The impact of trade openness on perceived corruption for OECD members and key partners between 1996-2016

The impact of trade openness on perceived

corruption for OECD members and key partners

between 1996-2016

University of Amsterdam

Amsterdam School of Economics

Faculty of Economics and Business

Name: Yannick Bisschops Student number: 10175865

Contact: yannick_bisschops@hotmail.com Date: 15-07-2018

Number of words: 14854

Course: Master Thesis Economics

Track: International Economics and Globalisation Supervisor: drs. N.J. Leefmans

Statement of originality

This document is written by student Yannick Bisschops who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.

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

Abstract

This thesis investigates the impact of trade openness on perceived corruption for all OECD countries and key partners for the time period 1996-2016. Perceived corruption is measured by the Corruption Perception Index (CPI). Six fixed effect regressions have been conducted using panel data to test three hypotheses. First, three regressions are conducted that assume a linear relationship between the explanatory variables of interest and corruption. The first regression estimates the impact of trade on perceived corruption. The second regression estimates the impact of exports on perceived corruption. The impact of imports on corruption is estimated by the third regression. When relaxing the assumption of linearity, the square of these explanatory variables is added to the fourth, fifth and sixth regressions. This thesis controls for reverse causality between corruption and trade, imports and exports by using the lagged values of these variables. The six regression models show that neither trade, nor exports, or imports have a significant impact on corruption. From the descriptive analyses it is shown that the levels of perceived corruption decreases over time, independently of international trade, export and import as a share of GDP.

Table of content

1. Introduction 5

2. Literature review 9

2.1 Literature Review on Corruption and its effect on imports and exports 9

2.2 Hypotheses 14

3. Methodology 15

3.1 Models 15

3.1.1 The regression models 15

3.1.2 The model 16

3.2 Data 17

3.2.1 Corruption 18

3.2.2 Explanatory variables of interest 20

3.2.3 Control variables 21

3.2.4 Descriptive analyses 23

3.3 Assumptions 29

3.4 Conclusion 31

4. Results 31

4.1 Results regression 1 and 4 32

4.2 Results regression 2 and 5 37

4.3 Results regression 3 and 6 38

4.4 Robustness Check 38 5. Conclusions 40 6. Discussion 42 7. References 43 8. Appendix A 45 Appendix B 46 Appendix C 46 Appendix D 49 Appendix E 58 Appendix F 59 Appendix G 63 Appendix H 69

1. Introduction

Corruption is known all over the world. In some countries it is more extensive than in others, but there are corrupt activities worldwide, for instance, a bouncer at a nightclub is paid off to skip a line, a police officer that receives a bribe to disregard a violation, or when customs officials receive a bribe. The basic element of corruption is large discretionary power to officials (Jain, 2001). Corruption decreases economic growth rates and is linked with distortions in the composition of government expenditures (Mauro, 1998). Figure 1 is constructed to show the average scores of perceived corruption for members of the Organisation of Economic Cooperation and Development (OECD). Figure 2 shows the average score of perceived corruption of the five key partners of the OECD. Figure 1 and 2, measured by the Corruption Perception Index (CPI) show average perceived corruption scores per country (see appendix A for the specific values). CPI uses a scale of 0 to 100, where 100 is very clean and 0 is highly corrupt. Due to this, high values correspond with low perceived corruption and low values correspond with high perceived corruption. Figure 1 shows that even members of the OECD face perceived corruption. Although most OECD countries have high average perceived corruption scores, there are also countries with lower perceived corruption scores (e.g. Greece, Mexico and Turkey). However, the five key partners of the OECD all show low perceived corruption scores, indicating that the level of perceived corruption is higher in comparison to the OECD countries.

The reason that OECD members have mainly low perceived corruption levels might be because these countries are highly involved in international trade. There has been excessive research into the impact of corruption on international trade and mostly a negative impact is found, however the research of the reverse relationship, namely; the impact of international trade on corruption is limited. The way international trade has an impact on corruption is described in prior literature as follows. If a tariff is present, foreign firms serving the domestic market have an incentive to evade this tariff. If trade barriers to international transactions are high, private agents are encouraged to bribe public officials (Gatti, 1999), while corruption will decrease when trade barriers disappear. Gatti (2004) shows two effects of trade barriers on corruption. The first is the foreign competition effect, which is measured by the share of imports in GDP. The presence of foreign competition can put pressure on the domestic sector and squeeze the margins for rent-seeking, as a consequence reducing corruption.

Figure 1. Average perceived corruption scores for all OECD countries from 1996-2016 given by CPI.

Source: Figure constructed by Y.A. Bisschops using data from Transparency International (2017)

Figure 2. Average perceived corruption scores for all Key partners from 1996-2016 given by CPI.

Second, the direct policy effect, is measured by using two indicators: the trade tariff average level, and the percentage of imports goods subject to quota

restrictions. Such trade restrictions might create opportunities for collusive interactions between public officials and private agents, which might lead to favourable treatments in exchange for bribes. Gatti (2004) finds that this effect is twice as large as the foreign competition effect.

Moreover, trade liberalisation of the market of goods and services increases competition, transparency and quality of public sector services; that is policies appropriate to fight against the elements that strengthen public sector corruption. Due to this, corruption levels might decrease when countries open up to globalisation (Lalountas, Manolas & Vavouras, 2011).

A structural argument could be bribes as a tax on the importer or domestic producer. The importer sets the imports price above the international price if the market is closed. A bribe-taking official could now share in the monopoly profits. When the market is open to trade, the bribery tax forces the price to decrease below the level prevailing in the market, so the taxed producers will drop out (Sandholtz & Koetzle, 2000). The increased competition will penalize bribery. Additionally, freer trade removes certain administrative goods (e.g. licences and permits) from officials that could be exchanged for private rewards. A cultural argument is given through socialisation. Corruption can be decreased through a higher involvement in international trade by strengthening economic competition, which leads to diminishing opportunities for corruption, and by socializing actors into the especially Western norms of the international economy (Sandholtz & Koetzle, 2000).

The studies that already investigated the impact of international trade on corruption are the following: Lalountas, Manolas & Vavouras (2011); Das and DiRienzo (2009); Sandholtz and Koetzle (2000); Dutt (2009); Gatti (2004); Treisman (2000), and most of these studies focus on the high- and middle-income countries. To my knowledge there has been no research that investigates the impact of international trade on corruption using the most recent data, since the most recent study uses data until 2008. Furthermore, some of the prior studies account for only one year, or limited years: Lalountas, Manolas & Vavouras (2011); Das and DiRienzo (2009); Sandholtz and Koetzle (2000). Therefore, this thesis will investigate the impact of international trade on corruption using a longer time period (1996-2016) including more recent data by using fixed effect regression models. Moreover, some of these studies do not account for reverse causality, stationarity, and none of these studies investigate if there is a different effect of imports and exports on perceived corruption, which is what this thesis will add. Additionally,

a non-linear relationship between the explanatory variables of interest and perceived corruption will be tested. Furthermore, this thesis will also compare the results when reverse causality and stationarity are neglected.

The research question of this thesis is the following: Does an increase in trade openness lead to lower perceived corruption levels for members of the Operation of Economic Cooperation and Development (OECD) and five key partners for 1996-2016? The prior research has already investigated the relationship for high- and middle- income countries, however as stated above most of these studies do not account for a unit root and some do not account for reverse causality. For this reason, this thesis focuses on the OECD countries. The OECD strengthened cooperation with Brazil, India, Indonesia, the People’s Republic of China and South Africa through “Enhanced Engagement” programmes since May 2007. These key partners commit to the OECD’s work in a constant and comprehensive manner and will be included in this thesis (OECD, 2018). Due to the fact that OECD countries have low corruption levels, the inclusion of the five key partners is done to create some variation in the corruption levels.

This thesis investigates the impact of international trade on public corruption by using fixed effect regression models. This thesis will also study if the impact of imports on perceived corruption differs from the impact of exports on perceived corruption. Furthermore, this thesis will also show if linearity of trade, exports, and imports has a significant effect on perceived corruption.

The corruption level of a country will decrease when competition from imports rises (Treisman, 2000). Leite and Weidmann (1999) show that economies that are abundant in natural resources, thus exporting a lot, have higher levels of corruption. This might not only be the case in countries that are abundant in natural resources, but also in countries that are exporting a lot of goods and services. These results show that there might be a different effect of imports and exports on perceived corruption. This will therefore also be investigated in this thesis.

Prior literature showed reverse causality between corruption and international trade (Ades and Di Tella, 1999; Damania, Fredriksson & List, 2003; Dutt, 2009). To account for this, the lagged value of international trade, exports, and imports will be used. Panel data is used for the OECD countries and five key partners for a period between 1996-2016 to fixed effect regressions.

Transparency International (TI) provides the data for perceived corruption using the Corruption Perception index (CPI). Data for trade, exports, imports, Gross Domestic Product (GDP) per capita, tariff, and subsidies comes from the World Bank. Political rights and civil-liberties are

both indicators for democracy and the data used is from Freedom House. Economic freedom data comes from Heritage Foundation.

This thesis is structured as follows. Section two presents a literature review of the prior research that has been done about trade, globalisation and the impact on corruption. Thereafter, the hypotheses will be presented. Section three will give the methodology, data, definitions of specific concepts, descriptive data analyses, and the model used will be specified. Section four presents the results. Furthermore, a robustness check is conducted to check for the plausibility and robustness of the estimates. Section five concludes and finally section six provides a discussion.

2. Literature review

In this section prior research on the effect of trade on corruption will be described. All these studies use corruption as the dependent variable. After the discussion of these papers, a short summary with the main conclusions will be given. At the end of this section hypotheses will be presented.

2.1. The effect of international trade on corruption

This section describes the prior literature of the impact of international trade on corruption. The research question, used models, variables used, and the conclusions of each research will be given. Most of these studies focus on high- and middle-income countries which is controlled for by using dummies. This section will also indicate if the prior research controls for the reverse causality problem. Furthermore, when studies use panel data, this thesis will indicate if they account for a unit root.

The research of Lalountas, Manolas & Vavouras (2011) answers the question: Do countries with a high degree of globalisation also show a lower degree of corruption in relation to the countries which are less open to globalisation? In this research cross-section data of the year 2006 for 127 countries is used. The dependent variable is measured by the Corruption Perception Index (CPI). The independent variables are: globalisation (measured by the KOF index), GDP per capita, political rights and civil liberties (democracy variable measured by the index of US-based Freedom House), and Regulation of credit, labour, and business (measures the degree to which policies and institutions of countries are supportive of economic freedom). The KOF index of globalisation is introduced by Dreher in 2002 and consists of weighted variables of flows (referring to trade and investment) and restrictions (barriers to trade and

capital). Because GDP per capita and globalisation are endogenous, the instrumental variables used for globalisation are the human development index (HDI), which combines the measures of literacy, standard of living and life expectancy, and the lagged globalisation index (KOFLAG), which is the lagged KOF variable with 6 period lags. The impact of globalisation on corruption is estimated by both ordinary least squares (OLS) and two stage least squares (2SLS) methods. The impact of globalisation on corruption is found to be positive and significant for OLS. When 2SLS is used the effect of globalisation increases and remains significant, while the impact of GDP per capita is insignificant. Only under the hypothesis that the model is linear in endogenous variables, especially GDP per capita, this conclusion is valid. They also test a model that is non-linear in endogenous variables, using the instruments KOFLAG2 _{and HDI}2_{. The impact of globalisation on corruption is not significant when the} model is non-linear in endogenous variables. It is found that linearity is only present in middle- and high-income countries and not in low-income countries. To distinguish between high-, middle- and low-income countries, dummies have been used. Since this research only uses data of the year 2006, the results should be interpreted with caution.

The paper of Das and DiRienzo (2009) also relaxes the assumption of a linear relationship between globalisation and corruption. The hypothesis of the paper is: globalisation could have different effects on country corruption levels, which are dependent on the country’s level of globalisation. If a country has low levels of globalisation and is in transition to increase the level of globalisation, the country’s corruption levels could be affected. For this reason, the assumption of a linear relationship between globalisation and corruption is relaxed. This paper takes 113 countries into account with corruption measured by the CPI in 2008. The independent variable is globalisation, measured by the KOF index of Globalisation in 2005. The control variables are democracy, measured by Freedom House in 2005, Economic freedom measured by the 2005 Heritage Foundation’s index of Economic Freedom, and the Fractionalization index created by Alesina et al (2003) is used to control for ethnic, linguistic, and religious diversity within a country. The last control variable is Economic Development, which is measured by 2005 GDP per capita (measured in constant 2000 U.S. dollars). Two different regression models have been conducted.

The first model is a regression of globalisation and the control variables on corruption. Globalisation is found to have a significant and negative linear effect on corruption, as indicated by a positive sign for globalisation. All the control variables are significant except for democracy and religious diversity. Except for linguistic diversity, all the significant variables have the expected sign.

The second model adds squared globalisation to the regression. No significant changes have been found between model one and two for the control variables. The variables globalisation and squared globalisation are both significant. The most important change is that the sign of globalisation changes from positive to negative. Squared globalisationhas a positive sign.

These results show that first corruption levels increase when a country starts to globalise. New opportunities for corrupt practices have been created due to the formation of new trade relationships. Nevertheless, when countries keep integrating in the world economy, they face increased regulation by the anti-corruption policies of the supranationals. This decreases their corruption levels. Thus, countries that are transitioning to more globalisation are found to have the highest corruption levels.

Another research that takes only one year into account is that of Sandholtz and Koetzle (2000). They propose that corruption levels are higher, the lower the degree of integration in the world economy. They do not control for reverse causality between the explanatory variables and corruption and use data of the year 1996 for fifty countries. The hypothesis is tested by a set of multivariate regression models where the dependent variable is measured by the CPI. Model one shows that average income (GDP/cap), economic freedom (measured by the Freedom House Indicators of Economic Freedom), democratic years (the number of years since 1948 that countries have democratic rule), democracy (measured by the Freedom House index of political rights and civil liberties), and trade integration (the sum of imports and exports as a share of GDP) are all found to be significant and negatively related to perceived corruption. In the second model ln(GDP) is added as a control variable for the country size but is insignificant. All other variables remain significant. The third model adds the British colony but is insignificant. Furthermore, democratic years become insignificant. All other variables remain significant. Model four with democracy, trade, economic structure and Protestantism variables, explains the largest share of variance in perceived corruption. The most powerful predictors of corruption levels are trade and economic freedom. A higher involvement in international trade can decrease corruption by strengthening economic competition, diminishing the opportunities for corruption, and socializing actors into the especially Western norms of international economy. Because the authors neglect to account for the reverse causality problem and only use data from 1996, these estimates should also be interpreted with caution.

The research of Dutt (2009) does account for reverse causality by using the 2SLS method. Furthermore, GMM estimators have been used to account for the potential endogeneity of trade policy. The research question of this study is the following: Are countries with lower

policy-induced barriers to international trade less corrupt, after other relevant country characteristics are controlled for?

The data of the dependent variable is collected from the International Country Risk Guide (ICRG) from 1984 until 2004, and also the CPI of 2000 has been used. The independent variables used are: total imports duties collected as a percentage of total imports, an average tariff rate, coverage ratio for non-tariff barriers to trade (Quota), the average tariff rate and Quota are both from Barro and Lee (1993), number of years outside the GATT since 1948, democracy (Gastil index), schooling (World Development Indicators, 2000), and number of newspapers per thousand of the population (World Development Indicators, 2000), index of regulation that measures the licencing requirements faced by domestic firms in the economy (Heritage Foundation, 1998), index of wage and price regulation (Heritage Foundation, 1998), subsidies (World Development Indicators, 2000), and the ratio of civil-sector to manufacturing sector wages.

Methods used are panel data regression OLS, ordered probit, and 2SLS. From OLS using the CPI they find that countries with higher tariffs in 1980s and 1990s are also more corrupt. Non-tariff barriers in the 1990s also lead to higher corruption levels, while these barriers the for 1980s are insignificant. Import duties are also insignificant.

The ordered probit method shows all measures of trade policies are highly significant, apart from the quotas from the 1980s. Subsidies and licencing are strongly significant in all the regressions and have the predicted sign. Political liberty and number of newspapers are significant, but wage-price regulation and schooling are not significant.

Finally, 2SLS is used to deal with endogeneity of trade restrictions, with respect to corruption. The instrumental variables used are political ideology, ideology multiplied by country’s capital-labour ratio, and capital-labour ratio. They concluded that trade policies have a significant effect on corruption even if they account for endogeneity.

Dutt (2009) uses the generalized method of moments (GMM) difference estimator. Due to the use of differences, the data is stationary. However, Dutt (2009) only mentions the use of differences when GMM estimators are used, the paper does not mention the possibility of a unit root in all the other methods. Trade liberalization could decrease corruption, since a small consistent relationship between protectionist policies and corruption is shown.

The research of Gatti (2004) focuses on whether the presence of barriers to international trade and international capital flows is associated with higher levels of corruption. Gatti (2004) captures two effects namely; the foreign competition effect and the direct policy effect. The measure for foreign competition effect is given as a share of imports of GDP. The direct policy

effect is captured with the use of two indicators: the average level of trade tariff, and the percentage of imports goods subject to quota restrictions. The dependent variable is measured by ICRG for the years 1982-2000. Gatti (2004) does not account for an indirect policy effect. The variables used for the basic model are the following; ln(GDP per capita), democracy, ln(population), and a time trend is included.

OLS is used to conduct the estimations. The standard errors are corrected for clustering and heteroscedasticity since correlation across period is allowed for. Imports shares have a positive effect on corruption, however this is highly dependent of the inclusion of Singapore. Trade tariffs can affect corruption through two channels: by decreasing imports, and by creating incentives to collusive behaviour between firms and customs officials to circumvent these restrictions. Higher trade barriers are significantly associated with corruption even after imports shares are controlled for, suggesting that the impact of trade restrictions on corruption comes mainly through the direct policy effect. Democracy is negatively correlated with corruption. However, these results should be interpreted with caution, since this paper does not account for the possibility of a unit root in the data.

Two interesting hypotheses are offered by Treisman (2000): first, corruption will be lower when the exposure to competition from imports of a country increases. Second, countries with larger endowments of valuable natural resources will have higher corruption. Countries that are dependent on the export of raw materials, often centralize economic power, this may reduce democratic stability and therefore increase corruption (Treisman, 2000)

The dependent variable is measured by CPI of 1996, 1997, and 1998 and also uses the ICRG for the 1980s. The following explanatory variables are used: Common law system, former British colony or UK, Never a colony, percent protestant 1980, ethnolinguistic division, (fuel, metal, and minerals exports), log GDP per capita, federal, uninterrupted democracy 1950-1995, Imports/GNP %, state intervention, government wage, government turnover.

A weighted least squares method is used. Openness to trade is endogenous, more imports may reduce corruption, but the likelihood that corrupt officials create rent-generating barriers to trade increases. A convincing instrument to test with 2SLS is not found in this research. Treisman (2000) concludes that openness to foreign trade reduces corruption.

When a large share of exports consists of fuel, metals and minerals, countries used to have higher corruption levels in the 1980s, even when economic development and democracy are controlled for. This effect dampens in the regressions of the 1990s, due to the control for economic development and democracy. Poorer countries have a higher share of raw materials

exports and poverty increases corruption. The possibility of a unit root in the data is not mentioned by Treisman (2000).

In sum, the research of Lalountas et al. (2011) obtains a positive significant coefficient of globalisation on the perceived corruption score of the CPI only for middle and high-income countries, the impact for low-income countries is found to be insignificant. This means a negative effect has been found of globalisation on corruption for high- and middle-income countries. Das and DiRienzo (2009) conclude that globalisation has a positive as well as a negative effect on corruption, depending on the level of globalisation of the country. It is found that the highest level of corruption arises in countries transitioning to globalisation. Furthermore, Sandholtz and Koetzle (2006) find a negative impact of international trade on corruption. It is found that a higher involvement in international trade can decrease corruption by strengthening economic competition, diminishing the opportunities for corruption, and socializing actors into the especially Western norms of international economy. Gatti (2004) finds that the foreign competition effect is highly sensitive to the regression specification and shows the negative impact of trade restrictions on corruption comes mainly through the direct policy effect. Dutt (2009) and Treisman (2000) show that trade liberalization has a negative effect on corruption. Concluding, only Das and DiRienzo (2009) find that the highest levels of corruption arise in countries transitioning to globalisation, all other studies find that an increase of trade openness leads to lower levels of corruption. As indicated, many of these conclusions should, however be interpreted with caution since most of these studies do not account for reverse causality and the presence of a unit root

2.2 Hypotheses

This thesis tests three hypotheses regarding the OECD countries and its five key partners. The first hypothesis is: an increase in international trade reduces the perceived corruption level of a country. An increase in exports reduces the perceived corruption level of a country, is the second hypothesis. The last hypothesis is: an increase in imports reduces the perceived corruption level of a country. The methodology used to test these hypotheses is elaborated in the next section.

3. Methodology

This section will discuss the following. Firstly, the regression models are presented and discussed. Secondly, the origin of the data, the meaning of the variables and a descriptive analysis of the data will be elaborated. Thereafter, the assumptions to use the models are presented and tested. Lastly, conclusions concerning the model will be drawn.

3.1 Models

Below, the regression models and the choice of econometric model is elaborated.

3.1.1 The regression models

This thesis focuses on the dependent variable corruption. To test the three hypotheses given above, three regression models will be applied, when a linear relationship between the explanatory variables of interest; trade, exports, and imports, and the dependent variable; corruption, is assumed. The explanatory variables of interest are measured as a percentage of GDP. The prior literature shows that the variables GDP per capita, democracy,export subsidies, economic freedom, and tariffs often have a significant effect on corruption. For this reason, these explanatory variables have been included as control variables. Panel data are used for the following fixed effect regressions:

Model 1

L1corruptioni,t= a + wi+ gt + b1 l2dtradei,t-1 + b2 ln(l1GDPi,t) + b3 l1democracyi,t + b4l1subsidiesi,t + b5 l1EconomicFreedomi,t+b6 l1tariffi,t +eI,t

Model 2

L1corruptioni,t= a + wi+ gt + b1 l2dexportsi,t-1 + b2 ln(l1GDPi,t) + b3 l1democracyi,t + b4 l1subsidiesi,t + b5 l1EconomicFreedomi,t+b6 l1tariffi,t +eI,t

Model 3

L1corruptioni,t= a + wi+ gt + b1 l2dimportsi,t-1 + b2 ln(l1GDPi,t) + b3 l1democracyi,t + b4 l1subsidiesi,t +b5 l1EconomicFreedomi,t+b6 l1tariffi,t +eI,t

This thesis will next also relax the assumption of a linear relationship between the explanatory variables of interest and the dependent variable. These regressions will show if a non-linear relationship between the explanatory variables of interest and corruption has a significant effect on perceived corruption levels of a country. The squares of the variables of interest are added to the model. The following regression models are conducted when relaxing the assumption of linearity.

Model 4

L1corruptioni,t= a + wI+ gt + b1 l2dtradei,t-1 + l2dtradei,t-12 + b2 ln(l1GDPi,t) + b3 l1democracyi,t + b4 l1subsidiesi,t +b5 l1EconomicFreedomi,t+b6 l1tariffi,t + eI,t

Model 5

L1corruptioni,t= a + wi+ gt + b1 l2dexportsi,t-1 + l2dexportsi,t-12 + b2 ln(l1GDPi,t) + b3 b3 l1democracyi,t + b4 l1subsidiesi,t + b5 l1EconomicFreedomi,t+b6 l1tariffi,t + eI,t

Model 6

L1corruptioni,t= a + wi+ gt + b1 l2dimportsi,t-1 + l2dimportsi,t-12 + b2 ln(l1GDPi,t) + b3

b3 l1democracyi,t + b4 l1subsidiesi,t +b5 l1EconomicFreedomi,t+b6 l1tariffi,t +eI,t

Where i = 1, 2, …,40 and indicates the country variable, t = 1, 2, …, 21 and indicates the year in the period 1996-2016, a indicates a constant, wI + gt, indicate the country specific and time specific interceptions for the given countries and time period. The error terms that has been used for the models is eI,t.

3.1.2 The model

To investigate if a random- or fixed effect model should be used, a Hausman test is conducted. The difference of the estimates of the coefficients in the random effect model and the fixed model is compared with this test.

The null hypothesis of the Hausman test is that the random effect model should be used. The random effect model assumes exogeneity of all regressors, but the regressors are allowed to correlate with the random individual effects. This limitation of the Hausman test should be taken into account when choosing between the two models. The fixed effect model is adopted if the null hypothesis is rejected. The fixed effect model allows for endogeneity of all regressors.

The random effect model and the fixed effect model should be estimated first to conduct the Hausman-test. Stata then compares these two models. For every model stated above, the Hausman-test is conducted. The results of these tests are shown in appendix C. The first model has a chi2 _{of 255.04 and a p-value of 0.000, implying that the null hypothesis is rejected, since} the p-value is smaller than 0.05, which means an adoption of the fixed effect model. The chi2 of the second model is 259.48 and the p-value is 0.000. The third model has a chi2 _{of 247.36} with a p-value of 0.000. The fourth model has a chi2 _{of 265.20 and a p-value of 0.000. The fifth} model has a chi2 _{of 297.99 and a p-value of 0.000. The last model has chi}2 _{of 347.65 and a} p-value of 0.000. Hence, the null hypothesis for all models is rejected, which means an adoption of the fixed effects models for all the regressions models. Therefore, a country and time fixed effect are added to all the regression models.

3.2 Data

The data is collected for OECD countries and the five key partners, from 1996 until 2016. Different databases have been used to collect the data for this thesis. A balanced panel dataset has all observations for each country in each year, when this is not the case the panel is unbalanced (Stock and Watson, 2015). The panel used in this thesis is unbalanced, due to missing data for some countries over different time periods.

The data of corruption is collected from the Corruption Perceived Index published by Transparency International, since this index is recognised as the most robust and comprehensive index of corruption (Das and DiRienzo, 2009; Serra, 2006)). The used data for trade, imports, exports, GDP per capita, subsidies and tariffs are collected from the World Development Indicators of the World Bank. This data could be downloaded as panel data. Panel data is data for a number of different countries observed over different periods of time (Stock and Watson, 2015). All data regarding the democracy variable is collected from the Freedom House index. Furthermore, the data of economic freedom is retrieved from the Index of Economic Freedom, published by The Heritage Foundation. All other databases do not provide the option of

downloading data in panel data form. A reformatting of the data has been done in Excel to create panel data.

First, corruption, which is the dependent variable, is discussed. Thereafter, the explanatory variables of interest, trade, imports, and exports will be elaborated. Third, all the other explanatory control variables will be reviewed.

3.2.1 Corruption

In this thesis the dependent variable is corruption. Corruption is defined in many ways. In this thesis the most commonly used definition will be used and is given by Transparency International (2018) and states: “The abuse of entrusted power for private gain”. This definition is also supported by Das & DiRienzo (2009), Lalountas et al (2011), Treisman (2000), among others. Following this definition, corruption can be classified in three different types, namely political, petty, and grand. This depends on the amount of money lost and the sector in which it occurs. A manipulation of institutions, policies, and rules of procedure in the allocation of resources and financing by political decision makers, that abuse their position to sustain their power, wealth and status are seen as political corruption (Transparency International, 2018). Petty corruption is seen as everyday abuse of entrusted power by low-and mid-level public officials in their interactions with ordinary people, who are frequently trying to access basic goods or services in places like schools, hospitals, police departments and other agencies (Transparency International, 2018). Policies or the central function of the state that is distorted by acts committed at a high level of government, enabling leaders to benefit at the expense of the public good, is seen as grand corruption (Transparency International, 2018). From the above it is clear that the definition given by the Transparency International focuses on the public sector.

The Corruption Perception Index was first launched by Transparency International in 1995. This index is a “poll of polls” and represents the average scores, given by financial journalists and international businessmen, to individual countries when polled in a variety of contexts. In the paper of Husted (1999) the limitations of the CPI are given. First, The CPI does not reflect the activity of business people who engage in corrupt activities abroad but abstain from corrupt activities at home. Second, the index is not a measure of the level of corruption but measures the perception of corrupt activities. The CPI indeed attempts to estimate the level of perceived corruption by businessmen and is not an estimate of the actual corruption level in

any country. These perceptions may not be a true image on the state of affairs, but they are a reality. It is this reality that the CPI seeks to assess (Transparency International, 1995).

The CPI is based on ten different international surveys of perceived corruption of business people and country experts around the globe (Husted, 1999). Furthermore, the CPI uses different data sources. To qualify as a valid source for the CPI, it must fulfil the following criteria; quantifies perceptions of corruption in the public sector, be based on a reliable and valid methodology (which scores and ranks multiple countries on the same scale), performed by a credible institution, allow for sufficient variation of scores to distinguish between countries, gives ratings to a substantial number of countries, the rating is given by a country expert or business person, and the institution repeats their assessments at least every two years (Transparency International, 2017). At least three valid sources of a country are needed to be included in the CPI. If these requirements are met, the data is standardized. The CPI is designed by TI to measure the degree to which politicians and public officials are believed to accept improper payments in public procurement, or bribes, or misuse public funds, or commit offences. Although the CPI is a perception-based measure of corruption, it is the most comprehensive quantitative index of cross-country corruption available (Das and DiRienzo, 2009). Regardless of the limitations given by Husted (1999), the CPI has been used in many studies such as Das and DiRienzo (2009), Dutt (2009), Lalountas et al. (2011), Sandholtz and Koetzle (2000), Treisman (2000), among others. Moreover, the research of Serra (2006) states that no index or measure of corruption is perfect, however the CPI is robust, this in contrast of other measures of corruption that are based on individual sources. Therefore, this thesis uses the CPI as well.

The CPI uses a scale from 0 to 10 for the time period of 1995 until 2011, where 0 is seen as highly corrupt and 10 is very clean. For the time period of 2012 until 2016 a scale from 0 to 100 is used, where 0 is seen as highly corrupt and 100 is seen as very clean. The change of scale from 0-10 to 0-100 was necessary, because the CPI used to be based on the perceptions of corruption in each country. The rank position of each country in each data source was captured by the CPI, as a consequence, scores were highly dependent on the changes in scores of the countries around it in the index. Since 2012, the raw scores from each of the data sources have been used. This provides more transparency of how the CPI scores have been constructed and this way captures the changes overtime better (Transparency International, 2012). To create the same scale for the different time periods, the data retrieved from 1995 until 2011 are multiplied by a factor of 10 and rearranged to a scale of 0 to 100. All the CPI scores used in this thesis are

on the same scale of 0 to 100, where high corruption levels correspond with low scores and vice versa.

3.2.2 Explanatory Variables of interest

The explanatory variables of interest are trade, exports and imports. These variables are measured as a percentage of GDP. Trade openness is measured based on trade, exports and imports. The data for exports, imports, and trade are retrieved from the World Bank.

Exports of goods and services is measured by the World Bank (2018) as a percentage of GDP and represents the value of all goods and other market services provided to the rest of the world (including the value of merchandise, license fees, insurance, transport, travel, freight, royalties, and other services, such as communication, construction, financial, information, business, personal, and government services). Compensation of employees and investment income and transfer payments are excluded from exports. A weighted average is taken to construct annual data. GDP from the expenditure side is constructed of household final consumption expenditure, gross capital formation, net exports of goods and services and general government final consumption expenditure. Net taxes on products and purchaser prices are included in these expenditures.

Customs reports and balance of payments data provide the data on exports and imports. This provides reliable records of cross-border transactions, however illegal transactions are not captured by the balance of payments or customs data that occur in many countries.

Imports of goods and services are measured by the World Bank (2018) as a percentage of GDP and represents the value of all goods and other market services received from the rest of the world. A weighted average is taken to create annual data. The same values are included and excluded as given in the variable exports. Furthermore, the limitations stated for exports also apply for imports.

The definition given by the World bank (2018) for trade is the following: “Trade is the sum of exports and imports of goods and services measured as a share of gross domestic product”. This variable is created annually by using a weighted average. The trade statistic may further be distorted because travellers carry goods across borders legally, but unreported.

3.2.3 Control Variables

Control variables are regressors included into the regression model to hold constant the factors that, if neglected, could lead the estimated causal effect of interest to suffer from omitted variable bias (Stock and Watson, 2015). The control variables used in this thesis are the following; GDP per capita, subsidies, tariffs, economic freedom, and democracy.

The control variables GDP per capita, subsidies and tariffs are also retrieved from the database of the World Bank. Seldadyo and de Haan (2006) state that the most commonly known factor to explain corruption levels is income. It is found in the literature that countries with a higher income per capita have lower levels of corruption. For this reason, GDP per capita should be taken into account. In order to control for that, in this thesis the annual growth rate of GDP per capita is used and is based on constant local currency. The GDP per capita variable is an annual weighted average and created by dividing the gross domestic product by midyear population.

The control variables subsidies and tariffs can affect corruption, as they could increase the propensity to pay bribes to bureaucrats by both domestic firms and foreign exporters. The variable subsidies is measured by the World Bank the following way. Subsidies, grants, and other social benefits include all unrequited, nonrepayable transfers on the current account to public and private enterprises; and social assistance benefits, social security; grants to foreign governments, international organizations, and other government units and employer social benefits in cash and in kind and is measured in percentage of government expenses. Since the cross-country data of production and export subsidies is very limited, this thesis is constrained to use this measure. This measure is also used by Dutt (2009). However, the measurement of the World Bank is not without limitations. The data of central government finance has been solidified into one account for most countries. Nevertheless, other countries only have data available from budgetary central governments accounts. Budgetary accounts usually provide an incomplete view since they may not include all central government units. Furthermore, due to the fact that export subsidies are only a small part of total subsidies, the magnitude of the impact should be interpreted with caution.

Weighted mean applied tariff is used as the tariff control variable. This is the average of the effectively applied rates and is weighted by the product imports shares of each partner country. The Harmonized System of trade is used to classify the data. The ad valorem rates have been converted and included in the calculation of weighted mean tariffs. The United Nations Statistics Division’s Commodity Trade database is used to calculate the imports weights. The average for products for each commodity group is taken from the effectively

applied tariff rates. The most favoured nation rate is used if the effectively applied rate in unavailable.

Knack and Azfar (2003) and Seldadyo and de Haan (2006) found that limited economic freedom stimulates corruption, since this provides opportunities for bribery and other corrupt practices. The data for economic freedom is retrieved from the Index of Economic Freedom given by Heritage Foundation and defines this as; the fundamental right of every human to control his or her own property and labour. When a society is economically free, individuals are free to work, consume, produce, and invest in any way the wish. Furthermore, labour, capital and goods are allowed to move freely.

Economic freedom is measured by four broad categories: Rule of Law, Government size, regulatory efficiency, and open markets. These four categories are further divided into 12 quantitative and qualitative factors. Rule of law is divided into the following; property rights, government integrity, and judicial effectiveness. Government size is divided into government spending, tax burden, and fiscal health. Business freedom, labour freedom, and monetary freedom together form regulatory efficiency. The category open markets is formed by trade freedom, investment freedom, and financial freedom. All of these 12 factors within the four main categories are graded on a scale from 1 to 100, where higher scores correspond with more economic freedom. The overall score of a country is derived from the average of these 12 economic freedoms. Every component of economic freedom is seen as equally important to the Index of Economic Freedom in achieving the positive benefits and therefore, equally weighted. The Heritage Foundation’s Index of Economic Freedom is also used by researchers such as Goel and Nelson (2005), Quazi (2007), and Das and DiRienzo (2009), among others.

Democracy is the last control variable that is added to our regression model. Seldadyo and de Haan (2006) state that political liberties (i.e. more democracy) increase transparency and provide a system of checks and balances within a country’s political structure, this decreases the level of corruption. Emerson (2006) and Goel and Nelson (2005) find a negative relationship between democracy and corruption, which means, an increase in democracy leads to a decrease in corruption. However, Ades and Di Tella (1999) do not find a significant relationship between democracy and corruption. The democracy data is collected from the Freedom House database. A country receives two scores, one for political rights and one for civil liberties. The scores for political rights is constructed by using 10 political rights questions that can be awarded between 0 and 4. The maximum amount of points for Political rights is 40. The scores for civil liberties are constructed in the same way, only 15 civil liberties questions are used. The maximum amount of points for civil liberties is 60. A country is assigned one

rating for political rights and one rating for civil liberties between 1 and 7, the rating 1 corresponds with the greatest degree of freedom and 7 with the smallest. Each of these ratings correspond to a specific range of total scores, see appendix B. The average of the political rights and civil liberties ratings is taken to measure democracy. Several researchers use these variables to indicate democracy, such as, Emerson (2006) Das and DiRienzo (2009), Lalountas et al. (2011), Treisman (2000), and Sandholtz and Koetzle (2000), among others.

3.2.4 Descriptive data analyses

The prior literature given above shows that an increase of international trade has a negative impact on perceived corruption levels. From the retrieved data described above, this section will provide descriptive analyses, and shows the trends of the perceived corruption scores, trade as share of GDP, exports as share of GDP, and imports as share of GDP for all OECD countries and key partners for 1996-2016. The OECD countries and the key partners will be discussed separately.

Firstly, the OECD countries will be discussed. Figure 3 shows the perceived corruption scores retrieved from the CPI and trade as a percentage of GDP retrieved from the World Bank for all OECD countries from 1996-2016. The perceived corruption scores for these countries do not change much over time for all of these countries. Trade as a percentage of GDP is also found to be relatively constant over time. However, as figure 3 shows, there has been an increase of trade for the following countries; Czech Republic, Hungary, Ireland, Luxembourg, Slovakia, and Slovenia. Nevertheless, for these countries, the data do not show an increase of perceived corruption scores (i.e. a decrease in corruption), nor a decrease in the corruption scores. Figure 4 shows the perceived corruption scores together with export as a percentage of GDP and figure 5 shows the perceived corruption scores together with imports as a share of GDP. The data do not show a common trend between perceived corruption scores and trade, exports or imports. The figures show that perceived corruption scores are relatively stable over time for all OECD countries, even when trade, exports, and imports changes. However, it has to be taken into account that the perceived corruption scores of the OECD countries are already relatively high (i.e. corruption is low) and for most of these countries trade, imports and exports as share of GDP did not change much over time.

Secondly, the key partners will be discussed. Figure 6 shows the perceived corruption scores retrieved from the CPI and trade as a percentage of GDP retrieved from the World Bank for all key partners from 1996-2016. For Brazil the graph shows a small increase of international

trade over time. The perceived corruption score is low and has, similar to international trade, slightly increased since 1996. For China an increase of international trade is shown. The perceived corruption scores of China have been increasing over time as well, meaning that perceived corruption has decreased. International trade for India in the late 1990s is very high and after that mainly decreases until 2016. The perceived corruption scores first decrease, from the early 2000s the figure shows an increase of perceived corruption scores. The perceived corruption scores of Indonesia increase slowly from 1996 to 2016, indicating that perceived corruption level has decreased over this time period. International trade, however first increases rapidly until 2012 and thereafter decreases. Figure 6 shows a slow decrease of perceived corruption scores from 1996 to 2016 for South Africa, which means the perceived corruption has increased in the country. Furthermore, the data shows that South Africa’s international trade has been increasing. Figure 7 shows the perceived corruption scores and exports as a percentage of GDP. The export level of Brazil is low and relatively stable with a small increase from 1996 until 2016. Figure 7 shows a peak of exports for China in 2006, but this peak evaporates again in 2009, and the level of export in 2016 is almost the same as it was in 1996. Similar to international trade, India starts with a peak in exports, however figure 7 shows a decrease in exports as a share of GDP from the late 1990s to 2016. Indonesia shows an increasing export level as a share of GDP from 1996 until 2013, and the export level decreases from 2013 until 2016. The level of exports for South Africa has slightly increased overtime. Figure 8 shows the perceived corruption scores and imports as a share of GDP. The imports of Brazil have not changed much over time and remain on almost the same level. The imports for China first increases from 1996 until 2005 and thereafter decreases again to almost the same level in 2016 as it was in 1996. The import levels of India first increases in 1997 and thereafter decreases again. Indonesia has increasing import levels from 1996 until 2012 and thereafter decreases until 2016. The import levels of South Africa have slightly increased over time. The above indicates that almost all key partners have increasing perceived corruption scores, accept for South Africa, indicating that the perceived corruption of these countries has been decreasing over time. It further indicates that international trade has increased in most key partners as well. It increased in South Africa and Indonesia, and with a very small amount in Brazil and China over time. International trade has however, decreased in India. Exports and imports of the five key partners of the OECD have shown a more mixed picture. Exports have been increasing for Indonesia, South Africa, and by a very small amount Brazil. The exports for China first increase, but later on decrease again to the starting level. Exports for India decreases over time. Imports increases for Brazil, Indonesia and South Africa. For China imports first increases and

F igur e 3. C or rupt ion sc or e and T rade for al l O E C D c ount rie s f or 19 96 -2016 Sour ce : W or ld B an k and C or ru pt ion Pe rc ept ion Ind ex – Y .A . B iss ch ops c alc ula tions

Sour ce : W or ld B an k and C or ru pt ion Pe rc ept ion Ind ex – Y .A . B iss ch ops c alc ula tions F igur e 4. C or rupt ion sc or e and E xpor ts for al l O E C D c ount rie s f or 1996 -2016

Sour ce : W or ld B an k and C or ru pt ion Pe rc ept ion Ind ex – Y .A . B iss ch ops c alc ula tions F igur e 5. C or rupt ion sc or e and Im por ts for al l O E C D c ount rie s f or 1996 -2016

Figure 6. Corruption score and Trade for all Key Partners for 1996-2016

Source: World Bank and Corruption Perception Index – Y.A. Bisschops calculations

Figure 7. Corruption score and Exports for all Key Partners for 1996-2016

after 2005 decreases to the levels as before. The level of imports for India decreases over time. In sum, looking at the trends of international trade, export, import and perceived corruption scores, the relation of higher trade going along with less perceived corruption (i.e. higher perceived corruption scores) that was found in the previous literature is not really visible when looking at the trends in the data.

3.3 Assumptions

In order to get unbiased coefficients for our regressions, the following assumptions should be met.

• Stationarity • Homoskedasticity

• No perfect multicollinearity

Figure 8. Corruption score and Exports for all Key Partners for 1996-2016

The first assumption that is tested in this thesis is stationarity. In this thesis panel data has been used for 1996 until 2016. If there are big changes over time, data from the past is not reliable for predicting the future. If the time series do not change over time, which means a constant mean and variance, and there is autocorrelation through time, they are called stationary (Stock and Watson, 2015).

A Dickey-Fuller test has been done to test for stationarity. The data used in this thesis is unbalanced, hence the Fisher variant has been used. Every variable is tested, and the results are shown in appendix D. Choi (2001) states that the best trade-off between size and power is the inverse normal Z statistic. The variable corruption is not stationary (see appendix D.1), since the inverse normal Z statistic is larger than 0.05, the null-hypothesis cannot be rejected. However, the first lagged value of this variable is stationary. For this reason, the first lagged value of corruption will be used as the dependent variable in this thesis.

The variable trade, exports, and imports are also found to be non-stationary (see appendix D.3, D.5, and D.7). For this reason, the first differences have been conducted. Since the first lagged of corruption is the dependent variable, the second lagged of the first difference of trade, exports, and imports have been taken to account for reverse causality. When relaxing the assumption of a linear relationship, the squared of the second lagged of the first difference of trade, exports, and imports have been conducted and found to be stationary (see appendix D.16, D.17, and D.18)

Since the first lagged value of corruption is used, the first lagged value of the control variables will be used as well. The variable subsidies, is also found to be non-stationary (see appendix D.9). The first difference has been conducted. The variables democracy, economic freedom, GDP per capita, and tariffs are all found to be stationary.

The second assumption that needs to be validated is the one of homoskedasticity. The following definition is given by Stock and Watson (2015): The error term is homoscedastic if the variance of the conditional distribution of ui given by Xi is constant for i = 1, ..., n and in particular does not depend on Xi”. The error term is heteroskedastic if this is not the case. The likelihood-ratio test has been conducted to test for homoscedasticity, the results are shown in appendix E. This test requires that both homo-and heteroskedastic panels have to be estimated with generalized least squares regressions. The null hypothesis of the likelihood-ratio test is that the panel is homoscedastic. The results of the likelihood-ratio tests are displayed in appendix E. For the first model a chi2 _{of 226.53 and a p-value of 0.000 has been found. From} this result the null hypothesis is clearly rejected; hence the panel is heteroskedastic. The second model has a chi2 _{of 221.45 and a p-value of 0.000. The null hypothesis of the second model is,}

just like the first, clearly rejected. The third model has a chi2 _{of 210.23 and a p-value of 0.000.} The fourth model has a chi2 _{of 224.63 and a p-value of 0.000. The fifth model has a chi}2 _of 223.07 and a p-value of 0.000. The last model has a chi2 _{of 225.52 and a p-value of 0.000. From} the above it can be concluded that all panels are heteroskedastic. Since the panels are heteroskedastic, regular standard errors cannot be used. For this reason, robust standard errors are used, these errors also allow for serial correlation.

No perfect multicollinearity is the third assumption. To test for this a VIF test is used. There is perfect multicollinearity is the value of the VIF test is equal or above 10. The results of the VIF tests are given in appendix F. All the values are below 2, from these results it can be concluded that there is no perfect multicollinearity.

3.4 Conclusion

Three hypotheses, as formulated in section 2.2, will be tested using panel data. From the Hausman test, the conclusion is drawn that all models should follow the fixed effect model. The country and time fixed effects are taken into account to further account for omitted variable bias. A stationarity test has been conducted and all the variables used are stationary. Robust standard errors are used since the panel is heteroskedastic. Furthermore, a VIF test showed that there is no perfect multicollinearity in the panel data.

4. Results

In this section the test results of the fixed effect models will be discussed. The fixed effect model uses a t-statistic to estimate the significance level of the coefficients. In all the tables in this section presenting the regression results, the robust standard errors are presented between brackets underneath the coefficients. Table 1 contains the estimates of the linear regression models 1, 2, and 3, while table 2 presents the estimates of the nonlinear models 4, 5, and 6, including a squared term for either trade, exports, or imports. For each model, the table shows three results in three columns. Column x.1 shows the regression results for the model including the correction for non-stationarity and reverse causality. For each model this is the appropriate result. Column x.2 for each model shows the results where a correction for stationarity is not accounted for (but a correction for reverse causality is included). Finally, column x.3 shows the results where neither reverse causality nor non-stationarity is accounted for.

The reason to include the results in columns x.2 is that, as described in the literature review, in almost all of the prior research, stationarity is not accounted for. Therefore, for comparison of our results with the results in the literature, this thesis will also show the regression analyses without accounting for stationarity. Where stationarity is not corrected for, the first differences of the explanatory variables that are non-stationary is not used, neither is the first lagged value of corruption, and hence, also not the lagged values of the control variables. Furthermore, the second lagged values of the first difference of trade, exports and imports are substituted by the first lagged values to account for the reverse causality problem. Since non-stationarity is not accounted for, similar to the other control variables, the first difference of trade, exports, and imports are not used in this model.

The reason to include the results in columns x.3 is that some of the prior research does not account for the reverse causality between corruption and trade, exports and imports and does also not account for non-stationarity. This thesis will therefore also show the results if the reverse causality as well as non-stationarity are neglected. The first lagged values of trade, exports and imports will then be substitutes with current values of trade, exports and imports. Since non-stationarity is not accounted for, the first difference of trade, exports, and imports are not used in this model as well. These results will be compared to the results that account for both stationarity and reverse causality, and also to the results that only account for reverse causality and not for non-stationarity.

The structure of this section is as follows: firstly, the results of the first model are discussed and compared with the fourth model. These are the models where trade as a percentage of GDP is included both only linear (model 1) and also quadratic (model 4). Secondly, the results of the second model are given and compared with the fifth model. These are the models where export as a percentage of GDP is included both only linear (model 2) and also quadratic (model 5). Thereafter, the results of the third model will be discussed and compared with the sixth model. These are the models where import as a percentage of GDP is included both only linear (model 3) and also quadratic (model 6). Fourthly, a robustness check will be presented. lastly a conclusion will be drawn.

4.1 Results regression 1 and 4

In this section the results of the first model, that assumes a linear relationship between trade and corruption, and the fourth model, which relaxes the assumption of linearity and includes a

squared term for trade, will be discussed. The first hypothesis will be answered by evaluating these results

on the impact of trade on corruption. Firstly, the regression output will be given and explained. The output for model 1 (and 2 and 3) is given in table 1 and the output for model 4 (and 5 and 6) is given in table 2. A fixed effect model is used for these regressions. The coefficients of the country- and time-fixed effects are not given in these output tables but are shown in appendix G for better readability of the output tables. (Appendix G1 shows these coefficients for model 1 and model 4, G2 for models 2 and 5, and G3 for models 3 and 6.) Thereafter, the results concerning trade will be discussed. Furthermore, the results of the explanatory variables will be explained.

The first hypothesis, an increase in international trade reduces the perceived corruption level of a country, has been tested. The fixed effects coefficient for trade are given in appendix G.1.

As stated above, the prior literature shows a significant negative effect of international trade on corruption (Das and DiRienzo, 2009; Sandholtz and Koetzle, 2000; Dutt, 2009; Gatti, 2004). The coefficient found in this thesis for trade is, however, insignificant. The findings of such contrary results in comparison with the prior literature, can be attributed to the use of stationary variables in this thesis. If stationarity is neglected, the estimate of international trade almost doubles and is positive (0.072) with a significance level of one per cent. When stationarity and reverse causality are both neglected the estimate of international trade increases by a small amount to 0.074 and is still significant at a one level. Since corruption is measured by the CPI from 0 to 100, where 100 means a country is not corrupt, a positive estimate means that the perceived corruption level of a country decreases. This result does correspond with the prior literature. However, due to the use of panel data, stationarity is required. Furthermore, as stated above, prior literature shows a reverse causality between corruption and trade. For that reason, the results of the prior research should be interpreted with caution. Moreover, when relaxing the assumption of a linear relationship, the coefficient for trade remains the same and is still found to be insignificant. The variable trade2 _{is also found to be insignificant. Das and DiRienzo} (2009) have also found no significant change when the assumption of a linear relationship between globalisation and corruption is relaxed. This is in line with the results presented in this thesis. However, when stationarity and reverse causality are neglected, it is found that the sign and the significant level of trade remain the same, but the magnitude of the coefficient increases. Trade2 _{is found to have a small negative effect and is significant at a one per cent level.} However, these results are bias and not reliable.

The democracy variable has an estimate of -2.796 and is significant at a five per cent level. Since this variable is measured by the Freedom House database from 1 to 7, where 1 corresponds with the greatest degree of freedom and 7 with the smallest. A negative coefficient was expected. If there is more freedom in a country, the democracy variable decreases. This will lead to a higher score of perceived corruption, which means that the perceived level of corruption of a country decreases. Phrased differently, more freedom leads to less perceived corruption. This variable has the expected sign and corresponds with the results found by Treisman (2000), Sandholtz and Koetzle (2000) and Gatti (2004). The estimate of the democracy is relatively stable and significant at a five per cent level. The estimates increase by a small amount, but the significance level does not change if stationarity and reverse causality are not accounted for. Furthermore, when relaxing the assumption of linearity, the estimate of the coefficient decreases to -2.814 and remains significant at a five per cent level.

The variable Economic freedom has a positive estimate of 0.422 and is significant at a one per cent level. When stationarity and reverse causality is neglected, the estimate increases to 0.582 and 0.606 and remains significant at a one per cent level. Moreover, when relaxing the assumption of linearity, the estimates remain almost the same and are still significant at a one per cent level. An increase in the variable economic freedom will increase the score of perceived corruption. This means that the perceived corruption of a country decreases. This result is supported by Das and DiRienzo (2009), and Sandholtz and Koetzle (2000).

Tariff is found to be not significant. When stationarity and reverse causality are neglected as well as when relaxing the assumption of linearity, tariff is still found to be insignificant. This is in contrast with Dutt (2009) and Gatti (2004). The time span of those papers the former is from 1984-2004 and 1982-2000 for the latter. Due to the increase of globalisation through the years of the OECD members and five key partners, it might be that tariffs have become less significant.

The estimate for subsidies is negative and not significant. Dutt (2009) states that the lagged value of subsidies is not significant. Since this thesis uses the lagged value and the first differences of subsidies, it was expected that this estimate would be insignificant. If the lagged value is not used, the estimates of subsidies is positive and highly significant at a one per cent level. This corresponds with the finding of Dutt (2009). When the assumption of linearity is relaxed, the same results are found.

GDP per capita is shown to be insignificant. The same results are found when relaxing the assumption of linearity. This does not come as a surprise since the prior literature shows mixed results about the significance of this variable. Lalountas, Manolas & Vavouras (2011)