The impact of natural resources
on corruption in the public sector
An analysis of 164 countries from 2005 till 2014
Ahmad Mukhtar Kakar
10561641
Bachelor of Science
Faculty of Economics and Business
Specialisation Economics
Field Macroeconomics
Supervisor: R.E.F. (Ron) van Maurik MSc
Thesis coordinator: Dr. D.F. (Dirk) Damsma
Amsterdam, 31 January 2017
An analysis of 164 countries from 2005 till 2014
2
Abstract
Mostly, people assume that countries with a lot of natural resources are highly corrupt. The previous study shows this is true in case the quality of the democratic institutions is poor. In order to find out whether this finding also represents current situation, we use panel data for 164 countries covering the period 2005-2014. Besides the Entity and Time Fixed Effects Regression Model, we also use Two-Stage Least Squares (2SLS) Instrumental Variables Regression Model to mitigate the endogeneity bias. Taking among other things the democracy level and time & country effects into account, we find strong empirical evidence for the impact of hard and soft coal on corruption in the public sector. If we focus on the rough annual regressions, one can state there is high corruption in countries with natural resources. The main finding – an aggregate of natural resources do not affect corruption in the public sector - is robust to the use of a different measure of corruption
Statement of Originality
This document is written by Ahmad Mukhtar Kakar 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
Table of Contents
ABSTRACT ... 2
I. INTRODUCTION ... 4
II. LITERATURE REVIEW... 6
II.1 Definition of corruption and a measure of natural resources ... 6
II.2 The explanation of causation between natural resources and corruption ... 6
II.3 The support for the causation in the literature ... 8
III. DESCRIPTION OF THE DATA ... 10
III.1 Data summarised ... 10
III.2 Dependent Variable ... 10
III.3 Explanatory Variables ... 12
IV. METHODOLOGY ... 14
IV.1 Multicollinearity ... 15
IV.2 Endogeneity ... 16
IV.3 The validity of instrument ... 16
V. EMPIRICAL ESTIMATES ... 18
V.1 OLS Results ... 18 V.2 2SLS IV-Regression Results... 20 V.3 Robustness check ... 26VI. CONCLUSION ... 27
VI.1 Conclusion ... 27 VI.2 Discussion ... 28APPENDIX A: DESCRIPTION OF THE DATA ... 29
APPENDIX B: VARIANCE INFLATION FACTORS ... 33
APPENDIX C: NATURAL RESOURCES SPECIFIED: OIL, GAS, FOREST & MINERALS. . 34
APPENDIX D: ROBUST REGRESSION OUTPUTS ... 38
4
I. Introduction
There is a prevailing stereotype that countries with a lot of natural resources are highly corrupt. Examples are Venezuela, Qatar and Saudi Arabia. The relationship between corruption and the natural resources is socially relevant because all the citizens jointly own natural resources of a country and they want their government to spend its natural wealth in the most efficient way. There is academic relevance as well since researchers are curious to know more about the causal
relationship between these two factors. Corruption is widely studied with regard to other variables such as economic growth, investment rates (Mauro, 1995) and cultural norms (Fisman & Miguel, 2007). Whereas the connection between corruption and the natural resources is not studied that often. Bhattacharyya and Hodler (2010) investigate this relationship and conclude that, indeed, natural resources have an impact on corruption. They add this is only the case in countries with a poor quality of democratic institutions. Their sample contains 124 countries covering the period 1980-2004. Moreover, Vicente (2010) conducts a natural experiment in two African countries and investigates whether the unexpected oil discoveries affect corruption in the public sector. He confirms the positive association between corruption and natural resources as well.
In comparison with the paper of Bhattacharyya and Hodler, the sample used in this paper contains 40 countries more, 164 in total. This large number of countries mitigate the sample
selection bias. In addition, this thesis aims to add value to the literature by using panel data of more recent years 2005-2014. Furthermore, Bhattacharyya and Hodler use log per capita rent1 from forestry, energy and minerals as a measure for natural resources. Whereas the standard of natural resources used in this paper contains not only forest rents and mineral rents but also oil rents and coal2 rents. This makes the measure of natural resources slightly more representative than the one used by the Bhattacharyya and Hodler.
The research question is therefore: “Do natural resources affect corruption in the public
sector positively?” Two models are used to answer this question namely Entity and Time Fixed Effects
Regression Model and Two-Stage Least Squares (2SLS) Instrumental Variables Regression Model. We do not only use an aggregate of natural resources as an independent variable, but also specific natural resources which are oil, gas, forest, minerals and coal. According to our OLS- and 2SLS IV-results with time- and country effects, natural resources have almost no impact on corruption. Focusing on a specific measure of natural resources, we observe that only hard and soft coal affects
1 Retrieved from the World Bank’s adjusted net savings dataset. 2 Hard and soft coal.
5
corruption in the public sector. This significant impact of coal on corruption can be explained by the highly concentrated coal reservoirs in just a few countries.
In order to answer the research question, an elaborate review of literature will be provided in the next section. There are also specific definitions of the terms corruption and natural resources. Section III contains a description of the data. Section IV discusses methodology in which
multicollinearity, endogeneity and the validity of instruments are clarified. Then, there are empirical results in section V. The sixth and the last section of this paper contains conclusion and discussion. References and appendices are added at the end of the paper.
6
II. Literature Review
There are three subsections in this part of the paper. Corruption is defined in the first subsection. On top of this, there is a description of the measure of natural resources. In the second subsection, the causation between corruption and natural resources is explained. In the third and the last
subsection, the causation is empirically supported and the existing literature is reviewed.
II.1 Definition of corruption and a measure of natural resources
Generally, corruption can occur in forty different shapes (Kangle, 1972). These practices of secretly taking money that belongs to an organisation or the government are called “forty ways of
embezzlement” according to Kautiliya. This can be a politician who buys votes (Vicente, 2007), an illegal refugee who bribes immigration officials to enter a country or an underage citizen paying the security guard to enter a night club. Specifically, corruption can take place in the public and private sector. In this paper, we focus on economic corruption in the public sector. Bardhan (1997) defines corruption: “The use of public office for private gains”. Relating this definition to the principal-agent theory, an official corresponds with the agent and principal is the public. The agent should act in the interest of the principal, but this does not happen in a proper way because of private gains for the official. The public is not fully able to monitor the tasks of the officials.
In this paper, corruption is defined as: “The abuse of entrusted power for private gain” (Transparency International, 2014). Corruption and the data used in this paper will be specifically described in section III.
Another variable used in our analysis is natural resources. Resources can be divided into two groups: renewables such as the sun and renewable such as oil. We focus in this paper on non-renewable natural resources. These resources are measured annually by the World Bank. More details on this variable will be provided in the next section. Why we have chosen this pathway is discussed then as well.
II.2 The explanation of causation between natural resources and corruption
One of the main determinants of corruption is natural resources’ share of exports (Treisman, 2000). In other words, the export of raw materials is a core part of government revenues in poor countries. Poverty leads to high corruption in these countries. The dependence on raw materials may affect political stability negatively which in turn increases corruption.
7
Furthermore, Collier (2010) argues that countries with an abundance of natural assets are associated with a poor political system i.e. corruption in the public sector. So the political problems and natural assets are positively correlated. The main purpose of a government is providing public goods and services (Collier, 2010). He focuses on two public goods, namely security and
accountability3. Both of these public goods can be undermined if there are valuable natural resources in a country. He argues that a state with natural wealth has no external threat and therefore no need to build an effective military. Security is undermined. This can be in the form of high corruption or even worse: civil war. Countries with natural resources are significantly associated with civil wars (Lujala, Gleditsch, & Gilmore, 2005).
Accountability can also be undermined because there is no pressure to raise tax rates. The high amount of public money coming from natural resources makes taxes unnecessary. Potentially, not enough attention is paid to its fiscal capacity and its legal system. This leads to a vulnerable and ineffective legal and fiscal system which eventually results in high corruption (Collier, 2010). Indeed, the increased amount of vote buying confirms high corruption (Vicente, 2006). On the other hand, the politician is inclined to raise taxes if there are no natural resources. A stable legal system and an effective fiscal capacity are needed to provide public goods. There is an incentive to do so otherwise neither public goods will be provided nor the politicians will get paid. Stated differently, if there are no natural resources, almost no tax revenues and politicians do not raise taxes then there is no reason to be corrupt because there are simply no revenues (Collier & Hoeffler, 2009). Collier and Hoeffler (2009) conclude that resource rents are the cause of worse governance. Poor security and accountability and worse governance lead to a reduced and unfair supply of public goods. This makes it easy for employees in the public sector to be corrupt. Collier (2010) states: “Without security against violence, property rights are void; and without accountability, both property rights and the supply of other public goods depend upon the personal whim of the ruler.”
Moreover, there are three reasons why the availability of natural resources make corruption possible by worsening the governance (Collier, 2010). Firstly, if the natural resources are larger than 8 percent of the GDP, then natural wealth hinders the normal functioning of democracies (Collier and Hoeffler, 2009). The reason for undermining the democracy is that politicians are prone to own as much money as possible in order to maintain power. Collier and Hoeffler (2009) state this can happen by means of patronage i.e. vote buying and jobs in the public sector. Secondly, politicians in autocracies have different interests than those of the citizens. This is especially the case in a country where natural resources are present. Good governance is hampered by the resource rents. The third
8
reason is that natural resources might change a democracy into a dictatorship. Empirical evidence is provided by Ross (2001) who shows that countries with an abundance of natural resources are mostly autocratic.
Thus, corruption is caused by natural resources because an abundance of natural assets leads to deterioration of governance and political problems. The result is high corruption.
II.3 The support for the causation in the literature
Bhattacharyya and Hodler (2010) have studied whether natural resources affect corruption. They use a game-theoretic model which is based on the actions of politicians and people. Specifically, the economy consists of natural resources and a production sector. In other words, total income is equal to the sum of domestic production At and the natural resources rent Ωt. There are three parties namely a president, a challenger and the people. The researchers distinguish between ‘good’ politicians who care about the people and ‘bad’ politicians who mainly focus on their own revenues and neglect the social welfare of the people. People should decide whether the incumbent president can stay in office for his second term or not. This corresponds to voting and election process in real life. In addition, people want to maximize their welfare function Wt = W(ct) ≡ (1 – ct)[A(ct) + Ω]. Corruption is indicated by Ct. The politician’s utility can be derived from the corruption revenues ∏t= ∏(ct) ≡ ct[A(ct) + Ω]. This function is comparable to a Laffer curve4 which shows the relationship between tax rates and tax revenues (Tucker, 2010). Since politician’s utility is a function of corruption, ct can be seen as a tax rate for which no public good is provided.
Bhattacharyya and Hodler (2010) argue there are Perfect Bayesian Equilibria (hereafter PBE) in this game-theoretic model with incomplete information. This model provides important support for the causation between natural resources and corruption. Moreover, these two researchers conclude that the PBE are confirmed by their empirical analysis based on data from 124 countries covering the period 1980-2004. Therefore, this theoretical model and the empirical analysis of Bhattacharyya and Hodler shows there is high corruption in countries with a lot of natural resources indeed. This is only the case in countries with democratic institutions of relatively poor quality. It does not hold true for democratic countries.
Leite and Weidmann (1999) confirm the dependence of corruption on natural resources. Ross (2001) confirms the conclusion of Bhattacharyya and Hodler (2010) as well. The democracies are damaged in oil-rich countries (Ross, 2001). Ross (2010) emphasizes that the negative effect of oil
4 According to this curve the tax revenues are maximized at an intermediate tax rate. Theses revenues are zero if tax rate is either 0% or 100%.
9
does not only hold in the Middle East but also elsewhere. Mexico, Malaysia and Nigeria are good examples. Ross (2001) just focuses on oil and minerals in his paper.
Vicente (2010) studies the effect of natural resources on corruption by conducting a natural experiment in West African country Sao Tome and Principe. There has been a significant oil discovery in Sao Tome and Principe from 1997 till 1999. The researcher has used Cape Verde as a control country. Both of these countries are located in West Africa and they were former Portuguese colonies. Sao Tome & Principe and Cape Verde became independent during the African
decolonialisaton in 1975 (Vicente 2010). On top of this, these two island-countries are labeled as ‘low-income’ by the World Bank. This makes Cape Verde an appropriate control country with an almost identical history.
Corruption is the dependent variable in the paper of Vicente (2010) which is measured by household surveys. The questions asked in surveys are related to public services such as healthcare, police, electoral politics i.e. vote buying etcetera. The data is based on 841 interviews conducted in Sao Tome and Principe and 1066 interviews in Cape Verde. Difference-in-difference estimator has been used to quantify the effects of natural resources on corruption. The difference between Sao Tome & Principe and Cape Verde has been identified not only before the oil discovery but also after 1991-1997 when the discovery took place. The clearest changes in corruption are found in the allocation of undeserved subsidies with a coefficient of 14%. This is followed by courts with a coefficient of 10% in which politicians have to deal with accusations of corruption in recent years. The oil discovery makes it attractive for politicians to hold political power. This is confirmed by the largest increase in vote buying which has the highest coefficient namely 40%. This is statistically significant at the 1% significance level. Also, police and a number of scholarships show a significant coefficient of 30%.
This natural experiment conducted in West Africa is a representative example of the theoretical prediction that corruption is affected by the natural resources. Because there is a clear shock of oil discovery in Sao Tome and Principe. This shock is analysed by using Cape Verde as an ideal control country. The only weakness might be the validity of the subjective household surveys used to measure corruption. This weakness can be ignored by the fact that the findings of Vicente (2010) are confirmed by Vicente (2006) who specifically investigates the effect of the campaign on vote buying.
10
III. Description of the Data
Firstly, there is a description of the data in III.1. This is followed by a detailed description of the dependent variable, CPI. The third and the last subsection of this chapter contains an elaborate description of the explanatory variables.
III.1 Data summarised
Among other things the averages and standard deviations of the variables used in this paper are summarised in Table 1. The two main variables – in which we are mainly interested to investigate its relationship – are the corruption perceptions index and the natural resources. The investigation of this relationship covers the period 2005 till 2014. The years 2007 and 2011 have the lowest average CPI score namely 40. This means there was high corruption on average. The highest average CPI score is 43 in 2014. This indicates less corruption on average. Focusing on the whole period 2005-2014, there is an average of approximately 41 (Table 1).
According to our hypothesis, there is high corruption in countries with a lot of natural resources. To check roughly if this is the case according to our data, we plot the natural resources on the horizontal axis and CPI on the vertical axis (Figure 1a & 1b). These ten graphs per year show a negative relationship between corruption and the natural resources. It means the higher natural endowment in a country, the higher corruption in the public sector of that country. A detailed graph of the recent year 2014 is added in Appendix A (Figure 2). Of course, only graphs are insufficient. Elaborate OLS- and IV-Regression analyses are required to check whether there is indeed a positive relationship.
As one can see in Figure 3, average natural resources vary over time. Its highest average is 14,77% of GDP in 2008. This is reduced to 10,4% in 2009. The mean of natural resources in the panel data is 10,14% (Table 1). This is approximately the same as the annual average of 2013 which is 10,24%. The variation of natural resources over time is problematic. So, one should control for unobserved variables that vary through time, but do not differ across states. We also use country effects in order to control for the unobserved variables that differ from one country to the next but do not change over time.
III.2 Dependent Variable
The dependent variable is Corruption Perceptions Index (hereafter CPI) measured by the
Transparency International (hereafter TI) annually. TI is a global movement which defines corruption as: “The abuse of entrusted power for private gain.” Their core purpose is minimising corruption in
11
the world by informing people about it. This non-governmental organisation measures corruption by aggregating data from more than ten reliable sources such as Economist Intelligence Unit Country Risk Ratings, World Bank Country Policy & Institutional Assessment, and World Economic Forum Executive Opinion Survey (EOS). These data stem from the perceptions of country experts and business people regarding the corruption in the public sector. Firstly, TI selects trustworthy sources. Secondly, there must be at least three sources of a country to be included in the CPI. If that is the case, then data are standardised. Finally, the average is calculated and a measure of uncertainty is reported.
Specifically, data retrieved from TI cover the period 2005-2014. The CPI scores of 2005 until 2011 range from 0 to 10, whereas the CPI scores of 2012, 2013 and 2014 are on a scale of 0 to 100. The reason for changing the scale is making it possible to compare scores over time, which was not possible prior to 2012 (Transparency International, 2014). Though the changes in the methodology do not affect the interpretation of CPI scores significantly. In both cases, a score of 0 means highly
corrupt. Low corruption or ‘very clean’ is indicated by a CPI score of 10 and 100. For the sake of ease,
the CPI data of the years 2005-2011 are multiplied by a factor of ten and rearranged to a scale of 0-100. This makes them easily comparable with the scores of the period 2012-2014. So all the CPI scores used in this paper are on the same scale namely 0 - 100. High corruption is indicated by low scores here as well.
The advantage of using the CPI scores is that it suits our purpose because these scores are only based on corruption in the public sector and no other sectors. This makes CPI relevant for our research question. Treisman (2000) use CPI scores to investigate the causes of corruption and he finds natural resources as a determinant. It is also used by Smarzynska & Wei (2000) and Wei (1998) who investigates corruption in economic development. The correlation coefficient between the measure of corruption used by Kaufmann et al. (2005) and the measure of TI used in this paper is extremely high namely 0.97. This confirms the reliability of TI CPI scores.
But Bhattacharyya and Hodler (2010) use the corruption index from the Political Risk Services (hereafter PRS). There are two reasons why PRS indices are not used in this paper. Firstly, the PRS corruption indices are not available for the period 2005-2014. In the second place, the CPI of TI is more representative than the index of PRS because it covers a wide range of countries. This keeps sample selection bias restricted. In short, the reliability and relevance of the CPI scores regarding the research question make this measure of corruption suitable to use it in this paper.
12
III.3 Explanatory Variables
The variable of interest, natural resources rents, is measured by the World Bank as a percentage of GDP. This standard of the World Bank is an aggregate of four non-renewable natural resources: oil rents, hard and soft coal rents, mineral rents5 and forest rents. The World Bank’s method which is used to calculate these five resources is reported below: (1) Oil rents: “are the difference between the value of crude oil production at world prices and total costs of production.” (2) Coal rents: “are the difference between the value of both hard and soft coal production at world prices and their total costs of production.” (3) Mineral rents: “are the difference between the value of production for a stock of minerals at world prices and their total costs of production.” (4) Forest rents: “are
roundwood harvest times the product of average prices and a region-specific rental rate.”
The advantages of using this measure are twofold. Firstly, data on volume and price are only selected if they are regularly available (The Changing Wealth of Nations: Measuring Sustainable Development in the New Millennium, 2011). So data from one-off studies are not used. This
strengthens its representativeness. The second criterion used in the measure of natural resources is intended to minimise the selection bias by requiring data which are available for a large number of countries.
Furthermore, the standard of natural resources provided by the World Bank is often used in the literature. Bhattacharyya and Hodler (2010) use these estimates and they find that the natural endowment of a country indeed affects corruption. Treisman (2000) investigates the association between corruption and natural wealth by using only fuel, metals and minerals as a measure for natural resources. We have added coal rents and forest rents to the measure used by Treisman (2000). This way a number of non-renewable natural resources available in a country is represented in a proper way.
Other independent variables used as controls are the unemployment rate (hereafter U), democracy score indicated by Voice and Accountability (hereafter VA) and the gross school
enrollment ratio (hereafter GSER). The unemployment rate is a percentage of total labor force. This national estimate is annually published by the World Bank. It is defined as the share of the labor force that has no job but the unemployed person is available to work and he looks for a job. The second control variable is VA which is estimated by the Economist Intelligence Unit. This standard of democracy rate is an aggregate of four other indices: Vested interests, Accountability of Public Officials, Human Rights and Freedom of association. These four indices are equally weighted to the
5 Ten types of minerals are included in the calculation: gold, silver, phosphate, bauxite, nickel, lead, zinc, iron, copper and tin
13
democracy index alone. The result is a weighted average called VA. The third control variable is a ratio of total enrollment and represents only the primary education through which children can learn mathematics skills, reading and writing. The elementary parts of geography, art, music, natural science, social science and history are included as well.
14
IV. Methodology
At the beginning of this part, the research method is described and the hypotheses are formulated. In the first subsection IV.1, the condition of No Perfect Multicollinearity is discussed. Secondly, the endogeneity bias is clarified. In the last subsection III.5, we check the validity of the instrument used in the Instrumental Variables Regression Model.
The research question is: “Do natural resources affect corruption in the public sector positively?”. We expect to find a positive relationship between corruption in the public sector and natural resources, because – as mentioned in the literature review – it is mostly true that countries with abundant natural resources are highly corrupt. In this case, positive relationship between corruption and natural resources corresponds with a negative coefficient of beta. This can be explained by the counter-intuitive scale of Corruption Perceptions Index (hereafter CPI), which ranges from 0 highly corrupt to 100 very clean. Stated differently, a country with abundant natural resources should have a low CPI score to be labeled as corrupt. In short, low CPI scores and abundant natural resources can be considered as a positive relationship between these two factors and not negative.
Venezuela is a good example of such a relationship. This country’s share of natural resources was 25,7 percent of GDP in 2012, but in the same year, its CPI score was very low, namely 19. This South American country was ranked 165 in the list of Corruption Perceptions Index, which totally contains 176 countries. There are several other countries comparable with Venezuela such as Nigeria6, Papa New Guinea, Congo and Mongolia.
In order to answer the research question, we should test the related hypotheses. The null hypothesis is that there is no corruption in the public sector of a country which has natural resources (H0: β = 0). The alternative hypothesis is that countries with natural resources are highly corrupt (H1: β < 0). Two models are used to test whether we can reject the null hypothesis: Entity and Time Fixed Effects Regression Model and Two-Stage Least Squares (2SLS) Instrumental Variables Regression
6 An example of a country with an abundance of natural resources is Nigeria. Nigerian natural resources rents as a percentage of gross domestic product are 45,41 (The World Bank, 2005). This oil-rich country was labeled as highly corrupt because its score of Corruption Perceptions Index (hereafter CPI) was 19 in 2005, which is 22 points far from the average 41. The CPI scores range from 0 to 100 with lower values indicating high corruption. If corruption is high, the investment level will be low because economic stability is missing. So high corruption leads to declining economic growth (Mauro, 1995). Besides the high corruption in this African country, its GDP per capita growth was only 0,8% in 2005. Nigerian unemployment rate rose from 12.3% in 2006 to 23.9% in 2011 (World Bank, 2011). In spite of the large natural wealth in Nigeria, this country’s economic performance is not as good as expected. There are also other countries with comparable corrupt economies. Angola, Liberia, Gabon and Kuwait are good examples.
15
Model. In both models, there are totally three control variables. Except the fact that we add an interaction term to the OLS-model. Entity and year dummies are added as well.
The first model used in this paper is The Entity and Time Fixed Effects Regression Model which contains four control variables besides natural resources.
CPIit = β0 + β1 NRit + β2 Uit+ β3 VAit+ + β4 (VAit * NRit) + β5 GSERit + αi + λt + εit (1) The variables in this OLS-Model are Corruption Perceptions Index (CPI), Natural Resources (NR), Unemployment rate (U), the democracy score indicated by Voice and Accountability (VA) and Gross School Enrolment Ratio (GSER). The entity is indicated by i and t shows Time. The entity fixed effect is shown by αi in which n – 1 entities are included in the model because they are dummies. The time fixed effect is indicated by λt in which T – 1 time binary indicators are included because it is a binary variable. As we concluded in the literature review, high corruption can be caused by natural resources in case the quality of democratic institutions is poor (Bhattacharyya & Hodler, 2010). This suggests that corruption in a country would interact with the democracy level in such a way that the impact of natural resources on corruption would depend on the democracy level. This is why an interaction term (VAit * NRit) is added in our OLS model.
The second model, used in this paper, is Two-Stage Least Squares (2SLS) Instrumental Variables Regression Model. The formulated hypotheses are also tested by using 2SLS IV-Regression Model because there is a potential concern for the endogeneity bias (subsection IV.2) which make the OLS estimators biased. The first stage of 2SLS IV-Model begins with a regression linking gas and natural resources:
NRit = π0 + π 1 Git + vit (2)
This is the First Stage equation (2) where π0 is the intercept, π1 is the slope, G = gas used as an instrument and vit is the error term. In the second stage of 2SLS, we regress CPI on natural resources and other variables which are indicated below:
CPIit = β0 + β1 NRit + β2 Uit+ β3 VAit+ β4 GSERit + αi + λt + εit (3)
IV.1 Multicollinearity
The estimates are only properly specified and indicate a structural relationship if there is no multicollinearity problem. Multicollinearity makes the estimation results unreliable because of the interdependency between the explanatory variables (Farrar & Glauber, 1967). Farrar and Glauber (1967) state the collinearity must be smaller than r = 0.8 or 0.9 to have trustworthy estimates. The correlations between the explanatory variables used in this paper are smaller than five percent
16
(Table 2). This shows there is no multicollinearity problem. So the collinearity between the independent variables is acceptable and the related signs are credible.
The condition of No Perfect Multicollinearity is also satisfied when we focus on the Variance Inflation Factors (hereafter VIF) of the independent variables. O’Brien (2007) state: “The VIF is based on the proportion of variance the ith independent variable shares with the other independent variables in the model.” In other words, these factors show whether the ith independent variable is correlated with the other explanatory variables and how inflated the standard errors are. If there is perfect multicollinearity then the VIF is larger than 10 and the standard errors are highly inflated (O’Brien, 2007). An independent variable should be dropped to avoid this problem. In our analysis, all VIF are smaller than two, which indicates there is no perfect multicollinearity (Appendix B, Table 9).
IV.2 Endogeneity
A potential concern is the endogeneity bias which leads to biased OLS-estimates. This bias can be caused by various sources such as omitted variables and errors in variables.
Treisman (2000) investigates the causes of corruption in his cross-national study. He finds that traditions of a country, histories of British rule, economic growth level and import affect corruption. Also, the democracy rate of a country and the presence of a federal system determine the extent to which a country is corrupt. The most relevant and complete data of variables are included in this empirical analysis (Table 1). Some of the variables are omitted because they are not available yet or not sufficient enough for our large sample of 164 countries. We avoid this omitted variable bias by using panel data in which the same country is observed at different points in time. To address omitted variable bias as good as possible, an instrumental variable regression is used as well.
Another potential concern is the errors-in-variables bias which can arise because of the independent variable, natural resources. This variable is an aggregate of oil, gas, forest, minerals and coal. Hence, this is possibly an imprecise measure. The instrumental variable regression used in this paper is intended to mitigate errors-in-variables bias.
In short, there is endogeneity bias which is limited by using a 2SLS IV-regression.
IV.3 The validity of instrument
There are two conditions for a valid instrument to be used in the instrumental variables regression: instrument relevance and instrument exogeneity. In the IV-model of this paper gas (hereafter G) is used as an instrument for the natural resources (hereafter NR). This variable satisfies these two conditions. Firstly, the variation in the instrument G is related to the variation in NR. It means the more gas is available in a country, the higher a number of natural resources in a country. G is indeed
17
correlated with the natural endowment of a country: r = 0.2446. This shows there is a correlation. So, the first condition is satisfied: corr(Gi, NRi) ≠ 0. The second condition requires no correlation
between the instrument and the dependent variable, CPI. If there is a correlation between G and CPI, then it should be only via the NR. The correlation between G and CPI is negligible small -0.0645. This estimate is based on 1463 observations which make it fairly representative. The absence of large correlation between G and CPI can be explained by the randomness of the amount of gas available in a country. The availability of gas is naturally determined. So the instrument exogeneity condition is satisfied as well: corr(Gi, ui) = 0.
We perform the Stock-Yogo test to check whether this instrument is indeed a valid and strong instrument. The null hypothesis of this test is that the instruments are weak and the alternative hypothesis is that the instruments are not weak. Stock and Yogo (2001) state that
instruments are only strong if the bias of the Two-Stages Least Square Estimator is not more than ten percent of the bias of the Ordinary Least Squares Estimator. They focus at the First-stage F-statistic and state a critical value between 9.08 and 11.52. Stock and Watson (Introduction to Econometrics, 2015) introduce a rule of thumb of 10 which is based on the findings of Stock and Yogo (2001). In other words, the instrument is weak if the First-stage F-statistic is smaller than 10. In our case the First stage F-statistic is F(1, 928) = 39.66 and Prob > F = 0.0000. This clearly represents a strong instrument because F-statistic is larger than 10. Stock and Watson (Introduction to Econometrics, 2015) state that a large F-statistic means there is more information content provided by the
instrument. Hence, our instrument G contains relevant and sufficient information because of its large First-stage F-statistic.
Moreover, the index of ethnolinguistic fractionalization7 is used as an instrument by Mauro (1995). He investigates the relationship between economic growth and corruption and he uses total investment per GDP (hereafter I) as a measure of economic growth. The index of ethnolinguistic fractionalization satisfies the conditions for a valid instrument in his case because this index is correlated with the indices of corruption and uncorrelated with economic growth. In this paper, an instrument is required which is not correlated with corruption perceptions index as we explore a different topic than Mauro (1995). In short, the ethnolinguistic fractionalization is not a valid instrument in our case.
7 The index of ethnolinguistic fractionalization is defined by Mauro (1995) as a standard which shows the likelihood that two persons - randomly drawn from a population - will not belong to the same ethnolinguistic group. The reason for not using this fractionalization as an instrument in our paper is mentioned above. Moreover, World Bank has a comparable measure called Internally Displaced Persons (hereafter IDP). This estimate is defined by the World Bank as a group of people who must flee or leave their homes because of armed conflicts. IDP is an alternative for the ethnolinguistic fractionalization. However, we do not need it. If we would need IDP, then there are merely 127 observations covering the period 2005-2014. That is not enough.
18
V. Empirical Estimates
There are three subsections in this chapter of the paper. In the first subsection, the regression outputs of Entity and Time Fixed Effects Regression model are discussed. Secondly, there are the results of 2SLS IV-Regression model. In the third and last subsection of this chapter, we use corruption indices of the World Bank to check whether our results are robust.
V.1 OLS Results
In this subsection, we distinguish between three types of results. Firstly, Table 3 presents rough coefficient estimates based on annual regressions from 2005 till 2014. Then there are more representative estimates in Table 4 which show regression output based on panel data of all 164 countries covering the period 2005-2014. At the end of this subsection in Table 5, we use specific resources as an independent variable such as oil, gas, forest, minerals and coal.
All OLS results are summarised in Table 3, 4, 5, 10, 11, 12, 13 & 14. These tables show results for the regressions linking corruption in the public sector (i.e. CPI) with natural resources as a percentage of GDP. Each column of the table shows a different regression. Coefficient estimates of various variables are placed in rows.
First of all, Table 3 shows ten different regressions for recent years. Each one represents the impact of natural resources on corruption in the public sector. If we control for the unemployment rate, VA (Democracy Score) and gross school enrolment ratio, then the coefficient estimate in column 1 of Table 3 is -0.217 which is statistically significant at the 5% significance level. It means a 1% change in natural resources leads to a decrease in the corruption perceptions index with -0.217*100%=-21.7. This regression contains 99 countries in which 54,4% of the variation in CPI can be explained by the variation in natural resources. The average CPI score in 2005 is around 41 (Table 8). So, a 1% change in natural resources in 2005 lowers the average CPI score to 41 – 21.7 = 19.3. This shows that the natural resources have a large impact on corruption in the public sector. It is
disputable how representative this result is as this regression is only based on one year. Only 99 countries are included and there may be omitted variable bias by not taking time and country effects into consideration. Furthermore, it is not confirmed by the coefficient estimates of other years 2006-2014 (columns 2 till 10 in Table 3).
Secondly, regressions based on panel data of 164 countries covering the period 2005-2014 without the time and country effects show different results (column 1 in Table 4). The related coefficient estimate is -0.0549 which is statistically significant at the 5% significance level. The average of CPI scores in these panel data is approximately 41 (Table 1). So, a 1% change in natural
19
resources reduces the average CPI score in the period 2005-2014 to 41 – 5.49 = 35.51. This is less extreme in comparison with those indicated by the regressions in Table 3. In this case, only 0.3% of the variation in CPI scores can be explained by the variation in natural resources. However, when we add country effects, the related R-squared remains exactly the same but the coefficient estimate becomes insignificant. After adding time effects (column 3, Table 4) the related R-squared jumps from 0.3% to 12.9%. In other words, the time effects account for a relatively large amount of the variation in CPI scores. At the same time, its coefficient estimate is not statistically significant anymore as the standard error rises from 0.0279 to 0.0448. The omitted dummy variable of time is 2005.
Furthermore, additional determinants of corruption in the public sector are added in column 4, 5, 6 and 7. Column 4 in Table 4 shows high unemployment rates are associated with high
corruption as its coefficient estimate of -0.229 is statistically significant at the 1% significance level. This estimate reduces the CPI score with 22.9 which indicates high corruption. In other words, this confirms the finding of Mauro (1995). He argues that corruption affects economic development negatively and leads to low investment rates. The high explanatory power of unemployment rate for the corruption in the public sector is also present in column 5, 6 and 7. In these three columns of
Table 4, the democracy score (VA), an interaction term (VA * Natural Resources) and gross enrolment ratio are added as additional control variables respectively. However, the democracy score,
interaction term and gross school enrolment ratio are not statistically significant. The coefficient estimate of natural resources remains insignificant as well. Even though the coefficient estimates of natural resources are not significant, they are negative in column 4, 5, 6 and 7. This indicates that natural resources have a small insignificant effect on corruption in the public sector. Altogether, these results do not provide support for the supposition that corruption in the public sector is caused by natural resources.
In the third place, we investigate the relationship between corruption and the natural resources by using specific resources as independent variables such as oil, gas, forest, minerals and coal instead of an aggregate of natural resources. In Table 5, we use coal8 as a measure of natural resources which delivers significant results. In column 1 of Table 5, coefficient estimate of coal is -0.336 which is significant at the 10% significance level. This regression of Table 5 is without country and time effects. The significance level jumps from 10% to 1% by adding country effects (column 2, Table 5). When we also take time effects into account, the coefficient estimate becomes smaller -0.210 but it is still statistically significant at the 1% level. The coefficient estimate of coal remains
20
statistically significant at the 1% significance level after controlling for unemployment rate (column 4,
Table 5), democracy level indicated by VA (column 5), an interaction term (column 6) and gross school enrolment ratio (column 7). A country with a relatively large amount of hard and soft coal is confronted with high corruption because a 1% change in the amount of coal lowers its CPI score with 26 (column 5, Table 5). This can be explained by the fact that hard and soft coal are highly
concentrated in just a few countries. For example in 2011, the largest amount of coal rents as a percentage of GDP were in Mongolia 19.3%, South Africa 4.1%, Kazakhstan 3,6%, Ukraine 3,1%, China 2,9%, Zimbabwe 2,4%, Indonesia 2,4%, Bosnia and Herzegovina 2,3%, Colombia 1,5%, India 1,5% and a few other countries.
In addition, the detailed regression output of using oil (Table 10), gas (Table 11), forest (Table 12) and minerals (Table 13) as a measure of natural resources are placed in Appendix C. None of these variables have significant results. However, the sign of their coefficient estimate is sometimes negative, which indicates an insignificant small impact of these resources on corruption.
Despite the fact that there is no statistically significant impact of natural resources on corruption except using coal as an independent variable, one can cautiously state that this analysis is reasonably strong. Because important determinants of corruption were added as explanatory variables. In addition, time and country effects are included which softens the threat of omitted variable bias. Time effects control for the unobserved variables that do not change over time.
Country effects are also taken into account. These effects mitigate omitted variable bias in case there are unobserved variables that do not vary across countries.
V.2 2SLS IV-Regression Results
In this subsection, we look further to check whether the estimates of OLS-regressions are consistent. Therefore we use gas as an instrument in 2SLS IV-regression.
The results of 2SLS IV model confirm the findings of OLS-regressions. The coefficient estimate of natural resources is shrinking and remains insignificant after adding country effects (column 2), time effects (column 3), unemployment rate (column 4), democracy score (column 5) and gross school enrolment ratio (column 6) in Table 6. The unemployment rate is added as a control variable which is statistically significant at the 1% level. This stresses the fact that high unemployment rates are associated with the unstable economic development and high corruption (Mauro, 1995). In column 6 of Table 6, we add gross school enrolment ratio as a control variable. Its coefficient estimate 0.0963 is statistically significant at the 1% significance level. This confirms the finding of Mocan (2008) who states that average education is a determinant of corruption. Moreover, the coefficient estimate of natural resources is statistically significant in our OLS-regression without the
21
time and country effects, but the estimate -0.160 of 2SLS IV model is not statistically significant. However, its sign is negative which signals the positive association between natural resources and corruption in the public sector.
To sum up, there is a small negligible and an insignificant effect of natural resources on corruption in the public sector in the OLS-regressions without taking country effects and time effects into account. The estimator of natural resources remains insignificant after controlling for time and country fixed effects and various other relevant variables. However, the OLS results are completely significant when we use coal as a measure of natural resources instead of the aggregate of natural resources. The OLS results in Table 5 show that there is high corruption in countries with a relatively large amount of hard and soft coal. This can be explained by the fact that coal is highly concentrated in just a few countries.
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Table 3Dependent Variable: Corruption Perceptions Index (0=highly corrupt
100=very clean) (1) 2005 OLS (2) 2006 OLS (3) 2007 OLS (4) 2008 OLS (5) 2009 OLS (6) 2010 OLS (7) 2011 OLS (8) 2012 OLS (9) 2013 OLS (10) 2014 OLS Natural Resources -0.217** 0.272*** 0.173 0.236** 0.338** 0.223* 0.279*** 0.111 0.191 0.716*** (0.107) (0.101) (0.104) (0.0921) (0.149) (0.125) (0.106) (0.123) (0.138) (0.224) Unemployment Rate -0.975*** (0.347) -1.032*** (0.292) -0.609** (0.285) -0.375** (0.171) -0.503** (0.207) -0.296* (0.176) -0.687*** (0.207) -0.337* (0.172) -0.306* (0.158) -0.767*** (0.208) VA (Democracy Score) 0.504*** 0.932*** 0.874*** 0.906*** 0.903*** 0.916*** 0.940*** 0.782*** 0.879*** 0.947*** (0.0682) (0.0791) (0.0701) (0.0695) (0.0781) (0.0778) (0.0644) (0.0633) (0.0621) (0.0756) Gross School Enrolment
Ratio -0.156 -0.00381 0.0233 -0.0753 -0.177* -0.121 -0.268** -0.131 -0.0155 0.0350 (0.120) (0.0968) (0.0832) (0.0919) (0.0987) (0.0932) (0.122) (0.136) (0.0701) (0.188) Constant 49.29*** -2.034 -4.410 1.800 14.12 6.071 22.86 19.70 -0.119 -5.417 (14.87) (10.18) (9.204) (10.10) (11.07) (11.70) (13.93) (14.30) (8.847) (19.83) Observations (Countries) 99 97 109 108 104 101 102 103 99 57 R-squared 0.544 0.707 0.704 0.709 0.689 0.728 0.713 0.714 0.775 0.804 Adjusted R-squared 0.525 0.694 0.692 0.698 0.676 0.717 0.702 0.703 0.765 0.789 Robust Standard Errors? Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
23
Table 4
Dependent Variable: Corruption Perceptions Index Scores
(0=highly corrupt 100=very clean)
Model (1) OLS Model (2) OLS Model (3) OLS Model (4) OLS Model (5) OLS Model (6) OLS Model (7) OLS Natural Resources -0.0549** -0.0549 0.0153 -0.0117 -0.0125 -0.00608 -0.0626 (0.0279) (0.0450) (0.0448) (0.0516) (0.0512) (0.0765) (0.0835) Unemployment Rate -0.229*** -0.228*** -0.228*** -0.260*** (0.0793) (0.0795) (0.0797) (0.0948) VA (Democracy score) 0.00646 0.00742 0.00661 (0.0381) (0.0422) (0.0388) VA * Natural Resources -0.000222 0.00125 (0.00185) (0.00189)
Gross School Enrolment Ratio 0.0688
(0.0607) Constant 41.87*** (0.300) Observations 1,574 1,574 1,574 1,138 1,138 1,138 979 R-squared 0.003 0.003 0.129 0.114 0.114 0.114 0.129 Adjusted R-squared -0.113 0.00209 0.124 0.105 0.104 0.104 0.116 Number of countries 164 164 164 146 146 146 138 Years 2005-2014 2005-2014 2005-2014 2005-2014 2005-2014 2005-2014 2005-2014
Country Effects? No Yes Yes Yes Yes Yes Yes
Time Effects? No No Yes Yes Yes Yes Yes
Clustered Standard Errors? No Yes Yes Yes Yes Yes Yes
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
24
Table 5
Dependent Variable: Corruption Perceptions Index Scores
(0=highly corrupt 100=very clean)
Model (1) OLS Model (2) OLS Model (3) OLS Model (4) OLS Model (5) OLS Model (6) OLS Model (7) OLS
Coal (hard & soft) -0.336* -0.336*** -0.210*** -0.263*** -0.260*** -0.663* -0.793*
(0.173) (0.0596) (0.0802) (0.0953) (0.0979) (0.395) (0.436) Unemployment Rate -0.233*** -0.233*** -0.232*** -0.262*** (0.0792) (0.0794) (0.0793) (0.0947) VA (Democracy score) 0.00683 0.00591 0.00797 (0.0382) (0.0384) (0.0352) VA * Coal 0.00902 0.0102 (0.00959) (0.0102)
Gross School Enrolment Ratio 0.0789
(0.0614) Constant 41.39*** (0.100) Observations 1,550 1,550 1,550 1,122 1,122 1,122 977 R-squared 0.003 0.003 0.128 0.115 0.115 0.115 0.131 Adjusted R-squared -0.114 0.00206 0.122 0.106 0.105 0.105 0.119 Number of countries 162 162 162 144 144 144 136 Years 2005-2014 2005-2014 2005-2014 2005-2014 2005-2014 2005-2014 2005-2014
Country Effects? No Yes Yes Yes Yes Yes Yes
Time Effects? No No Yes Yes Yes Yes Yes
Clustered Standard Errors? No Yes Yes Yes Yes Yes Yes
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
25
Table 6
Dependent Variable: Corruption Perceptions Index Scores
(0=highly corrupt 100=very clean)
Model (1) 2SLS IV Model (2) 2SLS IV Model (3) 2SLS IV Model (4) 2SLS IV Model (5) 2SLS IV Model (6) 2SLS IV Natural Resources -0.160 -0.160 0.198 0.0260 0.0154 0.0346 (0.122) (0.122) (0.177) (0.177) (0.179) (0.141) Unemployment Rate -0.240*** -0.238*** -0.260*** (0.0544) (0.0547) (0.0539) VA (Democracy score) 0.0144 0.00739 (0.0174) (0.0179)
Gross School Enrolment Ratio 0.0963***
(0.0294) Constant 43.08*** (1.300) Observations 1,429 1,429 1,429 1,056 1,056 926 Number of countries 162 162 162 143 143 135 Adjusted R-squared 0.113 0.974 0.976 0.980 0.980 0.981 Years 2005-2014 2005-2014 2005-2014 2005-2014 2005-2014 2005-2014
Country Effects? No Yes Yes Yes Yes Yes
Time Effects? No No Yes Yes Yes Yes
Clustered Standard Errors? No Yes Yes Yes Yes Yes
Gas is used as an instrument for natural resources.
Standard errors in parentheses
V.3 Robustness check
To know whether the results obtained from the previous regressions are consistent and credible, we use the corruption perceptions scores of the World Bank as a dependent variable. For this purpose, we test the research question again by exploring OLS- and 2SLS IV-regressions.
According to the robust OLS-regression outputs in Table 14 (Appendix D), high corruption is indeed associated with a large natural endowment when we do not control for time and fixed effects. The coefficient estimate of natural resources -0.0835 in column 1 of Table 14 is statistically significant at the 1% significance level. This shows a relatively stronger impact than the coefficient estimate -0.0549 obtained from the OLS-regression in which the corruption indices of Transparency International are used. After adding country effects, the coefficient estimate remains statistically significant at the 5% level. The previous results are confirmed because the coefficient estimates of natural resources are also here insignificant when we add unemployment rate (column 4),
democracy score (column 5) and gross enrolment ratio (column 6) in Table 14. However, the coefficients of unemployment rates are not statistically significant when we use World Bank’s measure of corruption. At the same time, the coefficient estimate of the interaction term (Va * Natural Resources) becomes statistically significant at the 10% significance level when we control for the unemployment rate, democracy score and gross school enrolment ratio. This confirms the finding of Bhattacharyya & Hodler (2010) who state that natural resources affect corruption in case the democratic institutions are of poor quality.
Using specific natural resources – coal - as an independent variable, there is no impact of coal on corruption in the public sector. None of the coefficient estimates of coal in Table 15 is statistically significant.
The 2SLS IV-regression in Table 16 (Appendix D) partly confirms and partly contradicts the results of 2SLS IV-regression in Table 6. Columns 3, 5 and 6 of Table 16 confirm that natural resources do not affect corruption in the public sector. In these columns, country- and time effects, democracy rate and gross school enrolment ratio are added as control variables respectively. More importantly, the coefficient estimates of natural resources are not statistically significant in these three columns of Table 16. However, using World Bank’s measure of corruption provides larger support for our alternative hypothesis than the measure of Transparency International.
Altogether, the findings of regressions - in which corruption indices of Transparency International are used – are robust to the use of an alternative measure of corruption provided by the World Bank.
27
VI. Conclusion
VI.1 Conclusion
The main research question is: “Do natural resources affect corruption in the public sector positively?” According to the theoretical evidence presented in the literature review, natural
resources have an impact on corruption in the public sector. Bhattacharyya & Hodler (2010) propose an important game-theoretic model and argue that politicians may be inclined to become corrupt in case there are a lot of natural resources. In addition, Collier (2010) states that natural resources make politicians more interested in their own revenues instead of maximising social welfare. This leads to worse governance which in turn make corruption in the public sector possible.
Empirically, the OLS- and 2SLS IV regressions are used to test whether there is indeed high corruption in the public sector of a country with natural resources. Annual regressions without the time and country effects provide strong evidence for this claim (Table 3). It is confirmed by an OLS-regression using panel data with no country and time effects (Table 4). Using coal as a measure of natural resources also provide statistically significant estimates. The OLS results in Table 5 show that countries with hard and soft coal are highly corrupt. This can be explained by the fact that hard and soft coal are highly concentrated in just a few countries.
In contrast, the claim that corruption in the public sector is caused by the natural resources is neither supported by the complete Entity and Time Fixed Regression Model nor by the 2SLS IV-Regression Model. This evidence is also missing when we control for the effects of the
unemployment rate, democracy rate and gross school enrolment ratio. It is important to mention that the findings of Mauro (1995) are confirmed as the unemployment rates have a significant effect on corruption in the public sector (Table 4, 5 and 6). The support for the finding of Mocan (2008) is provided as well who states that average education level determines corruption (Table 6).
The coefficient estimates of natural resources are robust to the use of World Bank’s measure of corruption. Specifically, there is some evidence for our alternative hypothesis when we focus on OLS- and 2SLS IV-regressions in which World Bank’s standard of corruption is used. This support is limited since it only holds true in case we control for country effects. Specifically, there is no evidence for alternative hypothesis when we control for time effects and various other control variables. Stated differently, the results obtained from the OLS- and 2SLS IV-models are robust to the use of an alternative measure of corruption. However, we partly confirm the finding of Bhattacharyya & Hodler (2010) who state that natural resources affect corruption in less democratic countries. Their finding cannot be fully confirmed because the interaction term is not statistically significant when we use the corruption perceptions index scores provided by the Transparency International.
28
To put it briefly, there is strong empirical evidence that coal rents affect corruption in the public sector when we use coal as an independent variable. In contrast, there is no empirical evidence that all natural resources jointly affect corruption in the public sector positively. Further research is required to make sure this is true.
VI.2 Discussion
Based on the findings of our empirical analysis, there is no causation between corruption in the public sector and natural resources except coal. Before generalising these findings, there are some limitations. Although the corruption perceptions indices of Transparency International are often used in the literature, they are not an appropriate measure of corruption in the public sector. These indices are mainly based on surveys. Hence, they are subjective. An objective measure of corruption is needed to test this relationship in a proper way.
According to the results of Stock-Yogo test, gas is a valid and strong instrument. So we can use it in 2SLS IV-regression. However, finding a really valid instrument is a real challenge for all researchers. The exogeneity of the instrument used in this paper is absolutely debatable. The empirical analysis will be even more representative when another instrument is used which satisfies the condition of instrument exogeneity even strongly.
29
Appendix A: Description of the Data
Table 1: Descriptive Statistics of Regression Variables
Variable Observations Mean Std. Dev. Min Max
CPIa (0=highly corrupt
100=very clean) 1604 40.90524 20.94108 8 97 Natural Resources (% of GDP) 1609 10.14422 13.9293 0.0003721 76.528 Unemployment Rate (% of total labor force) 1159 8.346764 5.985941 0.1 47.5 VAb (0=autocratic 100=democratic) 1640 47.52375 24.38052 0 96.525 GSERc (% of total enrollment) 1349 103.4214 14.04327 29.2021 149.9517 Oil (% of GDP) 1458 6.3373 13.36362 0 73.33453 Gas (% of GDP) 1463 1.561274 5.069468 0 70.51028 Forest (% of GDP) 1580 2.488754 4.780937 0 43.69727 Mineral (% of GDP) 1599 1.785978 4.785812 0 44.64384 Coal (% of GDP) 1584 0.1857716 0.8600293 0 19.25892
World Bank’s CPI d
(0=highly corrupt 100=very clean)
1640 45.48659 22.70695 5 100
a CPI: Corruption Perceptions Index measure by Transparency International. b VA: Voice and Accountability indicate the democracy score
c GSER: Gross School Enrolment Ratio
d World Bank’s CPI: the measure of corruption estimated by the World Bank.
Table 2: Correlation Matrix (Obs=914) CPIa Natural
Resources Unemployment Rate VA
b GSERC Oil Gas Forest Mineral Coal WB
CPId CPIa 1.0000 Natural Resources -0.3537 1.0000 Unemployment Rate -0.0916 -0.0882 1.0000 VAb 0.7860 -0.5051 0.0486 1.0000 GSERc -0.0332 -0.0725 -0.0340 0.0261 1.0000 Oil -0.2301 0.8667 -0.0893 -0.3993 -0.0473 1.0000 Gas -0.1040 0.2446 -0.0827 -0.1254 -0.0073 0.3282 1.0000 Forest -0.2745 0.2951 -0.0964 -0.3126 -0.0809 -0.0865 -0.1000 1.0000 Mineral -0.1471 0.3278 0.0659 -0.1072 -0.0389 -0.0579 -0.0670 0.1662 1.0000 Coal -0.1157 0.1839 0.0604 -0.0766 0.0997 0.0183 -0.0177 -0.0551 0.3306 1.0000 WB CPId 0.9706 -0.3799 -0.0550 0.7984 -0.0248 -0.2574 -0.1121 -0.2805 -0.1422 -0.1249 1.0000
a CPI: Corruption Perceptions Index measure by Transparency International.
b VA: Voice and Accountability indicate the democracy score
c GSER: gross school enrolment ratio
30
Figure 1a: Annual Plots
Figure 1b: Fitted Values
31
Figure 2: The relationship between CPI and Natural Resources in 2014
32
Table 7: List of countries used in the empirical analysis (alphabetic order)
Nr. Country Nr. Country Nr. Country Nr. Country
1 Afghanistan 42 Djibouti 83 Kyrgyzstan 124 Russia
2 Albania 43 Dominican Republic 84 Latvia 125 Rwanda
3 Algeria 44 Ecuador 85 Lebanon 126 Sao Tome and Principe
4 Angola 45 Egypt 86 Lesotho 127 Saudi Arabia
5 Argentina 46 El Salvador 87 Liberia 128 Senegal
6 Armenia 47 Eritrea 88 Libya 129 Serbia
7 Australia 48 Estonia 89 Lithuania 130 Seychelles
8 Austria 49 Ethiopia 90 Luxembourg 131 Sierra Leone
9 Azerbaijan 50 Finland 91 Madagascar 132 Singapore
10 Bahrain 51 France 92 Malawi 133 Slovakia
11 Bangladesh 52 Gabon 93 Malaysia 134 Slovenia
12 Belarus 53 Gambia 94 Mali 135 Somalia
13 Belgium 54 Georgia 95 Malta 136 South Africa
14 Benin 55 Germany 96 Mauritania 137 Spain
15 Bhutan 56 Ghana 97 Mauritius 138 Sri Lanka
16 Bolivia 57 Greece 98 Mexico 139 Sudan
17 Bosnia and Herzegovina 58 Guatemala 99 Moldova 140 Suriname
18 Botswana 59 Guinea 100 Mongolia 141 Sweden
19 Brazil 60 Guinea-Bissau 101 Montenegro 142 Switzerland
20 Bulgaria 61 Guyana 102 Morocco 143 Syria
21 Burkina Faso 62 Haiti 103 Mozambique 144 Tajikistan
22 Burundi 63 Honduras 104 Myanmar 145 Tanzania
23 Cabo Verde 64 Hong Kong 105 Namibia 146 Thailand
24 Cambodia 65 Hungary 106 Nepal 147 Timor-Leste
25 Cameroon 66 Iceland 107 Netherlands 148 Togo
26 Canada 67 India 108 New Zealand 149 Trinidad and Tobago
27 Central African Republic 68 Indonesia 109 Nicaragua 150 Tunisia
28 Chad 69 Iran 110 Niger 151 Turkey
29 Chile 70 Iraq 111 Nigeria 152 Turkmenistan
30 China 71 Ireland 112 Norway 153 Uganda
31 Colombia 72 Israel 113 Oman 154 Ukraine
32 Comoros 73 Italy 114 Pakistan 155 United Arab Emirates
33 Congo, Dem. Rep. 74 Jamaica 115 Panama 156 United Kingdom
34 Congo, Rep. 75 Japan 116 Papua New Guinea 157 United States
35 Costa Rica 76 Jordan 117 Paraguay 158 Uruguay
36 Cote d'Ivoire 77 Kazakhstan 118 Peru 159 Uzbekistan
37 Croatia 78 Kenya 119 Philippines 160 Venezuela
38 Cuba 79 Korea (North) 120 Poland 161 Vietnam
39 Cyprus 80 Korea (South) 121 Portugal 162 Yemen
40 Czech Republic 81 Kosovo 122 Qatar 163 Zambia
33
Table 8: Averages* of Corruption Perceptions Index Scores per year 2005-2014
Year 2005 2006 2007 2008 2009
CPI Average Score 41.06711 40.93464 39.90741 40.01235 40.04969
Year 2010 2011 2012 2013 2014
CPI Average Score 40.11801 39.79878 42.54268 42 42.58537
*Note: These averages may differ from those published by Transparency International because our sample contains only 164 countries and thus not all the countries included by Transparency International. See Table 6 on the previous page to know which countries are included in this sample.
Appendix B: Variance Inflation Factors
Table 9: No perfect multicollinearity confirmed by Variance Inflation Factors (VIF)
Abbreviations, used in Table 8, are:
va is Voice and Accountibility which indicate democracy rate; natreszgas shows natural resources;
unem presents unemployment rate;
Appendix C: Natural Resources Specified: Oil, Gas, Forest & Minerals.
Table 10
Dependent Variable: Corruption Perceptions Index Scores
(0=highly corrupt 100=very clean)
Model (1) OLS Model (2) OLS Model (3) OLS Model (4) OLS Model (5) OLS Model (6) OLS Oil -0.0146 -0.0146 0.0990 0.0145 0.0131 -0.0304 (0.0330) (0.0590) (0.0614) (0.0739) (0.0734) (0.0786) Unemployment Rate -0.238*** -0.237*** -0.260** (0.0850) (0.0852) (0.0998) VA (Democracy score) 0.0126 0.00791 (0.0358) (0.0343)
Gross School Enrolment Ratio 0.0993
(0.0683) Constant 41.45*** (0.231) Observations 1,430 1,430 1,430 1,066 1,066 925 R-squared 0.000 0.000 0.120 0.122 0.123 0.125 Adjusted R-squared -0.127 -0.000546 0.114 0.113 0.113 0.113 Number of countries 161 161 161 144 144 136 Years 2005-2014 2005-2014 2005-2014 2005-2014 2005-2014 2005-2014
Country Effects? No Yes Yes Yes Yes Yes
Time Effects? No No Yes Yes Yes Yes
Clustered Standard Errors? No Yes Yes Yes Yes Yes
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
35
Table 11
Dependent Variable: Corruption Perceptions Index Scores
(0=highly corrupt 100=very clean)
Model (1) OLS Model (2) OLS Model (3) OLS Model (4) OLS Model (5) OLS Model (6) OLS Gas -0.0482 -0.0482 0.0416 0.00911 0.00535 0.0233 (0.0365) (0.0575) (0.0411) (0.0655) (0.0639) (0.109) Unemployment Rate -0.237*** -0.236*** -0.259*** (0.0823) (0.0826) (0.0983) VA (Democracy score) 0.0146 0.00785 (0.0361) (0.0346)
Gross School Enrolment Ratio 0.0967
(0.0676) Constant 41.45*** (0.112) Observations 1,429 1,429 1,429 1,056 1,056 926 R-squared 0.001 0.001 0.104 0.108 0.109 0.112 Adjusted R-squared -0.126 0.000677 0.0975 0.0987 0.0986 0.0993 Number of countries 162 162 162 143 143 135 Years 2005-2014 2005-2014 2005-2014 2005-2014 2005-2014 2005-2014
Country Effects? No Yes Yes Yes Yes Yes
Time Effects? No No Yes Yes Yes Yes
Clustered Standard Errors? No Yes Yes Yes Yes Yes
