is that the effect of the government policy on the income distribution is not important for their specific research. Since this thesis focuses on OECD countries and more reliable data for these countries is available, variables on the government can be included. Furthermore, one of the main objectives of this research is to distinguish how governments may influence how trade globalization affects the income distribution. For this reason, the tax revenue as a share of GDP and the progressiveness of tax will be used as variables to measure the will and power of the government to redistribute income. These indicators are elaborated on below in 4.3.1 and 4.3.2 respectively.
4.3.1 Tax revenue as a percentage of GDP
The first government variable used to measure the ability and willingness of governments to equalize the income distribution if trade indeed increases the income inequality, is the total tax received by the government as a percentage of GDP. Tax revenue is defined as “the revenues
collected from taxes on income and profits, social security contributions, taxes levied on goods and services, payroll taxes, taxes on the ownership and transfer of property, and other taxes.” Total tax revenue as a percentage of GDP indicates the share of a country's output that is collected by the government through taxes (OECD, 2018b). All data on this variable is retrieved from OECD Stat12. The tax revenue as a percentage of GDP is very similar to the variable used by Milanovic (2005) and Lim & McNelis (2014) – government expenditure as a percentage of GDP – and should be the same over an infinite time horizon due to the No-Ponzi constraint on the government budget.
Korpi and Palme (1998, p. 676) find that a higher relative size of the government budget leads to more redistributive policies by the government and ultimately to a more equal distribution of income. For this reason, taxation as a percentage of GDP is a good proxy for the willingness of the government to redistribute income and to thus protect the poor from the harm trade globalisation may inflict.
4.3.2 Progressiveness of Tax
A measure of how much the government exactly redistributes is given by the progressiveness of tax. That is, the tax rate in the highest bracket of the income tax. A higher marginal top rate indicates the taxes are more progressive and the government thus redistributes more. Autor (2014) already proved this variable to be of significance for the rising inequality in the U.S. It serves as a second measure for the power and willingness of the government to redistribute income, next to taxes as a percentage of GPD.
Data for this variable has only been measured by OECD (2018a) since 2000. For this reason, regressions incorporating this data will only run from 2000 to 2018. This variable will therefore mostly be used for robustness checks, next to the taxes to GDP ratio.
4.4 Technological progress
As mentioned in the literature review, according to Lall et al. (2007) the main driver for the increase in income inequality over the last 30 years has been technological progress.To estimate the effect of trade on income inequality, it is therefore important to control for technological progress. Lall et al. (2007) used the share of ICT capital in total capital. Unfortunately, the data they used on this indicator is not publicly available. As a different way
to measure technological progress, the Total Factor Productivity (TFP) – or multifactor productivity (MFP) – can be used. According to Ray (1998, pp. 117-123), total factor productivity is an accurate measure for technological progress. Total factor productivity growth can be measured by the difference between actual output growth and the predicted output growth by the inputs capital and labour, which is exactly the way TFP is measured in this thesis. TFP is defined as reflecting the overall efficiency with which labour and capital inputs are used together in the production process.
The source used includes all OECD countries. It is retrieved from the Total Economy Database (TED) by The Conference Board (2018). The raw data is not ready to use. Change in TFP is constructed by taking the percentage growth in GDP for a given country in a given year and subtracting the contribution of both labour and capital to economic growth, so that only the GDP growth unexplained by labour and capital – the residual – is left. This residual is the change in TFP and represents the increase in GDP due to technological progress. Subsequently, the change in TFP is converted to an indexed level variable where 2016 is standardized at 100. Even though TFP is one of the most widely used estimators for technological progress, the approach has some shortcomings. Firstly, it is hard to obtain reliable data on all the variables needed. A variable like GDP growth is relatively easy to measure, but one also needs to reliably measure the quantity of labour, the quality of labour and the quantity of capital (Diewert, 2000). While an increase in labour statistically can be divided into growth in quantity and quality, such a distinction does not take place in the change of the capital stock. If the capital stock increases, but does not decrease with the same type of capital but with a more advanced type of capital, every economist would say this is technological progress. The TFP however, does not change. GDP will rise with the amount of extra capital, but to measure TFP, the rise in the capital stock is deducted from the rise in GDP so that TFP stays exactly the same. For this reason, TFP might not always be the best proxy for technological progress. On the other hand, as a concept of technological progress is very hard to quantify, there are few other reliable measures and because of the widespread usage it is the proxy for technological progress used in this thesis.
4.5 GDP per capita
It is common practice to include this variable in the regression, as richer countries on average tend to be more equal in terms of income distribution. If this is not included, the effect of a change in income inequality due to a change in income might falsely be contributed to a change in the trade globalization of a country. Among others, Spilimbergo et al. (1998) and Li et al. (1999) have incorporated GDP per capita as a control variable.
The GDP per capita data is retrieved from OECD Stat (2018a). Because an international comparison takes place in this thesis, GDP per capita levels are converted to a common currency using Purchasing Power Parities (PPPs). This corrects for international price differences in non-tradables and measures the prices of the same basket of consumption goods in different countries. Since the development of GDP per capita over the years as well as across different countries is looked at, 2010 constant prices are used so that the effect of inflation on GDP per capita is ruled out. Summarizing, GDP per capita in 2010 dollars (PPP) is used.
Since GDP per capita tends to increase/decrease with a constant percentage rather than a constant number, in all regressions the natural logarithm of the GDP per capita is taken. A change in this variable therefore implies a percentage change instead of an absolute change.
4.6 Education
The level of education is measured by the gross enrolment rate in tertiary education. This is the total number of people enrolled in post-secondary, including both public and private universities, colleges, technical training institutes and vocational school (World Bank, 2017). To obtain the gross enrolment rate, the total number of people enrolled in tertiary education regardless of age is to be divided by the age group to which the level of education responds.
According to Autor (2014, p. 2), in the U.S. between 1980 and 2005, two thirds of the rise in the earnings gap is accounted for by the growing premium to postsecondary education. In 1979, the income gap between the median college educated and the median high-school educated worker in the U.S. was $17,411 in 2012 constant dollars. This gap had risen to $34,969 in 2012. This proves that the education gap is widening. The more people have access to tertiary education, the more equal the educational opportunities are. Furthermore, higher access to education means that the skill premium is accruing more widely over the population. People with a better level of education are also better able to respond to labour market shocks.
For example, retraining an already well educated person is easier than retraining someone without proper education.
Data is obtained from the World Bank (2018).The time series show some missing observations for a number of countries, including Canada, Austria, the U.S. and New-Zealand. Since the gross enrolment rate of tertiary education follows a very linear process, the missing years are linearly interpolated.
5. Data description
Before we proceed to the empirical testing of the relationship between the independent variables and the dependent variable, I will elaborate on the final dataset that is used for the econometric tests. In appendix A, the descriptive statistics – the mean, standard deviation and the number of observations – are reported. The variables are sorted per indicator. The first two indicators are the dependent variables used in this thesis. From the third variable onwards, all are independent variables. For each variable, the trends will be discussed for different regions by regressing the variable over time in geographically grouped regressions.
As one can see, there are two measures for the Gini-coefficient. The data from the OECD is highly reliable, but the availability is limited. Only 307 data points of the coefficient are available through the database of the OECD, while the data for the Gini-coefficient by SWIDD (2016) contains 1,355 observations.
Ideally, both Gini-coefficients measure the exact same, meaning that the correlation would be equal to 1. When this is true, is does not matter which of the two variables is used and the one with the most observations can be used. To investigate whether or not the variables behave in the same way, the simple correlation coefficient between both Gini-coefficients is taken.
The correlation coefficient between the Gini-coefficient as measured by the SWIDD and the Gini-coefficient as measured by the OECD is 0.974, almost perfectly positive. In other words, the two variables behave in almost exactly the same way. It therefore does not really matter which of the two variables is chosen to be used and preference can be given to the variable with the most available data, which is the SWIDD data.
In appendix B.1, the graphs for the development of income inequality according to the SWIDD data for different countries are presented. Every country has its own graph, and this procedure will be followed for the remainder of the variables. As discussed in the introduction, there are vast differences among countries. In figure 5, the mean Gini-coefficient for all countries is portrayed. Taking all countries together, the income inequality decreased from 1970-1987 and increased thereafter. Using years as the independent variable, the coefficient on the income inequality is positive and significant at the 1% level. Figure 6 represents the development of the mean Gini-coefficient for all countries over time.
Figure 6 shows that the mean trade as a percentage of GDP has been growing relatively steadily, with some dips during the internet bubble in the early 2000s and the financial crisis of ‘08. This average trend does not, of course, hold for all separate countries. The development
of trade in all countries separately is presented in appendix B.2. Trade over GDP is on average rises with about 1%-point every year.
In figure 7, one can see that overall the tax base as a percentage of GDP shows a growing trend, especially during 1970-1990. There are some spikes visible. The country specific developments of the tax as a percentage of GDP in appendix B.3 show a similar image. There are few countries with real outliers and the growth of tax over GDP seems to be quite smooth. These observations are supported by a regression. For the sample as a whole the tax over GDP ratio has significantly increased from year to year.
Because the tax as a percentage of GDP is not constant over time for countries, as it ideally would be, the countries have been separated into two groups: high tax countries with above average levels of taxation and the low tax countries with below average levels of taxation. This distinction into high and low taxation countries is the dummy variable called Above average taxation dummy, as represented in the econometric model in section 3.4. This dummy is equal to 1 if the country collects above average taxation and 0 if the country collects below average taxation. The group every country is assigned to, as well as the average level of taxation, is presented in appendix C. There are 18 high taxation countries and 17 low taxation countries in the sample.
The division of countries into two groups allows a stricter distinction between countries and obtain a clearer image of what taxation does to income inequality when interacting with trade. For every country, the average taxation level is computed including all available tax data on the country. Subsequently, the average level of taxation of the country is compared to the average in the sample for the available years. For example, the average taxation level in
29 30 31 32 1960 1980 2000 2020 Mean Gini-coefficient Pe rce n ta g e Year 40 60 80 100 1960 1980 2000 2020
The mean trade as a percentage of GDP
Pe rce n ta g e Year Graphs by Country Figure 5: the mean Gini-coefficient as calculated by the
Slovakia from 1995 – the first year I have data for – to 2016 was 32.18%, but for the sample as a whole it was 33.42% for 1995-2016, so Slovakia is assigned as a low taxation country. The average taxation level in Luxembourg from 1970-2016 was 34.64%, for the full sample it was 32.15% for 1970-2016, so Luxembourg is considered a high taxation country.
In figure 8, the mean TFP by TED (The Conference Board, 2018) is shown. Here, an increase in technological capabilities is visible, with a spike prior to the financial crisis of 2008. In appendix B.4, the country specific trend in TFP over time is presented. These graphs show a similar image, except Chile. Chile is an enormous outlier from the rest, as the total factor productivity shows a drop almost every year. The consequence is that for the full sample, there is no proof of an increase in technological progress. If Chile is dropped from the full sample, the TFP does increase significantly on average every year. For the results of the regressions as a whole, presented in section 6, the inclusion or exclusion of Chile does not make a difference: the directions of the coefficients as well as most of the significance levels stay the same.
The next variable, mean GDP per capita, shows a clear increase over time, as can be seen in figure 9 for the mean GDP of all countries. Appendix B.5 contains the development of GDP per capita over time for all separate countries. For separate countries, a clear increase over time can be seen despite a dip in some countries around the 2008 financial crisis– most prominently Greece and Ireland. From 1970-2016, GDP grows by about 2.1% per year on average in OECD countries.
26 28 30 32 34 1960 1980 2000 2020 Tax as a percentage GDP Pe rce n ta g e Year 85 90 95 100 105 1960 1980 2000 2020
Total Factor Productivity
2 0 1 6 =1 0 0 Year Figure 7: the mean tax over GDP as calculated by the author with data of OECD (2018a) Figure 8: the mean TFP over GDP as calculated by the author with data of TED (2018)
