The impact of economic growth on income inequality

(1)

Faculty of Economics and Business

The impact of economic growth on income inequality

By: Max de Lorijn

Student number: 10445838

Bachelor Thesis 2015-2016 semester 1 Supervisor: Gabriele Ciminelli

(2)

1

Statement of originality

This document is written by Student Max de Lorijn who declares to take full responsibility for the contents of this document.

I certify that the intellectual content of this thesis is the product of my own work and that all the assistance received in preparing this thesis and sources have been acknowledged.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

2

Abstract

Using the income share of the richest 10 % as a proxy for income inequality, this paper tests Kuznets’ (1955) hypothesis, which implies that in the course of economic growth income inequality will decline. The latter has been done using a newly assembled panel of 5 countries (Sweden, USA, UK , Ireland and Japan) over 1980-2010. The empirical evidence suggests that a 1 percentage point increase in economic growth leads to a 2.696807 percentage points decrease in the income share of the top decile, indicating a decline in income inequality. This research has therefore found support for the hypothesis that economic growth is associated with a decline in income inequality.

(4)

3

Introduction

In 1955 Simon Kuznets laid the foundation for research regarding the relation between income distribution and economic growth in his well-known paper “Economic Growth and Income Inequality”, where he suggests that in the course of economic growth, at first income inequality would worsen and would than move towards more equality in later stages (known as the Kuznets curve). The supposed ‘U-shape’ has resulted in an excessive development in the literature concerning the impact of economic growth on income inequality. Previous papers confirm Kuznets’ (1955) theorem, however recent findings published by the OECD (2015) suggest the contrary. They state that in the 1980s the top decile of a OECD country’s income distribution earned 7 times the income of the lowest 10% income group, rising to 8:1 in the 1990s and 9:1 in the 2000s. These results are not in line with Kuznets’ (1955) proposition, since the economy in the OECD area has been growing over the past 30 years (Scarpetta, Bassanini, Pilat and Schreyer, 2000). This paper will therefore test Kuznets’ hypothesis, that income inequality will decline in the course of economic growth.

In order to test this hypothesis, an empirical research is done by analysing 5 OECD countries over 1980-2010. While the relationship studied in this paper has been examined before, the distinctive contribution of this paper lies in the data covering a long time period until recent times. Furthermore, the random effects regression model used in this paper allows to take all time-invariant characteristics into account, as well as the country specifications into account. At last, this paper made use of a number of methods eliminating endogeneity problems that can occur in panel data analyses.

The setup of this paper is as follows, section 2 provides a literature review giving an overview of a notable time period (1914-1945) regarding income inequality, and empirical work on the relationship between economic growth and income inequality. Furthermore, section 3 gives a description of the used data and will present the methodology used in this paper. Section 4, will summarize the research results that will be linked to the literature discussed in section 2. Lastly, section 5 (conclusion) will summarize the main findings of this paper.

2 Literature review

(6)

5

A discussion about the relationship between economic growth and income distribution will follow and at last, a review will be given about other trends influencing the secular income distribution.

2.1 Income inequality history

This part of the paper will discuss the World War time period (1914-1945) which is not obtained in this research, but stands out as a notable time period regarding income inequality. The latter is due to the fact that it was a period of sharp contraction in income inequality. Although it is known that during the World War time period technological innovations raised the income of the lower income groups due the increase in relative demand for unskilled workers (Frydman and Molloy, 2011), little is known about the causes that influenced the top income groups. It is therefore important to elaborate on these causes, also since Atkinson, Piketty and Saez (2009) state that the impact on the top income groups was way more severe.

In general, inequality in income distribution decreased due to the fact that the rich experienced negative shocks regarding their capital holdings during the World War time period (1914-1945) such as the great depression (1930s), inflation, destructions. Moreover, Frydman and Molloy (2011) argue that although these factor were at play regarding the income share of the top group, they cannot account for the contraction in the income share of the top decile since they found a more significant role for changes regarding firm and industry characteristics. Firstly, they found evidence that the salaries of the executives, who are considered to be in the top income group, declined due to the fact that power of labour unions to restrict executive bonuses strengthened. Secondly, they signify that the rate of return of big companies, in which the top income holders had a big share, declined excessively from 1940 to 1942.

Furthermore, when decomposing the top income decile into the top percentile and the next 4%, and the bottom 5%, Piketty and Saez (2006) describe that the decline in the income share of the top percentile was far worse than for the next two groups. Also, Piketty and Saez (2006) found evidence that the latter two recovered relatively more quick from the above described World War time period shocks than the top percentile.

The slow-recovery of the top percentile capital holders seem hard to explain. A possible explanation given by Piketty and Saez (2006), is that the shocks that took place

(7)

6

between 1914 and 1945 had a persistent effect due to the change from almost no tax progressivity to progressive income and estate taxation which made it nearly impossible for the top capital holders to recover.

In conclusion, the findings above suggest that during the World War time period (1914-1945) income inequality dropped due negative income shocks experienced by the top income group and technological innovations raising the income of the lower income groups. Also, the adopted progressive income and estate taxation seems to have disabled the top percentile to recover in the same pace as lowest 9% in the top 10% income group.

2.2 Review of empirical research

Now that one of the most influential historical events regarding income inequality is discussed, this part of the literature review will further elaborate on some influential papers regarding the relationship between economic growth and inequality in income distribution. 2.2.1 Economic growth and income distribution

The correlation between income distribution and economic growth is one of the oldest subjects within economic enquiry. The first real influential paper on this matter is the paper published by Simon Kuznets (1955). The hypothesis of the Kuznets curve resulted in an excessive development in the literature regarding the impact of economic growth on income inequality. Before elaborating on some influential papers covering this matter, Kuznets (1955) findings will first be discussed.

The central theme of Kuznets’ (1955) paper is the relationship between inequality in the distribution of income and a country’s economic growth. In order to analyse the gap between the rich and poor, Kuznets (1955) estimated equations describing the correlation between five different percentile groups (the top 20%, the next 40%, the lowest 40% and the lowest 20%). Kuznets (1955) found evidence that the relative distribution of income, as measured by annual income incidence in rather broad classes, has been moving towards equality. Kuznets (1955) therefore supposed a ‘U-shaped curve’ which holds that as the economy grows, at first the inequality in the distribution of income would worsen and would than move towards more equality in later stages. Furthermore, Kuznets (1955) also argued that if income distribution worsened due to growth, poverty might not alleviate. The findings

(8)

7

of Adelman and Morris (1973) are somewhat in line with those of Kuznets (1955), suggesting that a decline in the relative income shares of the lower income groups will appear in the early stages and then increase in later stages. Like Adelman and Morris (1973), Chenery and Syrquin (1975) point out that the decline is most significant for the lowest income groups in the early stages of growth.

Ahluwalia (1976), who looks at the relationship between per capita GNP and income inequality across 60 countries, agrees that the higher income recipients benefit more in the early stages of economic growth, with a reversal of this propensity in later stages. However, he does not support that the decrease of income inequality reflects a decrease in average absolute incomes in lower income groups . He states that the average absolute incomes of the upper income groups rise faster due to growth than those of the lower income groups, but that he found no significant evidence for impoverishment in the course of economic growth. Moreover, Roemer and Gugerty (1997) found similar evidence, who measured poverty using income distribution by analysing the lowest 20% and 40% income groups across several countries in Asia, America, Europe, Africa and Oceania. Their paper shows that an increase in the rate of GDP growth results in a direct increase in growth of average incomes of the lowest 40%.

The discussion above shows that first income distribution will widen in the early stages of development at the expense of the lower income groups. However, the shares of the lower income groups increase sufficiently in the later stages of growth.

2.2.2. Other factors influencing income distribution

Additionally to the effect of growth on income distribution and thus the gap between the rich and poor, this paper attempts to explain further about trends in the secular income distribution. This section will therefore elaborate on previous results regarding growth, income distribution and other factors that may have been at play.

Kuznets (1955) suggest that industrialization, thus the shift from agriculture and the country side to industry and the city, has a relationship with income distribution. He set up two groups, each representing a sector: agriculture (A) and non-agriculture (B). For each sector, he used a cumulative frequency distribution representing income shares. Kuznets (1955) argued that in the early stages of industrialization the gap between the rich and poor

(9)

8

would increase, indicating more widening in income distribution. However, he found evidence that in the later stages of industrialization the rise in shares of the lower income groups increased significantly more in the non-agriculture sector, indicating a decrease in the inequality of income distribution. Kuznets (1955) proposition regarding the shift from agriculture and countryside to industry and the city is consistent with the findings of Ahluwalia (1976). Ahluwalia (1976) found that the share of agriculture in GDP and the urbanization trend are both significantly related to inequality in income distribution. Moreover, he found that as the share of agriculture in GDP declines in the early stages of development, growth is associated with a relative shift of income away from the middle group and towards the top income group. However, he states that this development also results in a shift of population towards the city, the latter appears to benefit the poor at the expense of the top income groups in later stages of development.

Gregorio and Lee (2002), whose results confirms Kuznets ‘U-shape’ hypothesis regarding growth and income inequality, used cross-country data so as to be able to analyse the impact of years of schooling on the level of earnings. They found evidence that an increase in educational inequality unambiguously causes income distribution to widen. Ahluwalia (1976), whose findings are broadly in line with those of Gregorio and Lee (2002), analysed the relationship between education and income inequality in terms of 3 explanatory variables; literacy rate, the primary school enrolment ratio and the secondary school enrolment ratio. Aware of the high correlation among these variables, he has chosen literacy rate as a measure for the basic education level of a countries population and the secondary school enrolment rate as a measure of the degree of education beyond this basic level of education. Furthermore, he found that without adding GNP to the regression equation the secondary schooling variable is significantly associated with income shifting from the top 20% to all other groups except the lowest 20%. The literacy variable was not significant. However, after adding GNP to the regression equation the literacy variable had a positive significant effect on the 3 lowest income groups but the effect of the secondary schooling variable remained the same.

Additionally to the trends discussed above, it appears that the level of openness of an economy is also involved in measuring income distribution, growth and inequality. Sachs and Warner (1995) examined this using their openness variable, that contains observation of the parallel market exchange rate premium, quantitative import restrictions, the number of

(10)

9

export restrictions and socialist economic management. For a country considered being open, a country must have a low score on all four criteria. In their regression regarding the annual growth rate of per capita GDP from 1970 to 1989 containing data for 114 countries, they found highly significant and negative coefficients for the openness variable. Furthermore, Sachs and Warner (1995) found that on average, open economies have a 2.8 percentage points higher growth rate of per capita GDP than closed economies. Gugerty and Roemer (1997), who used the openness variable developed by Sachs and Warner (1995), ran two sets of regression to test the proposition that more open economies result in more rapid growth of per capita GDP which in turn benefits the lower income groups. In their first two regressions, where they regressed growth of average income on just the openness variable, they found that in a more open economy the lowest 20% and 40% of the population experienced more growth in their income, indicating a more equal income distribution. However, they found in their first two regression that the R-squared was a little less than 5% and that the effect of the openness variable on income growth disappeared when they added GDP growth. In their second regression set GDP was highly significant, in contrast to the openness variable which turned out to be insignificant but still positive. Furthermore, openness is occasionally measured by trade as the sum of exports and imports in a countries GDP (Birdsall and Hamoudi, 2002). However, Birsdall and Gamoudi (2002) state that the level or the change in a country’s trade to GDP ratio cannot serve as an indicator for the level of openness of a country’s policy, since this ratio may not decline because of a more closed economy but due to for instance balance of payments problems. Milanovic (2005), who examined the relationship between income distribution and trade for the years 1988, 1993 and 1998 based on data for 129 countries (82 countries being a balanced panel), agrees with this criticism. He therefore used the trade to GDP ratio to measure the impact of trade on income distribution. He performed two types of estimations: a simultaneous decile estimation and an instrumental variable generalized method of moments estimation, instrumenting openness as a share of GDP by their lagged values and country’s population. For both regressions, an increase in trade caused the income shares of the bottom six income groups to decrease. However, Milanovic found evidence that middle and low income groups started benefiting whenever a countries relative average income level rises above $7500. When he added regional dummy variables, the income share of the lower seven income groups decreased and that of the top two increased. The new average

(11)

10

income level for the middle and low income groups to benefit was approximately $8000. The papers above suggest that besides the impact of growth, income distribution is also influenced by industrialization, education, trade and the level of openness of the economy. This paper will broadly incorporate the factors above to test for the effect of growth on a country’s income distribution while controlling for other factors that might have be at play.

3 Methodology and Data description

3.1 Data description

Data on income inequality is available in different forms, but this research will focus on the change of the income share of the top 10% income share, so as to be able to also partially address the direction of the possible change in income distribution. Data on the top 10% income share is retrieved from the World Wealth and Income Database (2015). The latter provides the most extensive available data series on the world distribution of income and wealth, especially on top income shares. It contains data on 33 countries over 1810-2014, however this paper will only focus on 5 countries (Sweden, Ireland, Japan, United States and United Kingdom) covering the period 1980-2010 due to the lack of data. Figure 2 in the appendix, which displays the change in the income share of the top decile over 1980-2010, shows a peak during the late 1980s and early 1990s for the majority of the countries. A possible explanation can be the early 1990s recession, which resulted into high unemployment rates in the manufacturing industry in which the majority of the employees are low income holders. However, Japan experienced a big drop in the income share held by the top decile. A possible explanation for this matter could be the asset price bubble in Japan that lasted from the late 1980s until the early 1990s in which capital holdings lost their value due to high inflation. Lastly, Figure 2 also shows a big drop for all countries during the late 2000s. The most logical explanation would be the financial crisis starting in 2008, in which a lot of capital in the real estate market lost its value.

Economic growth is measured by analysing the change in real GDP per capita, an inflation adjusted measure of all goods and services produced in a country for a given year, which is a variable commonly used to analyse economic growth. Real GDP per capita is preferred over real GDP level, since you can take the effect of the population size into

(12)

11

account. Data on real GDP per capita is obtained from EconStats (2015), which is a highly trusted source for economic data. This database provides data for real GDP for 184 countries from 1970 until 2011. However, this paper will only examine the time period 1980-2010. Analysing Figure 2 (shows economic growth over 1980-2010) in the appendix, the thing that stands out the most is the big drop in the change of per capita real GDP during the late 2000s. The most logical explanation for this matter would be the financial crisis starting in 2008.

As mentioned in the literature review, the trade to GDP ratio (sum of total export and import as percentage of GDP) is occasionally used as an openness variable. However, Birdsall and Hamoudi (2002) argue that the trade to GDP ratio is not a good openness variable since it is not necessarily influenced by the level of openness of a country’s policy. Aware of this problem, this paper will incorporate the trade to GDP ratio variable the same way as Milanovic (2005), thus solely using trade to GDP to measure the impact of trade on income distribution. Moreover, as discussed in section 2.2.2 it seems that a country’s policy regarding growth is involved in explaining changes in income distribution. This research will therefore measure this using grass capital formation (% of GDP), which are basically public investments made by a country’s government that are associated with economic growth. Data on the trade to GDP ratio and gross capital formation (% of GDP) is retrieved from the

World Bank (2015), which provides data covering the time period 1980-2014. As mentioned

before this paper will only cover the time period 1980-2010.

The effect of education is measured in terms of 3 explanatory variables; gross primary enrolment ratio, gross secondary enrolment ratio and gross tertiary enrolment ratio. The total enrolment ratio regarding these 3 education levels, regardless of age, are expressed as percentage of the population of required education age. Data on gross primary enrolment ratio, gross secondary enrolment ratio and gross tertiary enrolment ratio is retrieved from the World Bank (2015), which provides data for education covering the time period 1980-2014. As mentioned before this paper will only incorporate the time period 1980-2010.

Lastly, as suggested by Kuznets (1955) and Ahluwalia (1976) industrialization is involved in explaining inequality in income distribution. Ahluwalia (1976) examined this effect by using the share of agriculture in GDP. Due to the lack of data on the latter, the variable used in this paper to represent the effect of industrialization is a modified version of

(13)

12

the method used by Kuznets (1955), namely arable land (% of land). The latter is a good alternative since its land specially assigned to the agricultural sector and it is therefore expected to replicate the same change as the change in the agricultural sector. Data on arable land is retrieved form the World Bank (2015), which provides data covering the time period 1980-2014. However, this paper will only examine the time period 1980-2010.

3.2 Methodology

As mentioned, the approach of this paper is to research whether economic growth is associated with a decrease in income inequality. This relationship will be estimated by the following random effects regression model:

(1) ∆𝑡𝑜𝑝 10% 𝑖𝑛𝑐𝑜𝑚𝑒 𝑠ℎ𝑎𝑟𝑒𝑖𝑡 = 𝑐 + 𝛽1∗ ∆log (𝑟𝑒𝑎𝑙 𝐺𝐷𝑃𝑡−1) 𝑖𝑡+ 𝑢𝑡+ 𝜀𝑖𝑡

The dependent variable represents the change in the income share of the top decile at time t, which can be interpreted the following way; an increase in the income share of the top decile implies a decrease in the income share of the remaining 9 deciles. So it therefore serves as a proxy for income inequality. The explanatory variable, representing economic growth, is displayed by the change in the logarithm of real GDP at time t-1. This variable can be interpreted as follows; if the coefficient of real GDP growth is 2, a 1 percentage point increase in real GDP growth increases the income share of the top decile two percentage points. Moreover, the first difference method described above is adopted to control for serial autocorrelation. Due to the fact that the latter does not control for all endogeneity problems, Wooldridge’s cluster (csid) command was adopted while running the regressions in STATA.

Furthermore, in this paper I considered two techniques to analyse my data set, namely a fixed effects regression model and a random effects regression model. The latter has been chosen after the Hausman test was performed, which will be explained in section 4.3. Moreover, 𝑢𝑡 and 𝜀𝑖𝑡 are included due to the random regression method adopted in this paper, which allows for between (𝑢𝑡) and within (𝜀𝑖𝑡 ) country differences.

Additional to the random effects regression model described above, multiple control variables were added to control for the income share of the top decile and economic growth. Namely, arable land, gross primary enrolment ratio, gross secondary enrolment

(14)

13

ratio, gross tertiary enrolment ratio, gross capital form and trade to GDP ratio. The latter leads to the estimation of the following equation:

(2) ∆𝑡𝑜𝑝 10% 𝑖𝑛𝑐𝑜𝑚𝑒 𝑠ℎ𝑎𝑟𝑒𝑖𝑡 = 𝑐 + 𝛽1∗ ∆log (𝑟𝑒𝑎𝑙 𝐺𝐷𝑃𝑡−1) 𝑖𝑡+ 𝛽2∗ 𝑎𝑟𝑎𝑏𝑙𝑒 𝑙𝑎𝑛𝑑𝑖𝑡+ 𝛽3∗ 𝑔𝑟𝑜𝑠𝑠 𝑐𝑎𝑝𝑖𝑡𝑎𝑙 𝑓𝑜𝑟𝑚𝑖𝑡+ 𝛽4∗ 𝑔𝑟𝑜𝑠𝑠 𝑝𝑟𝑖𝑚𝑎𝑟𝑦 𝑒𝑛𝑟𝑜𝑙𝑙 𝑟𝑎𝑡𝑖𝑜𝑖𝑡+ 𝛽5∗

𝑔𝑟𝑜𝑠𝑠 𝑠𝑒𝑐𝑜𝑛𝑑𝑎𝑟𝑦 𝑒𝑛𝑟𝑜𝑙𝑙 𝑟𝑎𝑡𝑖𝑜𝑖𝑡+ 𝛽6∗ 𝑔𝑟𝑜𝑠𝑠 𝑡𝑒𝑟𝑡𝑖𝑎𝑟𝑦 𝑒𝑛𝑟𝑜𝑙𝑙 𝑟𝑎𝑡𝑖𝑜𝑖𝑡+ 𝛽7∗ 𝑡𝑟𝑎𝑑𝑒/𝑔𝑑𝑝𝑖𝑡+ 𝑢𝑡+ 𝜀𝑖𝑡

Moreover, looking at Figure 1 (display of the relationship between the change in the income share of the top decile and economic growth) in the appendix, one can expect that economic growth is associated with a decrease in the income share of the top decile, indicating a decrease in income inequality.

Lastly, a robustness check will be conducted on the regression results by modifying the sample in several ways. The latter will be outlined in section 4.3 along with all tests and models regarding autocorrelation and other endogeneity problems.

4 Results

In this part of the paper, the results from the panel regressions described above will be reported. Firstly, the results regarding the baseline will be discussed. A discussion about the results from regression (2) will follow and lastly, the robustness check results will be outlined along with all tests and models regarding endogeneity problems.

4.1 Basic regression

As mentioned before, the approach of this paper is to research whether economic growth is associated with a change in a developed country’s income distribution by analysing 5 OECD countries (Sweden, Ireland, Japan, United States and United Kingdom) covering the period 1980-2010. Table 2 in de appendix presents the results from the baseline for this matter. The baseline is formulated by the following equation:

(1) ∆𝑡𝑜𝑝 10% 𝑖𝑛𝑐𝑜𝑚𝑒 𝑠ℎ𝑎𝑟𝑒𝑖𝑡 = 𝑐 + 𝛽1∗ ∆log (𝑟𝑒𝑎𝑙 𝐺𝐷𝑃𝑡−1) 𝑖𝑡+ 𝑢𝑡+ 𝜀𝑖𝑡

Several interesting results can be derived from Table 2. At first, one can see that there is a strong negative correlation between economic growth and the income share of the top decile. The coefficient of economic growth is negative and significant (p-value=0.012),

(15)

14

suggesting that economic growth is associated with a decrease in the income share of the top, resulting in a more equal income distribution. Quantitatively, the estimated effects suggest that a 1 percentage point increase in the real GDP per capita will cause the income share of the top decile to decrease by 2.823497 percentage points. These results are in line with those of Kuznets (1955 ) and other researchers discussed in section 2.2.1., that in the course of economic growth the low income group will be the one benefitting at the expense of the rich resulting a more equal income distribution.

4.2 Regression with added control variables

In order to get a better understanding of the effect of economic growth on the share of the top decile, the section below will discuss the results regarding the regression with the included control variables provided in Table 2. The latter is formulated by the following equation:

(2) ∆𝑡𝑜𝑝 10% 𝑖𝑛𝑐𝑜𝑚𝑒 𝑠ℎ𝑎𝑟𝑒𝑖𝑡 = 𝑐 + 𝛽1∗ ∆log (𝑟𝑒𝑎𝑙 𝐺𝐷𝑃𝑡−1) 𝑖𝑡+ 𝛽2∗ 𝑎𝑟𝑎𝑏𝑙𝑒 𝑙𝑎𝑛𝑑𝑖𝑡+ 𝛽3∗ 𝑔𝑟𝑜𝑠𝑠 𝑐𝑎𝑝𝑖𝑡𝑎𝑙 𝑓𝑜𝑟𝑚𝑖𝑡+ 𝛽4∗ 𝑔𝑟𝑜𝑠𝑠 𝑝𝑟𝑖𝑚𝑎𝑟𝑦 𝑒𝑛𝑟𝑜𝑙𝑙 𝑟𝑎𝑡𝑖𝑜𝑖𝑡+ 𝛽5∗

𝑔𝑟𝑜𝑠𝑠 𝑠𝑒𝑐𝑜𝑛𝑑𝑎𝑟𝑦 𝑒𝑛𝑟𝑜𝑙𝑙 𝑟𝑎𝑡𝑖𝑜𝑖𝑡+ 𝛽6∗ 𝑔𝑟𝑜𝑠𝑠 𝑡𝑒𝑟𝑡𝑖𝑎𝑟𝑦 𝑒𝑛𝑟𝑜𝑙𝑙 𝑟𝑎𝑡𝑖𝑜𝑖𝑡+ 𝛽7∗ 𝑡𝑟𝑎𝑑𝑒/𝑔𝑑𝑝𝑖𝑡+ 𝑢𝑡+ 𝜀𝑖𝑡

A number of interesting results can be derived from this regression. Starting with economic growth, one can clearly see that the effect on the top income share with respect to regression 1 does not change sign. However, the p-value dropped from 0.012 to 0.007 indicating a more significant effect. Quantitatively, the estimated effects suggest that a 1% increase in the real GDP per capita will cause the income share of the top decile to decrease by 2.696807 percentage points.

Looking at the results regarding industrialization, on can conclude that the effect is highly significant and positive with a p-value equalling 0.000. As mentioned in the literature review, the effect of the shift from agriculture to industry is divided into two stages, namely that the top income group benefits more in the early stages of industrialization with a reversal of this propensity in later stages. The most logical reason for the found results in this paper regarding industrialization is that the impact in the early stages between 1980 and 2010 has been more extreme and is therefore dominating the effect of the later stages.

(16)

15

appendix shows that trade has a highly negative significant effect with a p-value equalling 0.001. These results are in line with those from Milanovic (2005), that an increase in trade causes the share of the lower income groups to increase and that of the top income groups to decrease resulting in a more unequal income distribution. However, his results also show that the low an middle income groups start experience negative shocks to their income share whenever their country’s average income level would be below approximately $7500-$8000 (this is below the level of Sweden, USA, UK, Japan and Ireland).

The results on education are in contrast with those discussed in the literature in section 2.2.2, since the coefficients for neither gross enrolment rates of tertiary or secondary nor tertiary show any significant effect. Nonetheless, Gregorio and Lee (2002) and Ahluwalia (1976) used different education indicators, which might be the reason for the contrasting results. Moreover, the results regarding gross capital formation displayed by Table 2 in the appendix indicate no significant effect. These results are with those from Roemer and Gugerty’s (1997) openness variable since it turned out to be insignificant and positive in their second regression where they added GDP growth.

At last, when comparing the outcome of 𝑅2_{from regression (2) with that of} regression (1), one can clearly see that the 𝑅2_{has increased from 0.0119 to 0.0473,} indicating an increase in explanatory power regarding the variance of the top 10% income share.

4.3 Robustness check

The dataset used in this paper is an example of a panel data set (also known as longitudinal data or cross-sectional time series data). The latter refers to data for different entities observed at different time periods (Stock and Watson, 2012). As mentioned in the methodology I considered two techniques to analyse my data set, namely a fixed effects regression model and a random effects regression model. The fixed effects regression model is a method that controls for all time-invariant differences between entities, the latter will eliminate the bias regarding omitted time-invariant characteristics periods (Stock and Watson, 2012). However, a fixed effects regression models has its side effect. Namely, that it cannot incorporate the possible time-invariant characteristics of the dependent variable. In contrast, the random effects regression model incorporates the effects from time-invariant characteristics as explanatory variables since it assumes zero correlation between the error

(17)

16

term and the predictors. To determine which one to apply to the dataset used in this paper, a Hausman test has been performed. The Hausman test basically tests whether the error tem is correlated with the regressors. Examining the results provided in Table 1 in the appendix, one can clearly see that the null hypothesis (zero correlation between the error term and the predictors) is not rejected since the p-value equals 0.6488. It is for this reason that the random effects model is adopted in this paper.

Moreover, serial correlation in the error term can be expected when using panel data. It is for this reason that the first difference method has been applied which relies on the assumption that the first differences of the error term are uncorrelated (Roin, Vlachos and Waldenström, 2009). If the latter does not hold, one will have a biased estimator which causes the results to be less efficient (Drukker, 2003). It is therefore critical to test for serial autocorrelation in the idiosyncratic error term. A commonly used method is the LM-test derived in Baltagi and Li (1995), which Drukker (2003) argues to be good within a class of tests. However, Drukker (2003) argues that the Wooldridge test is more attractive since it should be more robust and it has relatively few assumptions. Given the latter, Drukker (2003) still verifies that the Wooldridge has good size and power properties. For these reasons, a Wooldridge test has been performed. Step 1 of the test contains the following: the coefficient of the main explanatory variable will be computed by regressing the change of the dependent variable on the change of the main explanatory variable and obtaining the residuals. Considering the latter, Wooldridge (2002) found evidence that if the residuals are not serially correlated, then the correlation between the change of the residuals at time t and the change of the residuals at time t-1 equals -0.5. It is for this reason that the test will regress the obtained residuals in step 1 on its lagged values and see if this equals -0.5 . Aware of possible within-panel correlation in the regression of the residuals (obtained in step 1) on its lagged values, “the variance-covariance matrix (VCE) is adjusted for clustering at the panel level” (Drukker, 2003). The latter makes the test robust against conditional heteroscedasticity. Furthermore, when applying the Wooldridge test to the data set used in this paper, one can clearly see in table 2 that the null hypothesis (no first-order autocorrelation) is not rejected since the p-value equals 0.3056 in regression (1) and 0.2070 in (2). Furthermore, controlling for serial correlation does not imply that all the possible endogeneity problems are solved. It is for this reason that Wooldridge’s cluster(csid) command is adopted running the regressions in STATA, the resulting standard errors are

(18)

17

therefore fully robust to any form of heteroscedasticity and serial correlation. Moreover, there could also be reverse causality between the change in top 10% income share and change in economic growth. To resolve this threat, the change of the top 10% income share at time t has been regressed on the change of the logarithm of real GDP at time t-1 instead of time t.

Additional to the tests above, a robustness check on the regression results has been conducted so as to be able to see whether the core regression coefficient shows structural validity. While still using the change of the log of real GDP at time t-1, the change in the income share of the top 10% has been replaced by the change in the Gini coefficient. The latter is another indicator that is widely used to measure income inequality. The main reasons for replacing the initial dependent variable by the Gini coefficient are the fluctuations in the income share of the top decile shown in Figure 2 in the appendix. Furthermore, the data for 2009 and 2010 have been omitted due to the financial crisis starting in 2008, during which the drop in economic growth rates was quit extreme. The above changes will result in the following random effects regression:

(3) ∆𝐺𝑖𝑛𝑖 𝑐𝑜𝑒𝑓𝑓𝑖𝑐𝑖𝑒𝑛𝑡𝑖𝑡 = 𝑐 + 𝛽1∗ ∆log (𝑟𝑒𝑎𝑙 𝐺𝐷𝑃𝑡−1) 𝑖𝑡+ 𝑢𝑡+ 𝜀𝑖𝑡

The results from the above robustness test reported in Table 3 show some reason for doubts regarding the structural validity of the economic growth variable used this paper, since the effect on income distribution is not significant anymore when the GINI coefficient is used as inequality indicator. Although it does confirm the earlier results regarding the negative sign of the economic growth coefficient used in regression (1) and (2), it is now questionable to what degree the hypothesis is explained by the empirical evidence found in this paper.

5 Conclusion

This paper aimed to empirically analyse the correlation between income inequality and economic growth while controlling for other factors. It does so by; analysing 5 OECD countries over 1980-2010 using a random effects regression model which takes all time-invariant characteristics, as well as the country specifications into account. Furthermore, this paper made use of several methods eliminating endogeneity problems that can occur in panel data analyses.

(19)

18

in line with those of Kuznets (1955) and other researchers (Adelman an Morris 1973; Chenery and Syrquin 1975; and Ahluwalia 1975) discussed in the literature review (section 2.2.1). They all argue that in the course of economic growth the low income groups will be the one benefitting at the expense of the rich resulting in a more equal income distribution. Furthermore, the inclusion of control variables resulted in a number of interesting findings, shown in Table 2. First of all, the explanatory power regarding the variance of the top 10% income share has increased (𝑅2_{increased from 0.0119 to 0.0473). Secondly, the} effect of economic growth became more significant (p-value dropped from 0.012 to 0.07) but did not change sign and is therefore still in line with the findings of previous papers discussed in section 2.2.1. Thirdly, the effect of industrialization on income inequality turned out to be positive and highly significant. The results regarding the latter are in line with those of Kuznets (1955) and Ahluwalia (1976), who argue that income inequality will decrease in the course of industrialization. However, they are supposing a ‘U-shape’ since they argue that the effect of industrialization on income inequality is negative in the early stages. It is therefore most likely that the impact of industrialization in the early stages between 1980 and 2010 was more extreme and is therefore dominating the effect experienced in the later stages. The results on trade, suggest that an increase in trade would decrease the income share of the top decile indicating a more equal income distribution. These findings confirm those from Milanovic (2005). Nonetheless, his results also show a contradistinction, since they suggest that the low and middle income groups experience a negative effect from trade growth whenever their country’s average income level would be below approximately $7500-$8000 (does not apply for; Ireland, USA, UK, Sweden and Japan). It might therefore be an interesting topic for future research, so as to be able to get a better understanding on the possible other factors causing this difference. As for the results for the education indicators, this paper found no evidence for any significant impact. The latter is not in line with the findings from Gregorio and Lee (2002) and Ahluwalia (1976). However, they used different education indicators, which might be the cause for the contrasting results. Moreover, the results regarding the openness variable (gross capital formation) used in this paper seem to be in line with those of Roemer and Gugerty (1997), that open policy regarding growth has no significant effect on income inequality.

In conclusion, this research has found support for the hypothesis that economic growth is associated with a decrease in income inequality. It can therefore be concluded that

(20)

19

Kuznets’ proposition still holds. Nonetheless, it must be said that this paper has left some stones unturned. For example, the robustness check test results showed structural validity regarding direction of the coefficient of economic growth on income inequality, but the effect of economic growth did not turn out to be significant anymore. Given the latter, I hope that my work will stimulate future research on this matter so as to be able to improve our understanding of the determinants of income inequality.

(21)

20

6 Bibliography

Adelman, I., Morris C.T., (1973). Economic growth and social equity in developing countries. Stanford: Stanford University Press.

Ahluwalia, M.S., (1976). Inequality, poverty and development. Journal of Development

Economics, 3(4), 307-342.

Ahluwalia, M.S., Carter, N.G., Chenery H.B., (1979). Growth and poverty in developing countries. Journal of Development Economics, 6(3), 299-341.

Birdsall, N., Hamoudi, A., (2002). Commodity Dependence, Trade, and Growth: When ”openness” is not enough. Center for Global Development, Working Paper No. 7. Chenery, H.B., Syrquin, M., (1975). Patterns of development, 1950-1970.

London: Oxford University Press.

Drukker, D.M., (2003). Testing for serial correlation in linear panel-data models. The Stata

Journal, 3(2), 168-177

Frydman, C., & Molloy, R. (2011). The compression in top income inequality during the 1940s. Working paper, MIT.

Gregorio De, J., Lee, J.W., (2002). Education and income inequality: new evidence from cross-country data. Review of Income and Wealth, 48(3), 395-416.

Gugerty, M.K., Roemer, M., (1997). Does economic growth reduce poverty? Harvard

Institute for International Development, No.5.

Kuznets, S., (1955). Economic growth and Income inequality. The American Economic

Review, 45(1), 1-28.

Milanovic, B., (2005). Can We Discern the Effect of Globalization on Income Distribution? Evidence from Household Surveys. The World Bank Economic Review, 19(1), 21-44.

(22)

21

Piketty, T., Saez, E., (2006). The Evolution of Top Incomes: A Historical and International Perspetive. National Bureau of Economic Research, Working paper No. 11955

Roine, J., Vlachos, J., Waldenström, D., (2009). The long-run determinants of inequality: What can we learn from top income data?. Journal of Public economics, 93(7-8), 974-988.

Sachs, J., Warner, A., (1995). Economic Convergence and Economic Policies. Harvard

Institute for International Development, Development Discussion Paper, 502.

Scarpetta, S., Bassanini, A., Pilat, D., & Schreyer, P. (2000). Economic growth in the OECD area: recent trends at the aggregate and sectoral level. OECD Economics

Department, Working Paper No. 248

Stock, J.H. & Watson, M.M., (2012). Introduction to Econometrics (third edition). Harlow: Pearson Education Limited

The Organization for Economic Co-operation and Development (2015) In it together: Why

Less Benefits All. Retrieved from:

https://data.oecd.org/inequality/income-inequality.htm

The World Bank (2015). Indicators. Retrieved from: http://data.worldbank.org/indicator The World Wealth and Income database (2015) The database. Retrieved from:

http://www.wid.world/#Database

Yitzhaki, S., (1983). On an extension of the GINI inequality index. International Economic

(23)

22

Appendix

Table 1- Hausman test results

Hausman test statistics

Chi2 0.21

Prob>Chi2 0.6488

Table 2 – Random effects regression results

Change top 10% income share (1) (2)

D.logrealgdp80 -2.823 -2.697 (0.012)* (0.007)** arableland 0.020 (0.000)*** grosscapitalformgdp 0.008 (0.363) grossprimaryenrolmentratio -0.006 (0.775) grosssecondaryenrolmentratio 0.004 (0.338) grosstertiaryenrolmentratio -0.001 (0.534) tradegdpratio -0.002 (0.001)*** _cons 0.368 0.122 (0.000)** (0.946) N Random effects R2 (overall) Wooldridge (p-value) 145 Yes 0.0119 0.3056 133 Yes 0.0473 0.2070

(24)

23 Table 3 – Robustness test results

Change Gini coefficient (3)

D.logrealgdp80 -1.435 (0.632) _cons 0.216 (0.067)* N Random effects R2 (overall) Wooldridge (p-value) 137 Yes 0.0067 0.4713

* p<0.05; ** p<0.01, *** p<0.001. (p-value between parentheses)

Figure 1 – Change (𝒕𝒐𝒑 𝟏𝟎% 𝒊𝒏𝒄𝒐𝒎𝒆 𝒔𝒉𝒂𝒓𝒆𝒕)𝒕 vs. Change 𝒍𝒐𝒈(𝒓𝒆𝒂𝒍 𝑮𝑫𝑷𝒕−𝟏)𝒕

-2 -1 0 1 2 3 C ha ng e ( to p 10 % i nc o m e s ha re_ t) _t -.1 -.05 0 .05 .1 Change Log(real GDP_t-1)_t

(25)

24 Figure 2 – Plot change (𝒕𝒐𝒑 𝟏𝟎% 𝒊𝒏𝒄𝒐𝒎𝒆 𝒔𝒉𝒂𝒓𝒆𝒕)𝒕

Figure 3 – Plot Change 𝒍𝒐𝒈(𝒓𝒆𝒂𝒍 𝑮𝑫𝑷𝒕−𝟏)𝒕

-2 0 2 4 -2 0 2 4 1980 1990 2000 2010 1980 1990 2000 20101980 1990 2000 2010

Ireland Japan Sweden

UK USA T o p 10 % i n c om e s h. (c ha ng e) Year Graphs by Country -.1 -.0 5 0 .0 5 .1 -.1 -.0 5 0 .0 5 .1 1980 1990 2000 2010 1980 1990 2000 20101980 1990 2000 2010

Ireland Japan Sweden

UK USA C ha ng e Lo g( re a l G D P _t -1) _ t Year Graphs by Country

The impact of economic growth on income inequality

Faculty of Economics and Business