Can Equality in Education Be a New Anti-Corruption Tool? Cross-Country Evidence ( )

Transcription

1 127 Can Equality in Education Be a New Anti-Corruption Tool? Cross-Country Evidence ( ) Abstract Kraiyos Patrawart* Cross-country evidence shows that corruption could be controlled with support from the education, free press and independent judicial systems, yet the theoretical foundation for such a connection is somewhat limited. This paper investigates the mechanisms behind the anti-corruption effect of education through civic engagement. We argue that equal universal access to education and the free press is a crucial tool for the majority of citizens to acquire the correct information needed to succeed in their anti-corruption initiatives. A simple reduced-form theoretical model, which allows for heterogeneity in educational attainment among agents, is used to explain the link between education equality and corruption. Evidence from cross-national panel data estimation between 1990 and 2005 shows * Teaching Fellow and PhD candidate, Royal Holloway University of London, Egham, Surrey, United, Kingdom and Visiting Fellow, Department of Government and Institute for Quantitative Social Science (IQSS), Harvard University, Cambridge, Massachusetts, United States. k.patrawart@rhul.ac.uk; kpatraw@fas.harvard.edu The author is indebted to Cecilia Testa for precious discussions throughout the course of writing this paper, and would like to express gratitude to Michael Spagat, Jonathan Wadsworth, Peter Dolton, Arnaud Chevalier, Anne Booth, Srunya Nopsuwanwong and seminar participants at the Department of Economics, Royal Holloway, Faculty of Economics Thammasat University, and the NACC s Conference on Evidence-Based Anti-Corruption Policy, Bangkok Thailand for helpful comments. Financial support from Royal Holloway and NACC is acknowledged with gratitude. Finally, the author would like to thank Amparo Castello, Rafael Domenech and Vinod Thomas for providing data on education distribution. The views expressed herein reflect those of the author and no one else. JEL Classification Code: D72, I21, O15 robust support for the relationship. Education equality has independent and complimentary anti-corruption effects with press freedom and the duration of democracy. Keywords: political economy, corruption, distribution of education, factor analysis 1. Introduction In the fight against corruption, citizens can no longer be seen as passive recipients; they are main actors and strategic partners rather than targeted groups. The principles of empowerment, transparency, participation and accountability are at the core of the civic-based anti-corruption initiative. (UNDP, 2004, p.6.) As one of the leading international organizations actively engaged in international anti-corruption programs, the United Nations Development Programme (UNDP) has developed a strong strategic vision for its anti-corruption initiatives. 1 That vision clearly values the new civic-based approach in combating corruption; the approach has received growing support from international anti-corruption initiatives. Svensson (2005) argues that the government-led anti-corruption programs that are aimed at reducing the size of government and regulation have had little success in practice, and only limited support from empirical studies. On the other hand, the civic-based 2 anti-corruption programs that are aimed at promoting socio-economic factors which encourage civic monitoring have received great support from both theoretical and empirical standpoints. 3 For 1 The Organaisation for Economic Co-operation and Development (OECD) and the United Nations Educational, Scientific and Cultural Organization (UNESCO) have also made similar policy recommendations. 2 Comparable to an external mechanism (outside the bureaucratic system) in Brunetti and Weder (2003). 3 For instance, Glaeser and Saks (2006) and Rikka and Svennsson (2006).

2 128 instance, Keen (2000), in association with the Human Rights Education Association (HREA), recommended that public education programs should incorporate anti-corruption elements as the prime objective of the national curriculum in order to reduce the likelihood of corruption in society. Such public education programs should cover the broad spectrum of activities which promote the dissemination of information and increase awareness of corruption. These programs should also change the perceptions and attitudes toward corruption and pass on the new skills and abilities needed to counter corruption. In practice, many governments have already adopted anti-corruption education in the actual compulsory curriculum. For example, Cameroon started its pilot program, Fighting against corruption through schools, to teach students and parents to identify and act against dishonesty in their schools and the rest of their local communities. Hence, education can be an effective channel to instill the correct awareness and perception toward corruption, which makes more apparent to people the perceived benefits of engaging in anti-corruption activities. D e s p i t e t h e s i g n i fi c a n t p o l i c y recommendations and extensive literature that support the roles of socio-economic factors in promoting civic-based anti-corruption initiatives, the empirical evidence and the theoretical foundation for the role of education are limited; no single study has examined the role of equality in education. On one hand, Magnus, Polterovich, Danilov and Savvateev (2002) found that education can increase the intolerance of people in society against the corrupt behavior of others and can stop them from performing corruptive activities and encourage them to engage in anti-corruption initiatives. Furthermore, as schooling raises interpersonal and cognitive skills, an increase in the duration of schooling reduces the interpersonal cost to individuals of engaging in anti-corruption initiatives and also improves people s productivity in doing so (Glaeser, Ponzetto, & Shleifer, 2007). On the other, education can promote corruption through various channels. Since corruption is an illegal and secret activity, politicians and bureaucratic officers have incentives to make their corrupt activities complicated and unnoticeable. Educated officials can be more effective in making corruption sophisticated. When corruption becomes more complicated and very well concealed, it is more difficult and costly for the media and citizens to challenge the corrupt acts (Ahrend, 2002). As a more specific example, an expansion of education creates larger potential rents that corrupted agents can extract (Eicher, García-Peñalosa, & Ypersele, 2007; Frechette, 2006). As a result, the relationship between education and corruption is non-monotonic, which could be the underlying reason behind the ambiguous relationship between education and corruption in the literature. Nonetheless, the academic literature has paid very limited attention to this issue. The main argument of this paper is that citizens who are potential monitors of corruption, depending on a personal stock of human capital, have heterogeneous attitudes toward corruption and have varied anti-corruption skills gained through civic participation. Several studies support this argument. For instance, Magnus, Polterovich, Danilov and Savvateev (2002) used cross-national evidence to show that the level of intolerance to cheating increases by years of schooling. 4 Dwivedi (1967) used evidence from the Indian Public Opinion Survey to show that differences in educational attainment among the participants could explain the heterogeneity in civic engagement and political knowledge, as well as the perspective toward honesty, of government officials. Hence, equality in the distribution of education should reduce the scale of disruptive heterogeneity and allow the monitoring agents to coordinate and make a monitoring threat against corrupted officials. This credible anti-corruption initiative could potentially control corruption. Therefore, what matters in constraining corruption through a civic-based anti-corruption mechanism is not just the absolute stock of human capital in 4 Magnus conducted the experiment with 885 students (high school, undergraduate and post graduate) from the Russian Federation, Israel, the Netherlands and the United States.

3 129 society, but rather the relative stock of human capital across the population. Consequently, the theoretical argument that assumes homogeneous stocks of human capital across agents or empirical analysis which employs only the average enrollment or attainment rates of schooling across the total population to identify the relationship between education and corruption, is likely to find inaccurate predictions and imprecise estimations of the relationship between education and corruption. This paper tests the hypothesis that education, measured by averaged attainment measures and equality measures, has either insignificant or undesirable effects on corruption against the alternative hypothesis that education, either the averaged attainment or the equality measures or both, has anti-corruption effects in the society. We focus on the cross-country evidence between 1990 and 2005 and identify the causal relationship between equality in education and the perceived level of corruption by using new measures of education equality that have not yet been employed in corruption studies. These measures are calculated from the educational attainment dataset of Barro and Lee (2000). Overall, the estimations suggest that the past condition of equality in education distribution affects the current perceived level of corruption independently and complimentarily with other socio-economic factors of civic participation. The anti-corruption effects of education equality remain robust through various specification changes. The following section reviews the literature that is relevant to our research agenda. Section 3 provides a simple theoretical model that allows for inequality in education among citizens and public officials. Section 4 illustrates the data set that we employ in the empirical analysis in sections 5-7 which, in turn, searches for empirical evidence for the theoretical predictions in section 3. Section 8 concludes the findings and describes the policy implications and research opportunities. 2. Literature Review This section reviews the relevant findings of the determinants of corruption in order to address the non-monotonic relationship between education and corruption. The review serves as the foundation of our main argument that the causal link between education and corruption could potentially connect through the equality of education among the relevant agents in the economy. 2.1 The Anti-Corruption Effects of Education The relationship between education and corruption is nothing new. Scholars have praised the anti-corruption role of education through civic participation and political accountability. Several empirical studies on the causes of corruption have found that education determines the perceived level of corruption. Ades and Di Tella (1999) formally investigated the various causes of corruption by using time-varying factors in their fixed-effect estimation. The identification strategies in their work also included the use of instrumental variables to overcome the endogeneity problem of country openness and corruption. Using the World Competitive Report and Business International Corruption Indices in the 1980s, the results showed that, other things being equal, in countries where domestic firms enjoy sheltered business and low competition, more rents could be extracted and thus more corruption could take place. More importantly, they were the first scholars to formally verify that civil society, measured by educational attainment, per capita income and a political rights index, could control corruption. However, these authors did not check the anti-corruption role of civil society in a fixed-effects model as they argued that there is no variation in schooling across time. Also, they did not include press freedom in their civil society analysis. However, this was due to data limitation problems. We address these shortcomings and try to overcome them in our empirical analysis by employing longer panel data of schooling and adding more relevant factors that support civic participation in controlling corruption, including press freedom.

4 130 Another important work in the study of the causes of corruption is that of Treisman (2000). His paper tested the broader ideas of factors that can determine the level of corruption across countries. As he extensively included all potential causes of corruption in his estimations (ordinary least squares - OLS and two-stage least squares - 2SLS), one may have to place greater emphasis on positive than negative results. Treisman focused on various determinants, ranging from religion and historical culture to current institutional and economic factors. Since his work employed generally time-invariant factors, the main sources of variation for his analysis came from cross-country differences. Primarily, he used the Corruption Perception Index for as well as Business International s Corruption Index for In addition, Treisman detected the potential endogeneity problem between per capita income and corruption in the empirical analysis where he proposed the time-invariant distance from the equator as an instrument for per capita GDP. Then, the income effect in reducing corruption still holds. However, Treisman did not include schooling and press freedom variables in his analysis. Although his fourth hypothesis intended to test the effects of democracy, a free press and civic association on corruption, the variables he actually employed, in contrast to the earlier attempt by Ades and Di Tella, were the duration of uninterrupted democracy and the political rights index from Freedom House. It is unclear for many reasons whether these variables appropriately capture the effect of civic society. The duration of democracy according to Alvarez et al., (1996) is measured by just a dummy variable that indicates whether a country has had an uninterrupted democratic regime in the period from 1950 to However, it is unclear whether continuity in democracy is the only determinant of civil society and a free press. Moreover, the differences in civil society and corruption among 23 countries with 40 years of uninterrupted democracy are substantial. 5 Also, 5 India, South Africa and Mexico have been democratic for 40 years as have Finland, Denmark, the United States and the United Kingdom. the political rights index from Freedom House was constructed mainly from the ratings of whether or not the country had a free, fair and competitive political system. Even though Freedom House produced an exclusive quantitative score for press freedom in 1994, Treisman did not employ this rating in his analysis for the fourth hypothesis. 6 As the anti-corruption effects of press freedom and education have been found in a number of studies on corruption, 7 the explanatory power of some factors in Treisman s analysis, which correlate with education and press freedom, may incorporate the effect of schooling and press freedom in its explanatory power. In other words, Treisman s estimations may have overestimated the anti-corruption effect of some variables that correlate with education and press freedom. Hence, we will contribute to this research gap by explicitly including the measures of schooling and a free press along with other significant factors found in Treisman (2000) to re-estimate the role of civic association and a free press on corruption. Glaeser and Saks (2006) studied corruption determinants within the United States by using average objective data on corruption convictions for the period , as documented in a Justice Department report, and the education and economic characteristics of various states. The main findings of their survey are that the states with more educated, richer and less unequal people have less corruption. Also, heterogeneity in ethnicity and earnings appears to promote corrupt practices. Moreover, states with more corruption convictions experienced slower growth in total output during the past two decades. However, there appears to be weak support on the negative impact of income and size of local government. Also, there is no evidence 6 The results of this hypothesis show that what matters to corruption is not the current status of democracy but the duration of uninterrupted democracy. 7 Brunetti and Weder (2003) and Chowdhury (2004) found the causal relationship between press freedom and corruption, while Ades and Di Tella (1999), Ahrend (2002), Glaeser and Saks (2006) and Svensson (2005) found the causal relationship between education and corruption.

5 131 that the degree of regulation brings about a higher level of corruption. These results reinforce the argument of Svensson (2005) mentioned in the introduction. Unlike the size of government or degree of regulation, the factors income and education are important because they raise the level of civic participation. Thus, the corrupted behavior of government officers is strongly motivated by the potential costs of being caught rather than by government reward. In other words, the external civic demand for greater accountability driven by education and income can effectively discipline public servants. This finding strongly supports the prominent study about crime and punishment by Becker (1968). Also, these findings strongly support the view that civic participation can help reduce corruption in the United States The Adverse Effects of Education on Corruption Ahrend (2002) found that education increases corruption when media freedom in the society is very limited. His reduced-form model predicted that an increase in human capital has two diverse effects: (a) increased corruption through a rise in the bureaucrat s productivity and the skills to make corruption sophisticated; and (b) reduced corruption through a rise in monitoring capacity, which can be seen as the effectiveness and the independence of monitoring institutions. Hence, the crucial factor which determines the nature of the net effect of education on corruption is the effectiveness of the monitoring institutions. Ahrend assumed that the determinants of the effectiveness and independence of monitoring institutions are a free press and the independence of the judicial system. He also found empirical evidence from the cross-national regressions that education reduces the perceived level of corruption only if the efficiency of the monitoring capacity in the society, measured by press freedom, is sufficiently high. 9 Similarly, Frechette (2006) 8 Reinikka and Svensson (2005) presented evidence in a least developed country that supports this argument. 9 Specifically, education reduces corruption only in countries with the free status of press freedom in the Freedom House index. found that the availability of rents driven by income and trade restrictions increased corruption. He also found that an increase in education, measured by the primary school enrollment rate of the total population, increased the perceived level of corruption in the society. We argue that Ahrend s theoretical conclusions were driven by the unusually strong assumptions in his theoretical model. For instance, he assumed an identical human capital stock among public officials and monitors, which is a very strong assumption. As we have discussed previously, the wide education gap between citizens and public officers can potentially create room for corruption; thus, this strong assumption prevents Ahrend s analysis from identifying the anti-corruption effect of equality in the distribution of education. Moreover, he concluded that freedom of the press and judicial independence determine the nature of education s role; however, these two factors do not exist in his theoretical model. We, thus, intend to contribute to Ahrend s work by relaxing the assumption of identical human capital stock and explicitly incorporating press freedom and judicial independence into the function of monitoring capacity in the subsequent theoretical section. From the empirical standpoint, we argue that the findings of Frechette (2006) and Ahrend (2002), that schooling increases corruption, are sensitive to their choice of education measures. Frechette (2006) claimed that an increase in education raises the availability of general rents and also increases bureaucrats skills in performing corrupt acts. The second part of the argument is unclear as he employed the measure of schooling, which is the current ratio of primary school enrollment, regardless of age, to the total population of the age group that officially relates to the primary school level. Ahrend (2002) and Frechette (2006) argued that this choice is preferred to that of Ades and Di Tella (1999) concerning secondary school attainment because it involves more variation. This choice of the schooling variable and its interpretation are quite ambiguous for several

6 132 reasons. An increase in the current share of total enrollment in primary school among the population should not affect a government officer s corruption skill for two reasons. First, there is no convincing reason for the connection between today s public officer s skill in corruption and today s enrollment ratio of the population of primary school age. Obviously, the age of public officers is substantially older than that of primary school students (6-13). In fact, a primary school qualification is typically insufficient if one is applying for bureaucratic jobs. Second, the enrollment rate is a weak proxy for human capital stock. When the school dropout rate is high, especially in developing countries, this indicator becomes severely misleading. The finding of Glaeser, Ponzetto and Shleifer (2007) supports our argument by showing that enrollment data conceptually reflect investment flows rather than stocks of human capital. To conclude this review, cross-country and within-country evidence suggests that education significantly determines the level of corruption. The measures of education that are found to be the determinants of corruption include the historical and current levels of educational attainments and the current enrollment rates across populations. Yet, no single study has examined the impact on corruption of equality in the distribution of education. Existing studies present two diverse types of education effects on corruption: promoting effects and controlling effects. Yet, those studies do not provide explicit theoretical argument or apply appropriate empirical treatment to study the non-monotonic relationship between education and corruption; instead, they either claim that there is only one type of relationship or argue that the type of relationship depends upon the condition of its complementary factor, such as press freedom. Consequently, the effects of education are found to be fragile 10 and highly sensitive to specification changes, as the typical chosen 10 Many studies, including the sensitivity analysis, e.g. Serra (2006), have found that the effects of education are insignificant. measures of aggregate human capital cannot explicitly account for the non-monotonicity in the relationship between education and corruption. These issues are investigated in the next section. 3. Theoretical Model This section illustrates the theoretical foundation for testing the hypothesis that education and, in particular, its distribution can play a vital role in determining the optimal level of corruption in society. We argue that education inequality and the limitation of press freedom create room for corruption in society. We begin by presenting the reduced-form model of rent-seeking, in which the bureaucrat optimizes his opportunity to extract rents from public goods without being noticed and prosecuted. This model is comparable to that of Ahrend (2002). However, as Ahrend s model does not allow heterogeneity in human capital among agents, we implement some necessary modifications following similar works by Persson and Tabellini (2000) and Gerrber and Green (1999). The refined model will show that the variations in distribution of human capital and media freedom are crucial in determining the bribery level chosen by opportunistic bureaucrats. The predictions from this reduced-form model will then be used as a foundation for empirical analysis in sections The Economy, Citizens and Bureaucrats The economy consists of two types of agent: citizens (C) who work and earn wages as an increasing function of their educational attainment,w c =E(e c ); and the government official (g) who authorizes the provision of public goods. Following Persson and Tabellini (2000), there are N citizens with identical preference given by equation 3.1: (3.1) where c, t, e c, G denote consumption, taxes, citizen s educational attainment and net public good, respectively, while H(G) is a concave and increasing function. The distribution of

7 133 educational attainment among citizens is predetermined and will be discussed shortly. Each citizen has a different view toward the bureaucrat s ethical standards, 11 which are normalized into a one-dimensional parameter represented by the bribe rate, b. All citizens receive the same public information regarding the bribery from the media. However, each citizen interprets the information from the media differently subject to his or her private human capital stock. We assume that the awareness and the intolerance toward information about corruption that the citizen extracts from the news are increasing in line with the citizen s private stock of human capital. This assumption is consistent with the empirical evidence in Dwivedi (1967) and Magnus, Polterovich, Danilov and Savvateev (2002). The levels of press freedom and education equality in the economy are exogenously predetermined; government officials cannot manipulate these determinants of civic society. This assumption portrays the findings in the literature that the variations in the stock of human capital, especially its distribution, and press freedom are highly persistent across time. 12 Moreover, the variations in theses variables rely on various exogenous factors outside the specific society. Furthermore, the appointment of a bureaucrat is normally on a short-term basis and non-tenurable. Hence, when the bureaucrat is assigned to a new post in any local economy, the only possible choice variable for him is the size of the bribe, b, that he plans to take, not the levels of education equality or media freedom in the local community. This design of the model enables us to study the variation in the government officials behavior given the changes in equality of educational attainment and media freedom. Our model is intended to explain why there are greater perceived levels of corruption in some economies than in others rather than examine the dynamics of corruption in one specific economy. 11 See Dwivedi (1967). 3.2 Education We relax the assumption of identical schooling among agents in Ahrend (2002) by assuming that citizens and bureaucrats attain discrete years of schooling, which is normally distributed according to the distribution function. There is an inequality (skewness) in the distribution of education whereby the median level of educational attainment (e m ) is below the mean, which is normalized to 1, Additionally, working with the government requires a substantial level of schooling; thus, we assume that the government official attains an education higher than the average and median schooling level of the society, 3.3 Corruption The government official works as the provider of public goods, G, which is financed by the flat income tax rate. There are two sources of revenue from holding public office: private benefits and public welfare. The private gains for the bureaucrat are his official wage, W g, and the expected return from taking bribes. The officer can divert a proportion of the public goods, in the form of a monetary bribe (B); thus, Hence, government spending and bribery are financed by flat income tax, where is the citizens average income and N is the total populationsize. Corruption is noticed and prosecuted with probability P, 0<P<1, which is a function of the distribution of education and media freedom in the economy. We depart from Ahrend s model by relaxing the assumption of perfect equality in schooling among agents, and explicitly include press freedom within P. To focus our analysis on the anti-corruption role of civic engagement, we assume certainty in the prosecution of corruption if it emerges. Therefore, the corrupted officer will be prosecuted with probability (P), which depends primarily on three main determinants of civic monitoring capacity: press freedom, which measures the efficiency of the media in minimizing the imprecision of available 12 See section 5.2.

8 134 information on the actual bribe free); the size of the bribe (b); and the equality in education. Making P an increasing function of b reflects the fact that, when the size of the bribe is large, it is more obvious and likely to be noticed and challenged. In addition, to capture an aspect of education inequality, we assume that P is determined by the relative human capital stocks of the median citizens and bureaucrats, relative to the mean schooling of the society. Alternatively, P can be seen as the observable amount of the bribe from the public point of view. Equation 3.2 shows that the ratio increases when the gap between the educational attainment of the median citizen and the bureaucrat decreases, letting. in anti-corruption initiatives. If caught extracting a bribe, the government official would face punishment, which, for simplicity, we assume to be a monetary cost, F. Nonetheless, when the political competition is tense, the public officer would be under political pressure from the central government to uphold the provision of G. The relative weight,, of the officer s concern over the public good provision to his private gain represents this political pressure. 13 Equation 3.3 describes the government official s original utility function while equation 3.4 substitutes G with the budget constraint. The next section derives the optimal bribery rate (b) of the bureaucrat and carries out some comparative static analysis using the main properties of interest. (3.2) Our testable hypothesis is that, as press freedom and education equality improve, a government officer will be more accountable for his or her rent-seeking activities. Although the bureaucrat cannot manipulate P directly, he can make corruption sophisticated and more difficult to track down as his educational attainment,, increases. The opportunistic bureaucrat takes advantage owing to his highly educated background by making corruption as sophisticated and secretive as possible in order to maximize his expected gain from corruption. As the government officer knows the quality of press freedom (I) and education equality in his service area, he therefore optimizes the bribe strategy (b) accordingly. When corruption becomes complicated, it reduces the chance for citizens to assimilate a revealing piece of information. This creates imprecision of information on the bribe rate observed by citizens who, on average, have a less-educated background relative to the government officer. The empirical evidence presented in Dwivedi (1967) shows that this group of citizens is likely to misevaluate downward the level of corruption and it has less incentive to participate 3.4 Optimal Bribe Level Equations show first order and second order conditions (FOC and SOC) of the government official s utility with respect to b. Equation 3.6 presents the optimal bribe rate, which is a function of political pressure (γ), monetary fine (F), press freedom (I) and education equality (σ), while equation 3.7 shows its concavity. We then calculate the optimal bribe level with respect to the determinants of interest. Equations 3.8 and 3.9 present the effects of changes in education inequality and press freedom, respectively. 13 Alternatively, γ may represent the type of officer, honest or corrupt.

9 135 credibility of the civic monitoring threat against corruption, which results in the persistence of the misuse of public office. The next section will seek empirical support for these theoretical arguments. Equations 3.8 and 3.9 represent the anti-corruption return to education equality and press freedom. Obviously, apart from its independent effects, both determinants of civic monitoring capacity work together in eliminating corruption. This finding contributes to that of Ahrend (2002), which implicitly shows that the anti-corruption role of education depends on monitoring capacity. Our finding explicitly shows that education equality and press freedom work together in controlling corruption through civic monitoring. Additionally, equations 3.8 and 3.9 prove that political pressure complementarily promotes the effects of education equality and press freedom. Moreover, from equations 3.10 and 3.11, political pressure (γ) and monetary fine (F) also have independent anti-corruption effects. The political pressure role increases with press freedom and education equality while the role of the monetary fine decreases with the size of the public good provision. This reflects the fact that what matters in constraining corruption is not the absolute value of punishment but the relative value of punishment and the value of the available rents. To conclude, this section illustrates the theoretical foundation for our research agenda. Predictions from the comparative analysis of the reduced-form model prove that improvements in education equality, press freedom, political pressure on bureaucrats and magnitude of punishment (I, σ, F, γ ) can reduce the optimal bribery level chosen by the government officials in the economy. Intuitively, when inequality in education and limitation in media freedom are substantial, they significantly reduce the 4. Data and Analysis 4.1 Subjective Corruption Indices We use cross-country subjective indices of the perceived level of corruption from three different sources: Transparency International s Corruption Perception Index (CPI), International Country Risk Guide s Corruption Index (ICRG) and Daniel Kaufmann s Control of Corruption (WB). All corruption indices are re-scaled on the basis of 0-10, where 10 stands for countries with the least corruption. These three corruption indices are very similar in design and variation; in fact, the correlations between them from 1995 to 2005 are well above 0.8. Table 1 summarizes all corruption indices by year. The average perceived level of corruption has been increasing across the globe, while its dispersion has been decreasing. The mean values of each index are very close to each other throughout time. All the indices will be employed as the dependent variable in three different regression specifications: cross-national OLS analysis, pooled OLS and the panel fixed effects - model. The first specification employs the averaged values of the indices between 1995 and In the pooled OLS and panel fixed-effects model, as all education variables are available on a five-year basis, the corruption indices will be an average value of the periods of four years around the time of the analysis, e.g. the average for the years represents the year As the CPI and WB indices are available between 1995 and 2005 while the ICRG index is available between 1984 and 2003, pooled OLS and fixed-effect estimations that use the CPI and WB indices as dependent variables will employ three periods of repeated cross-country data, whereas the ICRG regression will employ four periods of data between 1990 and 2005.

10 136 Table 1. Descriptive Statistics of Corruption Indices 4.2 New Measures of Education Equality Identifying education equality from the B&L data set is nothing new. Castello and Domenech (2002) and Thomas et al. (2003), using B&L education attainment data, have calculated the Gini coefficient of education distribution (Gh) and the ratio between the level of education attained by the lowest and the highest quintiles. In addition, the literature 14 on economic growth employs the standard deviation of educational attainment (ESD) as a proxy for education inequality. However, ESD is not suitable for identifying education equality in our framework. It measures primarily the absolute dispersion of human capital across the population, but does not control for the differences in the mean of the distributions. As some countries characterized by a low-educated population can have the same standard deviation in educational attainment as countries with a highly educated population, the interpretation of the anti-corruption effect of the distribution of education measured by ESD can be misleading. On the other hand, although Gh captures the information on education equality of interest, this indicator has two crucial methodological shortcomings. The Gini coefficient itself cannot precisely identify the relative distance between human capital stocks within the distribution. 14 Birdsall and Londono (1997), Lopezetal. (1998) and Park (2006) Rather, it identifies the distribution of human capital stock across the total population. In fact, the same value of the Gini coefficient can represent different shapes of Lorenz curves. Putting it differently, when the relative level of educational attainment between the median educated citizen and public officers (σ) varies, the value of the Gini coefficient does not necessarily identify this difference. Another shortcoming of the Gini coefficient is the limitation of the granularity of the measurements. For instance, using the same distribution of educational attainment, a Gini coefficient calculated from 10 levels of schooling, (high granularity) will often yield a higher value (less equal) than a Gini coefficient calculated from 5 levels of schooling (low granularity), and vice versa. Since the B&L data set on educational attainment contains only four different levels of schooling, it is likely that the Gini coefficient calculated from this data set would overestimate the true equality of the distribution in educational attainment. As our analysis focuses on the relative levels of schooling, we need a more precise indicator to identify σ from the B&L data set. To identify σ in the B&L data set, we calculate the years of schooling attained by the median citizen (median) and by public officers. To locate the median, we treat the B&L data set as simple group data where the percentage of the population that graduated in each level of schooling represents the frequency. For the public officer, we assume that the public officer s years of schooling are represented by the years of schooling attained by the 4 th quartile group (Q4) in the distribution of education. Table 2 shows summary statistics of the new variables. The mean years of schooling of the 4 th quartile was below 5 years in 1960; it increased to about 10 years in Apparently, 5-10 years of schooling is equivalent to a normal secondary school qualification, which is the typical criterion for the employment of public servants. This evidence supports the external validity of the assumption that uses the 4 th quartile as a proxy of public servant schooling. In addition, on average the median had been lower than the mean in the education distribution throughout the 40 years

11 137 covered in the B&L data set. Therefore, statistically, the median is the superior measure of the central tendency of education distribution rather than the mean in the B&L data set. Nevertheless, it is important to note that about one-fifth of the countries in the B&L data set accounted for more than 50 per cent of the adult population with no formal schooling. 15 Undeniably, in this case, the schooling year of the median citizen is equal to 0. As a result, the typical minimum values of the median in Table 2 are 0. The subsequent empirical analysis uses median/q4 as the main indicator in the regression analysis to identify the anti-corruption effect of education equality while the coefficient of variation (CV) and the Gini coefficient of education (Gh) are used as a robustness check. Table 2. New Measures of Education Equality 4.3 Other Determinants of Civic Monitoring To measure freedom of the press (I), we employ the quantitative press freedom score compiled by Freedom House. These survey data evaluate the freedom of information in printed media from various influential sources, such as legal, political and economic authorities. Although a qualitative score has been available since the early 1980s, a quantitative score had not been available until Hence, in the panel data analysis, the press freedom score in 1994 represents freedom of the press in 1990, while the averaged value of this score for the periods , and represent freedom of the press in 1995, 2000 and 2005, respectively. Owing to the unavailability of a quantitative press freedom score in 1985, the panel data regressions employ four time periods from 1990 to Furthermore, as the theoretical model in section 3 predicts that press freedom and education equality jointly affect the corruption level, we add the interaction terms between press freedom and different measures of education equality into the regression specification. A measure of judicial independence from the Political Constraint Index (POLCON) is used to measure the magnitude of punishment (F) in the economy. This index was compiled by Henisz. 16 The independence of the judiciary is identified by using the information from various measures of judicial independence, including a polity score on executive constraint and a score on law and order in the previously mentioned International Country Risk Guide. This measure is continuous from 0 to 1 and available from the mid-nineteenth century until The higher value of this measure represents the strength and impartiality of the legal system and the likelihood that the judiciary would successfully constrain the decision of the executive authority. As the degree of independence in the court of justice increases, it is more likely that corrupt public officers would be fully accountable for their misuse of power. This likelihood of punishment can be seen in the theoretical model as an increase in F. Political pressure (γ) is measured by the degree of political competition and turnout data from poliarchy measures of democracy from Vanhanen (2003). Data are available from 1810 to 2002 in nearly all independent countries around the world. The measure of electoral 15 This pattern is highly persistent in some countries over time (e.g. in Algeria, the Central African Republic, the Islamic Republic of Iran, and Pakistan) POLCON/ContactInfo.html

12 138 competition represents the percentage of votes in parliamentary or presidential elections, or both, won by the largest party. Therefore, the smaller the measure is, the more likely it is that a candidate from a small party would win the election. For turnout, the turnout record measures the percentage of the population that voted in the same elections which used to measure the electoral competition. 4.4 Control Variables Controls for a country s characteristics are based specifically on the findings of global sensitivity analysis in Serra (2006). To control for the economic development and the economic structure of a country, per capita income at constant prices and degrees of openness are employed. These indicators were obtained from Penn World Table 6.2. For institutional and legal factors, following the argument in Treisman (2000), we use the dummy variable for colonial history, equal to 1 if the country is a former British colony, 0 otherwise, and the dummy variable for uninterrupted democracy, equal to 1 if the country had remained democratic between 1950 and 1995, 0 otherwise. We also add into the regression specification the interaction term between an uninterrupted democracy and education equality in order to test whether education equality works differently between the countries with a different establishment of democracy. In addition, to control for the factor of religion, we use the population share with a Protestant tradition from La Porta, Lopez-de-Silanes, Shleifer and Vishny (1999) as a proxy for Protestants in the regression analysis. Lastly, continental dummies are employed as the proxy for regional factors that could determine the perceived level of corruption. The dummies correspond to the division of regions in the World Bank s classification, which includes Latin America, Eastern Europe, the Middle East, Africa, South Asia, and Asia and the Pacific. The baseline category comprises the Western Europe and North American continents, which contain the typical least corrupt countries. The descriptive statistics are presented in Table 3. Table 3. Summary Statistics of Civic Monitoring Capacity and Control Variables 5. Regression Specifications 5.1 Ordinary Least Squares and Pooled OLS Cross-National Analysis Equation 5.1 formalizes the theoretical predictions into a typical cross-national OLS regression on the determinants of corruption literature, which was reviewed earlier. We regress corruption indices concerning the measures of education equality and the control variables described in the previous section. All variables represent the average values from 1995 to In addition, as education equality shows its supporting anti-corruption role with regard to other variables, such as press freedom, we look for empirical evidence of this argument by including in the regression analysis the interaction terms (k) between education equality and other variables. To the best of our knowledge, this is the very first attempt in determinants of corruption studies to incorporate into the analysis the interaction effects between education and press freedom. Also, one might argue that what determines corruption is the level of average years of schooling in the total population, or the proportion of the population that graduates at a particular level of schooling, even though the theoretical model shows that it is indeed an equality effect. Furthermore, one could also think that education determines corruption through an income effect. Therefore, we anticipate the arguments and test them by including those education variables (e) and per capita income

13 139 as control variables in the regression analysis in order to check the robustness of education equality. Additionally, we also control for a country s characteristics and other determinants of corruption (x) that had their robustness verified by the sensitivity analysis in Serra (2006). Hence, the averaged perceived level of corruption in country i between 1995 and 2005 is determined by the specification of the formula: To estimate a more precise effect of education equality on corruption and to incorporate into the analysis a within-country variation across the time dimension, we extend our investigation to the repeated cross-country estimation. In a similar fashion to equation 5.1, we estimate the pooled OLS regression model of cross-country data between 1990 and 2005 by regressing the ICRG corruption index on the determinants of civic monitoring capacity, the interactions terms (ĸ ij ), the control variables (x ij ) and the time-fixed effects, λ ij. Specifically, the observed level of corruption in country i at time j between 1990 and 2005 is determined by the specification of the formula: The empirical results of regressions 5.1 and 5.2 are discussed in section 6. Next, we discuss the regression specification of the fixed effects model Panel Data Fixed-Effects Model To anticipate the possibility of the omitting variable problem and the existence of unobservable heterogeneity, we employ a more advanced methodology in the empirical investigation, that is, the fixed-effects estimation. In a similar fashion to equation 5.2, we regress the ICRG index on the determinants of civic monitoring capacity (M), other education variables (e), the interaction term (ĸ), the control variables (x) and the time-fixed effects, λ ij. Additionally, we control for individual country-specific characteristics, Ψ ij. 17 Hence, the observed level of corruption in country i at time j is determined by the equation 5.3 below: If education equality has a causal relationship with the perceived level of corruption, the regression models presented here should be able to consistently identify significant causal relationships. Moreover, the results should remain robust through various specification changes. 6. Empirical Results 6.1 Descriptive Analysis This section gives a graphical description of the relationship between the dependent variable and its determinants. Figure 1 presents scatter diagrams, with the fitted regression lines located between the averaged values of the corruption indices and education equality measures between 1995 and Obviously, all three corruption indices show a strong relationship with education equality measures. On average, the higher is the median/q4 ratio (σ), the lower is corruption, whereas the higher is the Gini coefficient for educational attainment (less equal distribution), the more likely would corruption be observed in the society. 17 Most of the empirical literature on corruption employs the fixed-effect model rather than random-effect model, as it is generally believed that the country s specific effects correlate to some extent with the covariates, after using the Durbin-Wu- Housman test. The results support this argument well.

14 140 Figure 6.1 Scatter Diagrams of Corruption and Education Inequality the less likely it is for corruption to be perceived in that society. Nevertheless, it is worth noting that among three schooling levels, the percentage of secondary schooling attainment shows the strongest link with the corruption index, whereas primary schooling attainment shows a somewhat unclear association. Therefore, the theoretical predictions in section 3, that less equality in the distribution of education increases the likelihood of corrupt practices are well supported by the graphical evidence shown in the scatter diagrams. Nevertheless, all hypothetical observations here need to be verified by formal regression analysis in the next part, so that we will be able to see whether these associations are causal or just spurious relationships. Figure 2. Corruption, and Income and Press Freedom In addition, countries with higher per capita income and freedom of the press seem to have less corruption, as shown in Figure 2 Furthermore, Figure 3 presents scatter diagrams of the CPI corruption index and four different education measures: average years of schooling in the total population, percentage of the population attaining no schooling, and primary and secondary schooling, respectively. A country with higher years of average schooling and a smaller proportion of the population that had no schooling qualification tends to have less corruption in its society. Additionally, the larger the population that attained primary, secondary or tertiary education,

15 141 Figure 3. Corruption and Measures of Schooling 6.2 Cross-Country Ordinary Least Squares It is crucial to note that all education measures in our analysis are lagged variables. In OLS regressions, education variables are the average values between 1960 and 1980, while all the other variables are the average values between 1995 and 2005, whereas the pooled OLS estimations in section 6.3 employ 10-year lagged values of education measures. This identification strategy is used in order to prevent the endogeneity problem between corruption and education equality and to minimize the undesirable transitory shocks that may affect corruption in each country. Moreover, using lagged values of educational measures reflects a more realistic story. People influence the level of corruption in society for most of their lives, as educated citizens can produce lasting anti-corruption or corruption initiatives. This argument is in line with that of Glaeser, Ponzetto and Shleifer (2007) and Svensson (2005). 18 Appendix Table 1 presents the baseline results of OLS regressions, as specified by equations 5.1 and 5.2, respectively, without and with the vector of interaction terms (ĸ). Columns in Appendix Table 1 present the results of pooled OLS regressions with time-fixed effects, as specified by equation 5.2. that table presents the key estimations that use median/q4 as a 18 They both found empirical evidence to support the economic and human capital theories of institutional development.

16 142 measure of education equality, which has theoretical support from section 3, while Appendix Table 2 presents a robustness check by re-estimating equations 5.1 and 5.2 again, with alternative measures of education equality. (a) Education Equality and Press Freedom In columns 1-3 in Appendix Table 1, the estimates of the association between education equality and the corruption indices, as specified in equation 5.1, without an interaction term, yield no significant result. Thus, we follow the theoretical prediction by adding the interaction terms into the specification; the results are presented columns 4-15 of Appendix Table 1. The measures of education equality now show a significant relationship with the corruption indices. A possible explanation is that the equality in education distribution determines the corruption level individually and jointly with other variables. When we exclude the interaction terms from the regression specification, the interaction effect remains inside the error term, which then creates the problem of omitted variable bias. As most of the significant interaction terms (columns 4, 5, 7, 10-11, in Appendix Table 1) have opposite signs to the measures of education equality, there are two opposite forces determining the corruption level which need to be identified. Unable to identify such an effect, the regression specifications in columns 1-3 in Appendix Table 1 fail to reject the null hypothesis that the measures of education equality have no causal relationship with corruption indices. From the results in regressions 4, 5 and 7 in Appendix Table 1 everything else being equal, the countries with a smaller gap of schooling years between the median and the 4 th quartile in the distribution of education (higher median/q4) during the period were less likely to have been corrupt during the period More specifically, from regression 4 in Appendix Table 1, reducing 10 per cent of the years of schooling gap between the median and the 4 th quartile (σ) produces an increase of in the score of corruption indices. This increase is equivalent to the difference between the CPI index of Cameroon (2.3) and Argentina (2.9) in Nonetheless, as regression models in Appendix Table 1 contain two interactions of education equality measures with press freedom (PF) and continuity of democracy (AllDem), the interpretation of the anti-corruption effect of education equality measures needs to incorporate the supplementary effect of other determinants if the interaction effects are significantly different from 0. Otherwise, the interpretation can be inaccurate. In regressions 5, 7, and 13-14, the coefficients of the interaction term between education equality and press freedom are significantly different from 0. Consequently, the interpretation of the anti-corruption effect of education equality σ, needs to incorporate the supplementary effect of press freedom. As both factors of interaction are continuous, we need to calculate the net effect as follows: In letting X i represent other covariates, in Appendix Table 2 we regress as follows: To calculate for the main effect, we use the following formula: The anti-corruption effect of education equality (σ,) depends on the coefficients of education equality, press freedom (I), their interaction term and the level of press freedom. To make our interpretation more meaningful we choose the mean level of press freedom to interpret the result, which equals Thus, using equation 6.2, from regression 4 in Appendix Table 1 a reduction of 10 per cent in the years of schooling gap yields, on average, an increase in the CPI index of about This suggests that the OLS estimators in column 1 of Table 1 underestimate the effect of education equality on corruption due to the omitted variable problem. Also, the earlier interpretation, which does not incorporate interaction effects,

17 143 underestimates the true effect of education equality. Figure 4 presents graphically the effects of median/q4 (σ,) on corruption, depending on the levels of press freedom. To make this interpretation more intuitive, the upper panel in Figure 4 shows that, given a press freedom level below 50 (free press), as we move down to 1 along the median/q4 axis (more equality) or move up to 0 along the press freedom axis (more freedom), the CPI score increases. Put in other words. the marginal anti-corruption effect of education equality is positive. On the other hand, when press freedom is above 50, the slopes become negative. This shows that the marginal effect of an increase in education equality becomes negative when press freedom is limited. However, although the slopes become negative, the net effects still remain positive until the changes in median/q4 start to exceed 0.4. Hence, the negative net effects of the median/q4 level occur when two conditions are met: when the level of press freedom is above 50 and an increase in education equality is dramatic (over 0.3). However, in reality it is very difficult to see the net negative effect of education equality. From the median/q4 data between 1960 and 2000, which contain 832 observations, we calculate the first differences of this variable to see how likely any country in the data set has experienced a change of over 0.3 units in median/q4 within this five-year interval. We find that, out of 738 observations of the first differences, there are only 2 observations (0.28 per cent) that had values above 0.3. Hence, in principle a negative effect of education equality is possible, but in reality it is very unlikely to occur. This likelihood is far less than the findings of Ahrend (2002). Figure 4. Upper Panel: The Effects of Median/ Q4 and Press Freedom on the CPI Index (with Interactions); Lower Panel: Corruption Measured by Press Freedom and Gini Coefficient for Education The other measures of education distribution, coefficient of variation and the Gini coefficient of education, the results of which are presented in Appendix Table 2 show similar relationships between the corruption indices and median/q4. In the countries with less equal education distribution, the likelihood of observing the incidence of more corruption increases. From regression 2 in Appendix Table 2, the increase by 0.1 of the coefficient of variation in educational attainment can explain the reduction in the WB corruption index of about This is comparable to the average gap between the WB corruption index from 1996 to 2005 between Brazil (4.8) and South Africa (5.9). Likewise, from regression 10 in Appendix Table 3, an increase of 0.1 in the Gini coefficient for educational attainment can explain the reduction of about 2.47 in the average

18 144 CPI corruption index between 1995 and In other words, given all possible levels of press freedom, the more equal is the distribution of education, measured by the Gini coefficient and coefficient of variation, the less frequently corruption will be perceived in the society. Similarly, press freedom has both individual and joint relationships with education equality measures, especially with the Gini coefficient and coefficient of variation regressions in Appendix Table 2 consistently show that countries with more press freedom were less likely to witness corruption. However, the effect of press freedom was reduced by the inequality in education distribution. The visual interpretation of the press freedom effects is presented by the bottom panel of Figure 4, which depicts a negative relationship between press freedom and corruption from regression 10 in Appendix Table 2. Obviously, this link depends partially on the level of education equality measured by the Gini coefficient. The least corrupted society occurs at the top left of the plain, where the distribution of education is at the most equal point and press freedom is at its freest point. On the contrary, the most corrupted society occurs at the bottom right of the plain, with the most unequal distribution in education and fully limited freedom of the press. Hence, press freedom and education equality should be considered as complimentary tools in any anti-corruption campaign. (b) Education Equality and Democracy In Appendix Table 1 the coefficients of the interaction terms between median/q4 and uninterrupted democracy are insignificantly different from 0; hence, the anti-corruption effect of education equality in a country that has more than 40 years of uninterrupted democracy (Alldem=1) is insignificantly different from a country that has not had such stability in its political system. 19 Therefore, we do not need to interpret an anti-corruption effect of education 19 Although we exclude this interaction from the model, all the main results are still robust. equality in the country, separately with and without uninterrupted democracy. Besides, the duration of democracy shows no significant anti-corruption effect when median/q4 is used as the measure of education equality. Nevertheless, in Appendix Table 2 when the Gini coefficient and coefficient of variation are used as the measure of education equality, the effect of democratic stability becomes significantly positive. Based on the results in Appendix Table 2 the country with uninterrupted democracy is more likely to receive, on average, a corruption score 1-2 points higher compared with a country that has experienced such as interruption. A possible explanation for this finding is that countries with established democracy may have institutional factors or supporting mechanisms which promote accountability more effectively than countries with fragile democracies. Moreover, some interaction terms between education equality measures and uninterrupted democracy are significantly different from 0 (columns 1-3, 6-9, 10 and 16-18), which means that the impact of changes in education distribution, measured by Gini coefficient and coefficient of variation in a country that has not experienced 40 years of uninterrupted democracy (Alldem=0) is different from a country with 40 years of uninterrupted democracy. However, this finding is barely significant and is highly sensitive to specification changes; thus, we focus our attention primarily on the results from Appendix Table 1. (c) Other Determinants of Civic Monitoring Most of the other determinants of civic monitoring capacity show significant associations with corruption indices in regressions presented in Appendix Tables 1 and 2. In countries with higher judicial independence and competitive democracy, corruption was less likely to occur, ceteris paribus. Moreover, countries with fewer Protestants, lower income per capita and smaller degrees of openness were more prone to corruption. However, the measure of legal and institutional culture, British heritage, does not show any significant link to perceived levels of corruption around the globe. These findings

19 145 are consistent with those of Serra (2006). Additionally, the results in regressions 7-9 and in Appendix Table1 and regressions 4-6, and in Appendix Table 2 include regional dummies within the regression model. 20 Latin American countries tend to have more corruption than Western European and North American countries; specifically, they are about 1.5 points lower in the corruption indices. Moreover, the anti-corruption effects of all education equality measures increase when the regional dummies are added. This suggests that the regional factors do matter and cannot be left in the residual terms. Nonetheless, all main results discussed above still remain robust. 6.3 Pooled Ordinary Least Squares Estimation We re-estimate the OLS regression model using the repeated cross-country data set of the same set of countries between 1990 and 2005 and include time-fixed effects within the model. Columns in Appendix Table 1 and columns 7-9, in Appendix Table 2 present the results of pooled OLS regressions, as specified by equation 5.2. As the CPI and WB corruption indices are available only from 1995, the regressions, which employ these indices as dependent variables, will estimate the repeated cross-country data set between 1995 and All education measures are in 10-year lagged values. The main findings are highly consistent with the earlier OLS estimations. For instance, from regressions 10 and 11 in Appendix Table 1, reducing by 10 per cent the years of schooling gap between the median and the 4 th quartile produces a net increase of about and in the scores of the CPI and WB corruption indices, respectively. When including the regional dummies into the regression model, the anti-corruption effects of education equality in terms of the corruption indices score are slightly reduced to and 0.673, respectively, as presented in columns 13 and14 of the same table. Moreover, as was the case with the earlier findings, richer countries seem to have less 20 The coefficients of regional dummies are not presented in the tables but are available upon request. corruption: specifically, an increase of US$ 1,000 in per capita income can result in increases of about in the corruption indices scores. In addition, the anti-corruption effects of press freedom, share of Protestants in the population and duration of uninterrupted democracy are very much the same as in the cross-country OLS estimation. The effects of openness and political pressures are insignificant here. Interestingly, however, the income effect disappears when regional dummies are included within the regression specification while the effects of other explanatory variables remain unchanged. Moreover, we also add into the OLS specification in equation 5.2 the proportions of the total adult population that attained three different schooling levels: primary schooling, secondary schooling and tertiary schooling. The results are presented in Appendix Table 3. Interestingly, having a more educated population does not necessarily reduce the likelihood of corruption in the society. All measures of population share, in each schooling level, show significant and differential effects toward corruption. An increase in the proportion of the population attaining schooling can reduce corruption in the society only if the schooling is at the secondary level or above, see columns 2-3 and 5-6 in Appendix Table 3. This finding, however, contradicts that of Ahrend (2002, p. 14), which shows that only tertiary education can significantly control corruption while primary and secondary levels of education have neither positive nor negative significant effects. These different findings could potentially originate from the difference in the measure in education between our work and that of Ahrend (2002), as discussed previously. More importantly, the results in columns 1 and 4 highlight the alarming evidence that primary schooling is insufficient to control corruption. As discussed in the theoretical section, the higher share of adult population who graduated only from primary school increases the likelihood that the median population attains only primary schooling which critically undermines the capacity of civic monitoring to detect or to control corruption among public officers. This argument is reinforced by the positive and

20 146 significant coefficients of median/q4 in columns 1 and 4. Therefore, social planners who employ civic education in measures to fight corruption should incorporate this non-monotonic relationship between educational attainment and corruption. The greater the number of people who acquire post-primary education, the more effective monitors they can be in the fight against corruption. 7. Identification Problems and Strategy This section verifies the robustness of the results of the OLS estimations in section 6 by examining for endogeneity and omitted variable bias, which could lead to loss of identifiability in the parameters of interest in the OLS regression. We employ a fixed - effects model and instrumental variable to control for endogeneity and omitted variable bias. 7.1 Endogeneity Problem The reverse causality between education and corruption has been consistently shown in the literature that examines the effects of corruption on the public provision of education. Research suggests that, by reducing corruption today, the education system can be improved in the future by securing government funding for the targeted educational plans (Mauro, 1998; Reinikka & Svensson, 2005). Using the instrumental variable 21 estimation, Reinikka and Svensson (2005) showed that an increase in public information exposure is associated with an increase in government spending for local schools, which would have been extracted by rent-seeking activities. The estimations presented in sections 6.2 and 6.3 anticipate the endogenous relationship between education and corruption presented by the literature, and identify the anti-corruption effects of education equality by using the lagged values of education measures to prevent such a problem. The validity of this strategy depends upon the identifying assumption that the present value of corruption does not affect the past value of education measures. There are realistic arguments that support this assumption. First, there is no evidence to support the hypothesis that corruption today can cause education inequality in the past. Mauro (1998) and Rikka and Svensson (2005) present only evidence in which current levels of corruption affect future values of education spending, enrollment rates and academic performance. Second, as subjective corruption indices were conducted primarily by evaluating the perception of international businessmen toward corruption in their host country, there is no clear channel that their perceptions 22 toward the level of corruption today could possibly determine the situation of education inequality a decade ago. Additionally, we perform the Durbin-Wu-Hausman test for endogeneity to see how effective the current identification strategy is in preventing such a problem. The results of the test on OLS and pooled OLS estimations are presented in Appendix Tables 4-5. Surprisingly, the coefficients of the residual terms from the reduced-form regression from regressions 4-6, and in Appendix Table 4 and from regressions 3-4 in Appendix Table 5 are consistently significant, which means that the endogeneity problem still exists in the estimations. However, in recalling the argument about heterogeneity discussed previously, we suspect that the endogeneity problem detected by the Durbin-Wu-Hausman test does not originate from the reverse causality problem but from the unobservable heterogeneity problem. 7.2 Unobservable Heterogeneity T h e p r o b l e m o f u n o b s e r v a b l e heterogeneity in our context is similar to the individuals unobservable ability in the study on return to schooling. In our estimates, this problem could potentially originate from the correlation between educational regressors and unobservable anti-corruption capacity of civil society, and more specifically as follows: 21 They used the distance to the nearest newspaper outlet as an instrument of the teacher s knowledge about a grant program. 22 There is no question about the country education system in the past years in the questionnaire.

21 147 Equation 7.1 presents the hypothetical equality between corruption and its determinants where E and X stand for educational determinants and other determinants in equation 5.2, while z represents the unobservable anti-corruption ability of civil society. When z is correlated with E, the residual term (αz+v) will be associated with the regressor E, which causes the inconsistency in the OLS estimators. 23 In attempts to evaluate the civic return to education in cross-country analysis by Ades and Di Tella (1999) and Ahrend (2002), they overcome the unobservable heterogeneity problem by using the instrumental variable and fixed effects model to capture the unobservable country characteristic. We adopt both methods in our empirical investigation. If the unobservable anti-corruption ability is a time-invariant, the fixed effects estimation should provide consistent parameters of interest to us. However, if the ability is indeed a time-varying factor, we also need to employ the instrumental variable method to identify the causal relationships in the fixed effects estimation. On the other hand, for average cross-national OLS estimates, using the instrumental variable method is the only option we have. 7.3 Fixed-Effects Estimation Fixed-effects estimation is a demanding methodology; thus, owing to the limited time dimension of our panel data set, we start the fixed-effects estimation from the simplest specification as presented in equation 5.3 by regressing all corruption indices on education equality measures and income in regressions 1-9 in Appendix Table 6. The anti-corruption effect of education equality measures, namely median/ Q4 and the Gini coefficient, remain significant but their magnitude is substantially reduced compared with that of the earlier estimates. From 23 In fact, from the surveys and evidence provided by the India Institute of Public Opinion (1965), it is difficult to reject the hypothesis that Cov[E, z] 0. regressions 2-3, reducing by 10 per cent the years of schooling gap between the median and the 4 th quartile can increase by approximately the score of the corruption indices, while a 10 per cent reduction in the Gini coefficient increases by about 2 scores the WB corruption indices. The reduction in anti-corruption effect of median/ Q4 supports the hypothesis that Cov[E, z] > 0. Thus, the OLS estimator overestimates the anti-corruption effect of the education equality measure. Interestingly, the anti-corruption effect of mean/q4 is insignificant, which suggests that the relative years of schooling attained by the median is more important than the mean citizen in the anti-corruption context. 24 More importantly, in regressions 11-12, when including the measure of average years of schooling within the model, the coefficient of median/q4 is relatively unchanged from regressions 2-3, while the effect of averaged schooling years is insignificant. This result suggests that the significance of this variable is driven by the correlation with unobservable anti-corruption ability in the error term. 7.4 Instrumental Variable and Two-Stage Least-Squares (2SLS) Estimates T o c o n t r o l t h e u n o b s e r v a b l e heterogeneity in average cross-national OLS estimates, the method of instrumental variable is the most employable choice in the literature. In section 7.2. we have reviewed some of the efforts of Ades and DiTella (1999) and Glaeser and Saks (2006) to ease the endogeneity problem by instrumenting education indicators. Moreover, Cook (2002) and Moretti (2004) used demographic variables, including the population structure, sex ratio and expectation of life at birth, as the instruments, whereas Park (2006) also employed the global price of some commodities as an instrument. They argued that the violation in these variables had an influence on human capital formation, but not corruption and income. We follow this identification strategy in our analysis by using the share of the middle-age cohort (15-60 years of age) in 24 The result will be made available on request.

22 148 the total population and the sex ratio as the instruments for the measures of education equality. 25 The share of the middle-age cohort (m) is calculated by using the information of the young (y) and aging (o) cohorts shares in the total population, specifically, as follows: From equation 7.2, the share of the middle-age cohort in country i at time t is equal to 1 minus the shares of the young and old cohorts 10 years previously. Hence, when the shares of young and/or old cohorts increase, they raises the future share of the middle-age cohort. The sex ratio is also in 10-year lagged values; the value of 110 means that, on average, there are 110 males for every 100 females in the total population. To be valid instruments, these demographic variables need to be uncorrelated with the error terms in equations 5.1 and 5.2, in particular, the unobservable anti-corruption ability, and sufficiently correlate with the education equality measures. For the exogeneity criteria, the underlying argument is that a larger share of the middle-age cohort means that the old-age population with lower averaged years of schooling is leaving the population structure while the younger generation with higher averaged years of schooling enters the population. On the other hand, women have a better opportunity to obtain schooling nowadays than decades ago. The new generation of females should attain higher schooling than ancestoral generations did. Therefore, the changes in these instruments are mainly due to demographic variation, for which there is no obvious link to unobserved anti-corruption ability. Therefore, exogenous variations are created for the distribution of human capital across the population. For the second criterion, we find that a larger share of the middle-age cohort and female population correlate with the more equal distribution of education in the society The data are obtainable from the United Nations Population Division website: 26 The result will be made available on request. Nevertheless, the validity of this identification strategy still depends on the identifying assumption that these past demographic changes do not directly correlate with the current perceived level of corruption in the society. One concern may be that a larger middle-age population induces intolerance against corruption and civic movements. However, based on the evidence presented in Magnus (2002) and Glaeser, Ponzetto and Shleifer (2007), we argue that these anti-corruption effects of demographic changes work through the education system. In other words, there are no clear direct links between demographic changes and the perceived level of corruption except through the education channel. Hence, we have run the first stage regression according to the following specification: letting р ij stand for all predetermined variables while Z ij is a vector of instrumental variables. Appendix Tables 4-5 present the 2SLS results of the averaged cross-national estimate and repeated cross-national estimate accordingly. For a robustness check, all three measures of education equality are instrumented and we estimate the 2SLS regressions for all available corruption indices. From Appendix Table 4, regressions 1-3 present the reduced forms estimates as specified by equation 7.3 using different measures of education equality. The results show that the middle-age cohort can significantly explain the variation of education equality measures, while the sex ratio can barely explain the variation. Columns 7-9, and present second stage estimations. When the measures of education equality are instrumented, their anti-corruption effects become substantially larger. However, the F-statistic for the test of overall fit of first-stage regression at the bottom of Appendix Table 4 shows the values range from 1.3 to 5.5. As the anti-corruption effects of the education equality measure become considerably larger than in the OLS estimates in Appendix Tables 1-2, even the values of the F-statistic in

23 149 reduced form estimates are lower than 10; it is a rule of thumb that these are the signal of a weak instrument problem. If the instrument is to be a legitimate one, it should correct the inconsistency in the estimator of educational measure, not increase it. In this case, as Cov[E, z] > 0, the 2SLS estimator should be smaller than the OLS estimator, which is contaminated by the effect of the error term. A possible explanation for the weak instrument problem is the limitation in the variation of instruments, especially the sex ratio. The 2SLS estimates on repeated cross-section in Appendix Table 5 and fixed-effects estimations with the instrumental variable in Appendix Table 6 also present similar results and problems. From columns 5-6, 9-10 and in Appendix Table 5, the anti-corruption effects of median/q4 and the Gini coefficient in the 2SLS estimates are larger than the OLS estimations in Appendix Tables 1-2. However, this problem is improved in the repeated cross-country context, as there are more observations available compared with the averaged OLS estimations. Consequently, the F-statistics increase about 5-6 times and the coefficients of the second-stage regression are sizeably reduced. Therefore, for the cross-country analysis we need a better instrument or larger panel data to estimate the consistent relationship between education equality and corruption. We leave this opportunity for future research. Nonetheless, we have estimated the fixed effects model that controls the unobservable heterogeneity problem by the country fixed effects. As long as the assumption that an unobserved anti-corruption ability is time invariant, the identifiability of these results remains consistent. In fact, unlike the time-varying unobserved heterogeneity found in the work of Frechette (2006), which determines the availability of rents in the economy, there is no clear evidence that the anti-corruption ability of civil society around the globe has changed dramatically during the past 10 years. Hence, the result of the fixed-effect model in Appendix Table 6 supports our hypothesis that education equality can control corruption, while the results from OLS regressions should be treated with caution until legitimate instruments for education equality measures or a larger panel data set are available. 8. Conclusion The literature on the causes of corruption suggests that education, in general, is statistically irrelevant or positively correlated with corruption. However, these studies usually use a vague measure of human capital, mostly primary school enrollment and the unrealistic assumption of a homogenous stock of human capital in the theoretical model, to support the conclusion. This paper argues that this inappropriate identification strategy and modeling are the underlying factors of such findings. As education is like a two-edged sword in the conduct and control of corruption, understanding what each side is meant for should help policymakers and practitioners to employ education in a more accurate and efficient way in order to combat corruption. Our theoretical model internalizes the diverse roles of education on corruption in order to identify a non-monotonic relationship between education and corruption. The theoretical and empirical analyses consistently suggest that education equality can capture a more precise role of education on corruption. The findings support an alternative hypothesis: that the effect of education is differential in the levels of schooling. More importantly, the equality aspect of education plays a crucial role in determining the level of corruption in the cross-country database. Two lessons can be learned from this study. The distribution of human capital should receive more attention from anti-corruption initiatives as a supplementary tool to combat corruption, together with the existing anti-corruption toolkit. Furthermore, the formulation and assessment of the anti-corruption program, in which education has been involved, should treat the impact of education on corruption distinctively by its levels and the nature of measurement.

24 150 In addition, although education equality has direct anti-corruption effects at almost all quantitative levels of press freedom, it works more effectively in a country with freedom of the media and a stable democracy. Future research should develop further theoretical arguments to gain a better understanding of the non-monotonic association between educational attainment and corruption. The potential effects of human capital distribution across other socio-economic dimensions, such as gender and geographical areas, should also be examined. From the empirical standpoint, a longer panel data set, and a microbased data set, which are rich in both observations and duration of time, should enable researchers to obtain more precise estimates of the effect of education on corruption with a more demanding identification strategy. References Ades, A., & Tella, R.D. (1999). Rents, competition, and corruption, The American Economic Review, 89 (4), Ahrend, R. (2002). Press freedom, human capital and Corruption. DELTA Working Paper. 11. Alvarez, M., Cheibub, J.A., Limongi, F, & Przeworski, A. (1996). Classifying political regimes, Studies in Comparative International Development, 31:3-36. Barro, R. J., & Lee, J.W. (2001). International data on educational attainment: Updates and implications, Oxford Economic Papers, 53 (3), Brunetti, A., & Weder, B. (2003). A free press is bad news for corruption,. Journal of Public Economics, 87 (7-8), Castello, A., & Domenech, R. (2002). human capital inequality and economic growth: Some new evidence,. The Economic Journal, 2002, 112 (478), C187-C200. Cook, D. E. (2002). Education and growth: Instrumental variables estimates ECON Working Papers. Dee, T. S. (2004). Are there civic returns to education? Journal of Public Economics, 88 (9-10), Dwivedi, O. P (1967). Bureaucratic corruption in developing countries, Asian Survey, 7 (4), Eicher, To, García-Peñalosa, C., & Ypersele, T. (2006). Education, corruption, and the distribution of income,. monograph. Fréchette, G. R. (2006). Panel data analysis of the time-varying determinants of corruption, CIRANO Working Papers, 28. Gerber, A. &Green, D. (1992). Misperceptions about perceptual bias. Annual Review of Political Science, 2 (1), Glaeser, E. L., & Saks, R.E. (2006). Corruption in America., Journal of Public Economics, 90 (6-7), Glaeser, E.L., Ponzetto, G., & Shleifer, A. (2007). Why does democracy need education? Journal of Economic Growth, 12 (2), Indian Institute of Public Opinion (1965). The citizen and his society: Attitudes of the Indian people toward government, modernization and each other,. Public Opinion Surveys, Keen, E. (2000). Fighting corruption through education. Budapest: Open Society Institute. Koop, G., & Tobias, J. L. (2004). Learning about heterogeneity in returns to schooling,. Journal of Applied Econometrics, 19 (7), La Porta, R., Lopez-de-Silanes, F., Shleifer, A., & Vishny, R. (1999). The quality of government.journal of Law, Economics, and Organization, 15 (1), Lambsdorff, J. G (2007). The institutional economics of corruption and reform, Cambridge: Cambridge University Press. Magnus, J. R., Polterovich, V. M., Danilov, D. L., & Savvateev, A.V. (2002). Tolerance of cheating: An analysis across countries,. Journal of Economic Education, 33 (2), Mauro, P. (1998). Corruption and the composition of government expenditure,. Journal of Public Economics, 69 (2), Moretti,E. (2004). Estimating the social return to higher education: Evidence from longitudinal and repeated cross-sectional data,. Journal of Econometrics, 121 (1-2), Park, J. (2006). Dispersion of human capital and economic growth,. Journal of Macroeconomics, 28 (3), Persson, T., & Tabellini, G. (2000). Political economics: Explaining economic policy. Cambridge, MA: MIT Press. Persson, T., Tabellini, G., & Trebbi, F. (2003). Electoral rules and corruption, Journal of the European Economic Association, 1 (4),

25 151 Reinikka, R., & Svensson, J. (2005). Fighting corruption to improve schooling: Evidence from a newspaper campaign in Uganda,.Journal of the European Economic Association, 3 (2-3), Seldadyo, H., & de Haan, J. (2005). Determinants of corruption: A literature surrey and new evidence,. monograph. Serra, D. (2006). Empirical determinants of corruption: A sensitivity analysis,. Public Choice, 126, Svensson, J. (2005). Eight questions about corruption,. The Journal of Economic Perspectives,19, Thomas, V., Wang, Y.,.& Fan, X. (2003).Measuring education inequality: Gini coefficients of education,. Journal of Education Planning and Administration, 17 (1),5-33. Treisman, D. (2000). The causes of corruption: A cross-national study,. Journal of Public Economics, 76 (3), UNDP (2004). Anti-corruption: Practice note, Governance Practice Notes. Vanhanen, T. (2003). Democratization: A comparative analysis of 170 countries. New York: Routledge Waisman, G. (2005). Complementary controls of corruption, monograph.