An empirical research on unconditional convergence of GDP per capita levels between 2002 and 2014

(1)

An empirical research on unconditional

convergence of GDP per capita levels

between 2002 and 2014

Abstract

This paper studies unconditional convergence of GDP per capita levels. It uses the time period 2002 to 2014. Because the academic literature in ambiguous about the existence of unconditional convergence, it is interesting to research this for a recent time period. This paper uses the Solow model as its theoretical basis. Using panel data on an as large as possible sample of countries, this paper tests empirically whether we observed unconditional convergence, divergence or neither during this time period, by testing whether past GDP per capita levels are correlated with current GDP per capita growth after controlling for other potentially important determinants of growth. The fixed effects OLS estimator is used. This research finds strong evidence for a negative long-run cumulative dynamic multiplier on the logarithm of past GDP per capita levels and therefore strong evidence for unconditional convergence. This paper also finds significant correlations for some of the other included variables, such as capital growth, domestic capital market quality, inflation and unemployment. The strong evidence for unconditional convergence shows that countries having a low GDP per capita tend to grow more than countries having a high GDP per capita, ceteris paribus.

Keywords: Economic growth, panel data, Solow model, unconditional convergence. JEL Classification: O47.

Justin Baars BSc Thesis Economics and Business

10772278 Specialization Economics & Finance

June 26, 2017 Supervisor: Ron van Maurik

(2)

2

Statement of Originality

This document is written by student Justin Baars 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.

(3)

3

List of tables

Table 1. Variable definitions 35

Table 2. Summarizing statistics 36

Table 3. Outcomes of Fisher-type test on stationarity 37

Table 4. First stage regression on TAX 38

(5)

5

1. Introduction

A major question within the field of macroeconomics is what the determinants are of economic growth. This question arises naturally because we observe huge differences in wealth levels between different countries, while this often is not an obvious consequence of the intellectual potential of the population or of differences in natural recourses, since people of different nationalities are assumed to have the same intellectual potential and western countries (e.g. European countries, the USA or Japan) with the same natural recourses as non-western countries tend to be richer than non-western countries. Baroich (1981) shows that the ratio of GNP per capita between the richest and the poorest country before the start of the Industrial Revolution was in the range of 1.0-1.6, compared to a ratio of 45.2 today. If we look at the whole group of advanced capitalist countries versus non-advanced capital countries, we observe that advance capitals countries have 4.5 times as much purchasing power as other (non-advanced capitalist) countries (Pritchet, 1997, pp. 10-12). This shows that there has been divergence during this period; advanced countries have seen their purchasing power grown by approximately 800% during 1870-1990, while poor countries have seen growth rates over this 120 year period in the range of 60%-150% (Pritchet, 1997, p. 11). However, Pritchet (1997, pp. 13-14) also shows that some but far from all countries are able to show convergence towards richer countries, but that this convergence is often not stable over time, and that we even observe divergence for some of these countries. Now the question comes up what accounts for these differences, and whether it is possible that poorer countries will catch up with richer countries over time. This might be, for example, because of productivity spillovers (Gorg & Strobl, 2001, p. 723), where technology or knowledge from a multinational company spills over to employees of the affiliate or to competing firms in the less developed country. Gorg and Stroble (2001, pp. 729-737) find in a meta-analysis on the literature that, although there may be a publication bias, in general the evidence points towards productivity spillovers.

Because the differences in GDP per capita between countries are large and consistent over time, it is hard to believe that these differences are coincidental. Therefore, policymakers might be curious what causes higher growth rates, and whether there are variables that are relatively easy to change that are likely to increase economic growth. If this is the case, there would be a possibility to increase overall welfare and reduce poverty. According to World Bank (2013), poverty, as defined by an income of less than (PPP) a day, has been falling since 1981, but still more than 10 percent of world population lives in poverty, however. Half of the extreme poor people live in Sub-Saharan Africa. There are several ways to alleviate poverty, which can be split up by exogenous and endogenous measures. It seems to be most sustainable to let these economies grow endogenously, because in that way poorer countries will not be dependent on richer countries. Also for countries

(6)

6

not living in extreme poverty, economic growth is important to improve overall happiness, which can be accomplished by, among other factors, better health, a decent standard of living and better housing (Easterlin, 1995, p. 40). Therefore, it is relevant to research what determines economic growth.

The literature identifies some factors of possible influence on economic growth, which I will explore later. However, in general the literature does not agree on whether we observe unconditional convergence. Korotayev, Zinkina, Bogevolnov and Malkov (2011, p. 26) find in their literature review that in the 1950’s economists reasoned that under a few assumptions there would be unconditional convergence, where countries’ GDP per capita levels converge. However, Korotayev et al. (2011, pp. 27-29) also find that thereafter conflicting theories arose and that the empirical evidence generally was not consistent; some researchers actually find unconditional divergence, while others observe unconditional convergence. On the other hand, most papers find evidence for conditional convergence, in the sense that the further a country is from its own balanced growth path, the higher the rate of convergence towards its balanced growth path (Korotayev et al., 2011, p. 28). Besides that, they observe unconditional convergence between 1998 and 2008 but reason that these findings in fact do not differ much from previous results, but previous results rather insisted on conditional convergence (Korotayev et al., 2011, pp. 37-50).

Because the literature shows evidence for both unconditional convergence and divergence, and therefore is not consistent, in this paper I will research whether we observed unconditional convergence or divergence or neither worldwide recently. The literature provides arguments and evidence for both, but most papers do not include as many countries as possible nor consider a recent time period. Therefore, in this paper I will research the time period 2002 until 2014 and include as many countries as possible in the regression. Because it could be the case that there is unconditional convergence of per capita GDP levels, but that this convergence is actually the result of other factors influencing economic growth, I will look into the literature for possible factors causing economic growth. After controlling for those other factors, I will look whether there is still a case for unconditional convergence or divergence. The research question of this paper will be: did we observe unconditional convergence during the period 2002 to 2014?

To answer this question, I will first provide a literature review in the next section, summarizing economic theory on growth, suggesting factors of possible influence on growth and explaining conditional and unconditional convergence further. In the third section, I will explain my research method, explain how I collected the data and I will give variable definitions and some summarizing statistics. In the fourth section, I will analyze these data and I will look which variables significantly explained variation in growth factors. In particular, I will look whether we observed

(7)

7

unconditional convergence, divergence or neither. In the last section, I conclude and make some suggestions for further research.

2. Literature review

To be able to research unconditional convergence, I first need to investigate economic theory on this subject and define the most important concepts. I will start with the Solow model on economic growth and define the factor in this model that accounts for income differences. Because unconditional convergence says something about economic growth, it is important to look at other possible determinants of economic growth, so later on I can control for these other possibly important factors. After listing some important possible determinants of economic growth, I will give a short summary on the literature on convergence and state some reasons why we might expect unconditional convergence to occur.

2.1. Growth models

There are many models on economic growth. However, to keep things simple, I consider the Solow growth model (1956), sometimes referred to as the Solow-Swan growth model, which also acknowledges Swan’s (1956) contribution. This model assumes that output is dependent on three factors: two endogenous factors; capital and labor, and one exogenous factor, which is supposed to capture ‘technology’ or ‘effectiveness of labor’. The endogenous inputs capital and labor can easily be measured by looking at the amount that is paid to the suppliers of these inputs. However, this implicitly assumes that labor and capital markets are efficient, since only in efficient markets factors would earn their contribution to output. Because we are here interested in the determinants of growth, we can use the techniques of growth accounting, a technique developed by Solow (1957). Using this technique, we can calculate the contribution of growth of labor, growth of capital and a ‘residual’, which is assumed to capture growth of the ‘effectiveness of labor’ or of ‘efficiency’. This technique is explained in the research method of this paper.

Increasing the amount of labor (by stimulating population growth, for example), ceteris paribus, does not increase output per capita, but keeps production per unit of labor constant (trivial because of the ‘ceteris paribus’). Also, assuming efficient labor markets, employing more people is not profitable, so forcing higher employment than in equilibrium will at best keep production per unit of labor constant, but is likely to decrease output per unit of labor.

If we assume economic agents optimize and markets are rational, as is assumed by several extensions of the Solow-Swan model, such as the Ramsey-Cass-Koopmans model (Cass, 1965, p. 233) and the Diamond model (Diamond, 1965, p. 1127), a policy maker cannot increase utility by altering

(8)

8

capital used in production, because after optimization by economic agents the capital level will already equal its (optimal) golden rule level. The Ramsey-Cass-Koopmans model assumes that the community goes on forever, and does not change in number (Ramsey, 1928, p. 543). This implies that households live infinitely and therefore have infinite horizons. Furthermore, Cass (1965, pp. 233-234) assumes a central planner choosing the savings rate to maximize utility by maximizing the present value of all future utilities. In this way, the central planner effectively sets optimal capital allocation, since the model assumes a market without any imperfections and assumes households that optimize in a predictable way.

A similar model that uses different assumptions, is the model proposed by Diamond (1965). This model lets go the assumption of infinitely living generations by assuming generations that live for two periods; one period while people are young and one while they are old. However, the model still assumes that the community itself goes on forever (Diamond, 1965, pp. 1127-1128). As Phelps (1965, pp. 804-812) proofs, letting go the assumption of infinitely living generations can lead to dynamic inefficiency; thus an economy described by the Diamond model might be Pareto inefficient. Abel, Mankiw, Summers and Zeckhauser (1989, pp. 8-14) find a condition for dynamic efficiency: that the net capital income exceeds investment. They also show that this condition is satisfied for the United States in each year since 1929 (Abel et al., 1989, pp. 8-14), which is strong evidence that dynamic inefficiency does not occur in reality. Both the Ramsey-Cass-Koopmans model and the Diamond model therefore sketch economies that provide efficient capital allocation.

We have seen that increasing the amount of labor, ceteris paribus, will not increase the amount of GDP per capita, which is trivial because of the notion of ceteris paribus. Also, economies behaving as in famous growth models such as the Ramsey-Cass-Koopmans model and the Diamond model provide efficient capital allocations, if we assume that the condition for dynamic efficiency holds in reality. Furthermore, one of the most important theoretical results from the field of finance, the efficient market hypothesis, also implies efficient capital markets, and therefore efficient capital allocation, if true (Fama, 1970, pp. 383-388). In the same paper, Fama (1970, pp. 389-404) argues that the weak-from hypothesis holds. This hypothesis asserts that past stock price movements and past volume data do not have any effect on future stock prices. Fama (1970, pp. 404-409) also argues that the semi-strong form holds, that assumes that all publicly and in a cost-effective way available information is reflected in stock prices. Malkiel (2005, pp. 2-8) concludes from the inability of actively managed mutual funds to outperform index funds, both in the United States and abroad, that capital markets are at least approximately efficient; benefits to reap are considered to be small, because the benefits of hiring experts to manage a fund are smaller than these experts’ salaries. To conclude, both labor allocation and capital allocation cannot be altered in efficient markets to increase the output per capita.

(9)

9

However, there is one factor in the Solow-Swan model that might be altered by a policy maker to increase output per worker, which is the ‘effectiveness of labor’ or ‘technology’ in the model. The Solow-Swan model considers the growth of this factor to be exogenous, however (Solow, 1956, p. 85). Since only capital and labor are considered to be endogenous inputs, including an exogenous input allows other factors than capital and labor to contribute to output. Because this factor is assumed to be determined exogenously in the Solow model, this factor cannot be altered by changing labor or capital amounts. Note that because of the wide differences between GDP per capita among countries, the richest country (the United States) having an about 45 times as big GDP per capita as the poorest country (Chad) in 1990 (Pritchett, 1997, p. 11), there appear to be wide differences in effectiveness of labor.

2.2. The determinants of economic growth

There are several extensions of the Solow model which are designed to make the model more realistic. One of these extensions is the inclusion of human capital. The development of the Solow model including human capital is shown by Mankiw, Romer and Weil (1992, pp. 416-418). This extension to the model acknowledges that human capital might be important, but also faces several difficulties. For example, this extension assumes that forgone labor earnings by accumulating more human capital are the same for different workers; that accumulation of human capital takes place only in school; that all education is intended to yield productive human capital; and that accumulation of human capital is only dependent on the years of education (Mankiw et al., 1992, pp. 418-419).

This extension raises the predictive power of the Solow growth model (Mankiw et al., 1992, pp. 420-421). The ‘effectiveness of labor’ still remains an exogenous factor. Because raising capital and labor is, in efficient markets, not able to increase GDP per capita for reasons already explained, the Solow growth model gives us no clear view on the determinants of economic growth.

There are several theories about the determinants economic growth. As was shown by the increased significance of the Solow model including human capital, the amount of schooling seems to contribute to higher economic growth. Barro (1996, p. 70) suggests that maintenance of law, government consumption, inflation, level of democracy, life expectancy, schooling, tax distortions, income distribution, terms of trade and infrastructure are important factors for growth. However, as Sarel (1996, pp. 208-210) points out, the function that relates economic growth to inflation shows a structural break. This break appears to be most significant at around eight percent annually. Furthermore, basic economic theory states that deflation might lead to decreased economic growth. This is, however, generally not confirmed by statistical evidence (Atkeson & Kehoe, 2004). Alesina, Özler, Roubini and Swagel (1996, pp. 199-205) find that political instability reduced economic growth

(10)

10

significantly in their sampling period 1950 to 1982. Political stability thus seems to be an important explanatory variable for economic growth. Wurgler (2000, p. 188) argues that countries that are financially developed increase investment more in their growing industries and decrease investment more in their declining industries than less financially developed countries. He also finds that the extent of state ownership decreases efficient capital allocation, and in that way decreases economic growth (Wurgler, 2000, p. 189).

2.3. Convergence

Yet another factor that might influence economic growth is the past level of GDP per capita. There is an ongoing academic debate whether there is convergence of GDP per capita levels or not. As Xavier and Sala-i-Martin explain, there is a difference between unconditional convergence, which occurs if poor economies tend to grow faster than rich ones (1996, p. 1020), and conditional convergence, which means that countries tend to have higher (lower) growth the further they are below (above) their own steady state level (1996, p. 1027).

A first reason why one might expect convergence of GDP levels was already implied by the initial paper written by Solow. Here he states that any economy behaving as in the model will move towards a state of balanced growth at its natural rate (Solow, 1956, pp. 70-72). It might therefore be the case that countries that have a low GDP per capita are just below their balanced growth path. If this is the case, the Solow model predicts that these countries will grow faster than richer countries, which then are above their balanced growth paths. However, this reasoning assumes that countries have the same balanced growth paths. This might not be the case, since countries can have different fundamentals determining these paths, such as savings, depreciation of capital, growth of effectiveness of labor and population growth.

Since the Solow model uses a Cobb-Douglas production function, it assumes diminishing marginal products of inputs. Therefore, the model predicts that the more capital per unit of effective labor a country uses, the smaller its marginal product will be, and because the model assumes that inputs earn their marginal products, the model also predict that the more capital a country uses, the smaller its compensation will be. Developed markets typically are more capital-intensive than developing markets, so the Solow model implies a higher rate of return for capital for developing countries, and therefore a capital flow from developed to developing countries, which would cause convergence between GDP per capita levels. The existence of these capital flows is confirmed by empirical data, but these same data also show that investments in emerging markets tend to be riskier than these in developed countries (Salomons & Grootveld, 2003, pp. 129-140). If we assume investors are rational mean-variance optimizers, as the Capital Asset Pricing Model assumes (Sharpe, 1964), higher return and higher risk in emerging markets than in developed markets would not

(11)

11

create incentives to invest more in emerging markets above obvious diversification benefits. A capital outflow from developed to developing countries is therefore not obvious.

Keller (2004, pp. 775-776) finds that foreign technology sources account for 90 percent or more of domestic productivity growth for most countries and that this effect is larger for countries with smaller domestic R&D development. This shows that there appear to be significant technology and knowledge spillovers between countries over time. Again, the Solow model assumes that effectiveness of labor or technology is a factor that increases production, keeping labor and capital input constant. When there are technology and knowledge spillovers between countries over time, we should observe relatively large (small) growth rates in technology for countries having a low (high) initial technology . Because technology is a productive input in the Solow model, this would cause more economic growth for countries having a low initial value of than for countries having a high initial value of , ceteris paribus. This creates a possibility for unconditional convergence of GDP per capita levels.

An influential paper by Baumol (1986, pp. 1075-1077) suggests there is strong evidence for unconditional convergence, because he found a strong negative relationship between income levels in 1870 and growth rates between 1870 and 1979. However, he does not control for any other variables, so this convergence of GDP per capita levels might also be (partly) attributable to other factors in a more extended research. In a response to this paper, De Long (1988, pp. 1141-1145) points out that there does not seem to be strong evidence for convergence in the period 1870-1979, because the dataset used was biased; only for countries that are now the richest there are data available on per capita incomes as far back as 1870. By different sample selection, he is able to alleviate this bias, and finds less evidence for convergence.

Another factor that might contribute to convergence is economic integration. Kim (1998) finds evidence that economic integration provided income convergence in the United States between 1840 and 1987. He finds that convergence (divergence) in regional industrial structures within member states led to convergence (divergence) of regional incomes (Kim, 1998, p. 678). Similarly, Jones (2002, pp. 36-39) finds that for countries within the Economic Community Of West African States (ECOWAS) there was unconditional convergence of real GDP per capita between 1960 and 1990. While significant, however, Jones (2002, p. 40) also finds that this convergence actually occurs quite slowly, at a rate of around two percent per year. Bouvet (2010, pp. 334-339) shows that interregional income inequality in the Economic and Monetary Union (EMU) decreased between 1970 and 2003, while regional inequality exacerbated, however. These findings suggest that an increasing scope of economic integration might contribute to unconditional convergence of per capita GDP levels between countries.

(12)

12

2.4. Hypothesis

Generally, we do not observe strong evidence for unconditional convergence, and when we do, this convergence turns out to be quite slow, as mentioned above. Therefore, under the null hypothesis I assume that there is no unconditional convergence or divergence, so economic growth in not correlated with (the logarithm of) the level of lagged GDP per capita levels. However, we also see worldwide poverty declining, so unconditional convergence might be a recent development. Furthermore, I just identified several factors that might cause unconditional convergence of GDP per capita levels. I use a two-sided alternative hypothesis because divergence might also be possible. I set as the alternative hypothesis that we actually observe unconditional convergence or divergence, so that past GDP per capita explain variation in current GDP per capita growth. To this end, I need lagged data on per capita income and look whether, after controlling for other possibly significant factors as set out in this section, these lagged data are able to significantly explain variation in GDP per capita growth rates. If this is the case, I will observe a nonzero long-run cumulative dynamic multiplier, which is equal to the sum of the coefficients on the lagged (logarithm of) GDP per capita variables and therefore captures the total effect of past GDP per capita levels on current GDP per capita growth. Because convergence is believed to occur slowly if it occurs at all, I will include lags up to ten years. So we have:

∑ ∑

3. Research method

In this section I will first look into the Solow model and explain why I will not use this model directly to derive the dependent variable that I will use in this regression. In addition, I will define the dependent variable to test unconditional convergence. Next, I will define the explanatory variables included in this research. Thereafter I will explain my regression equation and estimation method.

3.1. The dependent variable

Using the techniques of growth accounting and starting with a Cobb-Douglas production function, Solow (1957, p. 312) finds the following expression:

(13)

13 ̇( ) ( ) ( ) ̇( ) ( ) ( ) ̇( ) ( ) ( ) ̇( )

The dot above a variable indicates the first derivative with respect to time. Using the fact that income must accrue to either capital or labor, where is the capital share of income and is the labor share of income, so , we can rewrite this expression (Romer & Chow, 1996, p. 30) to

̇( ) ( ) ̇( ) ( ) ( ) [ ̇( ) ( ) ̇( ) ( )] ( )

Using this expression, the Solow residual ( ) can be calculated, which is defined by the -elasticity of income times the relative growth of , where denotes the effectiveness of labor, or, mathematically: ( ) ( ) ( ) ( ) ( ) ̇( ) ( )

Therefore, this Solow residual ( ) is supposed to capture how much income changes due to changes in effectiveness of labor . This Solow residual measures all sources of growth other than capital and labor input. For capital and labor we assume they earn their marginal products, so their contributions to output are assumed to be same as their income share. Note that for the other factors to be a significant source of growth, we must observe both a change different from zero in the magnitude of this factor, and we must observe an elasticity of income with respect to this factor (or a factor captured by ) different from zero. So, mathematically, the next condition has to be satisfied: ̇( ) ( ) ( ) ( ) ( ) ( )

This condition makes sense because for technology to be a significant growth source, it must both change and be able to contribute to a change in output when it changes.

However, in this research, I will not be able look at output, capital and labor growth rates in continuous time, but these changes have to be approximated in discrete time, because data on output, capital and labor (and on the other variables) is not available in continuous time, but only annually. Thus if the annual Solow residual using data between 2002 and 2014 is calculated, the resulting number might differ from the ‘real’ Solow residual, which is derived using derivatives with respect to time, so assuming continuity of time. Because this paper focuses on whether we observed unconditional convergence of GDP per capita levels between 2002 and 2014, I can focus as well on growth in GDP per capita levels and control for labor and capital growth.

(14)

14

Another reason why I do not use the Solow residual as the dependent variable is because it is not possible to find reliable data for a large sample. This is because we need several other statistics to calculate the Solow Residual: the growth rates of GDP, labor force and capital between 2002 and 2014 and the capital share of income for each country in our sample. To calculate the capital share of income, we can use the labor share of income and the identity . Guerriero (2012) provides for a sample of countries a single estimate of the labor share of income for 1970 to 2010. Because generally income shares are considered to be constant, it would be possible to rely just on this single estimate. Guerriero uses six labor share estimators, LS1 to LS6, where LS1 is the least sophisticated and LS6 is the most sophisticated estimator. However, the less sophisticated estimators are not accurate (Guerriero, 2012, pp. 5-6), but on the other hand data on the most sophisticated estimator is only available for 89 countries and provides only one estimate for the whole time period. One way of dealing with this problem is using the ‘typical’ value for the capital share of income of one third (Mankiw et al., 2012, pp. 410-417) for all countries. This ‘typical’ value might also be subject to a serious bias, however, because we cannot just assume that capital shares in very different countries are the same. Guerriero’s data (2012, pp. 31-34) actually confirms that LS6 estimator for many countries is very different from this ‘typical’ value.

Because of the limited availability of accurate labor shares estimates and because time in my sample is discrete, I decide not to use the Solow residual as a dependent variable, but use growth of GDP per capita levels instead. I collected data for as many countries as possible on the dependent variable from World Bank, which provides data on the growth rate for real GDP per capita between 2002 and 2014. I call the dependent variable ‘GDP_G’. In addition, I include labor force and gross fixed capital formation growth rates as independent variables to look whether these variables are able to explain variation in GDP per capita growth rates. These variables are called ‘LAB_G’ and ‘CAP_G’, respectively.

3.2. Explanatory variables in the model

In this paper, I will research the determinants of growth of GDP per capita levels. Therefore I have to look for variables that might be of influence on economic growth. I am particularly interested in whether there is a case for unconditional convergence after controlling for possibly significant factors. If there is unconditional convergence, the coefficients on lagged per capita GDP variables will be negative (on average), so the higher GDP per capita in the past, the lower the current growth rates of GDP per capita will be. In that case we would observe a negative long-run cumulative dynamic multiplier.

In the literature review I found several other variables that might be of possible significance for explaining economic growth, which are given and explained below.

(15)

15

First, I include the independent variables I’m particularly interested in: the lagged GDP per capita levels, using lags up to ten years. These data are provided by World Bank. The lagged variables are just the GDP per capita levels in previous years. For example, the ten year lagged GDP per capita level in 2005 is the GDP per capita level in 1995 in current US dollars. Because I use the percentage growth in GDP per capita as a dependent variable, I will use logarithms (base ) of past GDP per capita levels, because it is more reasonable that we observe an effect on GDP per capita growth per percentage change in past GDP per capita levels than per dollar change in past GDP per capita levels. In this paper, I will call lagged variables the same as the original variables, but I will add ‘_Lx’ to the name, where is a positive integer and denotes the lag in years. So for example, ‘LN_GDP_L4’ is the logarithm of GDP per capita level with a lag of four years.

Second, human capital might be of importance for explaining economic growth. This variable is calculated by the average years of schooling completed, for the population of age at least 15. These data are collected from the UNESCO Institute of Statistics, which uses data from Barro and Lee. The data are not available for every country, and only for 2000, 2005 and 2010. This variable is called ‘SCHOOLING’. Because it might be the case that schooling has a rising marginal effect on productivity, since post-secondary education might yield more relevant skills for the labor market than primary education, for example, I will also try whether including the square of this variable, called ‘SCHOOLING2’ yields a higher explanatory power of the model.

Next, we look at several political variables that might be of importance. I will call this ‘the set of six government indicator variables’. There are data available by World Bank on estimated indexes covering the following six variables: (1) control of corruption, ‘CORR_CONT’, which captures perceptions of the extent to which public power is exercised for private gain, including both petty and grand forms of corruption, as well as ‘capture’ of the state by elites and private interests; (2) government effectiveness, ‘GOV_EFF’, which captures perceptions of the quality of public services, the quality of the civil service and the degree of its independence from political pressures, the quality of policy formulation and implementation, and the credibility of the government's commitment to such policies; (3) political stability and absence of violence/terrorism, ‘POL_STA’, which measures perceptions of the likelihood of political instability and/or politically-motivated violence, including terrorism; (4) regulatory quality, ‘REG_QUA’, which captures perceptions of the ability of the government to formulate and implement sound policies and regulations that permit and promote private sector development; (5) rule of law, ‘RULE_LAW’, which captures perceptions of the extent to which agents have confidence in and abide by the rules of society, and in particular the quality of contract enforcement, property rights, the police, and the courts, as well as the likelihood of crime and violence; (6) voice and accountability, ‘VOI_ACC’, which captures perceptions of the extent to which a country's citizens are able to participate in selecting their government, as well as freedom of

(16)

16

expression, freedom of association, and a free media. For all six variables, the description above is taken from World Bank, and the estimate gives the country's score on the aggregate indicator in units of a standard normal distribution, where higher is considered to be better. I collected data for every year between 2002 and 2014. Using these six government indicator variables, the variables political stability, maintenance of law and the level of stability, which were found to be of possible influence in the literature review, are supposed to be captured.

We also need data on inflation rates for as many countries as possible. Again, we can use data provided by World Bank. World Bank provides annual data on growth rates in consumer prices for most countries. I collected these data for every year between 2002 and 2014. This variable is called ‘INFL’. Because there might be a structural break in the function relating economic growth to inflation at eight percent, I also created a dummy variable called ‘INFL8’ equal to 1 if inflation is above eight percent, and equal to 0 if inflation is below or equal to eight percent.

The terms of trade for a country might also be important for economic growth. World Bank provides data giving the terms of trade for each country, using an index which is equal to 100 in 2000 for each country. The number is calculated by dividing export value by import value, while also taking into account quality of both exports and imports. Clearly, being able to receive more for exports or pay less for imports is favorable, so the higher the value of the index, the bigger the terms of trade improvement since 2000. This variable is called ‘TERMS’.

Another factor that might be a determinant of economic growth is life expectancy, which proxies for the quality of medical care in a country. Again, World Bank provides data on this variable, which is given by the number of years a newborn infant would live if prevailing patterns of mortality at the time of its birth were to stay the same throughout its life. I collected data on this variable for every year between 2002 and 2014 and called it ‘LIFE_EX’.

The next variable that might be of significance is the quality of capital markets within a country. However, it is difficult to find a good index for capital market quality that is available for the majority of countries. Therefore, I used data provided by World Bank on the domestic credit provided by the (domestic) financial sector to the private sector, trough loans, non-equity securities, trade credits and other accounts receivable. The data are given as a percentage of GDP. This makes sense because the more developed domestic capital markets are, the better the domestic financial sector is able to provide credit to the private sector, and the higher this variable will be. I collected data on this variable for every year between 2002 and 2014 and called it ‘DCTPS’.

Also income inequality is a factor that possibly has influence on economic growth. Income inequality can be measured by the Gini coefficient, which is equal to the area between the Lorentz curve and the 45-degree line through the origin of the graph, scaled down to take on values ranging from 0 to 100, where 0 means perfect income equality and 100 means perfect income inequality.

(17)

17

World Bank provides data on Gini coefficients, which I collected for the period between 2002 and 2014. I call the resulting variable ‘GINI’.

Government consumption might also be a factor of relevance for economic growth. Data on this variable are available from World Bank. The variable is calculated as government current expenditures for purchases of goods and services, including compensation of employees and most military expenditure excluding investment. The data are given as a percentage of GDP. I collected the data for 2002 until 2014 and called the variable ‘GOV_EX’.

Furthermore, the business cycle of a country might be an important determinant of economic growth. Typically we see high unemployment during a recession and low unemployment during a boom, so I use data on unemployment here to proxy for the state of the business cycle. Unemployment is calculated by taking the unemployed, defined as those individuals without work, seeking work in a recent past period, and currently available for work, including people who have lost their jobs or who have voluntarily left work, as a percentage of total labor force. Data on this variable are available from World Bank. I collected these data for 2002 until 2014 and the resulting variable is called ‘UNEMP’.

As a last variable, tax distortions might also have a significant influence on economic growth. To measure this effect, I include a variable that measures the average tax distortions compared to GDP by taking the amount of taxes collected as a percentage of GDP. For each country, I collect data for the period 2002 until 2014 and I call the resulting variable ‘TAX’.

There might be a problem of endogeneity for some of the included variables, however. I deal with this in section 4.2.

A summary of variable definitions is given in Table 1 in the appendix.

3.3. Regression equation and estimation method

I will estimate the coefficients on the possible explanatory variables for growth of GDP per capita levels using a multiple regression model. I will use data on the explanatory variables and on the dependent variable described in this section. I will try multiple specifications to look whether some variables might better be left out of the regression. To decide on whether inclusion of lags in necessary and on the proper length of these lags, I will perform a Fisher-type test based on the Dickey-Fuller test. I am using panel data, so I will use either a fixed effects or random effects model, on which I will decide after performing a Durbin-Wu-Hausman test in the next section. For the fixed effects model, I will use the following regression equation

∑ ∑

(18)

18

Here is the growth in GDP levels for country at time ; ∑

is

the sum of the logarithms of the lagged GDP per capita variables with corresponding coefficients for country at time ; ∑ is the sum of the other explanatory variables (possibly with lags)

including their coefficients, totaling variables, for country at time ; denotes country fixed effects; denotes time fixed effects and denotes the error term for country at time . Note

that , the number of explanatory variables besides the logarithm of lagged GDP per capita levels, depends on the specification used. Because I am interested whether the long-run cumulative dynamic multiplier is significant, I will use the equivalent regression equation

∑ ∑

Here denotes the change in the variable with respect to the previous observation one year back. The variable names are changed accordingly by starting with a ‘d’ now (e.g. ). In this modified regression equation, it can be shown that

is equal to ∑ in the non-modified regression equation, see for example Stock & Watson

(2011, p. 637). Stock & Watson do not prove this simple result themselves, so I prove a more general result, of which this is a special case: theorem 1 in the appendix. In this way we can test the long-run cumulative dynamic multiplier (∑

) that enters the null and alternative hypotheses

directly. Using the long-run cumulative dynamic multiplier, the effect of past GDP levels on current GDP growth is estimated, and therefore I test for unconditional convergence or divergence here. The inclusion of the control variables is motivated in the literature review. I thus constructed the regression equation myself and did not take it from one specific paper. The fixed effects model will be estimated by the fixed effects OLS estimator, using cluster and thus HAC variance estimators. Note that using this method every country is exactly one observation, which might yield quite different results than a research method in which the estimates are weighted, using for example GDP per capita or population size to determine weights.

Table 2 provided in the appendix shows some summarizing statistics for all included variables. I conclude from the summarizing statistics table that there are no outliers that appear to be impossible or highly unlikely, given that the whole population of observations, ignoring panels, for many variables is not normally distributed. Also, variables such as inflation might be way more volatile for some panels than for others. Table 2 therefore shows that the minimum and maximum values for some variables are many standard deviations away from the mean, which would be highly unlikely under a normal distribution. Many variables therefore contain far outliers, but because these

(19)

19

observations are not measurement errors or the like, these observations should be kept in the dataset. The total sample, ignoring the panels, might therefore not satisfy the fixed effects OLS assumptions (one assumption is that far outliers are unlikely). However, by using cluster variance estimators, we can instead use the assumption that far outliers within each panel are unlikely, so using cluster variance estimators the fixed effects OLS assumptions are not violated.

Besides the mean, standard deviation, minimum and maximum, Table 2 also shows the total number of observations, the total number of countries for which there is at least one observation and the average number of observations per country for each variable. Table 2 shows that for most variables data are available for almost all countries. World Bank (2017) list 217 countries, from which British Virgin Islands; Channel Islands; Curacao; Faroe Islands; French Polynesia; Gibraltar; Isle of Man; Northern Mariana Islands; Sint Maarten; St. Martin; and Turks and Caicos Islands are excluded because these were missing in the database for the six government indicator variables. Note that from Table 2 we observe that on the variables ‘GINI’ and ‘SCHOOLING’ (including its square) there is little data available when compared to the other variables. Because I want to keep the sample big and because too fewer observations yield less accurate results, I decide to keep ‘GINI’, ‘SCHOOLING’ and ‘SCHOOLING2’ out of the regression. From the average number of observations per country we also see that for countries for which there are data available at all there are often missing data, so this research uses an unbalanced panel. A look at the table for correlation coefficients, which is not included in this paper, shows that there is no problem of high multicollinearity in the dataset, except for the set of six government indicator variables. For more information, I refer to section 4.3.4.

In all tests in this research I will use a significance level of . Sometimes I will give the p-value, because this gives more information on the level of significance than just stating that something is significant using .

4. Results

In the previous section I explained how I collected the data, what the research method will be and presented some summarizing statistics on the included variables. In this section, I will perform regressions and look whether there is sufficient statistical evidence to conclude that the long-run cumulative dynamic multiplier is different from zero. If this is the case, a negative long-run cumulative dynamic multiplier indicates unconditional convergence, while a positive one indicates unconditional divergence. To do this, I first have to look whether the variables are stationary, because otherwise the regression results will be biased. Next, I will test whether the instrumental variables for the endogenous variable are strong, because if they are weak the regression results will also be biased. Thereafter, I will use multiple specifications and look whether the inclusion of more

(20)

20

variables yields more significant results. I will also look whether an IV regression yields structurally different results than an OLS regression. If this is the case, I will use the consistent IV regression. If not, I will use the more efficient OLS regression.

4.1. Fisher-type tests for stationarity

The data used is panel data running from 2002 to 2014 and economic data over several years often exhibits autocorrelation. For example, a high unemployment level at makes a high unemployment level at , using relative small discrete time periods, more likely than a low unemployment level. To avoid the resulting bias of autocorrelation and to estimate the coefficients on the relevant variables more accurately, I will include lagged variables when necessary. Adding lagged variables might be necessary because the regression I will perform assumes the data is stationary. Because this cannot just be assumed, I will perform a test and include lagged variables when the data turns out to be nonstationary. The augmented Dickey-Fuller test assumes under the null hypothesis that the data is nonstationary and tests whether there is sufficient statistical evidence to conclude that the data are stationary. When the null cannot be rejected, I will add lagged variables to the regression.

There is one problem with this test, however. The augmented Dickey-Fuller test assumes a balanced panel, in the sense that there are no missing data, which is not the case in my sample. Therefore I will use a Fisher-type test, based on the augmented Dickey Fuller test. This test also works for unbalanced panels. I perform the test for lags of one and two years and I allow a time trend because I will also try including time fixed effects in the regression equation. The results from this test are presented in Table 3 in the appendix.

Table 3 shows that all variables are stationary (even at ) when we look at lags of one year. However, for lags of two years we do not have sufficient statistical evidence for some variables to conclude that the data are stationary ( ). These variables are CAP_G, DCTPS, GOV_EX, TERMS and HEALTHEX. This means that the data on these variables might have a stochastic trend. Therefore, I include lagged data with lags up to two years for these variables to avoid biased results. Including these lags might make the estimates for the relevant coefficients very inaccurate because of high multicollinearity. There is high multicollinearity for DCTPS and its lags (estimated values between and ), GOV_EX and its lags (estimated values between and ) and TERMS and its lags (estimated values between and ), so the coefficients on these variables might be inaccurate and might thus have large standard deviations and therefore close to zero t-values. If would thus not be surprising if we would not observe significant coefficients on these variables, even if the literature finds that these variables are determinants of economic growth.

(21)

21

4.2. Endogeneity

There might be a problem of endogeneity in this regression, because of simultaneous causality: the unemployment level influences GDP growth, because the more people are employed, the more people produce on average and the higher the GDP per capita will be, but on the other hand the level of GDP per capita is believed to have a negative influence on unemployment, because higher growth leads to higher growth expectations and thus to higher employment by firms. Current GDP growth might influence current unemployment, but cannot influence past unemployment. So lagged unemployment, using a lag of one year, called ‘UNEMP_L1’, will only be correlated with GDP growth through UNEMP, and is therefore uncorrelated with the error term. The unemployment level evolves smoothly over time so UNEMP and UNEMP_L1 are likely to be correlated.

I also found in the literature review that it is likely that domestic credit to private sector (DCTPS), inflation (INFL) and government expenditure (GOV_EX) where possible determinants of GDP per capita growth. However, it might be the case that these variables are endogenous too, because of simultaneous causality. Higher GDP per capita growth might cause higher growth expectations and therefore more confidence in the future state of the economy, therefore more loans being written and therefore a higher value for DCTPS. Similarly, higher GDP might cause higher demand by consumers because consumers earn more money because they produce more, and therefore a higher value for inflation. Typically governments try to stabilize the economy, so the lower the GDP growth, the higher government expenditure tends to be. To take away this simultaneous causality bias, I also include the first lags of DCTPS, INFL and GOV_EX as instrumental variables and threat DCTPS, INFL and GOV_EX as endogenous. DCTPS, INFL and GOV_EX evolve smoothly over time these variables are likely to be correlated with their first lags. Current GDP growth cannot influence past domestic credit, inflation or government expenditure, so the lags of DCTPS, INFL and GOV_EX will therefore be uncorrelated with the error term.

The same applies for TAX, because in periods of high economic growth tax collections tend to be higher than in period of low economic growth. Therefore, to avoid the resulting bias, I include two instrumental variables to estimate the tax rate. Taxes tend to be higher for countries that spend more, so I include variables on military expenditure and on public health expenditure, both given as a percentage of GDP. These variables are called ‘MILEX’ and ‘HEALTHEX’, respectively. Because health and military expenditure are typically planned in advance, it is safe to assume that current GDP per capita growth does not cause different military and health expenditure. However, it is still possible that these instruments are weak, so I also include the first lag of TAX, called ‘TAX_L1’ as an instrumental variable. Current GDP growth cannot cause different taxes in the past, so it is safe to assume that this last instrument is only correlated with GDP per capita growth through TAX, and

(22)

22

therefore is uncorrelated with the error term. The tax level evolves smoothly over time so TAX and TAX_L1 are likely to be correlated.

Now I will look at the instrumental variables used for the variable TAX. I found two possibly valid non-lagged instrumental variables to run a two-stage regression and overcome the resulting bias of endogeneity. The two instrumental variables are MILEX and HEALTHEX (military and public health expenditure), including the two lags of HEALTHEX.

Before I can use these variables as instruments, I first have to check whether the instruments are weak. The first stage regression results are presented in Table 4 in the appendix. Regression 1 has MILEX as the explanatory variable, Regression 2 has HEALTHEX and its lags as explanatory variable and Regression 3 includes both. All three regressions use TAX as the dependent variable, are fixed effects regressions and use robust cluster and therefore HAC variance estimators.

From Table 4 we can see that the only regression that yields jointly significant results is Regression 1. This indicates that HEALTHEX (including its lags) is not able to significantly improve the model. A partial F-test, comparing Regression 3 to Regression 1 and testing whether the inclusion of HEALTHEX and its lags significantly improves the explanatory power of the model confirms this. The F-statistic of Regression 1 of , which is smaller than 10, indicates that also MILEX is a weak instrument and therefore yields biased results, even in large samples.

Because MILEX and HEALTHEX turned out to be weak, I try whether the first lag of TAX is a strong instrument. This first stage F-statistic of , being much larger then 10, yields strong evidence that this first lag is a strong instrument. Similarly, for UNEMP its first lag is the only instrumental variable I identified. The first stage F-statistic is equal to , so for UNEMP there is also strong evidence that its first lag is a strong instrumental variable. For GOV_EX and its first lag I observe a first stage F-statistic of , for DCPTS and its first lag I observe a first stage F-statistic of , and for INFL and its first lag I observe a first stage F-statistic of , so the first lags for these three variables are strong instruments. Furthermore, for these first lags it is also clear that they are exogenous since current GDP growth cannot influence past unemployment, tax collections, domestic credit to private sector, inflation and government expenditure.

Concluding, I will look whether using a 2SLS regression, using the first lags of UNEMP and TAX as instrumental variables, improves the model, by testing whether UNEMP and TAX are endogenous. I will do this by a Durbin-Wu-Hausman test comparing the model estimated by simple OLS to the one estimated using 2SLS. Under the null hypothesis, both estimates are consistent, but the 2SLS estimates are inefficient. However, under the alternative hypothesis the simple OLS estimates are inconsistent. Thus, if the null hypothesis is rejected by this test, I will use 2SLS estimates as the base specification. I will use the more efficient OLS estimates if the null is not rejected.

(23)

23

4.3. Regression results

Now the regression can be performed to test whether the long-run cumulative dynamic multiplier is significantly different from zero, as set out in the hypotheses. However, some variables that were found to be possible determinants of economic growth might actually not be able to significantly explain any variation in GDP per capita growth. In that case, these variables can better be left out of the regression, because otherwise the estimates of the coefficients on the other variables might become less accurate. To this end, I will try several regression specifications. In all regressions done in this paper I use cluster and thus HAC variance estimators.

First of all, I have to choose whether to use a fixed effects or a random effects model. To this end, I performed a Durbin-Wu-Hausman test using the full specification. That is: the specification using all variables, except the variables GINI, SCHOOLING and SCHOOLING2, since these were already left out of the sample due to too many missing data. This test yields if time fixed

effects are included and if time fixed effects are excluded, which is in both cases

overwhelming evidence that the random effects estimator is inconsistent ( ). I will therefore use the fixed effects estimator.

4.3.1. Naïve regression specification and time fixed effects

Regression 1 in Table 5 in the appendix is a first but naïve regression, which includes no other explanatory variables than (logarithms of) past GDP per capita levels. Like all other regressions in this paper, Regression 1 also uses country fixed effects. The dependent variable is GDP_G. To estimate the long-run cumulative dynamic multiplier, which is the sum of the coefficients on all lagged (logarithm of) GDP per capita variables, the first nine lags of the logarithm of GDP per capita are replaced by first differences, and the tenth lag remains in the regression equation. The coefficient on the tenth lag is now equal to the long-run cumulative dynamic multiplier. The regression allows for country fixed effects but not for time fixed effects since there are no time dummies included. Regression 1 in Table 5 now shows that the coefficient on the tenth lag, so the long-run cumulative dynamic multiplier, is negative and significant ( ). This is evidence for unconditional convergence, because the sum (put differently: the average level) of GDP per capita levels for the last ten years decreases current GDP per capita growth, ceteris paribus, according to these first regression results.

Regression 2 in Table 5 adds time fixed effects to Regression 1 by adding dummies for each year. Note that one (arbitrary) dummy, in this case the dummy on the year 2002, is left out of the regression to avoid the dummy variable trap. Because it is not certain whether adding the time dummies increases the explanatory power of the model, I perform an F-test, with the null hypothesis

(24)

24

that the coefficients on all time dummies are zero versus the alternative hypothesis that at least one of these coefficients is nonzero. This test yields overwhelming evidence that time fixed effects significantly increase the explanatory power of the model ( , ), so a regression with time effects yields more accurate results than one without. The long-run cumulative dynamic multiplier gets even more negative under this specification and stays highly significant ( ).

4.3.2. Including labor and capital growth

Regression 3 in Table 5 adds capital growth including its two lags, because the Fisher-type test (see Table 3) could not reject the null hypothesis that there was some stochastic trend in the data on capital growth. Regression 3 also includes labor growth in the regression. The Cobb-Douglas function that is used by the Solow model predicts that higher capital growth results in higher economic growth as given by higher GDP per capita growth. This is not the same for labor growth, however, because we are looking to GDP per capita here, and higher labor force growth has both a nominator and denominator effect on GDP per capita, so it is ambiguous what the effect on higher labor growth on GDP per capita growth should be. The prediction of the Cobb-Douglas function is confirmed by the coefficients on capital growth and its lags, which are all positive and significant. I also performed an F-test on whether the inclusion of capital growth including its lags and labor growth significantly increases the explanatory power of the model, using the null hypothesis that all coefficients on these variables are equal to zero, and the F-test rejects the null ( , ). The model will therefore be more accurate including these variables. Under this alternative specification, the long-run cumulative dynamic multiplier stays negative, significant and in the same order or magnitude as in Regression 2.

4.3.3. Including macroeconomic variables

Regression 4 in Table 5 adds some policy and macroeconomic variables to Regression 3 that might be important. These variables are: domestic credit to private sector, which should proxy for the quality of domestic capital markets; government expenditure; inflation plus a dummy indicating inflation higher than eight percent a year; life expectancy, which should proxy for the quality of health care in each country; terms of trade; taxes and unemployment. Because for domestic credit to private sector, government expenditure and terms of trade the Fisher-type test, for which the results are shown in Table 3, could not reject the null hypothesis that there was some stochastic trend in the data, lagged terms for these variables are also included. Table 5 shows that domestic credit to private sector had a significant influence on GDP per capita growth, while the coefficients on the variable itself and its second lag are negative and the coefficient on the first lag is positive. The

(25)

25

overall effect from a change in this variable is therefore ambiguous. The rate of inflation and GDP per capita are negatively correlated, which is in line with economic theory, given that deflation occurred only 228 times (so of the observations) in the complete sample. The coefficient on unemployment is negative and significant. This is also as expected since unemployment was included to proxy for the business cycle, and the more people are without a job, the less people produce, on average. An F-test, using the null hypothesis that the coefficients on all variables included that were not included in Regression 3 are zero, gives overwhelming evidence that the inclusion of this set of variables significantly improves the explanatory power of the model ( , ). The coefficient on the long-run cumulative dynamic multiplier is in the same order of magnitude under this regression specification and also still significant ( ).

4.3.4. Including government indicator variables

Regression 5 in Table 5 uses the specification of Regression 3 but adds the set of six government indicator variables CORR_CONT to VOI_ACC, which should proxy for political stability, maintenance of law and the level of stability. All these six variables have coefficients that are not significantly different from zero. However, all coefficients, while not significant, are positive. This set of six variables is supposed to capture the soundness and the quality of government policy, and therefore are highly correlated. The correlation coefficients between these variables range from to and most are bigger than , indicating high multicollinearity and leading to higher variances of the estimates of the coefficients and therefore lower significance. These government indicator variables also tend to have high multicollinearity with LN_GDP_L10, indicating that countries with a higher (past) GDP per capita tend to have a better-organized government. This might be the reason for non-significant coefficients on the government indicator variables. The coefficient on LN_GDP_L10 is still significant, however. To test whether the inclusion of these six variables significantly increases the explanatory power of the model, when compared to Regression 3, an F-test is required. This F-test yields strong evidence that the inclusion of these six government indicator variables significantly increase the explanatory power of the model ( , ). Therefore, I include this set of variables in the model. The long-run cumulative dynamic multiplier is still highly significant under this specification and has approximately the same value as in Regression 4.

4.3.5. Including both macroeconomic and government indicator variables

Regression 6 in Table 5 uses Regression 4 as base specification and adds the set of six government indicator variables to the model, or, put differently, uses Regression 5 as base specification and adds

(26)

26

the set of policy and macroeconomic variables. To test whether this bigger model significantly increases the explanatory power of the model, another F-test is needed. When we test the significance of the set of government indicator variables, we find strong evidence that including these variables increases the explanatory power of the model ( , ). When we test the significance of the set of variables on government and macroeconomic policy, we also find strong evidence that including these variables significantly increase the explanatory power of the model ( , ). I conclude that the most complete regression model, Regression 6, yields the most accurate results.

However, there might still be a problem of endogeneity, so the results obtained in Regression 6 might be biased. To determine whether this is the case and whether an alternative regression specification is needed, using instrumental variables and estimating using 2SLS, I perform another Durbin-Wu-Hausman test. In this test, I compare Regression 6 to an equivalent model, in which TAX, UNEMP, DCTPS, INFL and GOV_EX are considered to be endogenous and are therefore estimated in a first stage regression using the instrumental variables TAX_L1 and UNEMP_L1. This equivalent but differently estimated model yields the results of Regression 7. The 2SLS regression is consistent both when TAX, UNEMP, DCTPS, INFL and GOV_EX are endogenous and when they are exogenous, because the instrumental variables used are found to be strong. However, if TAX, UNEMP, DCTPS, INFL and GOV_EX are exogenous the OLS estimator is more efficient than the 2SLS estimator and should therefore be used. The Durbin-Wu-Hausman test looks whether the difference in the estimated coefficients from the two estimation methods is systematic. If there is sufficient statistical evidence to conclude that this difference is systematic, the consistent 2SLS estimator yields different results than the OLS estimator, and because we know that the 2SLS estimator is consistent, it follows that the OLS estimator is inconsistent. This test yields , so we do not have sufficient

statistical evidence to conclude that the OLS estimator is inconsistent ( ). Thus I am safe to conclude that the OLS estimator is consistent and efficient, and should therefore be used.

The null hypothesis that the OLS estimator is consistent could not be rejected. The OLS estimator is also efficient. The partial F-tests showed that including the set of government indicator variables, including the set of macroeconomic and policy variables and including all significantly improved the explanatory power of the model. Including time fixed effects next to the country fixed effects also led to a significant improvement of the model. This leads me to conclude that Regression 6 is the most complete model, without incorporating non-relevant variables, yielding consistent and the most efficient results. Therefore I use Regression 6 as the base specification.

This most complete regression model is able to explain of the variation within countries and of the variation between countries. The model is able to explain of overall variation.

An empirical research on unconditional convergence of GDP per capita levels between 2002 and 2014